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 07:40:48 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_hc2cdft.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hc2cfdft_20 -include hc2cf.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 286 FP additions, 188 FP multiplications,
|
Chris@10
|
32 * (or, 176 additions, 78 multiplications, 110 fused multiply/add),
|
Chris@10
|
33 * 174 stack variables, 5 constants, and 80 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hc2cf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hc2cfdft_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
42 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
43 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
44 {
|
Chris@10
|
45 INT m;
|
Chris@10
|
46 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 38, MAKE_VOLATILE_STRIDE(80, rs)) {
|
Chris@10
|
47 E T4X, T5i, T5k, T5e, T5c, T5d, T5j, T5f;
|
Chris@10
|
48 {
|
Chris@10
|
49 E T2E, T4W, T3v, T4k, T2M, T3w, T4V, T4j, T2p, T2T, T5a, T5A, T3D, T3o, T4b;
|
Chris@10
|
50 E T4B, T1Y, T2S, T5z, T57, T3h, T3C, T4A, T44, TH, T2P, T50, T5x, T3z, T32;
|
Chris@10
|
51 E T3P, T4D, T3V, T3U, T5w, T53, T2Q, T1o, T3A, T39;
|
Chris@10
|
52 {
|
Chris@10
|
53 E T1V, T9, T2w, Tu, T1, T6, T1R, T1U, T1T, T2Y, T5, T40, T2F, T10, T2C;
|
Chris@10
|
54 E TE, TX, T2m, T1y, T4g, TS, T33, TW, Tw, TB, T2y, T2B, TA, T3L, T2A;
|
Chris@10
|
55 E T3t, T1q, T1v, T2i, T2l, T2k, T3d, T1u, T48, Tm, Tr, T2s, T2v, T2u, T3J;
|
Chris@10
|
56 E Tq, T3r, T20, T1g, T23, T1l, T1h, T3S, T3k, T21, T2H, TL, T2K, TQ, TM;
|
Chris@10
|
57 E T35, T4h, T2I, T2f, T2g, T1I, T1D, T2c, T46, T2e, T3b, T1E, T28, T16, T29;
|
Chris@10
|
58 E T1b, T25, T3i, T27, T3Q, T17, T1O, T1P, Tj, T1M, Te, T1L, Tb, T3Y, TV;
|
Chris@10
|
59 E T1d, T1Z;
|
Chris@10
|
60 {
|
Chris@10
|
61 E T1S, T4, T7, T8;
|
Chris@10
|
62 T7 = Rp[WS(rs, 9)];
|
Chris@10
|
63 T8 = Rm[WS(rs, 9)];
|
Chris@10
|
64 {
|
Chris@10
|
65 E Ts, Tt, T2, T3;
|
Chris@10
|
66 Ts = Rp[WS(rs, 2)];
|
Chris@10
|
67 Tt = Rm[WS(rs, 2)];
|
Chris@10
|
68 T2 = Ip[WS(rs, 9)];
|
Chris@10
|
69 T1V = T7 + T8;
|
Chris@10
|
70 T9 = T7 - T8;
|
Chris@10
|
71 T2w = Ts - Tt;
|
Chris@10
|
72 Tu = Ts + Tt;
|
Chris@10
|
73 T3 = Im[WS(rs, 9)];
|
Chris@10
|
74 T1 = W[36];
|
Chris@10
|
75 T6 = W[37];
|
Chris@10
|
76 T1R = W[34];
|
Chris@10
|
77 T1S = T2 - T3;
|
Chris@10
|
78 T4 = T2 + T3;
|
Chris@10
|
79 T1U = W[35];
|
Chris@10
|
80 }
|
Chris@10
|
81 {
|
Chris@10
|
82 E TY, TZ, TC, TD;
|
Chris@10
|
83 TY = Ip[0];
|
Chris@10
|
84 T1T = T1R * T1S;
|
Chris@10
|
85 T2Y = T6 * T4;
|
Chris@10
|
86 T5 = T1 * T4;
|
Chris@10
|
87 T40 = T1U * T1S;
|
Chris@10
|
88 TZ = Im[0];
|
Chris@10
|
89 TC = Rp[WS(rs, 7)];
|
Chris@10
|
90 TD = Rm[WS(rs, 7)];
|
Chris@10
|
91 {
|
Chris@10
|
92 E T1w, T1x, TT, TU;
|
Chris@10
|
93 T1w = Rp[WS(rs, 1)];
|
Chris@10
|
94 T2F = TY - TZ;
|
Chris@10
|
95 T10 = TY + TZ;
|
Chris@10
|
96 T2C = TC - TD;
|
Chris@10
|
97 TE = TC + TD;
|
Chris@10
|
98 T1x = Rm[WS(rs, 1)];
|
Chris@10
|
99 TT = Rm[0];
|
Chris@10
|
100 TU = Rp[0];
|
Chris@10
|
101 TX = W[0];
|
Chris@10
|
102 T2m = T1w + T1x;
|
Chris@10
|
103 T1y = T1w - T1x;
|
Chris@10
|
104 T4g = TU + TT;
|
Chris@10
|
105 TV = TT - TU;
|
Chris@10
|
106 TS = W[1];
|
Chris@10
|
107 }
|
Chris@10
|
108 }
|
Chris@10
|
109 }
|
Chris@10
|
110 {
|
Chris@10
|
111 E T2j, T1t, T1r, T1s;
|
Chris@10
|
112 {
|
Chris@10
|
113 E Tx, Ty, T2z, Tz;
|
Chris@10
|
114 Tx = Ip[WS(rs, 7)];
|
Chris@10
|
115 Ty = Im[WS(rs, 7)];
|
Chris@10
|
116 T33 = TX * TV;
|
Chris@10
|
117 TW = TS * TV;
|
Chris@10
|
118 Tw = W[26];
|
Chris@10
|
119 T2z = Tx + Ty;
|
Chris@10
|
120 Tz = Tx - Ty;
|
Chris@10
|
121 TB = W[27];
|
Chris@10
|
122 T2y = W[28];
|
Chris@10
|
123 T2B = W[29];
|
Chris@10
|
124 TA = Tw * Tz;
|
Chris@10
|
125 T3L = TB * Tz;
|
Chris@10
|
126 T2A = T2y * T2z;
|
Chris@10
|
127 T3t = T2B * T2z;
|
Chris@10
|
128 }
|
Chris@10
|
129 T1r = Ip[WS(rs, 1)];
|
Chris@10
|
130 T1s = Im[WS(rs, 1)];
|
Chris@10
|
131 T1q = W[4];
|
Chris@10
|
132 T1v = W[5];
|
Chris@10
|
133 T2i = W[2];
|
Chris@10
|
134 T2j = T1r - T1s;
|
Chris@10
|
135 T1t = T1r + T1s;
|
Chris@10
|
136 T2l = W[3];
|
Chris@10
|
137 {
|
Chris@10
|
138 E T2t, Tp, Tn, To;
|
Chris@10
|
139 Tn = Ip[WS(rs, 2)];
|
Chris@10
|
140 T2k = T2i * T2j;
|
Chris@10
|
141 T3d = T1v * T1t;
|
Chris@10
|
142 T1u = T1q * T1t;
|
Chris@10
|
143 T48 = T2l * T2j;
|
Chris@10
|
144 To = Im[WS(rs, 2)];
|
Chris@10
|
145 Tm = W[6];
|
Chris@10
|
146 Tr = W[7];
|
Chris@10
|
147 T2s = W[8];
|
Chris@10
|
148 T2t = Tn + To;
|
Chris@10
|
149 Tp = Tn - To;
|
Chris@10
|
150 T2v = W[9];
|
Chris@10
|
151 {
|
Chris@10
|
152 E T1e, T1f, T1j, T1k;
|
Chris@10
|
153 T1e = Ip[WS(rs, 3)];
|
Chris@10
|
154 T2u = T2s * T2t;
|
Chris@10
|
155 T3J = Tr * Tp;
|
Chris@10
|
156 Tq = Tm * Tp;
|
Chris@10
|
157 T3r = T2v * T2t;
|
Chris@10
|
158 T1f = Im[WS(rs, 3)];
|
Chris@10
|
159 T1j = Rp[WS(rs, 3)];
|
Chris@10
|
160 T1k = Rm[WS(rs, 3)];
|
Chris@10
|
161 T1d = W[10];
|
Chris@10
|
162 T20 = T1e + T1f;
|
Chris@10
|
163 T1g = T1e - T1f;
|
Chris@10
|
164 T23 = T1j - T1k;
|
Chris@10
|
165 T1l = T1j + T1k;
|
Chris@10
|
166 T1Z = W[12];
|
Chris@10
|
167 T1h = T1d * T1g;
|
Chris@10
|
168 }
|
Chris@10
|
169 }
|
Chris@10
|
170 }
|
Chris@10
|
171 {
|
Chris@10
|
172 E T2d, T1A, TI, T2G, T26, T13;
|
Chris@10
|
173 {
|
Chris@10
|
174 E TJ, TK, TO, TP;
|
Chris@10
|
175 TJ = Ip[WS(rs, 5)];
|
Chris@10
|
176 T3S = T1d * T1l;
|
Chris@10
|
177 T3k = T1Z * T23;
|
Chris@10
|
178 T21 = T1Z * T20;
|
Chris@10
|
179 TK = Im[WS(rs, 5)];
|
Chris@10
|
180 TO = Rp[WS(rs, 5)];
|
Chris@10
|
181 TP = Rm[WS(rs, 5)];
|
Chris@10
|
182 TI = W[20];
|
Chris@10
|
183 T2H = TJ - TK;
|
Chris@10
|
184 TL = TJ + TK;
|
Chris@10
|
185 T2K = TO + TP;
|
Chris@10
|
186 TQ = TO - TP;
|
Chris@10
|
187 T2G = W[18];
|
Chris@10
|
188 TM = TI * TL;
|
Chris@10
|
189 }
|
Chris@10
|
190 {
|
Chris@10
|
191 E T1G, T1H, T1B, T1C;
|
Chris@10
|
192 T1G = Rm[WS(rs, 6)];
|
Chris@10
|
193 T35 = TI * TQ;
|
Chris@10
|
194 T4h = T2G * T2K;
|
Chris@10
|
195 T2I = T2G * T2H;
|
Chris@10
|
196 T1H = Rp[WS(rs, 6)];
|
Chris@10
|
197 T1B = Ip[WS(rs, 6)];
|
Chris@10
|
198 T1C = Im[WS(rs, 6)];
|
Chris@10
|
199 T2f = W[23];
|
Chris@10
|
200 T2g = T1H + T1G;
|
Chris@10
|
201 T1I = T1G - T1H;
|
Chris@10
|
202 T2d = T1B - T1C;
|
Chris@10
|
203 T1D = T1B + T1C;
|
Chris@10
|
204 T2c = W[22];
|
Chris@10
|
205 T1A = W[24];
|
Chris@10
|
206 T46 = T2f * T2d;
|
Chris@10
|
207 }
|
Chris@10
|
208 {
|
Chris@10
|
209 E T14, T15, T19, T1a;
|
Chris@10
|
210 T14 = Ip[WS(rs, 8)];
|
Chris@10
|
211 T2e = T2c * T2d;
|
Chris@10
|
212 T3b = T1A * T1I;
|
Chris@10
|
213 T1E = T1A * T1D;
|
Chris@10
|
214 T15 = Im[WS(rs, 8)];
|
Chris@10
|
215 T19 = Rp[WS(rs, 8)];
|
Chris@10
|
216 T1a = Rm[WS(rs, 8)];
|
Chris@10
|
217 T28 = W[32];
|
Chris@10
|
218 T16 = T14 - T15;
|
Chris@10
|
219 T29 = T14 + T15;
|
Chris@10
|
220 T1b = T19 + T1a;
|
Chris@10
|
221 T26 = T1a - T19;
|
Chris@10
|
222 T25 = W[33];
|
Chris@10
|
223 T13 = W[30];
|
Chris@10
|
224 T3i = T28 * T26;
|
Chris@10
|
225 }
|
Chris@10
|
226 {
|
Chris@10
|
227 E Th, Ti, Tc, Td;
|
Chris@10
|
228 Th = Rm[WS(rs, 4)];
|
Chris@10
|
229 T27 = T25 * T26;
|
Chris@10
|
230 T3Q = T13 * T1b;
|
Chris@10
|
231 T17 = T13 * T16;
|
Chris@10
|
232 Ti = Rp[WS(rs, 4)];
|
Chris@10
|
233 Tc = Ip[WS(rs, 4)];
|
Chris@10
|
234 Td = Im[WS(rs, 4)];
|
Chris@10
|
235 T1O = W[15];
|
Chris@10
|
236 T1P = Ti + Th;
|
Chris@10
|
237 Tj = Th - Ti;
|
Chris@10
|
238 T1M = Tc - Td;
|
Chris@10
|
239 Te = Tc + Td;
|
Chris@10
|
240 T1L = W[14];
|
Chris@10
|
241 Tb = W[16];
|
Chris@10
|
242 T3Y = T1O * T1M;
|
Chris@10
|
243 }
|
Chris@10
|
244 }
|
Chris@10
|
245 {
|
Chris@10
|
246 E T1N, T2W, Tf, T2L, T4i;
|
Chris@10
|
247 {
|
Chris@10
|
248 E T2x, T2D, T3s, T3u, T2J;
|
Chris@10
|
249 T2x = FNMS(T2v, T2w, T2u);
|
Chris@10
|
250 T1N = T1L * T1M;
|
Chris@10
|
251 T2W = Tb * Tj;
|
Chris@10
|
252 Tf = Tb * Te;
|
Chris@10
|
253 T2D = FNMS(T2B, T2C, T2A);
|
Chris@10
|
254 T3s = FMA(T2s, T2w, T3r);
|
Chris@10
|
255 T3u = FMA(T2y, T2C, T3t);
|
Chris@10
|
256 T2J = W[19];
|
Chris@10
|
257 T2E = T2x - T2D;
|
Chris@10
|
258 T4W = T2x + T2D;
|
Chris@10
|
259 T3v = T3s + T3u;
|
Chris@10
|
260 T4k = T3u - T3s;
|
Chris@10
|
261 T2L = FNMS(T2J, T2K, T2I);
|
Chris@10
|
262 T4i = FMA(T2J, T2H, T4h);
|
Chris@10
|
263 }
|
Chris@10
|
264 {
|
Chris@10
|
265 E T42, T43, T45, T4a, T3O, T3N;
|
Chris@10
|
266 {
|
Chris@10
|
267 E T2a, T3j, T47, T3l, T24, T2o, T3n, T49, T22, T2h, T2n;
|
Chris@10
|
268 T2a = FMA(T28, T29, T27);
|
Chris@10
|
269 T3j = FNMS(T25, T29, T3i);
|
Chris@10
|
270 T2M = T2F - T2L;
|
Chris@10
|
271 T3w = T2L + T2F;
|
Chris@10
|
272 T4V = T4g + T4i;
|
Chris@10
|
273 T4j = T4g - T4i;
|
Chris@10
|
274 T22 = W[13];
|
Chris@10
|
275 T2h = FNMS(T2f, T2g, T2e);
|
Chris@10
|
276 T2n = FNMS(T2l, T2m, T2k);
|
Chris@10
|
277 T47 = FMA(T2c, T2g, T46);
|
Chris@10
|
278 T3l = FMA(T22, T20, T3k);
|
Chris@10
|
279 T24 = FNMS(T22, T23, T21);
|
Chris@10
|
280 T2o = T2h - T2n;
|
Chris@10
|
281 T3n = T2h + T2n;
|
Chris@10
|
282 T49 = FMA(T2i, T2m, T48);
|
Chris@10
|
283 {
|
Chris@10
|
284 E T2b, T58, T3m, T59;
|
Chris@10
|
285 T2b = T24 - T2a;
|
Chris@10
|
286 T58 = T2a + T24;
|
Chris@10
|
287 T3m = T3j - T3l;
|
Chris@10
|
288 T45 = T3j + T3l;
|
Chris@10
|
289 T4a = T47 - T49;
|
Chris@10
|
290 T59 = T47 + T49;
|
Chris@10
|
291 T2p = T2b - T2o;
|
Chris@10
|
292 T2T = T2b + T2o;
|
Chris@10
|
293 T5a = T58 + T59;
|
Chris@10
|
294 T5A = T59 - T58;
|
Chris@10
|
295 T3D = T3m + T3n;
|
Chris@10
|
296 T3o = T3m - T3n;
|
Chris@10
|
297 }
|
Chris@10
|
298 }
|
Chris@10
|
299 {
|
Chris@10
|
300 E T1z, T3e, T1Q, T3c, T1J, T1W, T3Z, T41, T1F;
|
Chris@10
|
301 T1z = FNMS(T1v, T1y, T1u);
|
Chris@10
|
302 T3e = FMA(T1q, T1y, T3d);
|
Chris@10
|
303 T1F = W[25];
|
Chris@10
|
304 T4b = T45 + T4a;
|
Chris@10
|
305 T4B = T4a - T45;
|
Chris@10
|
306 T1Q = FNMS(T1O, T1P, T1N);
|
Chris@10
|
307 T3c = FNMS(T1F, T1D, T3b);
|
Chris@10
|
308 T1J = FMA(T1F, T1I, T1E);
|
Chris@10
|
309 T1W = FNMS(T1U, T1V, T1T);
|
Chris@10
|
310 T3Z = FMA(T1L, T1P, T3Y);
|
Chris@10
|
311 T41 = FMA(T1R, T1V, T40);
|
Chris@10
|
312 {
|
Chris@10
|
313 E T56, T3g, T55, T1K, T1X, T3f;
|
Chris@10
|
314 T56 = T1J + T1z;
|
Chris@10
|
315 T1K = T1z - T1J;
|
Chris@10
|
316 T3g = T1Q + T1W;
|
Chris@10
|
317 T1X = T1Q - T1W;
|
Chris@10
|
318 T55 = T3Z + T41;
|
Chris@10
|
319 T42 = T3Z - T41;
|
Chris@10
|
320 T1Y = T1K - T1X;
|
Chris@10
|
321 T2S = T1X + T1K;
|
Chris@10
|
322 T43 = T3c + T3e;
|
Chris@10
|
323 T3f = T3c - T3e;
|
Chris@10
|
324 T5z = T55 - T56;
|
Chris@10
|
325 T57 = T55 + T56;
|
Chris@10
|
326 T3h = T3f - T3g;
|
Chris@10
|
327 T3C = T3g + T3f;
|
Chris@10
|
328 }
|
Chris@10
|
329 }
|
Chris@10
|
330 {
|
Chris@10
|
331 E Ta, T2Z, T3K, T2X, Tk, TG, T31, T3M, Tg, Tv, TF;
|
Chris@10
|
332 Ta = FNMS(T6, T9, T5);
|
Chris@10
|
333 T4A = T42 - T43;
|
Chris@10
|
334 T44 = T42 + T43;
|
Chris@10
|
335 T2Z = FMA(T1, T9, T2Y);
|
Chris@10
|
336 Tg = W[17];
|
Chris@10
|
337 Tv = FNMS(Tr, Tu, Tq);
|
Chris@10
|
338 TF = FNMS(TB, TE, TA);
|
Chris@10
|
339 T3K = FMA(Tm, Tu, T3J);
|
Chris@10
|
340 T2X = FNMS(Tg, Te, T2W);
|
Chris@10
|
341 Tk = FMA(Tg, Tj, Tf);
|
Chris@10
|
342 TG = Tv - TF;
|
Chris@10
|
343 T31 = Tv + TF;
|
Chris@10
|
344 T3M = FMA(Tw, TE, T3L);
|
Chris@10
|
345 {
|
Chris@10
|
346 E Tl, T4Z, T30, T4Y;
|
Chris@10
|
347 Tl = Ta - Tk;
|
Chris@10
|
348 T4Z = Tk + Ta;
|
Chris@10
|
349 T30 = T2X - T2Z;
|
Chris@10
|
350 T3O = T2X + T2Z;
|
Chris@10
|
351 T3N = T3K - T3M;
|
Chris@10
|
352 T4Y = T3K + T3M;
|
Chris@10
|
353 TH = Tl - TG;
|
Chris@10
|
354 T2P = TG + Tl;
|
Chris@10
|
355 T50 = T4Y + T4Z;
|
Chris@10
|
356 T5x = T4Y - T4Z;
|
Chris@10
|
357 T3z = T31 + T30;
|
Chris@10
|
358 T32 = T30 - T31;
|
Chris@10
|
359 }
|
Chris@10
|
360 }
|
Chris@10
|
361 {
|
Chris@10
|
362 E T11, T34, T36, TR, T1i, T3R, T1c, TN, T18;
|
Chris@10
|
363 T11 = FMA(TX, T10, TW);
|
Chris@10
|
364 T34 = FNMS(TS, T10, T33);
|
Chris@10
|
365 TN = W[21];
|
Chris@10
|
366 T3P = T3N + T3O;
|
Chris@10
|
367 T4D = T3N - T3O;
|
Chris@10
|
368 T18 = W[31];
|
Chris@10
|
369 T36 = FMA(TN, TL, T35);
|
Chris@10
|
370 TR = FNMS(TN, TQ, TM);
|
Chris@10
|
371 T1i = W[11];
|
Chris@10
|
372 T3R = FMA(T18, T16, T3Q);
|
Chris@10
|
373 T1c = FNMS(T18, T1b, T17);
|
Chris@10
|
374 {
|
Chris@10
|
375 E T52, T12, T3T, T1m;
|
Chris@10
|
376 T52 = TR + T11;
|
Chris@10
|
377 T12 = TR - T11;
|
Chris@10
|
378 T3T = FMA(T1i, T1g, T3S);
|
Chris@10
|
379 T1m = FNMS(T1i, T1l, T1h);
|
Chris@10
|
380 {
|
Chris@10
|
381 E T37, T51, T38, T1n;
|
Chris@10
|
382 T3V = T36 + T34;
|
Chris@10
|
383 T37 = T34 - T36;
|
Chris@10
|
384 T51 = T3R + T3T;
|
Chris@10
|
385 T3U = T3R - T3T;
|
Chris@10
|
386 T38 = T1c + T1m;
|
Chris@10
|
387 T1n = T1c - T1m;
|
Chris@10
|
388 T5w = T51 - T52;
|
Chris@10
|
389 T53 = T51 + T52;
|
Chris@10
|
390 T2Q = T1n + T12;
|
Chris@10
|
391 T1o = T12 - T1n;
|
Chris@10
|
392 T3A = T38 + T37;
|
Chris@10
|
393 T39 = T37 - T38;
|
Chris@10
|
394 }
|
Chris@10
|
395 }
|
Chris@10
|
396 }
|
Chris@10
|
397 }
|
Chris@10
|
398 }
|
Chris@10
|
399 }
|
Chris@10
|
400 {
|
Chris@10
|
401 E T4l, T4m, T4n, T4w, T4u;
|
Chris@10
|
402 {
|
Chris@10
|
403 E T4L, T2O, T3W, T4K, T4I, T4G, T4S, T4U, T4J, T4z, T4H;
|
Chris@10
|
404 {
|
Chris@10
|
405 E T4C, T2N, T4R, T1p, T4E, T2q, T4Q;
|
Chris@10
|
406 T4L = T4A + T4B;
|
Chris@10
|
407 T4C = T4A - T4B;
|
Chris@10
|
408 T2N = T2E + T2M;
|
Chris@10
|
409 T2O = T2M - T2E;
|
Chris@10
|
410 T4R = T1o - TH;
|
Chris@10
|
411 T1p = TH + T1o;
|
Chris@10
|
412 T4E = T3U - T3V;
|
Chris@10
|
413 T3W = T3U + T3V;
|
Chris@10
|
414 T2q = T1Y + T2p;
|
Chris@10
|
415 T4Q = T2p - T1Y;
|
Chris@10
|
416 {
|
Chris@10
|
417 E T4y, T4x, T4F, T2r;
|
Chris@10
|
418 T4F = T4D - T4E;
|
Chris@10
|
419 T4K = T4D + T4E;
|
Chris@10
|
420 T4y = T1p - T2q;
|
Chris@10
|
421 T2r = T1p + T2q;
|
Chris@10
|
422 T4I = FMA(KP618033988, T4C, T4F);
|
Chris@10
|
423 T4G = FNMS(KP618033988, T4F, T4C);
|
Chris@10
|
424 T4S = FNMS(KP618033988, T4R, T4Q);
|
Chris@10
|
425 T4U = FMA(KP618033988, T4Q, T4R);
|
Chris@10
|
426 Im[WS(rs, 4)] = KP500000000 * (T2r - T2N);
|
Chris@10
|
427 T4x = FMA(KP250000000, T2r, T2N);
|
Chris@10
|
428 T4J = T4j - T4k;
|
Chris@10
|
429 T4l = T4j + T4k;
|
Chris@10
|
430 T4z = FMA(KP559016994, T4y, T4x);
|
Chris@10
|
431 T4H = FNMS(KP559016994, T4y, T4x);
|
Chris@10
|
432 }
|
Chris@10
|
433 }
|
Chris@10
|
434 {
|
Chris@10
|
435 E T2R, T4s, T4d, T4f, T4t, T2U, T4P, T4T;
|
Chris@10
|
436 {
|
Chris@10
|
437 E T3X, T4O, T4M, T4c, T4N;
|
Chris@10
|
438 T4m = T3P + T3W;
|
Chris@10
|
439 T3X = T3P - T3W;
|
Chris@10
|
440 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP951056516, T4G, T4z));
|
Chris@10
|
441 Ip[WS(rs, 3)] = KP500000000 * (FNMS(KP951056516, T4G, T4z));
|
Chris@10
|
442 Im[WS(rs, 8)] = -(KP500000000 * (FNMS(KP951056516, T4I, T4H)));
|
Chris@10
|
443 Im[0] = -(KP500000000 * (FMA(KP951056516, T4I, T4H)));
|
Chris@10
|
444 T4O = T4K - T4L;
|
Chris@10
|
445 T4M = T4K + T4L;
|
Chris@10
|
446 T4c = T44 - T4b;
|
Chris@10
|
447 T4n = T44 + T4b;
|
Chris@10
|
448 T2R = T2P + T2Q;
|
Chris@10
|
449 T4s = T2P - T2Q;
|
Chris@10
|
450 Rm[WS(rs, 4)] = KP500000000 * (T4J + T4M);
|
Chris@10
|
451 T4N = FNMS(KP250000000, T4M, T4J);
|
Chris@10
|
452 T4d = FMA(KP618033988, T4c, T3X);
|
Chris@10
|
453 T4f = FNMS(KP618033988, T3X, T4c);
|
Chris@10
|
454 T4t = T2S - T2T;
|
Chris@10
|
455 T2U = T2S + T2T;
|
Chris@10
|
456 T4P = FNMS(KP559016994, T4O, T4N);
|
Chris@10
|
457 T4T = FMA(KP559016994, T4O, T4N);
|
Chris@10
|
458 }
|
Chris@10
|
459 {
|
Chris@10
|
460 E T3H, T3G, T2V, T3I, T4e;
|
Chris@10
|
461 T2V = T2R + T2U;
|
Chris@10
|
462 T3H = T2R - T2U;
|
Chris@10
|
463 Rp[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T4S, T4P));
|
Chris@10
|
464 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T4S, T4P));
|
Chris@10
|
465 Rm[0] = KP500000000 * (FNMS(KP951056516, T4U, T4T));
|
Chris@10
|
466 Rm[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T4U, T4T));
|
Chris@10
|
467 Ip[WS(rs, 5)] = KP500000000 * (T2O + T2V);
|
Chris@10
|
468 T3G = FNMS(KP250000000, T2V, T2O);
|
Chris@10
|
469 T3I = FMA(KP559016994, T3H, T3G);
|
Chris@10
|
470 T4e = FNMS(KP559016994, T3H, T3G);
|
Chris@10
|
471 T4w = FNMS(KP618033988, T4s, T4t);
|
Chris@10
|
472 T4u = FMA(KP618033988, T4t, T4s);
|
Chris@10
|
473 Ip[WS(rs, 9)] = KP500000000 * (FMA(KP951056516, T4d, T3I));
|
Chris@10
|
474 Ip[WS(rs, 1)] = KP500000000 * (FNMS(KP951056516, T4d, T3I));
|
Chris@10
|
475 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP951056516, T4f, T4e)));
|
Chris@10
|
476 Im[WS(rs, 2)] = -(KP500000000 * (FMA(KP951056516, T4f, T4e)));
|
Chris@10
|
477 }
|
Chris@10
|
478 }
|
Chris@10
|
479 }
|
Chris@10
|
480 {
|
Chris@10
|
481 E T3y, T5O, T5Q, T5F, T5K, T5I;
|
Chris@10
|
482 {
|
Chris@10
|
483 E T5G, T5H, T3x, T4q, T5E, T5C, T3a, T5N, T4p, T5M, T3p, T5y, T5B, T4o;
|
Chris@10
|
484 T5G = T5x + T5w;
|
Chris@10
|
485 T5y = T5w - T5x;
|
Chris@10
|
486 T5B = T5z - T5A;
|
Chris@10
|
487 T5H = T5z + T5A;
|
Chris@10
|
488 T3y = T3w - T3v;
|
Chris@10
|
489 T3x = T3v + T3w;
|
Chris@10
|
490 T4q = T4m - T4n;
|
Chris@10
|
491 T4o = T4m + T4n;
|
Chris@10
|
492 T5E = FMA(KP618033988, T5y, T5B);
|
Chris@10
|
493 T5C = FNMS(KP618033988, T5B, T5y);
|
Chris@10
|
494 T3a = T32 + T39;
|
Chris@10
|
495 T5N = T39 - T32;
|
Chris@10
|
496 Rp[WS(rs, 5)] = KP500000000 * (T4l + T4o);
|
Chris@10
|
497 T4p = FNMS(KP250000000, T4o, T4l);
|
Chris@10
|
498 T5M = T3o - T3h;
|
Chris@10
|
499 T3p = T3h + T3o;
|
Chris@10
|
500 {
|
Chris@10
|
501 E T5u, T5t, T4r, T4v, T3q, T5D, T5v;
|
Chris@10
|
502 T4r = FMA(KP559016994, T4q, T4p);
|
Chris@10
|
503 T4v = FNMS(KP559016994, T4q, T4p);
|
Chris@10
|
504 T5u = T3p - T3a;
|
Chris@10
|
505 T3q = T3a + T3p;
|
Chris@10
|
506 Rp[WS(rs, 9)] = KP500000000 * (FNMS(KP951056516, T4u, T4r));
|
Chris@10
|
507 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T4u, T4r));
|
Chris@10
|
508 Rm[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T4w, T4v));
|
Chris@10
|
509 Rm[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T4w, T4v));
|
Chris@10
|
510 Im[WS(rs, 9)] = KP500000000 * (T3q - T3x);
|
Chris@10
|
511 T5t = FMA(KP250000000, T3q, T3x);
|
Chris@10
|
512 T5O = FNMS(KP618033988, T5N, T5M);
|
Chris@10
|
513 T5Q = FMA(KP618033988, T5M, T5N);
|
Chris@10
|
514 T5F = T4V - T4W;
|
Chris@10
|
515 T4X = T4V + T4W;
|
Chris@10
|
516 T5D = FNMS(KP559016994, T5u, T5t);
|
Chris@10
|
517 T5v = FMA(KP559016994, T5u, T5t);
|
Chris@10
|
518 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP951056516, T5C, T5v)));
|
Chris@10
|
519 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T5C, T5v));
|
Chris@10
|
520 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP951056516, T5E, T5D)));
|
Chris@10
|
521 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP951056516, T5E, T5D));
|
Chris@10
|
522 T5K = T5G - T5H;
|
Chris@10
|
523 T5I = T5G + T5H;
|
Chris@10
|
524 }
|
Chris@10
|
525 }
|
Chris@10
|
526 {
|
Chris@10
|
527 E T54, T5b, T5s, T5q, T5g, T5h, T3F, T5m, T5o, T5p, T5J, T5l, T5r, T5n;
|
Chris@10
|
528 T54 = T50 + T53;
|
Chris@10
|
529 T5o = T50 - T53;
|
Chris@10
|
530 T5p = T5a - T57;
|
Chris@10
|
531 T5b = T57 + T5a;
|
Chris@10
|
532 Rm[WS(rs, 9)] = KP500000000 * (T5F + T5I);
|
Chris@10
|
533 T5J = FNMS(KP250000000, T5I, T5F);
|
Chris@10
|
534 T5s = FMA(KP618033988, T5o, T5p);
|
Chris@10
|
535 T5q = FNMS(KP618033988, T5p, T5o);
|
Chris@10
|
536 {
|
Chris@10
|
537 E T5L, T5P, T3B, T3E;
|
Chris@10
|
538 T5L = FNMS(KP559016994, T5K, T5J);
|
Chris@10
|
539 T5P = FMA(KP559016994, T5K, T5J);
|
Chris@10
|
540 T3B = T3z + T3A;
|
Chris@10
|
541 T5g = T3z - T3A;
|
Chris@10
|
542 T5h = T3C - T3D;
|
Chris@10
|
543 T3E = T3C + T3D;
|
Chris@10
|
544 Rm[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T5O, T5L));
|
Chris@10
|
545 Rp[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T5O, T5L));
|
Chris@10
|
546 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP951056516, T5Q, T5P));
|
Chris@10
|
547 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T5Q, T5P));
|
Chris@10
|
548 T3F = T3B + T3E;
|
Chris@10
|
549 T5m = T3B - T3E;
|
Chris@10
|
550 }
|
Chris@10
|
551 Ip[0] = KP500000000 * (T3y + T3F);
|
Chris@10
|
552 T5l = FNMS(KP250000000, T3F, T3y);
|
Chris@10
|
553 T5i = FMA(KP618033988, T5h, T5g);
|
Chris@10
|
554 T5k = FNMS(KP618033988, T5g, T5h);
|
Chris@10
|
555 T5r = FNMS(KP559016994, T5m, T5l);
|
Chris@10
|
556 T5n = FMA(KP559016994, T5m, T5l);
|
Chris@10
|
557 Im[WS(rs, 3)] = -(KP500000000 * (FNMS(KP951056516, T5q, T5n)));
|
Chris@10
|
558 Ip[WS(rs, 4)] = KP500000000 * (FMA(KP951056516, T5q, T5n));
|
Chris@10
|
559 Im[WS(rs, 7)] = -(KP500000000 * (FNMS(KP951056516, T5s, T5r)));
|
Chris@10
|
560 Ip[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5s, T5r));
|
Chris@10
|
561 T5e = T54 - T5b;
|
Chris@10
|
562 T5c = T54 + T5b;
|
Chris@10
|
563 }
|
Chris@10
|
564 }
|
Chris@10
|
565 }
|
Chris@10
|
566 }
|
Chris@10
|
567 Rp[0] = KP500000000 * (T4X + T5c);
|
Chris@10
|
568 T5d = FNMS(KP250000000, T5c, T4X);
|
Chris@10
|
569 T5j = FNMS(KP559016994, T5e, T5d);
|
Chris@10
|
570 T5f = FMA(KP559016994, T5e, T5d);
|
Chris@10
|
571 Rm[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T5i, T5f));
|
Chris@10
|
572 Rp[WS(rs, 4)] = KP500000000 * (FNMS(KP951056516, T5i, T5f));
|
Chris@10
|
573 Rm[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T5k, T5j));
|
Chris@10
|
574 Rp[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5k, T5j));
|
Chris@10
|
575 }
|
Chris@10
|
576 }
|
Chris@10
|
577 }
|
Chris@10
|
578
|
Chris@10
|
579 static const tw_instr twinstr[] = {
|
Chris@10
|
580 {TW_FULL, 1, 20},
|
Chris@10
|
581 {TW_NEXT, 1, 0}
|
Chris@10
|
582 };
|
Chris@10
|
583
|
Chris@10
|
584 static const hc2c_desc desc = { 20, "hc2cfdft_20", twinstr, &GENUS, {176, 78, 110, 0} };
|
Chris@10
|
585
|
Chris@10
|
586 void X(codelet_hc2cfdft_20) (planner *p) {
|
Chris@10
|
587 X(khc2c_register) (p, hc2cfdft_20, &desc, HC2C_VIA_DFT);
|
Chris@10
|
588 }
|
Chris@10
|
589 #else /* HAVE_FMA */
|
Chris@10
|
590
|
Chris@10
|
591 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hc2cfdft_20 -include hc2cf.h */
|
Chris@10
|
592
|
Chris@10
|
593 /*
|
Chris@10
|
594 * This function contains 286 FP additions, 140 FP multiplications,
|
Chris@10
|
595 * (or, 224 additions, 78 multiplications, 62 fused multiply/add),
|
Chris@10
|
596 * 98 stack variables, 5 constants, and 80 memory accesses
|
Chris@10
|
597 */
|
Chris@10
|
598 #include "hc2cf.h"
|
Chris@10
|
599
|
Chris@10
|
600 static void hc2cfdft_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
601 {
|
Chris@10
|
602 DK(KP125000000, +0.125000000000000000000000000000000000000000000);
|
Chris@10
|
603 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
604 DK(KP279508497, +0.279508497187473712051146708591409529430077295);
|
Chris@10
|
605 DK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
Chris@10
|
606 DK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
Chris@10
|
607 {
|
Chris@10
|
608 INT m;
|
Chris@10
|
609 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 38, MAKE_VOLATILE_STRIDE(80, rs)) {
|
Chris@10
|
610 E T12, T2w, T4o, T4V, T2H, T3a, T4y, T4Y, T1z, T2v, T25, T2y, T2s, T2z, T4v;
|
Chris@10
|
611 E T4X, T4r, T4U, T3A, T3Z, T2X, T37, T3k, T41, T2M, T39, T3v, T3Y, T2S, T36;
|
Chris@10
|
612 E T3p, T42, Td, T4G, T33, T3N, Tw, T4H, T32, T3O;
|
Chris@10
|
613 {
|
Chris@10
|
614 E T3, T3L, T1x, T2V, Th, Tl, TC, T3g, Tq, Tu, TH, T3h, T7, Tb, T1q;
|
Chris@10
|
615 E T2U, TR, T2P, T1F, T3r, T23, T2K, T2f, T3y, T1k, T3m, T2q, T2E, T10, T2Q;
|
Chris@10
|
616 E T1K, T3s, T1U, T2J, T2a, T3x, T1b, T3l, T2l, T2D;
|
Chris@10
|
617 {
|
Chris@10
|
618 E T1, T2, T1s, T1u, T1v, T1w, T1r, T1t;
|
Chris@10
|
619 T1 = Ip[0];
|
Chris@10
|
620 T2 = Im[0];
|
Chris@10
|
621 T1s = T1 + T2;
|
Chris@10
|
622 T1u = Rp[0];
|
Chris@10
|
623 T1v = Rm[0];
|
Chris@10
|
624 T1w = T1u - T1v;
|
Chris@10
|
625 T3 = T1 - T2;
|
Chris@10
|
626 T3L = T1u + T1v;
|
Chris@10
|
627 T1r = W[0];
|
Chris@10
|
628 T1t = W[1];
|
Chris@10
|
629 T1x = FNMS(T1t, T1w, T1r * T1s);
|
Chris@10
|
630 T2V = FMA(T1r, T1w, T1t * T1s);
|
Chris@10
|
631 }
|
Chris@10
|
632 {
|
Chris@10
|
633 E Tf, Tg, Tz, Tj, Tk, TB, Ty, TA;
|
Chris@10
|
634 Tf = Ip[WS(rs, 2)];
|
Chris@10
|
635 Tg = Im[WS(rs, 2)];
|
Chris@10
|
636 Tz = Tf - Tg;
|
Chris@10
|
637 Tj = Rp[WS(rs, 2)];
|
Chris@10
|
638 Tk = Rm[WS(rs, 2)];
|
Chris@10
|
639 TB = Tj + Tk;
|
Chris@10
|
640 Th = Tf + Tg;
|
Chris@10
|
641 Tl = Tj - Tk;
|
Chris@10
|
642 Ty = W[6];
|
Chris@10
|
643 TA = W[7];
|
Chris@10
|
644 TC = FNMS(TA, TB, Ty * Tz);
|
Chris@10
|
645 T3g = FMA(TA, Tz, Ty * TB);
|
Chris@10
|
646 }
|
Chris@10
|
647 {
|
Chris@10
|
648 E To, Tp, TE, Ts, Tt, TG, TD, TF;
|
Chris@10
|
649 To = Ip[WS(rs, 7)];
|
Chris@10
|
650 Tp = Im[WS(rs, 7)];
|
Chris@10
|
651 TE = To - Tp;
|
Chris@10
|
652 Ts = Rp[WS(rs, 7)];
|
Chris@10
|
653 Tt = Rm[WS(rs, 7)];
|
Chris@10
|
654 TG = Ts + Tt;
|
Chris@10
|
655 Tq = To + Tp;
|
Chris@10
|
656 Tu = Ts - Tt;
|
Chris@10
|
657 TD = W[26];
|
Chris@10
|
658 TF = W[27];
|
Chris@10
|
659 TH = FNMS(TF, TG, TD * TE);
|
Chris@10
|
660 T3h = FMA(TF, TE, TD * TG);
|
Chris@10
|
661 }
|
Chris@10
|
662 {
|
Chris@10
|
663 E T5, T6, T1n, T9, Ta, T1p, T1m, T1o;
|
Chris@10
|
664 T5 = Ip[WS(rs, 5)];
|
Chris@10
|
665 T6 = Im[WS(rs, 5)];
|
Chris@10
|
666 T1n = T5 + T6;
|
Chris@10
|
667 T9 = Rp[WS(rs, 5)];
|
Chris@10
|
668 Ta = Rm[WS(rs, 5)];
|
Chris@10
|
669 T1p = T9 - Ta;
|
Chris@10
|
670 T7 = T5 - T6;
|
Chris@10
|
671 Tb = T9 + Ta;
|
Chris@10
|
672 T1m = W[20];
|
Chris@10
|
673 T1o = W[21];
|
Chris@10
|
674 T1q = FNMS(T1o, T1p, T1m * T1n);
|
Chris@10
|
675 T2U = FMA(T1m, T1p, T1o * T1n);
|
Chris@10
|
676 }
|
Chris@10
|
677 {
|
Chris@10
|
678 E TM, T1C, TQ, T1E;
|
Chris@10
|
679 {
|
Chris@10
|
680 E TK, TL, TO, TP;
|
Chris@10
|
681 TK = Ip[WS(rs, 4)];
|
Chris@10
|
682 TL = Im[WS(rs, 4)];
|
Chris@10
|
683 TM = TK + TL;
|
Chris@10
|
684 T1C = TK - TL;
|
Chris@10
|
685 TO = Rp[WS(rs, 4)];
|
Chris@10
|
686 TP = Rm[WS(rs, 4)];
|
Chris@10
|
687 TQ = TO - TP;
|
Chris@10
|
688 T1E = TO + TP;
|
Chris@10
|
689 }
|
Chris@10
|
690 {
|
Chris@10
|
691 E TJ, TN, T1B, T1D;
|
Chris@10
|
692 TJ = W[16];
|
Chris@10
|
693 TN = W[17];
|
Chris@10
|
694 TR = FNMS(TN, TQ, TJ * TM);
|
Chris@10
|
695 T2P = FMA(TN, TM, TJ * TQ);
|
Chris@10
|
696 T1B = W[14];
|
Chris@10
|
697 T1D = W[15];
|
Chris@10
|
698 T1F = FNMS(T1D, T1E, T1B * T1C);
|
Chris@10
|
699 T3r = FMA(T1D, T1C, T1B * T1E);
|
Chris@10
|
700 }
|
Chris@10
|
701 }
|
Chris@10
|
702 {
|
Chris@10
|
703 E T1Y, T2c, T22, T2e;
|
Chris@10
|
704 {
|
Chris@10
|
705 E T1W, T1X, T20, T21;
|
Chris@10
|
706 T1W = Ip[WS(rs, 1)];
|
Chris@10
|
707 T1X = Im[WS(rs, 1)];
|
Chris@10
|
708 T1Y = T1W + T1X;
|
Chris@10
|
709 T2c = T1W - T1X;
|
Chris@10
|
710 T20 = Rp[WS(rs, 1)];
|
Chris@10
|
711 T21 = Rm[WS(rs, 1)];
|
Chris@10
|
712 T22 = T20 - T21;
|
Chris@10
|
713 T2e = T20 + T21;
|
Chris@10
|
714 }
|
Chris@10
|
715 {
|
Chris@10
|
716 E T1V, T1Z, T2b, T2d;
|
Chris@10
|
717 T1V = W[4];
|
Chris@10
|
718 T1Z = W[5];
|
Chris@10
|
719 T23 = FNMS(T1Z, T22, T1V * T1Y);
|
Chris@10
|
720 T2K = FMA(T1Z, T1Y, T1V * T22);
|
Chris@10
|
721 T2b = W[2];
|
Chris@10
|
722 T2d = W[3];
|
Chris@10
|
723 T2f = FNMS(T2d, T2e, T2b * T2c);
|
Chris@10
|
724 T3y = FMA(T2d, T2c, T2b * T2e);
|
Chris@10
|
725 }
|
Chris@10
|
726 }
|
Chris@10
|
727 {
|
Chris@10
|
728 E T1f, T2n, T1j, T2p;
|
Chris@10
|
729 {
|
Chris@10
|
730 E T1d, T1e, T1h, T1i;
|
Chris@10
|
731 T1d = Ip[WS(rs, 3)];
|
Chris@10
|
732 T1e = Im[WS(rs, 3)];
|
Chris@10
|
733 T1f = T1d - T1e;
|
Chris@10
|
734 T2n = T1d + T1e;
|
Chris@10
|
735 T1h = Rp[WS(rs, 3)];
|
Chris@10
|
736 T1i = Rm[WS(rs, 3)];
|
Chris@10
|
737 T1j = T1h + T1i;
|
Chris@10
|
738 T2p = T1h - T1i;
|
Chris@10
|
739 }
|
Chris@10
|
740 {
|
Chris@10
|
741 E T1c, T1g, T2m, T2o;
|
Chris@10
|
742 T1c = W[10];
|
Chris@10
|
743 T1g = W[11];
|
Chris@10
|
744 T1k = FNMS(T1g, T1j, T1c * T1f);
|
Chris@10
|
745 T3m = FMA(T1c, T1j, T1g * T1f);
|
Chris@10
|
746 T2m = W[12];
|
Chris@10
|
747 T2o = W[13];
|
Chris@10
|
748 T2q = FNMS(T2o, T2p, T2m * T2n);
|
Chris@10
|
749 T2E = FMA(T2m, T2p, T2o * T2n);
|
Chris@10
|
750 }
|
Chris@10
|
751 }
|
Chris@10
|
752 {
|
Chris@10
|
753 E TV, T1H, TZ, T1J;
|
Chris@10
|
754 {
|
Chris@10
|
755 E TT, TU, TX, TY;
|
Chris@10
|
756 TT = Ip[WS(rs, 9)];
|
Chris@10
|
757 TU = Im[WS(rs, 9)];
|
Chris@10
|
758 TV = TT + TU;
|
Chris@10
|
759 T1H = TT - TU;
|
Chris@10
|
760 TX = Rp[WS(rs, 9)];
|
Chris@10
|
761 TY = Rm[WS(rs, 9)];
|
Chris@10
|
762 TZ = TX - TY;
|
Chris@10
|
763 T1J = TX + TY;
|
Chris@10
|
764 }
|
Chris@10
|
765 {
|
Chris@10
|
766 E TS, TW, T1G, T1I;
|
Chris@10
|
767 TS = W[36];
|
Chris@10
|
768 TW = W[37];
|
Chris@10
|
769 T10 = FNMS(TW, TZ, TS * TV);
|
Chris@10
|
770 T2Q = FMA(TW, TV, TS * TZ);
|
Chris@10
|
771 T1G = W[34];
|
Chris@10
|
772 T1I = W[35];
|
Chris@10
|
773 T1K = FNMS(T1I, T1J, T1G * T1H);
|
Chris@10
|
774 T3s = FMA(T1I, T1H, T1G * T1J);
|
Chris@10
|
775 }
|
Chris@10
|
776 }
|
Chris@10
|
777 {
|
Chris@10
|
778 E T1P, T27, T1T, T29;
|
Chris@10
|
779 {
|
Chris@10
|
780 E T1N, T1O, T1R, T1S;
|
Chris@10
|
781 T1N = Ip[WS(rs, 6)];
|
Chris@10
|
782 T1O = Im[WS(rs, 6)];
|
Chris@10
|
783 T1P = T1N + T1O;
|
Chris@10
|
784 T27 = T1N - T1O;
|
Chris@10
|
785 T1R = Rp[WS(rs, 6)];
|
Chris@10
|
786 T1S = Rm[WS(rs, 6)];
|
Chris@10
|
787 T1T = T1R - T1S;
|
Chris@10
|
788 T29 = T1R + T1S;
|
Chris@10
|
789 }
|
Chris@10
|
790 {
|
Chris@10
|
791 E T1M, T1Q, T26, T28;
|
Chris@10
|
792 T1M = W[24];
|
Chris@10
|
793 T1Q = W[25];
|
Chris@10
|
794 T1U = FNMS(T1Q, T1T, T1M * T1P);
|
Chris@10
|
795 T2J = FMA(T1Q, T1P, T1M * T1T);
|
Chris@10
|
796 T26 = W[22];
|
Chris@10
|
797 T28 = W[23];
|
Chris@10
|
798 T2a = FNMS(T28, T29, T26 * T27);
|
Chris@10
|
799 T3x = FMA(T28, T27, T26 * T29);
|
Chris@10
|
800 }
|
Chris@10
|
801 }
|
Chris@10
|
802 {
|
Chris@10
|
803 E T16, T2k, T1a, T2i;
|
Chris@10
|
804 {
|
Chris@10
|
805 E T14, T15, T18, T19;
|
Chris@10
|
806 T14 = Ip[WS(rs, 8)];
|
Chris@10
|
807 T15 = Im[WS(rs, 8)];
|
Chris@10
|
808 T16 = T14 - T15;
|
Chris@10
|
809 T2k = T14 + T15;
|
Chris@10
|
810 T18 = Rp[WS(rs, 8)];
|
Chris@10
|
811 T19 = Rm[WS(rs, 8)];
|
Chris@10
|
812 T1a = T18 + T19;
|
Chris@10
|
813 T2i = T19 - T18;
|
Chris@10
|
814 }
|
Chris@10
|
815 {
|
Chris@10
|
816 E T13, T17, T2h, T2j;
|
Chris@10
|
817 T13 = W[30];
|
Chris@10
|
818 T17 = W[31];
|
Chris@10
|
819 T1b = FNMS(T17, T1a, T13 * T16);
|
Chris@10
|
820 T3l = FMA(T13, T1a, T17 * T16);
|
Chris@10
|
821 T2h = W[33];
|
Chris@10
|
822 T2j = W[32];
|
Chris@10
|
823 T2l = FMA(T2h, T2i, T2j * T2k);
|
Chris@10
|
824 T2D = FNMS(T2h, T2k, T2j * T2i);
|
Chris@10
|
825 }
|
Chris@10
|
826 }
|
Chris@10
|
827 {
|
Chris@10
|
828 E T2g, T2r, T3n, T3o;
|
Chris@10
|
829 {
|
Chris@10
|
830 E TI, T11, T4m, T4n;
|
Chris@10
|
831 TI = TC - TH;
|
Chris@10
|
832 T11 = TR - T10;
|
Chris@10
|
833 T12 = TI - T11;
|
Chris@10
|
834 T2w = TI + T11;
|
Chris@10
|
835 T4m = T3g + T3h;
|
Chris@10
|
836 T4n = TR + T10;
|
Chris@10
|
837 T4o = T4m + T4n;
|
Chris@10
|
838 T4V = T4m - T4n;
|
Chris@10
|
839 }
|
Chris@10
|
840 {
|
Chris@10
|
841 E T2F, T2G, T4w, T4x;
|
Chris@10
|
842 T2F = T2D - T2E;
|
Chris@10
|
843 T2G = T2a + T2f;
|
Chris@10
|
844 T2H = T2F - T2G;
|
Chris@10
|
845 T3a = T2F + T2G;
|
Chris@10
|
846 T4w = T2l + T2q;
|
Chris@10
|
847 T4x = T3x + T3y;
|
Chris@10
|
848 T4y = T4w + T4x;
|
Chris@10
|
849 T4Y = T4x - T4w;
|
Chris@10
|
850 }
|
Chris@10
|
851 {
|
Chris@10
|
852 E T1l, T1y, T1L, T24;
|
Chris@10
|
853 T1l = T1b - T1k;
|
Chris@10
|
854 T1y = T1q - T1x;
|
Chris@10
|
855 T1z = T1l + T1y;
|
Chris@10
|
856 T2v = T1y - T1l;
|
Chris@10
|
857 T1L = T1F - T1K;
|
Chris@10
|
858 T24 = T1U - T23;
|
Chris@10
|
859 T25 = T1L - T24;
|
Chris@10
|
860 T2y = T1L + T24;
|
Chris@10
|
861 }
|
Chris@10
|
862 T2g = T2a - T2f;
|
Chris@10
|
863 T2r = T2l - T2q;
|
Chris@10
|
864 T2s = T2g - T2r;
|
Chris@10
|
865 T2z = T2r + T2g;
|
Chris@10
|
866 {
|
Chris@10
|
867 E T4t, T4u, T4p, T4q;
|
Chris@10
|
868 T4t = T3r + T3s;
|
Chris@10
|
869 T4u = T1U + T23;
|
Chris@10
|
870 T4v = T4t + T4u;
|
Chris@10
|
871 T4X = T4t - T4u;
|
Chris@10
|
872 T4p = T3l + T3m;
|
Chris@10
|
873 T4q = T1q + T1x;
|
Chris@10
|
874 T4r = T4p + T4q;
|
Chris@10
|
875 T4U = T4p - T4q;
|
Chris@10
|
876 }
|
Chris@10
|
877 {
|
Chris@10
|
878 E T3w, T3z, T2T, T2W;
|
Chris@10
|
879 T3w = T2D + T2E;
|
Chris@10
|
880 T3z = T3x - T3y;
|
Chris@10
|
881 T3A = T3w + T3z;
|
Chris@10
|
882 T3Z = T3z - T3w;
|
Chris@10
|
883 T2T = T1b + T1k;
|
Chris@10
|
884 T2W = T2U + T2V;
|
Chris@10
|
885 T2X = T2T + T2W;
|
Chris@10
|
886 T37 = T2T - T2W;
|
Chris@10
|
887 }
|
Chris@10
|
888 {
|
Chris@10
|
889 E T3i, T3j, T2I, T2L;
|
Chris@10
|
890 T3i = T3g - T3h;
|
Chris@10
|
891 T3j = T2Q - T2P;
|
Chris@10
|
892 T3k = T3i + T3j;
|
Chris@10
|
893 T41 = T3i - T3j;
|
Chris@10
|
894 T2I = T1F + T1K;
|
Chris@10
|
895 T2L = T2J + T2K;
|
Chris@10
|
896 T2M = T2I + T2L;
|
Chris@10
|
897 T39 = T2I - T2L;
|
Chris@10
|
898 }
|
Chris@10
|
899 {
|
Chris@10
|
900 E T3t, T3u, T2O, T2R;
|
Chris@10
|
901 T3t = T3r - T3s;
|
Chris@10
|
902 T3u = T2K - T2J;
|
Chris@10
|
903 T3v = T3t + T3u;
|
Chris@10
|
904 T3Y = T3t - T3u;
|
Chris@10
|
905 T2O = TC + TH;
|
Chris@10
|
906 T2R = T2P + T2Q;
|
Chris@10
|
907 T2S = T2O + T2R;
|
Chris@10
|
908 T36 = T2O - T2R;
|
Chris@10
|
909 }
|
Chris@10
|
910 T3n = T3l - T3m;
|
Chris@10
|
911 T3o = T2U - T2V;
|
Chris@10
|
912 T3p = T3n + T3o;
|
Chris@10
|
913 T42 = T3n - T3o;
|
Chris@10
|
914 {
|
Chris@10
|
915 E Tc, T3M, T4, T8;
|
Chris@10
|
916 T4 = W[18];
|
Chris@10
|
917 T8 = W[19];
|
Chris@10
|
918 Tc = FNMS(T8, Tb, T4 * T7);
|
Chris@10
|
919 T3M = FMA(T4, Tb, T8 * T7);
|
Chris@10
|
920 Td = T3 - Tc;
|
Chris@10
|
921 T4G = T3L + T3M;
|
Chris@10
|
922 T33 = Tc + T3;
|
Chris@10
|
923 T3N = T3L - T3M;
|
Chris@10
|
924 }
|
Chris@10
|
925 {
|
Chris@10
|
926 E Tm, T30, Tv, T31;
|
Chris@10
|
927 {
|
Chris@10
|
928 E Te, Ti, Tn, Tr;
|
Chris@10
|
929 Te = W[8];
|
Chris@10
|
930 Ti = W[9];
|
Chris@10
|
931 Tm = FNMS(Ti, Tl, Te * Th);
|
Chris@10
|
932 T30 = FMA(Ti, Th, Te * Tl);
|
Chris@10
|
933 Tn = W[28];
|
Chris@10
|
934 Tr = W[29];
|
Chris@10
|
935 Tv = FNMS(Tr, Tu, Tn * Tq);
|
Chris@10
|
936 T31 = FMA(Tr, Tq, Tn * Tu);
|
Chris@10
|
937 }
|
Chris@10
|
938 Tw = Tm - Tv;
|
Chris@10
|
939 T4H = Tm + Tv;
|
Chris@10
|
940 T32 = T30 + T31;
|
Chris@10
|
941 T3O = T31 - T30;
|
Chris@10
|
942 }
|
Chris@10
|
943 }
|
Chris@10
|
944 }
|
Chris@10
|
945 {
|
Chris@10
|
946 E T3C, T3E, Tx, T2u, T3d, T3e, T3D, T3f;
|
Chris@10
|
947 {
|
Chris@10
|
948 E T3q, T3B, T1A, T2t;
|
Chris@10
|
949 T3q = T3k - T3p;
|
Chris@10
|
950 T3B = T3v - T3A;
|
Chris@10
|
951 T3C = FMA(KP475528258, T3q, KP293892626 * T3B);
|
Chris@10
|
952 T3E = FNMS(KP293892626, T3q, KP475528258 * T3B);
|
Chris@10
|
953 Tx = Td - Tw;
|
Chris@10
|
954 T1A = T12 + T1z;
|
Chris@10
|
955 T2t = T25 + T2s;
|
Chris@10
|
956 T2u = T1A + T2t;
|
Chris@10
|
957 T3d = KP279508497 * (T1A - T2t);
|
Chris@10
|
958 T3e = FNMS(KP125000000, T2u, KP500000000 * Tx);
|
Chris@10
|
959 }
|
Chris@10
|
960 Ip[WS(rs, 5)] = KP500000000 * (Tx + T2u);
|
Chris@10
|
961 T3D = T3d - T3e;
|
Chris@10
|
962 Im[WS(rs, 2)] = T3D - T3E;
|
Chris@10
|
963 Im[WS(rs, 6)] = T3D + T3E;
|
Chris@10
|
964 T3f = T3d + T3e;
|
Chris@10
|
965 Ip[WS(rs, 1)] = T3f - T3C;
|
Chris@10
|
966 Ip[WS(rs, 9)] = T3f + T3C;
|
Chris@10
|
967 }
|
Chris@10
|
968 {
|
Chris@10
|
969 E T3H, T3T, T3P, T3Q, T3K, T3R, T3U, T3S;
|
Chris@10
|
970 {
|
Chris@10
|
971 E T3F, T3G, T3I, T3J;
|
Chris@10
|
972 T3F = T12 - T1z;
|
Chris@10
|
973 T3G = T25 - T2s;
|
Chris@10
|
974 T3H = FMA(KP475528258, T3F, KP293892626 * T3G);
|
Chris@10
|
975 T3T = FNMS(KP293892626, T3F, KP475528258 * T3G);
|
Chris@10
|
976 T3P = T3N + T3O;
|
Chris@10
|
977 T3I = T3k + T3p;
|
Chris@10
|
978 T3J = T3v + T3A;
|
Chris@10
|
979 T3Q = T3I + T3J;
|
Chris@10
|
980 T3K = KP279508497 * (T3I - T3J);
|
Chris@10
|
981 T3R = FNMS(KP125000000, T3Q, KP500000000 * T3P);
|
Chris@10
|
982 }
|
Chris@10
|
983 Rp[WS(rs, 5)] = KP500000000 * (T3P + T3Q);
|
Chris@10
|
984 T3U = T3R - T3K;
|
Chris@10
|
985 Rm[WS(rs, 6)] = T3T + T3U;
|
Chris@10
|
986 Rm[WS(rs, 2)] = T3U - T3T;
|
Chris@10
|
987 T3S = T3K + T3R;
|
Chris@10
|
988 Rp[WS(rs, 1)] = T3H + T3S;
|
Chris@10
|
989 Rp[WS(rs, 9)] = T3S - T3H;
|
Chris@10
|
990 }
|
Chris@10
|
991 {
|
Chris@10
|
992 E T44, T46, T2C, T2B, T3V, T3W, T45, T3X;
|
Chris@10
|
993 {
|
Chris@10
|
994 E T40, T43, T2x, T2A;
|
Chris@10
|
995 T40 = T3Y - T3Z;
|
Chris@10
|
996 T43 = T41 - T42;
|
Chris@10
|
997 T44 = FNMS(KP293892626, T43, KP475528258 * T40);
|
Chris@10
|
998 T46 = FMA(KP475528258, T43, KP293892626 * T40);
|
Chris@10
|
999 T2C = Tw + Td;
|
Chris@10
|
1000 T2x = T2v - T2w;
|
Chris@10
|
1001 T2A = T2y + T2z;
|
Chris@10
|
1002 T2B = T2x - T2A;
|
Chris@10
|
1003 T3V = FMA(KP500000000, T2C, KP125000000 * T2B);
|
Chris@10
|
1004 T3W = KP279508497 * (T2x + T2A);
|
Chris@10
|
1005 }
|
Chris@10
|
1006 Im[WS(rs, 4)] = KP500000000 * (T2B - T2C);
|
Chris@10
|
1007 T45 = T3W - T3V;
|
Chris@10
|
1008 Im[0] = T45 - T46;
|
Chris@10
|
1009 Im[WS(rs, 8)] = T45 + T46;
|
Chris@10
|
1010 T3X = T3V + T3W;
|
Chris@10
|
1011 Ip[WS(rs, 3)] = T3X - T44;
|
Chris@10
|
1012 Ip[WS(rs, 7)] = T3X + T44;
|
Chris@10
|
1013 }
|
Chris@10
|
1014 {
|
Chris@10
|
1015 E T49, T4h, T4a, T4d, T4e, T4f, T4i, T4g;
|
Chris@10
|
1016 {
|
Chris@10
|
1017 E T47, T48, T4b, T4c;
|
Chris@10
|
1018 T47 = T2y - T2z;
|
Chris@10
|
1019 T48 = T2w + T2v;
|
Chris@10
|
1020 T49 = FNMS(KP293892626, T48, KP475528258 * T47);
|
Chris@10
|
1021 T4h = FMA(KP475528258, T48, KP293892626 * T47);
|
Chris@10
|
1022 T4a = T3N - T3O;
|
Chris@10
|
1023 T4b = T41 + T42;
|
Chris@10
|
1024 T4c = T3Y + T3Z;
|
Chris@10
|
1025 T4d = T4b + T4c;
|
Chris@10
|
1026 T4e = FNMS(KP125000000, T4d, KP500000000 * T4a);
|
Chris@10
|
1027 T4f = KP279508497 * (T4b - T4c);
|
Chris@10
|
1028 }
|
Chris@10
|
1029 Rm[WS(rs, 4)] = KP500000000 * (T4a + T4d);
|
Chris@10
|
1030 T4i = T4f + T4e;
|
Chris@10
|
1031 Rm[WS(rs, 8)] = T4h + T4i;
|
Chris@10
|
1032 Rm[0] = T4i - T4h;
|
Chris@10
|
1033 T4g = T4e - T4f;
|
Chris@10
|
1034 Rp[WS(rs, 3)] = T49 + T4g;
|
Chris@10
|
1035 Rp[WS(rs, 7)] = T4g - T49;
|
Chris@10
|
1036 }
|
Chris@10
|
1037 {
|
Chris@10
|
1038 E T50, T52, T34, T2Z, T4R, T4S, T51, T4T;
|
Chris@10
|
1039 {
|
Chris@10
|
1040 E T4W, T4Z, T2N, T2Y;
|
Chris@10
|
1041 T4W = T4U - T4V;
|
Chris@10
|
1042 T4Z = T4X - T4Y;
|
Chris@10
|
1043 T50 = FNMS(KP293892626, T4Z, KP475528258 * T4W);
|
Chris@10
|
1044 T52 = FMA(KP293892626, T4W, KP475528258 * T4Z);
|
Chris@10
|
1045 T34 = T32 + T33;
|
Chris@10
|
1046 T2N = T2H - T2M;
|
Chris@10
|
1047 T2Y = T2S + T2X;
|
Chris@10
|
1048 T2Z = T2N - T2Y;
|
Chris@10
|
1049 T4R = FMA(KP500000000, T34, KP125000000 * T2Z);
|
Chris@10
|
1050 T4S = KP279508497 * (T2Y + T2N);
|
Chris@10
|
1051 }
|
Chris@10
|
1052 Im[WS(rs, 9)] = KP500000000 * (T2Z - T34);
|
Chris@10
|
1053 T51 = T4R - T4S;
|
Chris@10
|
1054 Ip[WS(rs, 2)] = T51 + T52;
|
Chris@10
|
1055 Im[WS(rs, 1)] = T52 - T51;
|
Chris@10
|
1056 T4T = T4R + T4S;
|
Chris@10
|
1057 Ip[WS(rs, 6)] = T4T + T50;
|
Chris@10
|
1058 Im[WS(rs, 5)] = T50 - T4T;
|
Chris@10
|
1059 }
|
Chris@10
|
1060 {
|
Chris@10
|
1061 E T5c, T5d, T53, T56, T57, T58, T5e, T59;
|
Chris@10
|
1062 {
|
Chris@10
|
1063 E T5a, T5b, T54, T55;
|
Chris@10
|
1064 T5a = T2M + T2H;
|
Chris@10
|
1065 T5b = T2S - T2X;
|
Chris@10
|
1066 T5c = FNMS(KP293892626, T5b, KP475528258 * T5a);
|
Chris@10
|
1067 T5d = FMA(KP475528258, T5b, KP293892626 * T5a);
|
Chris@10
|
1068 T53 = T4G - T4H;
|
Chris@10
|
1069 T54 = T4V + T4U;
|
Chris@10
|
1070 T55 = T4X + T4Y;
|
Chris@10
|
1071 T56 = T54 + T55;
|
Chris@10
|
1072 T57 = FNMS(KP125000000, T56, KP500000000 * T53);
|
Chris@10
|
1073 T58 = KP279508497 * (T54 - T55);
|
Chris@10
|
1074 }
|
Chris@10
|
1075 Rm[WS(rs, 9)] = KP500000000 * (T53 + T56);
|
Chris@10
|
1076 T5e = T58 + T57;
|
Chris@10
|
1077 Rp[WS(rs, 6)] = T5d + T5e;
|
Chris@10
|
1078 Rm[WS(rs, 5)] = T5e - T5d;
|
Chris@10
|
1079 T59 = T57 - T58;
|
Chris@10
|
1080 Rp[WS(rs, 2)] = T59 - T5c;
|
Chris@10
|
1081 Rm[WS(rs, 1)] = T5c + T59;
|
Chris@10
|
1082 }
|
Chris@10
|
1083 {
|
Chris@10
|
1084 E T4A, T4C, T35, T3c, T4j, T4k, T4B, T4l;
|
Chris@10
|
1085 {
|
Chris@10
|
1086 E T4s, T4z, T38, T3b;
|
Chris@10
|
1087 T4s = T4o - T4r;
|
Chris@10
|
1088 T4z = T4v - T4y;
|
Chris@10
|
1089 T4A = FNMS(KP475528258, T4z, KP293892626 * T4s);
|
Chris@10
|
1090 T4C = FMA(KP475528258, T4s, KP293892626 * T4z);
|
Chris@10
|
1091 T35 = T33 - T32;
|
Chris@10
|
1092 T38 = T36 + T37;
|
Chris@10
|
1093 T3b = T39 + T3a;
|
Chris@10
|
1094 T3c = T38 + T3b;
|
Chris@10
|
1095 T4j = FNMS(KP125000000, T3c, KP500000000 * T35);
|
Chris@10
|
1096 T4k = KP279508497 * (T38 - T3b);
|
Chris@10
|
1097 }
|
Chris@10
|
1098 Ip[0] = KP500000000 * (T35 + T3c);
|
Chris@10
|
1099 T4B = T4k + T4j;
|
Chris@10
|
1100 Ip[WS(rs, 4)] = T4B + T4C;
|
Chris@10
|
1101 Im[WS(rs, 3)] = T4C - T4B;
|
Chris@10
|
1102 T4l = T4j - T4k;
|
Chris@10
|
1103 Ip[WS(rs, 8)] = T4l + T4A;
|
Chris@10
|
1104 Im[WS(rs, 7)] = T4A - T4l;
|
Chris@10
|
1105 }
|
Chris@10
|
1106 {
|
Chris@10
|
1107 E T4O, T4P, T4I, T4J, T4F, T4K, T4Q, T4L;
|
Chris@10
|
1108 {
|
Chris@10
|
1109 E T4M, T4N, T4D, T4E;
|
Chris@10
|
1110 T4M = T36 - T37;
|
Chris@10
|
1111 T4N = T39 - T3a;
|
Chris@10
|
1112 T4O = FMA(KP475528258, T4M, KP293892626 * T4N);
|
Chris@10
|
1113 T4P = FNMS(KP293892626, T4M, KP475528258 * T4N);
|
Chris@10
|
1114 T4I = T4G + T4H;
|
Chris@10
|
1115 T4D = T4o + T4r;
|
Chris@10
|
1116 T4E = T4v + T4y;
|
Chris@10
|
1117 T4J = T4D + T4E;
|
Chris@10
|
1118 T4F = KP279508497 * (T4D - T4E);
|
Chris@10
|
1119 T4K = FNMS(KP125000000, T4J, KP500000000 * T4I);
|
Chris@10
|
1120 }
|
Chris@10
|
1121 Rp[0] = KP500000000 * (T4I + T4J);
|
Chris@10
|
1122 T4Q = T4K - T4F;
|
Chris@10
|
1123 Rp[WS(rs, 8)] = T4P + T4Q;
|
Chris@10
|
1124 Rm[WS(rs, 7)] = T4Q - T4P;
|
Chris@10
|
1125 T4L = T4F + T4K;
|
Chris@10
|
1126 Rp[WS(rs, 4)] = T4L - T4O;
|
Chris@10
|
1127 Rm[WS(rs, 3)] = T4O + T4L;
|
Chris@10
|
1128 }
|
Chris@10
|
1129 }
|
Chris@10
|
1130 }
|
Chris@10
|
1131 }
|
Chris@10
|
1132
|
Chris@10
|
1133 static const tw_instr twinstr[] = {
|
Chris@10
|
1134 {TW_FULL, 1, 20},
|
Chris@10
|
1135 {TW_NEXT, 1, 0}
|
Chris@10
|
1136 };
|
Chris@10
|
1137
|
Chris@10
|
1138 static const hc2c_desc desc = { 20, "hc2cfdft_20", twinstr, &GENUS, {224, 78, 62, 0} };
|
Chris@10
|
1139
|
Chris@10
|
1140 void X(codelet_hc2cfdft_20) (planner *p) {
|
Chris@10
|
1141 X(khc2c_register) (p, hc2cfdft_20, &desc, HC2C_VIA_DFT);
|
Chris@10
|
1142 }
|
Chris@10
|
1143 #endif /* HAVE_FMA */
|