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 07:40:04 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 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hf2_32 -include hf.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 * 181 stack variables, 7 constants, and 128 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hf2_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
40 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
41 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
42 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
43 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
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 T7d, T7a;
|
Chris@10
|
50 {
|
Chris@10
|
51 E T2, T8, T3, T6, Te, Tr, T18, T4, Ta, Tz, T1n, T10, Ti, T5, Tc;
|
Chris@10
|
52 T2 = W[0];
|
Chris@10
|
53 T8 = W[4];
|
Chris@10
|
54 T3 = W[2];
|
Chris@10
|
55 T6 = W[3];
|
Chris@10
|
56 Te = W[6];
|
Chris@10
|
57 Tr = T2 * T8;
|
Chris@10
|
58 T18 = T3 * T8;
|
Chris@10
|
59 T4 = T2 * T3;
|
Chris@10
|
60 Ta = T2 * T6;
|
Chris@10
|
61 Tz = T3 * Te;
|
Chris@10
|
62 T1n = T8 * Te;
|
Chris@10
|
63 T10 = T2 * Te;
|
Chris@10
|
64 Ti = W[7];
|
Chris@10
|
65 T5 = W[1];
|
Chris@10
|
66 Tc = W[5];
|
Chris@10
|
67 {
|
Chris@10
|
68 E T34, T31, T2X, T2T, Tq, T46, T8H, T98, TH, T97, T4b, T8D, TZ, T7g, T4j;
|
Chris@10
|
69 E T6t, T1g, T7f, T4q, T6u, T4z, T6y, T1J, T7j, T7m, T8e, T6x, T4G, T2k, T7o;
|
Chris@10
|
70 E T7r, T8d, T6B, T4O, T6A, T4V, T6P, T61, T7G, T3G, T6M, T5E, T8n, T7N, T6I;
|
Chris@10
|
71 E T5s, T7v, T2N, T6F, T55, T8i, T7C, T5L, T63, T43, T7O, T5S, T62, T7J, T8o;
|
Chris@10
|
72 E T2U, T2R, T2V, T58, T3a, T5h, T2Y, T32, T35;
|
Chris@10
|
73 {
|
Chris@10
|
74 E T1K, T23, T1N, T26, T2b, T1U, T3C, T3j, T3z, T3f, T1R, T29, TR, Th, T2J;
|
Chris@10
|
75 E T2F, Td, TP, T3r, T3n, T2w, T2s, T3Q, T3M, T1Z, T1V, T2g, T2c;
|
Chris@10
|
76 {
|
Chris@10
|
77 E T11, T1C, TM, Tb, TJ, T7, T1o, T19, T1w, T1F, T15, T1s, T1d, T1z, TW;
|
Chris@10
|
78 E TS, Ty, T48, TG, T4a;
|
Chris@10
|
79 {
|
Chris@10
|
80 E T1, TA, Ts, TE, Tw, Tn, Tj, T8G, Tk, To, T14;
|
Chris@10
|
81 T1 = cr[0];
|
Chris@10
|
82 TA = FMA(T6, Ti, Tz);
|
Chris@10
|
83 T1K = FNMS(T6, Ti, Tz);
|
Chris@10
|
84 T14 = T2 * Ti;
|
Chris@10
|
85 {
|
Chris@10
|
86 E T1r, TD, T1c, Tv;
|
Chris@10
|
87 T1r = T8 * Ti;
|
Chris@10
|
88 TD = T3 * Ti;
|
Chris@10
|
89 T11 = FNMS(T5, Ti, T10);
|
Chris@10
|
90 T1C = FMA(T5, Ti, T10);
|
Chris@10
|
91 TM = FMA(T5, T3, Ta);
|
Chris@10
|
92 Tb = FNMS(T5, T3, Ta);
|
Chris@10
|
93 TJ = FNMS(T5, T6, T4);
|
Chris@10
|
94 T7 = FMA(T5, T6, T4);
|
Chris@10
|
95 T1o = FMA(Tc, Ti, T1n);
|
Chris@10
|
96 T23 = FMA(T6, Tc, T18);
|
Chris@10
|
97 T19 = FNMS(T6, Tc, T18);
|
Chris@10
|
98 T1w = FNMS(T5, Tc, Tr);
|
Chris@10
|
99 Ts = FMA(T5, Tc, Tr);
|
Chris@10
|
100 T1c = T3 * Tc;
|
Chris@10
|
101 Tv = T2 * Tc;
|
Chris@10
|
102 T1F = FNMS(T5, Te, T14);
|
Chris@10
|
103 T15 = FMA(T5, Te, T14);
|
Chris@10
|
104 T1s = FNMS(Tc, Te, T1r);
|
Chris@10
|
105 T1N = FMA(T6, Te, TD);
|
Chris@10
|
106 TE = FNMS(T6, Te, TD);
|
Chris@10
|
107 {
|
Chris@10
|
108 E T1T, T3i, T3e, T1Q;
|
Chris@10
|
109 T1T = TJ * Tc;
|
Chris@10
|
110 T3i = TJ * Ti;
|
Chris@10
|
111 T3e = TJ * Te;
|
Chris@10
|
112 T1Q = TJ * T8;
|
Chris@10
|
113 {
|
Chris@10
|
114 E Tg, T2I, T2E, T9;
|
Chris@10
|
115 Tg = T7 * Tc;
|
Chris@10
|
116 T2I = T7 * Ti;
|
Chris@10
|
117 T2E = T7 * Te;
|
Chris@10
|
118 T9 = T7 * T8;
|
Chris@10
|
119 {
|
Chris@10
|
120 E T3q, T3m, T2v, T2r;
|
Chris@10
|
121 T3q = T19 * Ti;
|
Chris@10
|
122 T3m = T19 * Te;
|
Chris@10
|
123 T2v = T1w * Ti;
|
Chris@10
|
124 T2r = T1w * Te;
|
Chris@10
|
125 {
|
Chris@10
|
126 E T2W, T2S, T3P, T3L;
|
Chris@10
|
127 T2W = T23 * Ti;
|
Chris@10
|
128 T2S = T23 * Te;
|
Chris@10
|
129 T3P = Ts * Ti;
|
Chris@10
|
130 T3L = Ts * Te;
|
Chris@10
|
131 T26 = FNMS(T6, T8, T1c);
|
Chris@10
|
132 T1d = FMA(T6, T8, T1c);
|
Chris@10
|
133 T1z = FMA(T5, T8, Tv);
|
Chris@10
|
134 Tw = FNMS(T5, T8, Tv);
|
Chris@10
|
135 T2b = FNMS(TM, T8, T1T);
|
Chris@10
|
136 T1U = FMA(TM, T8, T1T);
|
Chris@10
|
137 T3C = FNMS(TM, Te, T3i);
|
Chris@10
|
138 T3j = FMA(TM, Te, T3i);
|
Chris@10
|
139 T3z = FMA(TM, Ti, T3e);
|
Chris@10
|
140 T3f = FNMS(TM, Ti, T3e);
|
Chris@10
|
141 T1R = FNMS(TM, Tc, T1Q);
|
Chris@10
|
142 T29 = FMA(TM, Tc, T1Q);
|
Chris@10
|
143 TR = FNMS(Tb, T8, Tg);
|
Chris@10
|
144 Th = FMA(Tb, T8, Tg);
|
Chris@10
|
145 T34 = FMA(Tb, Te, T2I);
|
Chris@10
|
146 T2J = FNMS(Tb, Te, T2I);
|
Chris@10
|
147 T31 = FNMS(Tb, Ti, T2E);
|
Chris@10
|
148 T2F = FMA(Tb, Ti, T2E);
|
Chris@10
|
149 Td = FNMS(Tb, Tc, T9);
|
Chris@10
|
150 TP = FMA(Tb, Tc, T9);
|
Chris@10
|
151 T2X = FNMS(T26, Te, T2W);
|
Chris@10
|
152 T2T = FMA(T26, Ti, T2S);
|
Chris@10
|
153 T3r = FNMS(T1d, Te, T3q);
|
Chris@10
|
154 T3n = FMA(T1d, Ti, T3m);
|
Chris@10
|
155 T2w = FNMS(T1z, Te, T2v);
|
Chris@10
|
156 T2s = FMA(T1z, Ti, T2r);
|
Chris@10
|
157 T3Q = FNMS(Tw, Te, T3P);
|
Chris@10
|
158 T3M = FMA(Tw, Ti, T3L);
|
Chris@10
|
159 {
|
Chris@10
|
160 E T1Y, T1S, T2f, T2a;
|
Chris@10
|
161 T1Y = T1R * Ti;
|
Chris@10
|
162 T1S = T1R * Te;
|
Chris@10
|
163 T2f = T29 * Ti;
|
Chris@10
|
164 T2a = T29 * Te;
|
Chris@10
|
165 {
|
Chris@10
|
166 E Tm, Tf, TV, TQ;
|
Chris@10
|
167 Tm = Td * Ti;
|
Chris@10
|
168 Tf = Td * Te;
|
Chris@10
|
169 TV = TP * Ti;
|
Chris@10
|
170 TQ = TP * Te;
|
Chris@10
|
171 T1Z = FNMS(T1U, Te, T1Y);
|
Chris@10
|
172 T1V = FMA(T1U, Ti, T1S);
|
Chris@10
|
173 T2g = FNMS(T2b, Te, T2f);
|
Chris@10
|
174 T2c = FMA(T2b, Ti, T2a);
|
Chris@10
|
175 Tn = FNMS(Th, Te, Tm);
|
Chris@10
|
176 Tj = FMA(Th, Ti, Tf);
|
Chris@10
|
177 TW = FNMS(TR, Te, TV);
|
Chris@10
|
178 TS = FMA(TR, Ti, TQ);
|
Chris@10
|
179 T8G = ci[0];
|
Chris@10
|
180 }
|
Chris@10
|
181 }
|
Chris@10
|
182 }
|
Chris@10
|
183 }
|
Chris@10
|
184 }
|
Chris@10
|
185 }
|
Chris@10
|
186 }
|
Chris@10
|
187 Tk = cr[WS(rs, 16)];
|
Chris@10
|
188 To = ci[WS(rs, 16)];
|
Chris@10
|
189 {
|
Chris@10
|
190 E Tt, Tx, Tu, T47, TB, TF, TC, T49;
|
Chris@10
|
191 {
|
Chris@10
|
192 E Tl, T8E, Tp, T8F;
|
Chris@10
|
193 Tt = cr[WS(rs, 8)];
|
Chris@10
|
194 Tx = ci[WS(rs, 8)];
|
Chris@10
|
195 Tl = Tj * Tk;
|
Chris@10
|
196 T8E = Tj * To;
|
Chris@10
|
197 Tu = Ts * Tt;
|
Chris@10
|
198 T47 = Ts * Tx;
|
Chris@10
|
199 Tp = FMA(Tn, To, Tl);
|
Chris@10
|
200 T8F = FNMS(Tn, Tk, T8E);
|
Chris@10
|
201 TB = cr[WS(rs, 24)];
|
Chris@10
|
202 TF = ci[WS(rs, 24)];
|
Chris@10
|
203 Tq = T1 + Tp;
|
Chris@10
|
204 T46 = T1 - Tp;
|
Chris@10
|
205 T8H = T8F + T8G;
|
Chris@10
|
206 T98 = T8G - T8F;
|
Chris@10
|
207 TC = TA * TB;
|
Chris@10
|
208 T49 = TA * TF;
|
Chris@10
|
209 }
|
Chris@10
|
210 Ty = FMA(Tw, Tx, Tu);
|
Chris@10
|
211 T48 = FNMS(Tw, Tt, T47);
|
Chris@10
|
212 TG = FMA(TE, TF, TC);
|
Chris@10
|
213 T4a = FNMS(TE, TB, T49);
|
Chris@10
|
214 }
|
Chris@10
|
215 }
|
Chris@10
|
216 {
|
Chris@10
|
217 E TT, TX, TO, T4f, TU, T4g;
|
Chris@10
|
218 {
|
Chris@10
|
219 E TK, TN, TL, T4e;
|
Chris@10
|
220 TK = cr[WS(rs, 4)];
|
Chris@10
|
221 TN = ci[WS(rs, 4)];
|
Chris@10
|
222 TH = Ty + TG;
|
Chris@10
|
223 T97 = Ty - TG;
|
Chris@10
|
224 T4b = T48 - T4a;
|
Chris@10
|
225 T8D = T48 + T4a;
|
Chris@10
|
226 TL = TJ * TK;
|
Chris@10
|
227 T4e = TJ * TN;
|
Chris@10
|
228 TT = cr[WS(rs, 20)];
|
Chris@10
|
229 TX = ci[WS(rs, 20)];
|
Chris@10
|
230 TO = FMA(TM, TN, TL);
|
Chris@10
|
231 T4f = FNMS(TM, TK, T4e);
|
Chris@10
|
232 TU = TS * TT;
|
Chris@10
|
233 T4g = TS * TX;
|
Chris@10
|
234 }
|
Chris@10
|
235 {
|
Chris@10
|
236 E T17, T4m, T1a, T1e, T4d, T4i;
|
Chris@10
|
237 {
|
Chris@10
|
238 E T12, T16, TY, T4h, T13, T4l;
|
Chris@10
|
239 T12 = cr[WS(rs, 28)];
|
Chris@10
|
240 T16 = ci[WS(rs, 28)];
|
Chris@10
|
241 TY = FMA(TW, TX, TU);
|
Chris@10
|
242 T4h = FNMS(TW, TT, T4g);
|
Chris@10
|
243 T13 = T11 * T12;
|
Chris@10
|
244 T4l = T11 * T16;
|
Chris@10
|
245 TZ = TO + TY;
|
Chris@10
|
246 T4d = TO - TY;
|
Chris@10
|
247 T7g = T4f + T4h;
|
Chris@10
|
248 T4i = T4f - T4h;
|
Chris@10
|
249 T17 = FMA(T15, T16, T13);
|
Chris@10
|
250 T4m = FNMS(T15, T12, T4l);
|
Chris@10
|
251 }
|
Chris@10
|
252 T4j = T4d - T4i;
|
Chris@10
|
253 T6t = T4d + T4i;
|
Chris@10
|
254 T1a = cr[WS(rs, 12)];
|
Chris@10
|
255 T1e = ci[WS(rs, 12)];
|
Chris@10
|
256 {
|
Chris@10
|
257 E T1m, T4u, T1H, T4E, T1x, T1A, T1u, T4w, T1y, T4B;
|
Chris@10
|
258 {
|
Chris@10
|
259 E T1D, T1G, T1E, T4D;
|
Chris@10
|
260 {
|
Chris@10
|
261 E T1f, T4o, T4k, T4p;
|
Chris@10
|
262 {
|
Chris@10
|
263 E T1j, T1l, T1b, T4n, T1k, T4t;
|
Chris@10
|
264 T1j = cr[WS(rs, 2)];
|
Chris@10
|
265 T1l = ci[WS(rs, 2)];
|
Chris@10
|
266 T1b = T19 * T1a;
|
Chris@10
|
267 T4n = T19 * T1e;
|
Chris@10
|
268 T1k = T7 * T1j;
|
Chris@10
|
269 T4t = T7 * T1l;
|
Chris@10
|
270 T1f = FMA(T1d, T1e, T1b);
|
Chris@10
|
271 T4o = FNMS(T1d, T1a, T4n);
|
Chris@10
|
272 T1m = FMA(Tb, T1l, T1k);
|
Chris@10
|
273 T4u = FNMS(Tb, T1j, T4t);
|
Chris@10
|
274 }
|
Chris@10
|
275 T1g = T17 + T1f;
|
Chris@10
|
276 T4k = T17 - T1f;
|
Chris@10
|
277 T7f = T4m + T4o;
|
Chris@10
|
278 T4p = T4m - T4o;
|
Chris@10
|
279 T1D = cr[WS(rs, 26)];
|
Chris@10
|
280 T1G = ci[WS(rs, 26)];
|
Chris@10
|
281 T4q = T4k + T4p;
|
Chris@10
|
282 T6u = T4k - T4p;
|
Chris@10
|
283 T1E = T1C * T1D;
|
Chris@10
|
284 T4D = T1C * T1G;
|
Chris@10
|
285 }
|
Chris@10
|
286 {
|
Chris@10
|
287 E T1p, T1t, T1q, T4v;
|
Chris@10
|
288 T1p = cr[WS(rs, 18)];
|
Chris@10
|
289 T1t = ci[WS(rs, 18)];
|
Chris@10
|
290 T1H = FMA(T1F, T1G, T1E);
|
Chris@10
|
291 T4E = FNMS(T1F, T1D, T4D);
|
Chris@10
|
292 T1q = T1o * T1p;
|
Chris@10
|
293 T4v = T1o * T1t;
|
Chris@10
|
294 T1x = cr[WS(rs, 10)];
|
Chris@10
|
295 T1A = ci[WS(rs, 10)];
|
Chris@10
|
296 T1u = FMA(T1s, T1t, T1q);
|
Chris@10
|
297 T4w = FNMS(T1s, T1p, T4v);
|
Chris@10
|
298 T1y = T1w * T1x;
|
Chris@10
|
299 T4B = T1w * T1A;
|
Chris@10
|
300 }
|
Chris@10
|
301 }
|
Chris@10
|
302 {
|
Chris@10
|
303 E T4A, T1v, T7k, T4x, T1B, T4C;
|
Chris@10
|
304 T4A = T1m - T1u;
|
Chris@10
|
305 T1v = T1m + T1u;
|
Chris@10
|
306 T7k = T4u + T4w;
|
Chris@10
|
307 T4x = T4u - T4w;
|
Chris@10
|
308 T1B = FMA(T1z, T1A, T1y);
|
Chris@10
|
309 T4C = FNMS(T1z, T1x, T4B);
|
Chris@10
|
310 {
|
Chris@10
|
311 E T1I, T4y, T4F, T7l;
|
Chris@10
|
312 T1I = T1B + T1H;
|
Chris@10
|
313 T4y = T1B - T1H;
|
Chris@10
|
314 T4F = T4C - T4E;
|
Chris@10
|
315 T7l = T4C + T4E;
|
Chris@10
|
316 T4z = T4x + T4y;
|
Chris@10
|
317 T6y = T4x - T4y;
|
Chris@10
|
318 T1J = T1v + T1I;
|
Chris@10
|
319 T7j = T1v - T1I;
|
Chris@10
|
320 T7m = T7k - T7l;
|
Chris@10
|
321 T8e = T7k + T7l;
|
Chris@10
|
322 T6x = T4A + T4F;
|
Chris@10
|
323 T4G = T4A - T4F;
|
Chris@10
|
324 }
|
Chris@10
|
325 }
|
Chris@10
|
326 }
|
Chris@10
|
327 }
|
Chris@10
|
328 }
|
Chris@10
|
329 }
|
Chris@10
|
330 {
|
Chris@10
|
331 E T5C, T3u, T5y, T7L, T60, T5V, T3F, T5A, T4P, T4U;
|
Chris@10
|
332 {
|
Chris@10
|
333 E T1P, T4J, T2i, T4T, T21, T4L, T28, T4R;
|
Chris@10
|
334 {
|
Chris@10
|
335 E T1L, T1O, T1W, T20;
|
Chris@10
|
336 T1L = cr[WS(rs, 30)];
|
Chris@10
|
337 T1O = ci[WS(rs, 30)];
|
Chris@10
|
338 {
|
Chris@10
|
339 E T2d, T2h, T1M, T4I, T2e, T4S;
|
Chris@10
|
340 T2d = cr[WS(rs, 22)];
|
Chris@10
|
341 T2h = ci[WS(rs, 22)];
|
Chris@10
|
342 T1M = T1K * T1L;
|
Chris@10
|
343 T4I = T1K * T1O;
|
Chris@10
|
344 T2e = T2c * T2d;
|
Chris@10
|
345 T4S = T2c * T2h;
|
Chris@10
|
346 T1P = FMA(T1N, T1O, T1M);
|
Chris@10
|
347 T4J = FNMS(T1N, T1L, T4I);
|
Chris@10
|
348 T2i = FMA(T2g, T2h, T2e);
|
Chris@10
|
349 T4T = FNMS(T2g, T2d, T4S);
|
Chris@10
|
350 }
|
Chris@10
|
351 T1W = cr[WS(rs, 14)];
|
Chris@10
|
352 T20 = ci[WS(rs, 14)];
|
Chris@10
|
353 {
|
Chris@10
|
354 E T24, T27, T1X, T4K, T25, T4Q;
|
Chris@10
|
355 T24 = cr[WS(rs, 6)];
|
Chris@10
|
356 T27 = ci[WS(rs, 6)];
|
Chris@10
|
357 T1X = T1V * T1W;
|
Chris@10
|
358 T4K = T1V * T20;
|
Chris@10
|
359 T25 = T23 * T24;
|
Chris@10
|
360 T4Q = T23 * T27;
|
Chris@10
|
361 T21 = FMA(T1Z, T20, T1X);
|
Chris@10
|
362 T4L = FNMS(T1Z, T1W, T4K);
|
Chris@10
|
363 T28 = FMA(T26, T27, T25);
|
Chris@10
|
364 T4R = FNMS(T26, T24, T4Q);
|
Chris@10
|
365 }
|
Chris@10
|
366 }
|
Chris@10
|
367 {
|
Chris@10
|
368 E T22, T7p, T4M, T4N, T2j, T7q;
|
Chris@10
|
369 T4P = T1P - T21;
|
Chris@10
|
370 T22 = T1P + T21;
|
Chris@10
|
371 T7p = T4J + T4L;
|
Chris@10
|
372 T4M = T4J - T4L;
|
Chris@10
|
373 T4N = T28 - T2i;
|
Chris@10
|
374 T2j = T28 + T2i;
|
Chris@10
|
375 T7q = T4R + T4T;
|
Chris@10
|
376 T4U = T4R - T4T;
|
Chris@10
|
377 T2k = T22 + T2j;
|
Chris@10
|
378 T7o = T22 - T2j;
|
Chris@10
|
379 T7r = T7p - T7q;
|
Chris@10
|
380 T8d = T7p + T7q;
|
Chris@10
|
381 T6B = T4M - T4N;
|
Chris@10
|
382 T4O = T4M + T4N;
|
Chris@10
|
383 }
|
Chris@10
|
384 }
|
Chris@10
|
385 {
|
Chris@10
|
386 E T3l, T5X, T3E, T3v, T3t, T3w, T3x, T5Z, T3A, T3B, T3D, T3y, T5z;
|
Chris@10
|
387 {
|
Chris@10
|
388 E T3g, T3k, T3h, T5W;
|
Chris@10
|
389 T3g = cr[WS(rs, 31)];
|
Chris@10
|
390 T3k = ci[WS(rs, 31)];
|
Chris@10
|
391 T3A = cr[WS(rs, 23)];
|
Chris@10
|
392 T6A = T4P + T4U;
|
Chris@10
|
393 T4V = T4P - T4U;
|
Chris@10
|
394 T3h = T3f * T3g;
|
Chris@10
|
395 T5W = T3f * T3k;
|
Chris@10
|
396 T3B = T3z * T3A;
|
Chris@10
|
397 T3D = ci[WS(rs, 23)];
|
Chris@10
|
398 T3l = FMA(T3j, T3k, T3h);
|
Chris@10
|
399 T5X = FNMS(T3j, T3g, T5W);
|
Chris@10
|
400 }
|
Chris@10
|
401 {
|
Chris@10
|
402 E T3o, T5B, T3s, T3p, T5Y;
|
Chris@10
|
403 T3o = cr[WS(rs, 15)];
|
Chris@10
|
404 T3E = FMA(T3C, T3D, T3B);
|
Chris@10
|
405 T5B = T3z * T3D;
|
Chris@10
|
406 T3s = ci[WS(rs, 15)];
|
Chris@10
|
407 T3p = T3n * T3o;
|
Chris@10
|
408 T3v = cr[WS(rs, 7)];
|
Chris@10
|
409 T5C = FNMS(T3C, T3A, T5B);
|
Chris@10
|
410 T5Y = T3n * T3s;
|
Chris@10
|
411 T3t = FMA(T3r, T3s, T3p);
|
Chris@10
|
412 T3w = TP * T3v;
|
Chris@10
|
413 T3x = ci[WS(rs, 7)];
|
Chris@10
|
414 T5Z = FNMS(T3r, T3o, T5Y);
|
Chris@10
|
415 }
|
Chris@10
|
416 T3u = T3l + T3t;
|
Chris@10
|
417 T5y = T3l - T3t;
|
Chris@10
|
418 T3y = FMA(TR, T3x, T3w);
|
Chris@10
|
419 T5z = TP * T3x;
|
Chris@10
|
420 T7L = T5X + T5Z;
|
Chris@10
|
421 T60 = T5X - T5Z;
|
Chris@10
|
422 T5V = T3E - T3y;
|
Chris@10
|
423 T3F = T3y + T3E;
|
Chris@10
|
424 T5A = FNMS(TR, T3v, T5z);
|
Chris@10
|
425 }
|
Chris@10
|
426 {
|
Chris@10
|
427 E T2L, T53, T4Z, T2z, T7A, T5q, T2D, T51;
|
Chris@10
|
428 {
|
Chris@10
|
429 E T2q, T5n, T2y, T2A, T2C, T5p, T2B, T50;
|
Chris@10
|
430 {
|
Chris@10
|
431 E T2G, T2K, T2n, T5m, T2t, T5o;
|
Chris@10
|
432 {
|
Chris@10
|
433 E T2o, T2p, T5D, T7M;
|
Chris@10
|
434 T2n = cr[WS(rs, 1)];
|
Chris@10
|
435 T6P = T60 + T5V;
|
Chris@10
|
436 T61 = T5V - T60;
|
Chris@10
|
437 T7G = T3u - T3F;
|
Chris@10
|
438 T3G = T3u + T3F;
|
Chris@10
|
439 T5D = T5A - T5C;
|
Chris@10
|
440 T7M = T5A + T5C;
|
Chris@10
|
441 T2o = T2 * T2n;
|
Chris@10
|
442 T2p = ci[WS(rs, 1)];
|
Chris@10
|
443 T6M = T5y + T5D;
|
Chris@10
|
444 T5E = T5y - T5D;
|
Chris@10
|
445 T8n = T7L + T7M;
|
Chris@10
|
446 T7N = T7L - T7M;
|
Chris@10
|
447 T5m = T2 * T2p;
|
Chris@10
|
448 T2q = FMA(T5, T2p, T2o);
|
Chris@10
|
449 }
|
Chris@10
|
450 T2G = cr[WS(rs, 25)];
|
Chris@10
|
451 T2K = ci[WS(rs, 25)];
|
Chris@10
|
452 T5n = FNMS(T5, T2n, T5m);
|
Chris@10
|
453 {
|
Chris@10
|
454 E T2x, T2u, T2H, T52;
|
Chris@10
|
455 T2t = cr[WS(rs, 17)];
|
Chris@10
|
456 T2H = T2F * T2G;
|
Chris@10
|
457 T52 = T2F * T2K;
|
Chris@10
|
458 T2x = ci[WS(rs, 17)];
|
Chris@10
|
459 T2u = T2s * T2t;
|
Chris@10
|
460 T2L = FMA(T2J, T2K, T2H);
|
Chris@10
|
461 T53 = FNMS(T2J, T2G, T52);
|
Chris@10
|
462 T5o = T2s * T2x;
|
Chris@10
|
463 T2y = FMA(T2w, T2x, T2u);
|
Chris@10
|
464 }
|
Chris@10
|
465 T2A = cr[WS(rs, 9)];
|
Chris@10
|
466 T2C = ci[WS(rs, 9)];
|
Chris@10
|
467 T5p = FNMS(T2w, T2t, T5o);
|
Chris@10
|
468 }
|
Chris@10
|
469 T4Z = T2q - T2y;
|
Chris@10
|
470 T2z = T2q + T2y;
|
Chris@10
|
471 T2B = T8 * T2A;
|
Chris@10
|
472 T50 = T8 * T2C;
|
Chris@10
|
473 T7A = T5n + T5p;
|
Chris@10
|
474 T5q = T5n - T5p;
|
Chris@10
|
475 T2D = FMA(Tc, T2C, T2B);
|
Chris@10
|
476 T51 = FNMS(Tc, T2A, T50);
|
Chris@10
|
477 }
|
Chris@10
|
478 {
|
Chris@10
|
479 E T3N, T3K, T3O, T5H, T41, T5Q, T3R, T3U, T3W;
|
Chris@10
|
480 {
|
Chris@10
|
481 E T3H, T3I, T3J, T3Y, T40, T5G, T3Z, T5P;
|
Chris@10
|
482 T3H = cr[WS(rs, 3)];
|
Chris@10
|
483 {
|
Chris@10
|
484 E T5r, T2M, T54, T7B;
|
Chris@10
|
485 T5r = T2D - T2L;
|
Chris@10
|
486 T2M = T2D + T2L;
|
Chris@10
|
487 T54 = T51 - T53;
|
Chris@10
|
488 T7B = T51 + T53;
|
Chris@10
|
489 T6I = T5q - T5r;
|
Chris@10
|
490 T5s = T5q + T5r;
|
Chris@10
|
491 T7v = T2z - T2M;
|
Chris@10
|
492 T2N = T2z + T2M;
|
Chris@10
|
493 T6F = T4Z + T54;
|
Chris@10
|
494 T55 = T4Z - T54;
|
Chris@10
|
495 T8i = T7A + T7B;
|
Chris@10
|
496 T7C = T7A - T7B;
|
Chris@10
|
497 T3I = T3 * T3H;
|
Chris@10
|
498 }
|
Chris@10
|
499 T3J = ci[WS(rs, 3)];
|
Chris@10
|
500 T3Y = cr[WS(rs, 11)];
|
Chris@10
|
501 T40 = ci[WS(rs, 11)];
|
Chris@10
|
502 T3N = cr[WS(rs, 19)];
|
Chris@10
|
503 T3K = FMA(T6, T3J, T3I);
|
Chris@10
|
504 T5G = T3 * T3J;
|
Chris@10
|
505 T3Z = Td * T3Y;
|
Chris@10
|
506 T5P = Td * T40;
|
Chris@10
|
507 T3O = T3M * T3N;
|
Chris@10
|
508 T5H = FNMS(T6, T3H, T5G);
|
Chris@10
|
509 T41 = FMA(Th, T40, T3Z);
|
Chris@10
|
510 T5Q = FNMS(Th, T3Y, T5P);
|
Chris@10
|
511 T3R = ci[WS(rs, 19)];
|
Chris@10
|
512 T3U = cr[WS(rs, 27)];
|
Chris@10
|
513 T3W = ci[WS(rs, 27)];
|
Chris@10
|
514 }
|
Chris@10
|
515 {
|
Chris@10
|
516 E T2O, T2P, T2Q, T37, T39, T57, T38, T5g;
|
Chris@10
|
517 {
|
Chris@10
|
518 E T3T, T5F, T5J, T3X, T5O, T7I, T5K;
|
Chris@10
|
519 T2O = cr[WS(rs, 5)];
|
Chris@10
|
520 {
|
Chris@10
|
521 E T3S, T5I, T3V, T5N;
|
Chris@10
|
522 T3S = FMA(T3Q, T3R, T3O);
|
Chris@10
|
523 T5I = T3M * T3R;
|
Chris@10
|
524 T3V = Te * T3U;
|
Chris@10
|
525 T5N = Te * T3W;
|
Chris@10
|
526 T3T = T3K + T3S;
|
Chris@10
|
527 T5F = T3K - T3S;
|
Chris@10
|
528 T5J = FNMS(T3Q, T3N, T5I);
|
Chris@10
|
529 T3X = FMA(Ti, T3W, T3V);
|
Chris@10
|
530 T5O = FNMS(Ti, T3U, T5N);
|
Chris@10
|
531 T2P = T29 * T2O;
|
Chris@10
|
532 }
|
Chris@10
|
533 T7I = T5H + T5J;
|
Chris@10
|
534 T5K = T5H - T5J;
|
Chris@10
|
535 {
|
Chris@10
|
536 E T42, T5M, T7H, T5R;
|
Chris@10
|
537 T42 = T3X + T41;
|
Chris@10
|
538 T5M = T3X - T41;
|
Chris@10
|
539 T7H = T5O + T5Q;
|
Chris@10
|
540 T5R = T5O - T5Q;
|
Chris@10
|
541 T5L = T5F - T5K;
|
Chris@10
|
542 T63 = T5F + T5K;
|
Chris@10
|
543 T43 = T3T + T42;
|
Chris@10
|
544 T7O = T42 - T3T;
|
Chris@10
|
545 T5S = T5M + T5R;
|
Chris@10
|
546 T62 = T5M - T5R;
|
Chris@10
|
547 T7J = T7H - T7I;
|
Chris@10
|
548 T8o = T7I + T7H;
|
Chris@10
|
549 T2Q = ci[WS(rs, 5)];
|
Chris@10
|
550 }
|
Chris@10
|
551 }
|
Chris@10
|
552 T37 = cr[WS(rs, 13)];
|
Chris@10
|
553 T39 = ci[WS(rs, 13)];
|
Chris@10
|
554 T2U = cr[WS(rs, 21)];
|
Chris@10
|
555 T2R = FMA(T2b, T2Q, T2P);
|
Chris@10
|
556 T57 = T29 * T2Q;
|
Chris@10
|
557 T38 = T1R * T37;
|
Chris@10
|
558 T5g = T1R * T39;
|
Chris@10
|
559 T2V = T2T * T2U;
|
Chris@10
|
560 T58 = FNMS(T2b, T2O, T57);
|
Chris@10
|
561 T3a = FMA(T1U, T39, T38);
|
Chris@10
|
562 T5h = FNMS(T1U, T37, T5g);
|
Chris@10
|
563 T2Y = ci[WS(rs, 21)];
|
Chris@10
|
564 T32 = cr[WS(rs, 29)];
|
Chris@10
|
565 T35 = ci[WS(rs, 29)];
|
Chris@10
|
566 }
|
Chris@10
|
567 }
|
Chris@10
|
568 }
|
Chris@10
|
569 }
|
Chris@10
|
570 }
|
Chris@10
|
571 {
|
Chris@10
|
572 E T7e, T8T, T7D, T7y, T7h, T8U, T6s, T9o, T9n, T6v, T6Q, T6N, T6J, T6G, T6o;
|
Chris@10
|
573 E T6r;
|
Chris@10
|
574 {
|
Chris@10
|
575 E T5c, T5t, T5j, T5u, T8s, T8v;
|
Chris@10
|
576 {
|
Chris@10
|
577 E T8c, T1i, T8A, T8z, T8O, T8J, T8N, T2l, T8L, T45, T8t, T8l, T8u, T8q, T3c;
|
Chris@10
|
578 E T8p, T8k, T8w, T2m;
|
Chris@10
|
579 {
|
Chris@10
|
580 E T8x, T8y, T8j, T8C, T8I;
|
Chris@10
|
581 {
|
Chris@10
|
582 E TI, T30, T56, T5a, T36, T5f, T1h, T7x, T5b;
|
Chris@10
|
583 TI = Tq + TH;
|
Chris@10
|
584 T7e = Tq - TH;
|
Chris@10
|
585 {
|
Chris@10
|
586 E T2Z, T59, T33, T5e;
|
Chris@10
|
587 T2Z = FMA(T2X, T2Y, T2V);
|
Chris@10
|
588 T59 = T2T * T2Y;
|
Chris@10
|
589 T33 = T31 * T32;
|
Chris@10
|
590 T5e = T31 * T35;
|
Chris@10
|
591 T30 = T2R + T2Z;
|
Chris@10
|
592 T56 = T2R - T2Z;
|
Chris@10
|
593 T5a = FNMS(T2X, T2U, T59);
|
Chris@10
|
594 T36 = FMA(T34, T35, T33);
|
Chris@10
|
595 T5f = FNMS(T34, T32, T5e);
|
Chris@10
|
596 T1h = TZ + T1g;
|
Chris@10
|
597 T8T = TZ - T1g;
|
Chris@10
|
598 }
|
Chris@10
|
599 T7x = T58 + T5a;
|
Chris@10
|
600 T5b = T58 - T5a;
|
Chris@10
|
601 {
|
Chris@10
|
602 E T3b, T5d, T7w, T5i;
|
Chris@10
|
603 T3b = T36 + T3a;
|
Chris@10
|
604 T5d = T36 - T3a;
|
Chris@10
|
605 T7w = T5f + T5h;
|
Chris@10
|
606 T5i = T5f - T5h;
|
Chris@10
|
607 T5c = T56 - T5b;
|
Chris@10
|
608 T5t = T56 + T5b;
|
Chris@10
|
609 T3c = T30 + T3b;
|
Chris@10
|
610 T7D = T30 - T3b;
|
Chris@10
|
611 T5j = T5d + T5i;
|
Chris@10
|
612 T5u = T5i - T5d;
|
Chris@10
|
613 T7y = T7w - T7x;
|
Chris@10
|
614 T8j = T7x + T7w;
|
Chris@10
|
615 T8c = TI - T1h;
|
Chris@10
|
616 T1i = TI + T1h;
|
Chris@10
|
617 }
|
Chris@10
|
618 }
|
Chris@10
|
619 T8p = T8n - T8o;
|
Chris@10
|
620 T8x = T8n + T8o;
|
Chris@10
|
621 T8y = T8i + T8j;
|
Chris@10
|
622 T8k = T8i - T8j;
|
Chris@10
|
623 T7h = T7f - T7g;
|
Chris@10
|
624 T8C = T7g + T7f;
|
Chris@10
|
625 T8I = T8D + T8H;
|
Chris@10
|
626 T8U = T8H - T8D;
|
Chris@10
|
627 T8A = T8y + T8x;
|
Chris@10
|
628 T8z = T8x - T8y;
|
Chris@10
|
629 T8O = T8I - T8C;
|
Chris@10
|
630 T8J = T8C + T8I;
|
Chris@10
|
631 }
|
Chris@10
|
632 {
|
Chris@10
|
633 E T8h, T8m, T3d, T44;
|
Chris@10
|
634 T8h = T2N - T3c;
|
Chris@10
|
635 T3d = T2N + T3c;
|
Chris@10
|
636 T44 = T3G + T43;
|
Chris@10
|
637 T8m = T3G - T43;
|
Chris@10
|
638 T8N = T1J - T2k;
|
Chris@10
|
639 T2l = T1J + T2k;
|
Chris@10
|
640 T8L = T44 - T3d;
|
Chris@10
|
641 T45 = T3d + T44;
|
Chris@10
|
642 T8t = T8h - T8k;
|
Chris@10
|
643 T8l = T8h + T8k;
|
Chris@10
|
644 T8u = T8m + T8p;
|
Chris@10
|
645 T8q = T8m - T8p;
|
Chris@10
|
646 }
|
Chris@10
|
647 T8w = T1i - T2l;
|
Chris@10
|
648 T2m = T1i + T2l;
|
Chris@10
|
649 {
|
Chris@10
|
650 E T8Q, T8R, T8P, T8S;
|
Chris@10
|
651 {
|
Chris@10
|
652 E T8r, T8M, T8K, T8g, T8B, T8f;
|
Chris@10
|
653 T8Q = T8q - T8l;
|
Chris@10
|
654 T8r = T8l + T8q;
|
Chris@10
|
655 T8B = T8e + T8d;
|
Chris@10
|
656 T8f = T8d - T8e;
|
Chris@10
|
657 cr[0] = T2m + T45;
|
Chris@10
|
658 ci[WS(rs, 15)] = T2m - T45;
|
Chris@10
|
659 ci[WS(rs, 7)] = T8w + T8z;
|
Chris@10
|
660 cr[WS(rs, 8)] = T8w - T8z;
|
Chris@10
|
661 T8M = T8J - T8B;
|
Chris@10
|
662 T8K = T8B + T8J;
|
Chris@10
|
663 T8g = T8c - T8f;
|
Chris@10
|
664 T8s = T8c + T8f;
|
Chris@10
|
665 T8R = T8O - T8N;
|
Chris@10
|
666 T8P = T8N + T8O;
|
Chris@10
|
667 ci[WS(rs, 23)] = T8L + T8M;
|
Chris@10
|
668 cr[WS(rs, 24)] = T8L - T8M;
|
Chris@10
|
669 ci[WS(rs, 31)] = T8A + T8K;
|
Chris@10
|
670 cr[WS(rs, 16)] = T8A - T8K;
|
Chris@10
|
671 cr[WS(rs, 4)] = FMA(KP707106781, T8r, T8g);
|
Chris@10
|
672 ci[WS(rs, 11)] = FNMS(KP707106781, T8r, T8g);
|
Chris@10
|
673 }
|
Chris@10
|
674 T8S = T8u - T8t;
|
Chris@10
|
675 T8v = T8t + T8u;
|
Chris@10
|
676 ci[WS(rs, 19)] = FMA(KP707106781, T8Q, T8P);
|
Chris@10
|
677 cr[WS(rs, 28)] = FMS(KP707106781, T8Q, T8P);
|
Chris@10
|
678 ci[WS(rs, 27)] = FMA(KP707106781, T8S, T8R);
|
Chris@10
|
679 cr[WS(rs, 20)] = FMS(KP707106781, T8S, T8R);
|
Chris@10
|
680 }
|
Chris@10
|
681 }
|
Chris@10
|
682 {
|
Chris@10
|
683 E T6c, T4s, T9c, T4X, T9h, T9b, T9i, T6f, T5l, T6h, T6m, T6q, T6a, T66, T5v;
|
Chris@10
|
684 {
|
Chris@10
|
685 E T6d, T4H, T4W, T6e, T99, T9a, T4c, T4r, T5T, T64;
|
Chris@10
|
686 T6s = T46 + T4b;
|
Chris@10
|
687 T4c = T46 - T4b;
|
Chris@10
|
688 T4r = T4j + T4q;
|
Chris@10
|
689 T9o = T4q - T4j;
|
Chris@10
|
690 T6d = FNMS(KP414213562, T4z, T4G);
|
Chris@10
|
691 T4H = FMA(KP414213562, T4G, T4z);
|
Chris@10
|
692 ci[WS(rs, 3)] = FMA(KP707106781, T8v, T8s);
|
Chris@10
|
693 cr[WS(rs, 12)] = FNMS(KP707106781, T8v, T8s);
|
Chris@10
|
694 T6c = FMA(KP707106781, T4r, T4c);
|
Chris@10
|
695 T4s = FNMS(KP707106781, T4r, T4c);
|
Chris@10
|
696 T4W = FNMS(KP414213562, T4V, T4O);
|
Chris@10
|
697 T6e = FMA(KP414213562, T4O, T4V);
|
Chris@10
|
698 T9n = T98 - T97;
|
Chris@10
|
699 T99 = T97 + T98;
|
Chris@10
|
700 T9a = T6t - T6u;
|
Chris@10
|
701 T6v = T6t + T6u;
|
Chris@10
|
702 T9c = T4H + T4W;
|
Chris@10
|
703 T4X = T4H - T4W;
|
Chris@10
|
704 T9h = FNMS(KP707106781, T9a, T99);
|
Chris@10
|
705 T9b = FMA(KP707106781, T9a, T99);
|
Chris@10
|
706 T6Q = T5S - T5L;
|
Chris@10
|
707 T5T = T5L + T5S;
|
Chris@10
|
708 T64 = T62 - T63;
|
Chris@10
|
709 T6N = T63 + T62;
|
Chris@10
|
710 {
|
Chris@10
|
711 E T6k, T5U, T6l, T65, T5k;
|
Chris@10
|
712 T6J = T5j - T5c;
|
Chris@10
|
713 T5k = T5c + T5j;
|
Chris@10
|
714 T9i = T6e - T6d;
|
Chris@10
|
715 T6f = T6d + T6e;
|
Chris@10
|
716 T6k = FMA(KP707106781, T5T, T5E);
|
Chris@10
|
717 T5U = FNMS(KP707106781, T5T, T5E);
|
Chris@10
|
718 T6l = FMA(KP707106781, T64, T61);
|
Chris@10
|
719 T65 = FNMS(KP707106781, T64, T61);
|
Chris@10
|
720 T5l = FNMS(KP707106781, T5k, T55);
|
Chris@10
|
721 T6h = FMA(KP707106781, T5k, T55);
|
Chris@10
|
722 T6m = FNMS(KP198912367, T6l, T6k);
|
Chris@10
|
723 T6q = FMA(KP198912367, T6k, T6l);
|
Chris@10
|
724 T6a = FNMS(KP668178637, T5U, T65);
|
Chris@10
|
725 T66 = FMA(KP668178637, T65, T5U);
|
Chris@10
|
726 T5v = T5t + T5u;
|
Chris@10
|
727 T6G = T5t - T5u;
|
Chris@10
|
728 }
|
Chris@10
|
729 }
|
Chris@10
|
730 {
|
Chris@10
|
731 E T68, T4Y, T9j, T9l, T6i, T5w;
|
Chris@10
|
732 T68 = FNMS(KP923879532, T4X, T4s);
|
Chris@10
|
733 T4Y = FMA(KP923879532, T4X, T4s);
|
Chris@10
|
734 T9j = FMA(KP923879532, T9i, T9h);
|
Chris@10
|
735 T9l = FNMS(KP923879532, T9i, T9h);
|
Chris@10
|
736 T6i = FMA(KP707106781, T5v, T5s);
|
Chris@10
|
737 T5w = FNMS(KP707106781, T5v, T5s);
|
Chris@10
|
738 {
|
Chris@10
|
739 E T9g, T9f, T9d, T9e;
|
Chris@10
|
740 {
|
Chris@10
|
741 E T6g, T6p, T69, T5x, T6n, T6j;
|
Chris@10
|
742 T6o = FNMS(KP923879532, T6f, T6c);
|
Chris@10
|
743 T6g = FMA(KP923879532, T6f, T6c);
|
Chris@10
|
744 T6j = FNMS(KP198912367, T6i, T6h);
|
Chris@10
|
745 T6p = FMA(KP198912367, T6h, T6i);
|
Chris@10
|
746 T69 = FNMS(KP668178637, T5l, T5w);
|
Chris@10
|
747 T5x = FMA(KP668178637, T5w, T5l);
|
Chris@10
|
748 T6n = T6j + T6m;
|
Chris@10
|
749 T9g = T6m - T6j;
|
Chris@10
|
750 T9f = FNMS(KP923879532, T9c, T9b);
|
Chris@10
|
751 T9d = FMA(KP923879532, T9c, T9b);
|
Chris@10
|
752 {
|
Chris@10
|
753 E T6b, T9k, T9m, T67;
|
Chris@10
|
754 T6b = T69 + T6a;
|
Chris@10
|
755 T9k = T69 - T6a;
|
Chris@10
|
756 T9m = T66 - T5x;
|
Chris@10
|
757 T67 = T5x + T66;
|
Chris@10
|
758 ci[0] = FMA(KP980785280, T6n, T6g);
|
Chris@10
|
759 cr[WS(rs, 15)] = FNMS(KP980785280, T6n, T6g);
|
Chris@10
|
760 ci[WS(rs, 4)] = FNMS(KP831469612, T6b, T68);
|
Chris@10
|
761 cr[WS(rs, 11)] = FMA(KP831469612, T6b, T68);
|
Chris@10
|
762 ci[WS(rs, 28)] = FMA(KP831469612, T9k, T9j);
|
Chris@10
|
763 cr[WS(rs, 19)] = FMS(KP831469612, T9k, T9j);
|
Chris@10
|
764 ci[WS(rs, 20)] = FMA(KP831469612, T9m, T9l);
|
Chris@10
|
765 cr[WS(rs, 27)] = FMS(KP831469612, T9m, T9l);
|
Chris@10
|
766 cr[WS(rs, 3)] = FMA(KP831469612, T67, T4Y);
|
Chris@10
|
767 ci[WS(rs, 12)] = FNMS(KP831469612, T67, T4Y);
|
Chris@10
|
768 T9e = T6q - T6p;
|
Chris@10
|
769 T6r = T6p + T6q;
|
Chris@10
|
770 }
|
Chris@10
|
771 }
|
Chris@10
|
772 ci[WS(rs, 16)] = FMA(KP980785280, T9e, T9d);
|
Chris@10
|
773 cr[WS(rs, 31)] = FMS(KP980785280, T9e, T9d);
|
Chris@10
|
774 ci[WS(rs, 24)] = FMA(KP980785280, T9g, T9f);
|
Chris@10
|
775 cr[WS(rs, 23)] = FMS(KP980785280, T9g, T9f);
|
Chris@10
|
776 }
|
Chris@10
|
777 }
|
Chris@10
|
778 }
|
Chris@10
|
779 }
|
Chris@10
|
780 {
|
Chris@10
|
781 E T88, T90, T8Z, T8b;
|
Chris@10
|
782 {
|
Chris@10
|
783 E T7K, T7W, T7i, T7P, T8a, T86, T91, T8V, T8W, T7t, T7U, T7F, T92, T7Z, T89;
|
Chris@10
|
784 E T83;
|
Chris@10
|
785 {
|
Chris@10
|
786 E T7X, T7n, T7s, T7Y, T84, T85;
|
Chris@10
|
787 T7K = T7G - T7J;
|
Chris@10
|
788 T84 = T7G + T7J;
|
Chris@10
|
789 cr[WS(rs, 7)] = FMA(KP980785280, T6r, T6o);
|
Chris@10
|
790 ci[WS(rs, 8)] = FNMS(KP980785280, T6r, T6o);
|
Chris@10
|
791 T7W = T7e + T7h;
|
Chris@10
|
792 T7i = T7e - T7h;
|
Chris@10
|
793 T85 = T7O - T7N;
|
Chris@10
|
794 T7P = T7N + T7O;
|
Chris@10
|
795 T7X = T7j - T7m;
|
Chris@10
|
796 T7n = T7j + T7m;
|
Chris@10
|
797 T8a = FMA(KP414213562, T84, T85);
|
Chris@10
|
798 T86 = FNMS(KP414213562, T85, T84);
|
Chris@10
|
799 T91 = T8U - T8T;
|
Chris@10
|
800 T8V = T8T + T8U;
|
Chris@10
|
801 T7s = T7o - T7r;
|
Chris@10
|
802 T7Y = T7o + T7r;
|
Chris@10
|
803 {
|
Chris@10
|
804 E T81, T82, T7z, T7E;
|
Chris@10
|
805 T81 = T7v + T7y;
|
Chris@10
|
806 T7z = T7v - T7y;
|
Chris@10
|
807 T7E = T7C - T7D;
|
Chris@10
|
808 T82 = T7C + T7D;
|
Chris@10
|
809 T8W = T7n - T7s;
|
Chris@10
|
810 T7t = T7n + T7s;
|
Chris@10
|
811 T7U = FNMS(KP414213562, T7z, T7E);
|
Chris@10
|
812 T7F = FMA(KP414213562, T7E, T7z);
|
Chris@10
|
813 T92 = T7Y - T7X;
|
Chris@10
|
814 T7Z = T7X + T7Y;
|
Chris@10
|
815 T89 = FMA(KP414213562, T81, T82);
|
Chris@10
|
816 T83 = FNMS(KP414213562, T82, T81);
|
Chris@10
|
817 }
|
Chris@10
|
818 }
|
Chris@10
|
819 {
|
Chris@10
|
820 E T7S, T7u, T93, T95, T7T, T7Q;
|
Chris@10
|
821 T7S = FNMS(KP707106781, T7t, T7i);
|
Chris@10
|
822 T7u = FMA(KP707106781, T7t, T7i);
|
Chris@10
|
823 T93 = FMA(KP707106781, T92, T91);
|
Chris@10
|
824 T95 = FNMS(KP707106781, T92, T91);
|
Chris@10
|
825 T7T = FMA(KP414213562, T7K, T7P);
|
Chris@10
|
826 T7Q = FNMS(KP414213562, T7P, T7K);
|
Chris@10
|
827 {
|
Chris@10
|
828 E T80, T87, T8X, T8Y;
|
Chris@10
|
829 T88 = FNMS(KP707106781, T7Z, T7W);
|
Chris@10
|
830 T80 = FMA(KP707106781, T7Z, T7W);
|
Chris@10
|
831 {
|
Chris@10
|
832 E T7V, T94, T96, T7R;
|
Chris@10
|
833 T7V = T7T - T7U;
|
Chris@10
|
834 T94 = T7U + T7T;
|
Chris@10
|
835 T96 = T7Q - T7F;
|
Chris@10
|
836 T7R = T7F + T7Q;
|
Chris@10
|
837 ci[WS(rs, 5)] = FMA(KP923879532, T7V, T7S);
|
Chris@10
|
838 cr[WS(rs, 10)] = FNMS(KP923879532, T7V, T7S);
|
Chris@10
|
839 ci[WS(rs, 29)] = FMA(KP923879532, T94, T93);
|
Chris@10
|
840 cr[WS(rs, 18)] = FMS(KP923879532, T94, T93);
|
Chris@10
|
841 ci[WS(rs, 21)] = FMA(KP923879532, T96, T95);
|
Chris@10
|
842 cr[WS(rs, 26)] = FMS(KP923879532, T96, T95);
|
Chris@10
|
843 cr[WS(rs, 2)] = FMA(KP923879532, T7R, T7u);
|
Chris@10
|
844 ci[WS(rs, 13)] = FNMS(KP923879532, T7R, T7u);
|
Chris@10
|
845 T87 = T83 + T86;
|
Chris@10
|
846 T90 = T86 - T83;
|
Chris@10
|
847 }
|
Chris@10
|
848 T8Z = FNMS(KP707106781, T8W, T8V);
|
Chris@10
|
849 T8X = FMA(KP707106781, T8W, T8V);
|
Chris@10
|
850 T8Y = T8a - T89;
|
Chris@10
|
851 T8b = T89 + T8a;
|
Chris@10
|
852 ci[WS(rs, 1)] = FMA(KP923879532, T87, T80);
|
Chris@10
|
853 cr[WS(rs, 14)] = FNMS(KP923879532, T87, T80);
|
Chris@10
|
854 ci[WS(rs, 17)] = FMA(KP923879532, T8Y, T8X);
|
Chris@10
|
855 cr[WS(rs, 30)] = FMS(KP923879532, T8Y, T8X);
|
Chris@10
|
856 }
|
Chris@10
|
857 }
|
Chris@10
|
858 }
|
Chris@10
|
859 {
|
Chris@10
|
860 E T6Y, T6w, T9w, T6D, T9v, T9p, T9q, T71, T6O, T76;
|
Chris@10
|
861 {
|
Chris@10
|
862 E T70, T6Z, T6z, T6C;
|
Chris@10
|
863 ci[WS(rs, 25)] = FMA(KP923879532, T90, T8Z);
|
Chris@10
|
864 cr[WS(rs, 22)] = FMS(KP923879532, T90, T8Z);
|
Chris@10
|
865 cr[WS(rs, 6)] = FMA(KP923879532, T8b, T88);
|
Chris@10
|
866 ci[WS(rs, 9)] = FNMS(KP923879532, T8b, T88);
|
Chris@10
|
867 T70 = FNMS(KP414213562, T6x, T6y);
|
Chris@10
|
868 T6z = FMA(KP414213562, T6y, T6x);
|
Chris@10
|
869 T6C = FNMS(KP414213562, T6B, T6A);
|
Chris@10
|
870 T6Z = FMA(KP414213562, T6A, T6B);
|
Chris@10
|
871 T6Y = FNMS(KP707106781, T6v, T6s);
|
Chris@10
|
872 T6w = FMA(KP707106781, T6v, T6s);
|
Chris@10
|
873 T9w = T6z - T6C;
|
Chris@10
|
874 T6D = T6z + T6C;
|
Chris@10
|
875 T9v = FNMS(KP707106781, T9o, T9n);
|
Chris@10
|
876 T9p = FMA(KP707106781, T9o, T9n);
|
Chris@10
|
877 T9q = T70 + T6Z;
|
Chris@10
|
878 T71 = T6Z - T70;
|
Chris@10
|
879 T6O = FMA(KP707106781, T6N, T6M);
|
Chris@10
|
880 T76 = FNMS(KP707106781, T6N, T6M);
|
Chris@10
|
881 }
|
Chris@10
|
882 {
|
Chris@10
|
883 E T6U, T9u, T79, T6X, T9s, T9t, T9r, T72;
|
Chris@10
|
884 {
|
Chris@10
|
885 E T6E, T78, T6V, T6S, T75, T6W, T6L, T9x, T9z, T9y, T6T, T9A;
|
Chris@10
|
886 {
|
Chris@10
|
887 E T7c, T7b, T77, T6R;
|
Chris@10
|
888 T6U = FNMS(KP923879532, T6D, T6w);
|
Chris@10
|
889 T6E = FMA(KP923879532, T6D, T6w);
|
Chris@10
|
890 T77 = FNMS(KP707106781, T6Q, T6P);
|
Chris@10
|
891 T6R = FMA(KP707106781, T6Q, T6P);
|
Chris@10
|
892 {
|
Chris@10
|
893 E T73, T6H, T74, T6K;
|
Chris@10
|
894 T73 = FNMS(KP707106781, T6G, T6F);
|
Chris@10
|
895 T6H = FMA(KP707106781, T6G, T6F);
|
Chris@10
|
896 T74 = FNMS(KP707106781, T6J, T6I);
|
Chris@10
|
897 T6K = FMA(KP707106781, T6J, T6I);
|
Chris@10
|
898 T78 = FMA(KP668178637, T77, T76);
|
Chris@10
|
899 T7c = FNMS(KP668178637, T76, T77);
|
Chris@10
|
900 T6V = FMA(KP198912367, T6O, T6R);
|
Chris@10
|
901 T6S = FNMS(KP198912367, T6R, T6O);
|
Chris@10
|
902 T75 = FNMS(KP668178637, T74, T73);
|
Chris@10
|
903 T7b = FMA(KP668178637, T73, T74);
|
Chris@10
|
904 T6W = FNMS(KP198912367, T6H, T6K);
|
Chris@10
|
905 T6L = FMA(KP198912367, T6K, T6H);
|
Chris@10
|
906 }
|
Chris@10
|
907 T9x = FMA(KP923879532, T9w, T9v);
|
Chris@10
|
908 T9z = FNMS(KP923879532, T9w, T9v);
|
Chris@10
|
909 T7d = T7b - T7c;
|
Chris@10
|
910 T9y = T7b + T7c;
|
Chris@10
|
911 }
|
Chris@10
|
912 T9u = T6S - T6L;
|
Chris@10
|
913 T6T = T6L + T6S;
|
Chris@10
|
914 T9A = T78 - T75;
|
Chris@10
|
915 T79 = T75 + T78;
|
Chris@10
|
916 ci[WS(rs, 18)] = FNMS(KP831469612, T9y, T9x);
|
Chris@10
|
917 cr[WS(rs, 29)] = -(FMA(KP831469612, T9y, T9x));
|
Chris@10
|
918 cr[WS(rs, 1)] = FMA(KP980785280, T6T, T6E);
|
Chris@10
|
919 ci[WS(rs, 14)] = FNMS(KP980785280, T6T, T6E);
|
Chris@10
|
920 cr[WS(rs, 21)] = FMS(KP831469612, T9A, T9z);
|
Chris@10
|
921 ci[WS(rs, 26)] = FMA(KP831469612, T9A, T9z);
|
Chris@10
|
922 T6X = T6V - T6W;
|
Chris@10
|
923 T9s = T6W + T6V;
|
Chris@10
|
924 }
|
Chris@10
|
925 T7a = FNMS(KP923879532, T71, T6Y);
|
Chris@10
|
926 T72 = FMA(KP923879532, T71, T6Y);
|
Chris@10
|
927 T9t = FNMS(KP923879532, T9q, T9p);
|
Chris@10
|
928 T9r = FMA(KP923879532, T9q, T9p);
|
Chris@10
|
929 ci[WS(rs, 6)] = FMA(KP980785280, T6X, T6U);
|
Chris@10
|
930 cr[WS(rs, 9)] = FNMS(KP980785280, T6X, T6U);
|
Chris@10
|
931 ci[WS(rs, 2)] = FMA(KP831469612, T79, T72);
|
Chris@10
|
932 cr[WS(rs, 13)] = FNMS(KP831469612, T79, T72);
|
Chris@10
|
933 ci[WS(rs, 30)] = FMA(KP980785280, T9s, T9r);
|
Chris@10
|
934 cr[WS(rs, 17)] = FMS(KP980785280, T9s, T9r);
|
Chris@10
|
935 ci[WS(rs, 22)] = FMA(KP980785280, T9u, T9t);
|
Chris@10
|
936 cr[WS(rs, 25)] = FMS(KP980785280, T9u, T9t);
|
Chris@10
|
937 }
|
Chris@10
|
938 }
|
Chris@10
|
939 }
|
Chris@10
|
940 }
|
Chris@10
|
941 }
|
Chris@10
|
942 }
|
Chris@10
|
943 cr[WS(rs, 5)] = FMA(KP831469612, T7d, T7a);
|
Chris@10
|
944 ci[WS(rs, 10)] = FNMS(KP831469612, T7d, T7a);
|
Chris@10
|
945 }
|
Chris@10
|
946 }
|
Chris@10
|
947 }
|
Chris@10
|
948
|
Chris@10
|
949 static const tw_instr twinstr[] = {
|
Chris@10
|
950 {TW_CEXP, 1, 1},
|
Chris@10
|
951 {TW_CEXP, 1, 3},
|
Chris@10
|
952 {TW_CEXP, 1, 9},
|
Chris@10
|
953 {TW_CEXP, 1, 27},
|
Chris@10
|
954 {TW_NEXT, 1, 0}
|
Chris@10
|
955 };
|
Chris@10
|
956
|
Chris@10
|
957 static const hc2hc_desc desc = { 32, "hf2_32", twinstr, &GENUS, {236, 98, 252, 0} };
|
Chris@10
|
958
|
Chris@10
|
959 void X(codelet_hf2_32) (planner *p) {
|
Chris@10
|
960 X(khc2hc_register) (p, hf2_32, &desc);
|
Chris@10
|
961 }
|
Chris@10
|
962 #else /* HAVE_FMA */
|
Chris@10
|
963
|
Chris@10
|
964 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hf2_32 -include hf.h */
|
Chris@10
|
965
|
Chris@10
|
966 /*
|
Chris@10
|
967 * This function contains 488 FP additions, 280 FP multiplications,
|
Chris@10
|
968 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
|
Chris@10
|
969 * 158 stack variables, 7 constants, and 128 memory accesses
|
Chris@10
|
970 */
|
Chris@10
|
971 #include "hf.h"
|
Chris@10
|
972
|
Chris@10
|
973 static void hf2_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
974 {
|
Chris@10
|
975 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
976 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
977 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
978 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
979 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
980 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
981 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
982 {
|
Chris@10
|
983 INT m;
|
Chris@10
|
984 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
|
985 E T2, T5, T3, T6, T8, TM, TO, Td, T9, Te, Th, Tl, TD, TH, T1y;
|
Chris@10
|
986 E T1H, T15, T1A, T11, T1F, T1n, T1p, T2q, T2I, T2u, T2K, T2V, T3b, T2Z, T3d;
|
Chris@10
|
987 E Tu, Ty, T3l, T3n, T1t, T1v, T2f, T2h, T1a, T1e, T32, T34, T1W, T1Y, T2C;
|
Chris@10
|
988 E T2E, Tg, TR, Tk, TS, Tm, TV, To, TT, T1M, T21, T1P, T22, T1Q, T25;
|
Chris@10
|
989 E T1S, T23;
|
Chris@10
|
990 {
|
Chris@10
|
991 E Ts, T1d, Tx, T18, Tt, T1c, Tw, T19, TB, T14, TG, TZ, TC, T13, TF;
|
Chris@10
|
992 E T10;
|
Chris@10
|
993 {
|
Chris@10
|
994 E T4, Tc, T7, Tb;
|
Chris@10
|
995 T2 = W[0];
|
Chris@10
|
996 T5 = W[1];
|
Chris@10
|
997 T3 = W[2];
|
Chris@10
|
998 T6 = W[3];
|
Chris@10
|
999 T4 = T2 * T3;
|
Chris@10
|
1000 Tc = T5 * T3;
|
Chris@10
|
1001 T7 = T5 * T6;
|
Chris@10
|
1002 Tb = T2 * T6;
|
Chris@10
|
1003 T8 = T4 + T7;
|
Chris@10
|
1004 TM = T4 - T7;
|
Chris@10
|
1005 TO = Tb + Tc;
|
Chris@10
|
1006 Td = Tb - Tc;
|
Chris@10
|
1007 T9 = W[4];
|
Chris@10
|
1008 Ts = T2 * T9;
|
Chris@10
|
1009 T1d = T6 * T9;
|
Chris@10
|
1010 Tx = T5 * T9;
|
Chris@10
|
1011 T18 = T3 * T9;
|
Chris@10
|
1012 Te = W[5];
|
Chris@10
|
1013 Tt = T5 * Te;
|
Chris@10
|
1014 T1c = T3 * Te;
|
Chris@10
|
1015 Tw = T2 * Te;
|
Chris@10
|
1016 T19 = T6 * Te;
|
Chris@10
|
1017 Th = W[6];
|
Chris@10
|
1018 TB = T3 * Th;
|
Chris@10
|
1019 T14 = T5 * Th;
|
Chris@10
|
1020 TG = T6 * Th;
|
Chris@10
|
1021 TZ = T2 * Th;
|
Chris@10
|
1022 Tl = W[7];
|
Chris@10
|
1023 TC = T6 * Tl;
|
Chris@10
|
1024 T13 = T2 * Tl;
|
Chris@10
|
1025 TF = T3 * Tl;
|
Chris@10
|
1026 T10 = T5 * Tl;
|
Chris@10
|
1027 }
|
Chris@10
|
1028 TD = TB + TC;
|
Chris@10
|
1029 TH = TF - TG;
|
Chris@10
|
1030 T1y = TZ + T10;
|
Chris@10
|
1031 T1H = TF + TG;
|
Chris@10
|
1032 T15 = T13 + T14;
|
Chris@10
|
1033 T1A = T13 - T14;
|
Chris@10
|
1034 T11 = TZ - T10;
|
Chris@10
|
1035 T1F = TB - TC;
|
Chris@10
|
1036 T1n = FMA(T9, Th, Te * Tl);
|
Chris@10
|
1037 T1p = FNMS(Te, Th, T9 * Tl);
|
Chris@10
|
1038 {
|
Chris@10
|
1039 E T2o, T2p, T2s, T2t;
|
Chris@10
|
1040 T2o = T8 * Th;
|
Chris@10
|
1041 T2p = Td * Tl;
|
Chris@10
|
1042 T2q = T2o + T2p;
|
Chris@10
|
1043 T2I = T2o - T2p;
|
Chris@10
|
1044 T2s = T8 * Tl;
|
Chris@10
|
1045 T2t = Td * Th;
|
Chris@10
|
1046 T2u = T2s - T2t;
|
Chris@10
|
1047 T2K = T2s + T2t;
|
Chris@10
|
1048 }
|
Chris@10
|
1049 {
|
Chris@10
|
1050 E T2T, T2U, T2X, T2Y;
|
Chris@10
|
1051 T2T = TM * Th;
|
Chris@10
|
1052 T2U = TO * Tl;
|
Chris@10
|
1053 T2V = T2T - T2U;
|
Chris@10
|
1054 T3b = T2T + T2U;
|
Chris@10
|
1055 T2X = TM * Tl;
|
Chris@10
|
1056 T2Y = TO * Th;
|
Chris@10
|
1057 T2Z = T2X + T2Y;
|
Chris@10
|
1058 T3d = T2X - T2Y;
|
Chris@10
|
1059 Tu = Ts + Tt;
|
Chris@10
|
1060 Ty = Tw - Tx;
|
Chris@10
|
1061 T3l = FMA(Tu, Th, Ty * Tl);
|
Chris@10
|
1062 T3n = FNMS(Ty, Th, Tu * Tl);
|
Chris@10
|
1063 }
|
Chris@10
|
1064 T1t = Ts - Tt;
|
Chris@10
|
1065 T1v = Tw + Tx;
|
Chris@10
|
1066 T2f = FMA(T1t, Th, T1v * Tl);
|
Chris@10
|
1067 T2h = FNMS(T1v, Th, T1t * Tl);
|
Chris@10
|
1068 T1a = T18 - T19;
|
Chris@10
|
1069 T1e = T1c + T1d;
|
Chris@10
|
1070 T32 = FMA(T1a, Th, T1e * Tl);
|
Chris@10
|
1071 T34 = FNMS(T1e, Th, T1a * Tl);
|
Chris@10
|
1072 T1W = T18 + T19;
|
Chris@10
|
1073 T1Y = T1c - T1d;
|
Chris@10
|
1074 T2C = FMA(T1W, Th, T1Y * Tl);
|
Chris@10
|
1075 T2E = FNMS(T1Y, Th, T1W * Tl);
|
Chris@10
|
1076 {
|
Chris@10
|
1077 E Ta, Tf, Ti, Tj;
|
Chris@10
|
1078 Ta = T8 * T9;
|
Chris@10
|
1079 Tf = Td * Te;
|
Chris@10
|
1080 Tg = Ta - Tf;
|
Chris@10
|
1081 TR = Ta + Tf;
|
Chris@10
|
1082 Ti = T8 * Te;
|
Chris@10
|
1083 Tj = Td * T9;
|
Chris@10
|
1084 Tk = Ti + Tj;
|
Chris@10
|
1085 TS = Ti - Tj;
|
Chris@10
|
1086 }
|
Chris@10
|
1087 Tm = FMA(Tg, Th, Tk * Tl);
|
Chris@10
|
1088 TV = FNMS(TS, Th, TR * Tl);
|
Chris@10
|
1089 To = FNMS(Tk, Th, Tg * Tl);
|
Chris@10
|
1090 TT = FMA(TR, Th, TS * Tl);
|
Chris@10
|
1091 {
|
Chris@10
|
1092 E T1K, T1L, T1N, T1O;
|
Chris@10
|
1093 T1K = TM * T9;
|
Chris@10
|
1094 T1L = TO * Te;
|
Chris@10
|
1095 T1M = T1K - T1L;
|
Chris@10
|
1096 T21 = T1K + T1L;
|
Chris@10
|
1097 T1N = TM * Te;
|
Chris@10
|
1098 T1O = TO * T9;
|
Chris@10
|
1099 T1P = T1N + T1O;
|
Chris@10
|
1100 T22 = T1N - T1O;
|
Chris@10
|
1101 }
|
Chris@10
|
1102 T1Q = FMA(T1M, Th, T1P * Tl);
|
Chris@10
|
1103 T25 = FNMS(T22, Th, T21 * Tl);
|
Chris@10
|
1104 T1S = FNMS(T1P, Th, T1M * Tl);
|
Chris@10
|
1105 T23 = FMA(T21, Th, T22 * Tl);
|
Chris@10
|
1106 }
|
Chris@10
|
1107 {
|
Chris@10
|
1108 E TL, T6f, T8c, T8q, T3F, T5t, T7I, T7W, T2y, T6B, T6y, T7j, T4k, T5G, T4B;
|
Chris@10
|
1109 E T5J, T3h, T6H, T6O, T7o, T4L, T5Q, T52, T5N, T1i, T7V, T6i, T7D, T3K, T5u;
|
Chris@10
|
1110 E T3P, T5v, T1E, T6k, T6n, T7f, T3W, T5z, T41, T5y, T29, T6p, T6s, T7e, T47;
|
Chris@10
|
1111 E T5C, T4c, T5B, T2R, T6z, T6E, T7k, T4v, T5K, T4E, T5H, T3y, T6P, T6K, T7p;
|
Chris@10
|
1112 E T4W, T5O, T55, T5R;
|
Chris@10
|
1113 {
|
Chris@10
|
1114 E T1, T7G, Tq, T7F, TA, T3C, TJ, T3D, Tn, Tp;
|
Chris@10
|
1115 T1 = cr[0];
|
Chris@10
|
1116 T7G = ci[0];
|
Chris@10
|
1117 Tn = cr[WS(rs, 16)];
|
Chris@10
|
1118 Tp = ci[WS(rs, 16)];
|
Chris@10
|
1119 Tq = FMA(Tm, Tn, To * Tp);
|
Chris@10
|
1120 T7F = FNMS(To, Tn, Tm * Tp);
|
Chris@10
|
1121 {
|
Chris@10
|
1122 E Tv, Tz, TE, TI;
|
Chris@10
|
1123 Tv = cr[WS(rs, 8)];
|
Chris@10
|
1124 Tz = ci[WS(rs, 8)];
|
Chris@10
|
1125 TA = FMA(Tu, Tv, Ty * Tz);
|
Chris@10
|
1126 T3C = FNMS(Ty, Tv, Tu * Tz);
|
Chris@10
|
1127 TE = cr[WS(rs, 24)];
|
Chris@10
|
1128 TI = ci[WS(rs, 24)];
|
Chris@10
|
1129 TJ = FMA(TD, TE, TH * TI);
|
Chris@10
|
1130 T3D = FNMS(TH, TE, TD * TI);
|
Chris@10
|
1131 }
|
Chris@10
|
1132 {
|
Chris@10
|
1133 E Tr, TK, T8a, T8b;
|
Chris@10
|
1134 Tr = T1 + Tq;
|
Chris@10
|
1135 TK = TA + TJ;
|
Chris@10
|
1136 TL = Tr + TK;
|
Chris@10
|
1137 T6f = Tr - TK;
|
Chris@10
|
1138 T8a = TA - TJ;
|
Chris@10
|
1139 T8b = T7G - T7F;
|
Chris@10
|
1140 T8c = T8a + T8b;
|
Chris@10
|
1141 T8q = T8b - T8a;
|
Chris@10
|
1142 }
|
Chris@10
|
1143 {
|
Chris@10
|
1144 E T3B, T3E, T7E, T7H;
|
Chris@10
|
1145 T3B = T1 - Tq;
|
Chris@10
|
1146 T3E = T3C - T3D;
|
Chris@10
|
1147 T3F = T3B + T3E;
|
Chris@10
|
1148 T5t = T3B - T3E;
|
Chris@10
|
1149 T7E = T3C + T3D;
|
Chris@10
|
1150 T7H = T7F + T7G;
|
Chris@10
|
1151 T7I = T7E + T7H;
|
Chris@10
|
1152 T7W = T7H - T7E;
|
Chris@10
|
1153 }
|
Chris@10
|
1154 }
|
Chris@10
|
1155 {
|
Chris@10
|
1156 E T2e, T4x, T2w, T4i, T2j, T4y, T2n, T4h;
|
Chris@10
|
1157 {
|
Chris@10
|
1158 E T2c, T2d, T2r, T2v;
|
Chris@10
|
1159 T2c = cr[WS(rs, 1)];
|
Chris@10
|
1160 T2d = ci[WS(rs, 1)];
|
Chris@10
|
1161 T2e = FMA(T2, T2c, T5 * T2d);
|
Chris@10
|
1162 T4x = FNMS(T5, T2c, T2 * T2d);
|
Chris@10
|
1163 T2r = cr[WS(rs, 25)];
|
Chris@10
|
1164 T2v = ci[WS(rs, 25)];
|
Chris@10
|
1165 T2w = FMA(T2q, T2r, T2u * T2v);
|
Chris@10
|
1166 T4i = FNMS(T2u, T2r, T2q * T2v);
|
Chris@10
|
1167 }
|
Chris@10
|
1168 {
|
Chris@10
|
1169 E T2g, T2i, T2l, T2m;
|
Chris@10
|
1170 T2g = cr[WS(rs, 17)];
|
Chris@10
|
1171 T2i = ci[WS(rs, 17)];
|
Chris@10
|
1172 T2j = FMA(T2f, T2g, T2h * T2i);
|
Chris@10
|
1173 T4y = FNMS(T2h, T2g, T2f * T2i);
|
Chris@10
|
1174 T2l = cr[WS(rs, 9)];
|
Chris@10
|
1175 T2m = ci[WS(rs, 9)];
|
Chris@10
|
1176 T2n = FMA(T9, T2l, Te * T2m);
|
Chris@10
|
1177 T4h = FNMS(Te, T2l, T9 * T2m);
|
Chris@10
|
1178 }
|
Chris@10
|
1179 {
|
Chris@10
|
1180 E T2k, T2x, T6w, T6x;
|
Chris@10
|
1181 T2k = T2e + T2j;
|
Chris@10
|
1182 T2x = T2n + T2w;
|
Chris@10
|
1183 T2y = T2k + T2x;
|
Chris@10
|
1184 T6B = T2k - T2x;
|
Chris@10
|
1185 T6w = T4x + T4y;
|
Chris@10
|
1186 T6x = T4h + T4i;
|
Chris@10
|
1187 T6y = T6w - T6x;
|
Chris@10
|
1188 T7j = T6w + T6x;
|
Chris@10
|
1189 }
|
Chris@10
|
1190 {
|
Chris@10
|
1191 E T4g, T4j, T4z, T4A;
|
Chris@10
|
1192 T4g = T2e - T2j;
|
Chris@10
|
1193 T4j = T4h - T4i;
|
Chris@10
|
1194 T4k = T4g + T4j;
|
Chris@10
|
1195 T5G = T4g - T4j;
|
Chris@10
|
1196 T4z = T4x - T4y;
|
Chris@10
|
1197 T4A = T2n - T2w;
|
Chris@10
|
1198 T4B = T4z - T4A;
|
Chris@10
|
1199 T5J = T4z + T4A;
|
Chris@10
|
1200 }
|
Chris@10
|
1201 }
|
Chris@10
|
1202 {
|
Chris@10
|
1203 E T31, T4H, T3f, T50, T36, T4I, T3a, T4Z;
|
Chris@10
|
1204 {
|
Chris@10
|
1205 E T2W, T30, T3c, T3e;
|
Chris@10
|
1206 T2W = cr[WS(rs, 31)];
|
Chris@10
|
1207 T30 = ci[WS(rs, 31)];
|
Chris@10
|
1208 T31 = FMA(T2V, T2W, T2Z * T30);
|
Chris@10
|
1209 T4H = FNMS(T2Z, T2W, T2V * T30);
|
Chris@10
|
1210 T3c = cr[WS(rs, 23)];
|
Chris@10
|
1211 T3e = ci[WS(rs, 23)];
|
Chris@10
|
1212 T3f = FMA(T3b, T3c, T3d * T3e);
|
Chris@10
|
1213 T50 = FNMS(T3d, T3c, T3b * T3e);
|
Chris@10
|
1214 }
|
Chris@10
|
1215 {
|
Chris@10
|
1216 E T33, T35, T38, T39;
|
Chris@10
|
1217 T33 = cr[WS(rs, 15)];
|
Chris@10
|
1218 T35 = ci[WS(rs, 15)];
|
Chris@10
|
1219 T36 = FMA(T32, T33, T34 * T35);
|
Chris@10
|
1220 T4I = FNMS(T34, T33, T32 * T35);
|
Chris@10
|
1221 T38 = cr[WS(rs, 7)];
|
Chris@10
|
1222 T39 = ci[WS(rs, 7)];
|
Chris@10
|
1223 T3a = FMA(TR, T38, TS * T39);
|
Chris@10
|
1224 T4Z = FNMS(TS, T38, TR * T39);
|
Chris@10
|
1225 }
|
Chris@10
|
1226 {
|
Chris@10
|
1227 E T37, T3g, T6M, T6N;
|
Chris@10
|
1228 T37 = T31 + T36;
|
Chris@10
|
1229 T3g = T3a + T3f;
|
Chris@10
|
1230 T3h = T37 + T3g;
|
Chris@10
|
1231 T6H = T37 - T3g;
|
Chris@10
|
1232 T6M = T4H + T4I;
|
Chris@10
|
1233 T6N = T4Z + T50;
|
Chris@10
|
1234 T6O = T6M - T6N;
|
Chris@10
|
1235 T7o = T6M + T6N;
|
Chris@10
|
1236 }
|
Chris@10
|
1237 {
|
Chris@10
|
1238 E T4J, T4K, T4Y, T51;
|
Chris@10
|
1239 T4J = T4H - T4I;
|
Chris@10
|
1240 T4K = T3a - T3f;
|
Chris@10
|
1241 T4L = T4J - T4K;
|
Chris@10
|
1242 T5Q = T4J + T4K;
|
Chris@10
|
1243 T4Y = T31 - T36;
|
Chris@10
|
1244 T51 = T4Z - T50;
|
Chris@10
|
1245 T52 = T4Y + T51;
|
Chris@10
|
1246 T5N = T4Y - T51;
|
Chris@10
|
1247 }
|
Chris@10
|
1248 }
|
Chris@10
|
1249 {
|
Chris@10
|
1250 E TQ, T3H, T1g, T3N, TX, T3I, T17, T3M;
|
Chris@10
|
1251 {
|
Chris@10
|
1252 E TN, TP, T1b, T1f;
|
Chris@10
|
1253 TN = cr[WS(rs, 4)];
|
Chris@10
|
1254 TP = ci[WS(rs, 4)];
|
Chris@10
|
1255 TQ = FMA(TM, TN, TO * TP);
|
Chris@10
|
1256 T3H = FNMS(TO, TN, TM * TP);
|
Chris@10
|
1257 T1b = cr[WS(rs, 12)];
|
Chris@10
|
1258 T1f = ci[WS(rs, 12)];
|
Chris@10
|
1259 T1g = FMA(T1a, T1b, T1e * T1f);
|
Chris@10
|
1260 T3N = FNMS(T1e, T1b, T1a * T1f);
|
Chris@10
|
1261 }
|
Chris@10
|
1262 {
|
Chris@10
|
1263 E TU, TW, T12, T16;
|
Chris@10
|
1264 TU = cr[WS(rs, 20)];
|
Chris@10
|
1265 TW = ci[WS(rs, 20)];
|
Chris@10
|
1266 TX = FMA(TT, TU, TV * TW);
|
Chris@10
|
1267 T3I = FNMS(TV, TU, TT * TW);
|
Chris@10
|
1268 T12 = cr[WS(rs, 28)];
|
Chris@10
|
1269 T16 = ci[WS(rs, 28)];
|
Chris@10
|
1270 T17 = FMA(T11, T12, T15 * T16);
|
Chris@10
|
1271 T3M = FNMS(T15, T12, T11 * T16);
|
Chris@10
|
1272 }
|
Chris@10
|
1273 {
|
Chris@10
|
1274 E TY, T1h, T6g, T6h;
|
Chris@10
|
1275 TY = TQ + TX;
|
Chris@10
|
1276 T1h = T17 + T1g;
|
Chris@10
|
1277 T1i = TY + T1h;
|
Chris@10
|
1278 T7V = TY - T1h;
|
Chris@10
|
1279 T6g = T3M + T3N;
|
Chris@10
|
1280 T6h = T3H + T3I;
|
Chris@10
|
1281 T6i = T6g - T6h;
|
Chris@10
|
1282 T7D = T6h + T6g;
|
Chris@10
|
1283 }
|
Chris@10
|
1284 {
|
Chris@10
|
1285 E T3G, T3J, T3L, T3O;
|
Chris@10
|
1286 T3G = TQ - TX;
|
Chris@10
|
1287 T3J = T3H - T3I;
|
Chris@10
|
1288 T3K = T3G + T3J;
|
Chris@10
|
1289 T5u = T3G - T3J;
|
Chris@10
|
1290 T3L = T17 - T1g;
|
Chris@10
|
1291 T3O = T3M - T3N;
|
Chris@10
|
1292 T3P = T3L - T3O;
|
Chris@10
|
1293 T5v = T3L + T3O;
|
Chris@10
|
1294 }
|
Chris@10
|
1295 }
|
Chris@10
|
1296 {
|
Chris@10
|
1297 E T1m, T3X, T1C, T3U, T1r, T3Y, T1x, T3T;
|
Chris@10
|
1298 {
|
Chris@10
|
1299 E T1k, T1l, T1z, T1B;
|
Chris@10
|
1300 T1k = cr[WS(rs, 2)];
|
Chris@10
|
1301 T1l = ci[WS(rs, 2)];
|
Chris@10
|
1302 T1m = FMA(T8, T1k, Td * T1l);
|
Chris@10
|
1303 T3X = FNMS(Td, T1k, T8 * T1l);
|
Chris@10
|
1304 T1z = cr[WS(rs, 26)];
|
Chris@10
|
1305 T1B = ci[WS(rs, 26)];
|
Chris@10
|
1306 T1C = FMA(T1y, T1z, T1A * T1B);
|
Chris@10
|
1307 T3U = FNMS(T1A, T1z, T1y * T1B);
|
Chris@10
|
1308 }
|
Chris@10
|
1309 {
|
Chris@10
|
1310 E T1o, T1q, T1u, T1w;
|
Chris@10
|
1311 T1o = cr[WS(rs, 18)];
|
Chris@10
|
1312 T1q = ci[WS(rs, 18)];
|
Chris@10
|
1313 T1r = FMA(T1n, T1o, T1p * T1q);
|
Chris@10
|
1314 T3Y = FNMS(T1p, T1o, T1n * T1q);
|
Chris@10
|
1315 T1u = cr[WS(rs, 10)];
|
Chris@10
|
1316 T1w = ci[WS(rs, 10)];
|
Chris@10
|
1317 T1x = FMA(T1t, T1u, T1v * T1w);
|
Chris@10
|
1318 T3T = FNMS(T1v, T1u, T1t * T1w);
|
Chris@10
|
1319 }
|
Chris@10
|
1320 {
|
Chris@10
|
1321 E T1s, T1D, T6l, T6m;
|
Chris@10
|
1322 T1s = T1m + T1r;
|
Chris@10
|
1323 T1D = T1x + T1C;
|
Chris@10
|
1324 T1E = T1s + T1D;
|
Chris@10
|
1325 T6k = T1s - T1D;
|
Chris@10
|
1326 T6l = T3X + T3Y;
|
Chris@10
|
1327 T6m = T3T + T3U;
|
Chris@10
|
1328 T6n = T6l - T6m;
|
Chris@10
|
1329 T7f = T6l + T6m;
|
Chris@10
|
1330 }
|
Chris@10
|
1331 {
|
Chris@10
|
1332 E T3S, T3V, T3Z, T40;
|
Chris@10
|
1333 T3S = T1m - T1r;
|
Chris@10
|
1334 T3V = T3T - T3U;
|
Chris@10
|
1335 T3W = T3S + T3V;
|
Chris@10
|
1336 T5z = T3S - T3V;
|
Chris@10
|
1337 T3Z = T3X - T3Y;
|
Chris@10
|
1338 T40 = T1x - T1C;
|
Chris@10
|
1339 T41 = T3Z - T40;
|
Chris@10
|
1340 T5y = T3Z + T40;
|
Chris@10
|
1341 }
|
Chris@10
|
1342 }
|
Chris@10
|
1343 {
|
Chris@10
|
1344 E T1J, T43, T27, T4a, T1U, T44, T20, T49;
|
Chris@10
|
1345 {
|
Chris@10
|
1346 E T1G, T1I, T24, T26;
|
Chris@10
|
1347 T1G = cr[WS(rs, 30)];
|
Chris@10
|
1348 T1I = ci[WS(rs, 30)];
|
Chris@10
|
1349 T1J = FMA(T1F, T1G, T1H * T1I);
|
Chris@10
|
1350 T43 = FNMS(T1H, T1G, T1F * T1I);
|
Chris@10
|
1351 T24 = cr[WS(rs, 22)];
|
Chris@10
|
1352 T26 = ci[WS(rs, 22)];
|
Chris@10
|
1353 T27 = FMA(T23, T24, T25 * T26);
|
Chris@10
|
1354 T4a = FNMS(T25, T24, T23 * T26);
|
Chris@10
|
1355 }
|
Chris@10
|
1356 {
|
Chris@10
|
1357 E T1R, T1T, T1X, T1Z;
|
Chris@10
|
1358 T1R = cr[WS(rs, 14)];
|
Chris@10
|
1359 T1T = ci[WS(rs, 14)];
|
Chris@10
|
1360 T1U = FMA(T1Q, T1R, T1S * T1T);
|
Chris@10
|
1361 T44 = FNMS(T1S, T1R, T1Q * T1T);
|
Chris@10
|
1362 T1X = cr[WS(rs, 6)];
|
Chris@10
|
1363 T1Z = ci[WS(rs, 6)];
|
Chris@10
|
1364 T20 = FMA(T1W, T1X, T1Y * T1Z);
|
Chris@10
|
1365 T49 = FNMS(T1Y, T1X, T1W * T1Z);
|
Chris@10
|
1366 }
|
Chris@10
|
1367 {
|
Chris@10
|
1368 E T1V, T28, T6q, T6r;
|
Chris@10
|
1369 T1V = T1J + T1U;
|
Chris@10
|
1370 T28 = T20 + T27;
|
Chris@10
|
1371 T29 = T1V + T28;
|
Chris@10
|
1372 T6p = T1V - T28;
|
Chris@10
|
1373 T6q = T43 + T44;
|
Chris@10
|
1374 T6r = T49 + T4a;
|
Chris@10
|
1375 T6s = T6q - T6r;
|
Chris@10
|
1376 T7e = T6q + T6r;
|
Chris@10
|
1377 }
|
Chris@10
|
1378 {
|
Chris@10
|
1379 E T45, T46, T48, T4b;
|
Chris@10
|
1380 T45 = T43 - T44;
|
Chris@10
|
1381 T46 = T20 - T27;
|
Chris@10
|
1382 T47 = T45 - T46;
|
Chris@10
|
1383 T5C = T45 + T46;
|
Chris@10
|
1384 T48 = T1J - T1U;
|
Chris@10
|
1385 T4b = T49 - T4a;
|
Chris@10
|
1386 T4c = T48 + T4b;
|
Chris@10
|
1387 T5B = T48 - T4b;
|
Chris@10
|
1388 }
|
Chris@10
|
1389 }
|
Chris@10
|
1390 {
|
Chris@10
|
1391 E T2B, T4m, T2G, T4n, T4l, T4o, T2M, T4q, T2P, T4r, T4s, T4t;
|
Chris@10
|
1392 {
|
Chris@10
|
1393 E T2z, T2A, T2D, T2F;
|
Chris@10
|
1394 T2z = cr[WS(rs, 5)];
|
Chris@10
|
1395 T2A = ci[WS(rs, 5)];
|
Chris@10
|
1396 T2B = FMA(T21, T2z, T22 * T2A);
|
Chris@10
|
1397 T4m = FNMS(T22, T2z, T21 * T2A);
|
Chris@10
|
1398 T2D = cr[WS(rs, 21)];
|
Chris@10
|
1399 T2F = ci[WS(rs, 21)];
|
Chris@10
|
1400 T2G = FMA(T2C, T2D, T2E * T2F);
|
Chris@10
|
1401 T4n = FNMS(T2E, T2D, T2C * T2F);
|
Chris@10
|
1402 }
|
Chris@10
|
1403 T4l = T2B - T2G;
|
Chris@10
|
1404 T4o = T4m - T4n;
|
Chris@10
|
1405 {
|
Chris@10
|
1406 E T2J, T2L, T2N, T2O;
|
Chris@10
|
1407 T2J = cr[WS(rs, 29)];
|
Chris@10
|
1408 T2L = ci[WS(rs, 29)];
|
Chris@10
|
1409 T2M = FMA(T2I, T2J, T2K * T2L);
|
Chris@10
|
1410 T4q = FNMS(T2K, T2J, T2I * T2L);
|
Chris@10
|
1411 T2N = cr[WS(rs, 13)];
|
Chris@10
|
1412 T2O = ci[WS(rs, 13)];
|
Chris@10
|
1413 T2P = FMA(T1M, T2N, T1P * T2O);
|
Chris@10
|
1414 T4r = FNMS(T1P, T2N, T1M * T2O);
|
Chris@10
|
1415 }
|
Chris@10
|
1416 T4s = T4q - T4r;
|
Chris@10
|
1417 T4t = T2M - T2P;
|
Chris@10
|
1418 {
|
Chris@10
|
1419 E T2H, T2Q, T6C, T6D;
|
Chris@10
|
1420 T2H = T2B + T2G;
|
Chris@10
|
1421 T2Q = T2M + T2P;
|
Chris@10
|
1422 T2R = T2H + T2Q;
|
Chris@10
|
1423 T6z = T2H - T2Q;
|
Chris@10
|
1424 T6C = T4q + T4r;
|
Chris@10
|
1425 T6D = T4m + T4n;
|
Chris@10
|
1426 T6E = T6C - T6D;
|
Chris@10
|
1427 T7k = T6D + T6C;
|
Chris@10
|
1428 }
|
Chris@10
|
1429 {
|
Chris@10
|
1430 E T4p, T4u, T4C, T4D;
|
Chris@10
|
1431 T4p = T4l + T4o;
|
Chris@10
|
1432 T4u = T4s - T4t;
|
Chris@10
|
1433 T4v = KP707106781 * (T4p - T4u);
|
Chris@10
|
1434 T5K = KP707106781 * (T4p + T4u);
|
Chris@10
|
1435 T4C = T4t + T4s;
|
Chris@10
|
1436 T4D = T4l - T4o;
|
Chris@10
|
1437 T4E = KP707106781 * (T4C - T4D);
|
Chris@10
|
1438 T5H = KP707106781 * (T4D + T4C);
|
Chris@10
|
1439 }
|
Chris@10
|
1440 }
|
Chris@10
|
1441 {
|
Chris@10
|
1442 E T3k, T4S, T3p, T4T, T4R, T4U, T3t, T4N, T3w, T4O, T4M, T4P;
|
Chris@10
|
1443 {
|
Chris@10
|
1444 E T3i, T3j, T3m, T3o;
|
Chris@10
|
1445 T3i = cr[WS(rs, 3)];
|
Chris@10
|
1446 T3j = ci[WS(rs, 3)];
|
Chris@10
|
1447 T3k = FMA(T3, T3i, T6 * T3j);
|
Chris@10
|
1448 T4S = FNMS(T6, T3i, T3 * T3j);
|
Chris@10
|
1449 T3m = cr[WS(rs, 19)];
|
Chris@10
|
1450 T3o = ci[WS(rs, 19)];
|
Chris@10
|
1451 T3p = FMA(T3l, T3m, T3n * T3o);
|
Chris@10
|
1452 T4T = FNMS(T3n, T3m, T3l * T3o);
|
Chris@10
|
1453 }
|
Chris@10
|
1454 T4R = T3k - T3p;
|
Chris@10
|
1455 T4U = T4S - T4T;
|
Chris@10
|
1456 {
|
Chris@10
|
1457 E T3r, T3s, T3u, T3v;
|
Chris@10
|
1458 T3r = cr[WS(rs, 27)];
|
Chris@10
|
1459 T3s = ci[WS(rs, 27)];
|
Chris@10
|
1460 T3t = FMA(Th, T3r, Tl * T3s);
|
Chris@10
|
1461 T4N = FNMS(Tl, T3r, Th * T3s);
|
Chris@10
|
1462 T3u = cr[WS(rs, 11)];
|
Chris@10
|
1463 T3v = ci[WS(rs, 11)];
|
Chris@10
|
1464 T3w = FMA(Tg, T3u, Tk * T3v);
|
Chris@10
|
1465 T4O = FNMS(Tk, T3u, Tg * T3v);
|
Chris@10
|
1466 }
|
Chris@10
|
1467 T4M = T3t - T3w;
|
Chris@10
|
1468 T4P = T4N - T4O;
|
Chris@10
|
1469 {
|
Chris@10
|
1470 E T3q, T3x, T6I, T6J;
|
Chris@10
|
1471 T3q = T3k + T3p;
|
Chris@10
|
1472 T3x = T3t + T3w;
|
Chris@10
|
1473 T3y = T3q + T3x;
|
Chris@10
|
1474 T6P = T3q - T3x;
|
Chris@10
|
1475 T6I = T4N + T4O;
|
Chris@10
|
1476 T6J = T4S + T4T;
|
Chris@10
|
1477 T6K = T6I - T6J;
|
Chris@10
|
1478 T7p = T6J + T6I;
|
Chris@10
|
1479 }
|
Chris@10
|
1480 {
|
Chris@10
|
1481 E T4Q, T4V, T53, T54;
|
Chris@10
|
1482 T4Q = T4M + T4P;
|
Chris@10
|
1483 T4V = T4R - T4U;
|
Chris@10
|
1484 T4W = KP707106781 * (T4Q - T4V);
|
Chris@10
|
1485 T5O = KP707106781 * (T4V + T4Q);
|
Chris@10
|
1486 T53 = T4R + T4U;
|
Chris@10
|
1487 T54 = T4P - T4M;
|
Chris@10
|
1488 T55 = KP707106781 * (T53 - T54);
|
Chris@10
|
1489 T5R = KP707106781 * (T53 + T54);
|
Chris@10
|
1490 }
|
Chris@10
|
1491 }
|
Chris@10
|
1492 {
|
Chris@10
|
1493 E T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B;
|
Chris@10
|
1494 {
|
Chris@10
|
1495 E T1j, T2a, T7C, T7J;
|
Chris@10
|
1496 T1j = TL + T1i;
|
Chris@10
|
1497 T2a = T1E + T29;
|
Chris@10
|
1498 T2b = T1j + T2a;
|
Chris@10
|
1499 T7x = T1j - T2a;
|
Chris@10
|
1500 T7C = T7f + T7e;
|
Chris@10
|
1501 T7J = T7D + T7I;
|
Chris@10
|
1502 T7K = T7C + T7J;
|
Chris@10
|
1503 T7M = T7J - T7C;
|
Chris@10
|
1504 }
|
Chris@10
|
1505 {
|
Chris@10
|
1506 E T2S, T3z, T7y, T7z;
|
Chris@10
|
1507 T2S = T2y + T2R;
|
Chris@10
|
1508 T3z = T3h + T3y;
|
Chris@10
|
1509 T3A = T2S + T3z;
|
Chris@10
|
1510 T7L = T3z - T2S;
|
Chris@10
|
1511 T7y = T7o + T7p;
|
Chris@10
|
1512 T7z = T7j + T7k;
|
Chris@10
|
1513 T7A = T7y - T7z;
|
Chris@10
|
1514 T7B = T7z + T7y;
|
Chris@10
|
1515 }
|
Chris@10
|
1516 ci[WS(rs, 15)] = T2b - T3A;
|
Chris@10
|
1517 cr[WS(rs, 24)] = T7L - T7M;
|
Chris@10
|
1518 ci[WS(rs, 23)] = T7L + T7M;
|
Chris@10
|
1519 cr[0] = T2b + T3A;
|
Chris@10
|
1520 cr[WS(rs, 8)] = T7x - T7A;
|
Chris@10
|
1521 cr[WS(rs, 16)] = T7B - T7K;
|
Chris@10
|
1522 ci[WS(rs, 31)] = T7B + T7K;
|
Chris@10
|
1523 ci[WS(rs, 7)] = T7x + T7A;
|
Chris@10
|
1524 }
|
Chris@10
|
1525 {
|
Chris@10
|
1526 E T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T;
|
Chris@10
|
1527 E T5X, T5w, T89;
|
Chris@10
|
1528 T5w = KP707106781 * (T5u + T5v);
|
Chris@10
|
1529 T5x = T5t - T5w;
|
Chris@10
|
1530 T5Z = T5t + T5w;
|
Chris@10
|
1531 T89 = KP707106781 * (T3K - T3P);
|
Chris@10
|
1532 T8d = T89 + T8c;
|
Chris@10
|
1533 T8j = T8c - T89;
|
Chris@10
|
1534 {
|
Chris@10
|
1535 E T5A, T5D, T67, T68;
|
Chris@10
|
1536 T5A = FMA(KP923879532, T5y, KP382683432 * T5z);
|
Chris@10
|
1537 T5D = FNMS(KP923879532, T5C, KP382683432 * T5B);
|
Chris@10
|
1538 T5E = T5A + T5D;
|
Chris@10
|
1539 T88 = T5A - T5D;
|
Chris@10
|
1540 T67 = T5N + T5O;
|
Chris@10
|
1541 T68 = T5Q + T5R;
|
Chris@10
|
1542 T69 = FNMS(KP980785280, T68, KP195090322 * T67);
|
Chris@10
|
1543 T6d = FMA(KP980785280, T67, KP195090322 * T68);
|
Chris@10
|
1544 }
|
Chris@10
|
1545 {
|
Chris@10
|
1546 E T5I, T5L, T60, T61;
|
Chris@10
|
1547 T5I = T5G - T5H;
|
Chris@10
|
1548 T5L = T5J - T5K;
|
Chris@10
|
1549 T5M = FMA(KP831469612, T5I, KP555570233 * T5L);
|
Chris@10
|
1550 T5W = FNMS(KP831469612, T5L, KP555570233 * T5I);
|
Chris@10
|
1551 T60 = FNMS(KP382683432, T5y, KP923879532 * T5z);
|
Chris@10
|
1552 T61 = FMA(KP382683432, T5C, KP923879532 * T5B);
|
Chris@10
|
1553 T62 = T60 + T61;
|
Chris@10
|
1554 T8i = T61 - T60;
|
Chris@10
|
1555 }
|
Chris@10
|
1556 {
|
Chris@10
|
1557 E T64, T65, T5P, T5S;
|
Chris@10
|
1558 T64 = T5G + T5H;
|
Chris@10
|
1559 T65 = T5J + T5K;
|
Chris@10
|
1560 T66 = FMA(KP195090322, T64, KP980785280 * T65);
|
Chris@10
|
1561 T6c = FNMS(KP195090322, T65, KP980785280 * T64);
|
Chris@10
|
1562 T5P = T5N - T5O;
|
Chris@10
|
1563 T5S = T5Q - T5R;
|
Chris@10
|
1564 T5T = FNMS(KP555570233, T5S, KP831469612 * T5P);
|
Chris@10
|
1565 T5X = FMA(KP555570233, T5P, KP831469612 * T5S);
|
Chris@10
|
1566 }
|
Chris@10
|
1567 {
|
Chris@10
|
1568 E T5F, T5U, T8h, T8k;
|
Chris@10
|
1569 T5F = T5x + T5E;
|
Chris@10
|
1570 T5U = T5M + T5T;
|
Chris@10
|
1571 ci[WS(rs, 12)] = T5F - T5U;
|
Chris@10
|
1572 cr[WS(rs, 3)] = T5F + T5U;
|
Chris@10
|
1573 T8h = T5X - T5W;
|
Chris@10
|
1574 T8k = T8i + T8j;
|
Chris@10
|
1575 cr[WS(rs, 19)] = T8h - T8k;
|
Chris@10
|
1576 ci[WS(rs, 28)] = T8h + T8k;
|
Chris@10
|
1577 }
|
Chris@10
|
1578 {
|
Chris@10
|
1579 E T8l, T8m, T5V, T5Y;
|
Chris@10
|
1580 T8l = T5T - T5M;
|
Chris@10
|
1581 T8m = T8j - T8i;
|
Chris@10
|
1582 cr[WS(rs, 27)] = T8l - T8m;
|
Chris@10
|
1583 ci[WS(rs, 20)] = T8l + T8m;
|
Chris@10
|
1584 T5V = T5x - T5E;
|
Chris@10
|
1585 T5Y = T5W + T5X;
|
Chris@10
|
1586 cr[WS(rs, 11)] = T5V - T5Y;
|
Chris@10
|
1587 ci[WS(rs, 4)] = T5V + T5Y;
|
Chris@10
|
1588 }
|
Chris@10
|
1589 {
|
Chris@10
|
1590 E T63, T6a, T87, T8e;
|
Chris@10
|
1591 T63 = T5Z - T62;
|
Chris@10
|
1592 T6a = T66 + T69;
|
Chris@10
|
1593 ci[WS(rs, 8)] = T63 - T6a;
|
Chris@10
|
1594 cr[WS(rs, 7)] = T63 + T6a;
|
Chris@10
|
1595 T87 = T69 - T66;
|
Chris@10
|
1596 T8e = T88 + T8d;
|
Chris@10
|
1597 cr[WS(rs, 31)] = T87 - T8e;
|
Chris@10
|
1598 ci[WS(rs, 16)] = T87 + T8e;
|
Chris@10
|
1599 }
|
Chris@10
|
1600 {
|
Chris@10
|
1601 E T8f, T8g, T6b, T6e;
|
Chris@10
|
1602 T8f = T6d - T6c;
|
Chris@10
|
1603 T8g = T8d - T88;
|
Chris@10
|
1604 cr[WS(rs, 23)] = T8f - T8g;
|
Chris@10
|
1605 ci[WS(rs, 24)] = T8f + T8g;
|
Chris@10
|
1606 T6b = T5Z + T62;
|
Chris@10
|
1607 T6e = T6c + T6d;
|
Chris@10
|
1608 cr[WS(rs, 15)] = T6b - T6e;
|
Chris@10
|
1609 ci[0] = T6b + T6e;
|
Chris@10
|
1610 }
|
Chris@10
|
1611 }
|
Chris@10
|
1612 {
|
Chris@10
|
1613 E T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v;
|
Chris@10
|
1614 {
|
Chris@10
|
1615 E T7d, T7g, T7O, T7P;
|
Chris@10
|
1616 T7d = TL - T1i;
|
Chris@10
|
1617 T7g = T7e - T7f;
|
Chris@10
|
1618 T7h = T7d - T7g;
|
Chris@10
|
1619 T7t = T7d + T7g;
|
Chris@10
|
1620 T7O = T1E - T29;
|
Chris@10
|
1621 T7P = T7I - T7D;
|
Chris@10
|
1622 T7Q = T7O + T7P;
|
Chris@10
|
1623 T7S = T7P - T7O;
|
Chris@10
|
1624 }
|
Chris@10
|
1625 {
|
Chris@10
|
1626 E T7i, T7l, T7n, T7q;
|
Chris@10
|
1627 T7i = T2y - T2R;
|
Chris@10
|
1628 T7l = T7j - T7k;
|
Chris@10
|
1629 T7m = T7i + T7l;
|
Chris@10
|
1630 T7u = T7i - T7l;
|
Chris@10
|
1631 T7n = T3h - T3y;
|
Chris@10
|
1632 T7q = T7o - T7p;
|
Chris@10
|
1633 T7r = T7n - T7q;
|
Chris@10
|
1634 T7v = T7n + T7q;
|
Chris@10
|
1635 }
|
Chris@10
|
1636 {
|
Chris@10
|
1637 E T7s, T7R, T7w, T7N;
|
Chris@10
|
1638 T7s = KP707106781 * (T7m + T7r);
|
Chris@10
|
1639 ci[WS(rs, 11)] = T7h - T7s;
|
Chris@10
|
1640 cr[WS(rs, 4)] = T7h + T7s;
|
Chris@10
|
1641 T7R = KP707106781 * (T7v - T7u);
|
Chris@10
|
1642 cr[WS(rs, 20)] = T7R - T7S;
|
Chris@10
|
1643 ci[WS(rs, 27)] = T7R + T7S;
|
Chris@10
|
1644 T7w = KP707106781 * (T7u + T7v);
|
Chris@10
|
1645 cr[WS(rs, 12)] = T7t - T7w;
|
Chris@10
|
1646 ci[WS(rs, 3)] = T7t + T7w;
|
Chris@10
|
1647 T7N = KP707106781 * (T7r - T7m);
|
Chris@10
|
1648 cr[WS(rs, 28)] = T7N - T7Q;
|
Chris@10
|
1649 ci[WS(rs, 19)] = T7N + T7Q;
|
Chris@10
|
1650 }
|
Chris@10
|
1651 }
|
Chris@10
|
1652 {
|
Chris@10
|
1653 E T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R;
|
Chris@10
|
1654 E T6V;
|
Chris@10
|
1655 {
|
Chris@10
|
1656 E T6o, T6t, T6A, T6F;
|
Chris@10
|
1657 T6j = T6f - T6i;
|
Chris@10
|
1658 T7X = T7V + T7W;
|
Chris@10
|
1659 T83 = T7W - T7V;
|
Chris@10
|
1660 T6X = T6f + T6i;
|
Chris@10
|
1661 T6o = T6k + T6n;
|
Chris@10
|
1662 T6t = T6p - T6s;
|
Chris@10
|
1663 T6u = KP707106781 * (T6o + T6t);
|
Chris@10
|
1664 T7U = KP707106781 * (T6o - T6t);
|
Chris@10
|
1665 {
|
Chris@10
|
1666 E T75, T76, T6Y, T6Z;
|
Chris@10
|
1667 T75 = T6O + T6P;
|
Chris@10
|
1668 T76 = T6H + T6K;
|
Chris@10
|
1669 T77 = FMA(KP382683432, T75, KP923879532 * T76);
|
Chris@10
|
1670 T7b = FNMS(KP923879532, T75, KP382683432 * T76);
|
Chris@10
|
1671 T6Y = T6k - T6n;
|
Chris@10
|
1672 T6Z = T6p + T6s;
|
Chris@10
|
1673 T70 = KP707106781 * (T6Y + T6Z);
|
Chris@10
|
1674 T82 = KP707106781 * (T6Z - T6Y);
|
Chris@10
|
1675 }
|
Chris@10
|
1676 T6A = T6y - T6z;
|
Chris@10
|
1677 T6F = T6B - T6E;
|
Chris@10
|
1678 T6G = FMA(KP382683432, T6A, KP923879532 * T6F);
|
Chris@10
|
1679 T6U = FNMS(KP923879532, T6A, KP382683432 * T6F);
|
Chris@10
|
1680 {
|
Chris@10
|
1681 E T72, T73, T6L, T6Q;
|
Chris@10
|
1682 T72 = T6B + T6E;
|
Chris@10
|
1683 T73 = T6y + T6z;
|
Chris@10
|
1684 T74 = FNMS(KP382683432, T73, KP923879532 * T72);
|
Chris@10
|
1685 T7a = FMA(KP923879532, T73, KP382683432 * T72);
|
Chris@10
|
1686 T6L = T6H - T6K;
|
Chris@10
|
1687 T6Q = T6O - T6P;
|
Chris@10
|
1688 T6R = FNMS(KP382683432, T6Q, KP923879532 * T6L);
|
Chris@10
|
1689 T6V = FMA(KP923879532, T6Q, KP382683432 * T6L);
|
Chris@10
|
1690 }
|
Chris@10
|
1691 }
|
Chris@10
|
1692 {
|
Chris@10
|
1693 E T6v, T6S, T81, T84;
|
Chris@10
|
1694 T6v = T6j + T6u;
|
Chris@10
|
1695 T6S = T6G + T6R;
|
Chris@10
|
1696 ci[WS(rs, 13)] = T6v - T6S;
|
Chris@10
|
1697 cr[WS(rs, 2)] = T6v + T6S;
|
Chris@10
|
1698 T81 = T6V - T6U;
|
Chris@10
|
1699 T84 = T82 + T83;
|
Chris@10
|
1700 cr[WS(rs, 18)] = T81 - T84;
|
Chris@10
|
1701 ci[WS(rs, 29)] = T81 + T84;
|
Chris@10
|
1702 }
|
Chris@10
|
1703 {
|
Chris@10
|
1704 E T85, T86, T6T, T6W;
|
Chris@10
|
1705 T85 = T6R - T6G;
|
Chris@10
|
1706 T86 = T83 - T82;
|
Chris@10
|
1707 cr[WS(rs, 26)] = T85 - T86;
|
Chris@10
|
1708 ci[WS(rs, 21)] = T85 + T86;
|
Chris@10
|
1709 T6T = T6j - T6u;
|
Chris@10
|
1710 T6W = T6U + T6V;
|
Chris@10
|
1711 cr[WS(rs, 10)] = T6T - T6W;
|
Chris@10
|
1712 ci[WS(rs, 5)] = T6T + T6W;
|
Chris@10
|
1713 }
|
Chris@10
|
1714 {
|
Chris@10
|
1715 E T71, T78, T7T, T7Y;
|
Chris@10
|
1716 T71 = T6X + T70;
|
Chris@10
|
1717 T78 = T74 + T77;
|
Chris@10
|
1718 cr[WS(rs, 14)] = T71 - T78;
|
Chris@10
|
1719 ci[WS(rs, 1)] = T71 + T78;
|
Chris@10
|
1720 T7T = T7b - T7a;
|
Chris@10
|
1721 T7Y = T7U + T7X;
|
Chris@10
|
1722 cr[WS(rs, 30)] = T7T - T7Y;
|
Chris@10
|
1723 ci[WS(rs, 17)] = T7T + T7Y;
|
Chris@10
|
1724 }
|
Chris@10
|
1725 {
|
Chris@10
|
1726 E T7Z, T80, T79, T7c;
|
Chris@10
|
1727 T7Z = T77 - T74;
|
Chris@10
|
1728 T80 = T7X - T7U;
|
Chris@10
|
1729 cr[WS(rs, 22)] = T7Z - T80;
|
Chris@10
|
1730 ci[WS(rs, 25)] = T7Z + T80;
|
Chris@10
|
1731 T79 = T6X - T70;
|
Chris@10
|
1732 T7c = T7a + T7b;
|
Chris@10
|
1733 ci[WS(rs, 9)] = T79 - T7c;
|
Chris@10
|
1734 cr[WS(rs, 6)] = T79 + T7c;
|
Chris@10
|
1735 }
|
Chris@10
|
1736 }
|
Chris@10
|
1737 {
|
Chris@10
|
1738 E T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57;
|
Chris@10
|
1739 E T5b, T3Q, T8p;
|
Chris@10
|
1740 T3Q = KP707106781 * (T3K + T3P);
|
Chris@10
|
1741 T3R = T3F - T3Q;
|
Chris@10
|
1742 T5d = T3F + T3Q;
|
Chris@10
|
1743 T8p = KP707106781 * (T5v - T5u);
|
Chris@10
|
1744 T8r = T8p + T8q;
|
Chris@10
|
1745 T8x = T8q - T8p;
|
Chris@10
|
1746 {
|
Chris@10
|
1747 E T42, T4d, T5l, T5m;
|
Chris@10
|
1748 T42 = FNMS(KP923879532, T41, KP382683432 * T3W);
|
Chris@10
|
1749 T4d = FMA(KP923879532, T47, KP382683432 * T4c);
|
Chris@10
|
1750 T4e = T42 + T4d;
|
Chris@10
|
1751 T8o = T4d - T42;
|
Chris@10
|
1752 T5l = T52 + T55;
|
Chris@10
|
1753 T5m = T4L + T4W;
|
Chris@10
|
1754 T5n = FNMS(KP195090322, T5m, KP980785280 * T5l);
|
Chris@10
|
1755 T5r = FMA(KP980785280, T5m, KP195090322 * T5l);
|
Chris@10
|
1756 }
|
Chris@10
|
1757 {
|
Chris@10
|
1758 E T4w, T4F, T5e, T5f;
|
Chris@10
|
1759 T4w = T4k - T4v;
|
Chris@10
|
1760 T4F = T4B - T4E;
|
Chris@10
|
1761 T4G = FNMS(KP555570233, T4F, KP831469612 * T4w);
|
Chris@10
|
1762 T5a = FMA(KP831469612, T4F, KP555570233 * T4w);
|
Chris@10
|
1763 T5e = FMA(KP382683432, T41, KP923879532 * T3W);
|
Chris@10
|
1764 T5f = FNMS(KP382683432, T47, KP923879532 * T4c);
|
Chris@10
|
1765 T5g = T5e + T5f;
|
Chris@10
|
1766 T8w = T5e - T5f;
|
Chris@10
|
1767 }
|
Chris@10
|
1768 {
|
Chris@10
|
1769 E T5i, T5j, T4X, T56;
|
Chris@10
|
1770 T5i = T4B + T4E;
|
Chris@10
|
1771 T5j = T4k + T4v;
|
Chris@10
|
1772 T5k = FMA(KP195090322, T5i, KP980785280 * T5j);
|
Chris@10
|
1773 T5q = FNMS(KP980785280, T5i, KP195090322 * T5j);
|
Chris@10
|
1774 T4X = T4L - T4W;
|
Chris@10
|
1775 T56 = T52 - T55;
|
Chris@10
|
1776 T57 = FMA(KP555570233, T4X, KP831469612 * T56);
|
Chris@10
|
1777 T5b = FNMS(KP831469612, T4X, KP555570233 * T56);
|
Chris@10
|
1778 }
|
Chris@10
|
1779 {
|
Chris@10
|
1780 E T4f, T58, T8v, T8y;
|
Chris@10
|
1781 T4f = T3R + T4e;
|
Chris@10
|
1782 T58 = T4G + T57;
|
Chris@10
|
1783 cr[WS(rs, 13)] = T4f - T58;
|
Chris@10
|
1784 ci[WS(rs, 2)] = T4f + T58;
|
Chris@10
|
1785 T8v = T5b - T5a;
|
Chris@10
|
1786 T8y = T8w + T8x;
|
Chris@10
|
1787 cr[WS(rs, 29)] = T8v - T8y;
|
Chris@10
|
1788 ci[WS(rs, 18)] = T8v + T8y;
|
Chris@10
|
1789 }
|
Chris@10
|
1790 {
|
Chris@10
|
1791 E T8z, T8A, T59, T5c;
|
Chris@10
|
1792 T8z = T57 - T4G;
|
Chris@10
|
1793 T8A = T8x - T8w;
|
Chris@10
|
1794 cr[WS(rs, 21)] = T8z - T8A;
|
Chris@10
|
1795 ci[WS(rs, 26)] = T8z + T8A;
|
Chris@10
|
1796 T59 = T3R - T4e;
|
Chris@10
|
1797 T5c = T5a + T5b;
|
Chris@10
|
1798 ci[WS(rs, 10)] = T59 - T5c;
|
Chris@10
|
1799 cr[WS(rs, 5)] = T59 + T5c;
|
Chris@10
|
1800 }
|
Chris@10
|
1801 {
|
Chris@10
|
1802 E T5h, T5o, T8n, T8s;
|
Chris@10
|
1803 T5h = T5d + T5g;
|
Chris@10
|
1804 T5o = T5k + T5n;
|
Chris@10
|
1805 ci[WS(rs, 14)] = T5h - T5o;
|
Chris@10
|
1806 cr[WS(rs, 1)] = T5h + T5o;
|
Chris@10
|
1807 T8n = T5r - T5q;
|
Chris@10
|
1808 T8s = T8o + T8r;
|
Chris@10
|
1809 cr[WS(rs, 17)] = T8n - T8s;
|
Chris@10
|
1810 ci[WS(rs, 30)] = T8n + T8s;
|
Chris@10
|
1811 }
|
Chris@10
|
1812 {
|
Chris@10
|
1813 E T8t, T8u, T5p, T5s;
|
Chris@10
|
1814 T8t = T5n - T5k;
|
Chris@10
|
1815 T8u = T8r - T8o;
|
Chris@10
|
1816 cr[WS(rs, 25)] = T8t - T8u;
|
Chris@10
|
1817 ci[WS(rs, 22)] = T8t + T8u;
|
Chris@10
|
1818 T5p = T5d - T5g;
|
Chris@10
|
1819 T5s = T5q + T5r;
|
Chris@10
|
1820 cr[WS(rs, 9)] = T5p - T5s;
|
Chris@10
|
1821 ci[WS(rs, 6)] = T5p + T5s;
|
Chris@10
|
1822 }
|
Chris@10
|
1823 }
|
Chris@10
|
1824 }
|
Chris@10
|
1825 }
|
Chris@10
|
1826 }
|
Chris@10
|
1827 }
|
Chris@10
|
1828
|
Chris@10
|
1829 static const tw_instr twinstr[] = {
|
Chris@10
|
1830 {TW_CEXP, 1, 1},
|
Chris@10
|
1831 {TW_CEXP, 1, 3},
|
Chris@10
|
1832 {TW_CEXP, 1, 9},
|
Chris@10
|
1833 {TW_CEXP, 1, 27},
|
Chris@10
|
1834 {TW_NEXT, 1, 0}
|
Chris@10
|
1835 };
|
Chris@10
|
1836
|
Chris@10
|
1837 static const hc2hc_desc desc = { 32, "hf2_32", twinstr, &GENUS, {376, 168, 112, 0} };
|
Chris@10
|
1838
|
Chris@10
|
1839 void X(codelet_hf2_32) (planner *p) {
|
Chris@10
|
1840 X(khc2hc_register) (p, hf2_32, &desc);
|
Chris@10
|
1841 }
|
Chris@10
|
1842 #endif /* HAVE_FMA */
|