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 07:42:09 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 -sign 1 -n 20 -dif -name hc2cbdft2_20 -include hc2cb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 286 FP additions, 148 FP multiplications,
|
Chris@10
|
32 * (or, 176 additions, 38 multiplications, 110 fused multiply/add),
|
Chris@10
|
33 * 122 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hc2cb.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hc2cbdft2_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(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
43 {
|
Chris@10
|
44 INT m;
|
Chris@10
|
45 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
|
46 E T5s, T5v, T5t, T5z, T5q, T5y, T5u, T5A, T5w;
|
Chris@10
|
47 {
|
Chris@10
|
48 E T3T, T27, T2o, T41, T2p, T40, TU, T15, T2Q, T1N, T2L, T1w, T59, T4n, T5e;
|
Chris@10
|
49 E T4A, T2m, T24, T2Z, T2h, T4J, T3P, T3Y, T3W, T2d, TJ, T3H, T2c, TD, T52;
|
Chris@10
|
50 E T3G, T1E, T4f, T5I, T4e, T4w, T5L, T4v, T1J, T1H;
|
Chris@10
|
51 {
|
Chris@10
|
52 E T1A, T3, T25, TI, TF, T6, T26, T1D, TO, T47, T3z, Te, T1S, T3M, T1e;
|
Chris@10
|
53 E T4k, TZ, T4a, T3C, Tt, T1Z, T3J, T1p, T4h, T14, T4b, T3D, TA, T22, T3K;
|
Chris@10
|
54 E T1u, T4i, Ti, T1f, Th, T1T, TS, Tj, T1g, T1h;
|
Chris@10
|
55 {
|
Chris@10
|
56 E T4, T5, T1B, T1C;
|
Chris@10
|
57 {
|
Chris@10
|
58 E T1, T2, TG, TH;
|
Chris@10
|
59 T1 = Rp[0];
|
Chris@10
|
60 T2 = Rm[WS(rs, 9)];
|
Chris@10
|
61 TG = Ip[0];
|
Chris@10
|
62 TH = Im[WS(rs, 9)];
|
Chris@10
|
63 T4 = Rp[WS(rs, 5)];
|
Chris@10
|
64 T1A = T1 - T2;
|
Chris@10
|
65 T3 = T1 + T2;
|
Chris@10
|
66 T25 = TG - TH;
|
Chris@10
|
67 TI = TG + TH;
|
Chris@10
|
68 T5 = Rm[WS(rs, 4)];
|
Chris@10
|
69 T1B = Ip[WS(rs, 5)];
|
Chris@10
|
70 T1C = Im[WS(rs, 4)];
|
Chris@10
|
71 }
|
Chris@10
|
72 {
|
Chris@10
|
73 E Tq, T1l, Tp, T1X, TY, Tr, T1m, T1n;
|
Chris@10
|
74 {
|
Chris@10
|
75 E Tb, T1a, Ta, T1Q, TN, Tc, T1b, T1c;
|
Chris@10
|
76 {
|
Chris@10
|
77 E T8, T9, TL, TM;
|
Chris@10
|
78 T8 = Rp[WS(rs, 4)];
|
Chris@10
|
79 TF = T4 - T5;
|
Chris@10
|
80 T6 = T4 + T5;
|
Chris@10
|
81 T26 = T1B - T1C;
|
Chris@10
|
82 T1D = T1B + T1C;
|
Chris@10
|
83 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
84 TL = Ip[WS(rs, 4)];
|
Chris@10
|
85 TM = Im[WS(rs, 5)];
|
Chris@10
|
86 Tb = Rp[WS(rs, 9)];
|
Chris@10
|
87 T1a = T8 - T9;
|
Chris@10
|
88 Ta = T8 + T9;
|
Chris@10
|
89 T1Q = TL - TM;
|
Chris@10
|
90 TN = TL + TM;
|
Chris@10
|
91 Tc = Rm[0];
|
Chris@10
|
92 T1b = Ip[WS(rs, 9)];
|
Chris@10
|
93 T1c = Im[0];
|
Chris@10
|
94 }
|
Chris@10
|
95 {
|
Chris@10
|
96 E Tn, To, TW, TX;
|
Chris@10
|
97 Tn = Rp[WS(rs, 8)];
|
Chris@10
|
98 {
|
Chris@10
|
99 E TK, Td, T1R, T1d;
|
Chris@10
|
100 TK = Tb - Tc;
|
Chris@10
|
101 Td = Tb + Tc;
|
Chris@10
|
102 T1R = T1b - T1c;
|
Chris@10
|
103 T1d = T1b + T1c;
|
Chris@10
|
104 TO = TK + TN;
|
Chris@10
|
105 T47 = TN - TK;
|
Chris@10
|
106 T3z = Ta - Td;
|
Chris@10
|
107 Te = Ta + Td;
|
Chris@10
|
108 T1S = T1Q + T1R;
|
Chris@10
|
109 T3M = T1Q - T1R;
|
Chris@10
|
110 T1e = T1a - T1d;
|
Chris@10
|
111 T4k = T1a + T1d;
|
Chris@10
|
112 To = Rm[WS(rs, 1)];
|
Chris@10
|
113 }
|
Chris@10
|
114 TW = Ip[WS(rs, 8)];
|
Chris@10
|
115 TX = Im[WS(rs, 1)];
|
Chris@10
|
116 Tq = Rm[WS(rs, 6)];
|
Chris@10
|
117 T1l = Tn - To;
|
Chris@10
|
118 Tp = Tn + To;
|
Chris@10
|
119 T1X = TW - TX;
|
Chris@10
|
120 TY = TW + TX;
|
Chris@10
|
121 Tr = Rp[WS(rs, 3)];
|
Chris@10
|
122 T1m = Im[WS(rs, 6)];
|
Chris@10
|
123 T1n = Ip[WS(rs, 3)];
|
Chris@10
|
124 }
|
Chris@10
|
125 }
|
Chris@10
|
126 {
|
Chris@10
|
127 E Tx, T1q, Tw, T20, T13, Ty, T1r, T1s;
|
Chris@10
|
128 {
|
Chris@10
|
129 E Tu, Tv, T11, T12;
|
Chris@10
|
130 Tu = Rm[WS(rs, 7)];
|
Chris@10
|
131 {
|
Chris@10
|
132 E TV, Ts, T1Y, T1o;
|
Chris@10
|
133 TV = Tq - Tr;
|
Chris@10
|
134 Ts = Tq + Tr;
|
Chris@10
|
135 T1Y = T1n - T1m;
|
Chris@10
|
136 T1o = T1m + T1n;
|
Chris@10
|
137 TZ = TV + TY;
|
Chris@10
|
138 T4a = TY - TV;
|
Chris@10
|
139 T3C = Tp - Ts;
|
Chris@10
|
140 Tt = Tp + Ts;
|
Chris@10
|
141 T1Z = T1X + T1Y;
|
Chris@10
|
142 T3J = T1X - T1Y;
|
Chris@10
|
143 T1p = T1l + T1o;
|
Chris@10
|
144 T4h = T1l - T1o;
|
Chris@10
|
145 Tv = Rp[WS(rs, 2)];
|
Chris@10
|
146 }
|
Chris@10
|
147 T11 = Im[WS(rs, 7)];
|
Chris@10
|
148 T12 = Ip[WS(rs, 2)];
|
Chris@10
|
149 Tx = Rm[WS(rs, 2)];
|
Chris@10
|
150 T1q = Tu - Tv;
|
Chris@10
|
151 Tw = Tu + Tv;
|
Chris@10
|
152 T20 = T12 - T11;
|
Chris@10
|
153 T13 = T11 + T12;
|
Chris@10
|
154 Ty = Rp[WS(rs, 7)];
|
Chris@10
|
155 T1r = Im[WS(rs, 2)];
|
Chris@10
|
156 T1s = Ip[WS(rs, 7)];
|
Chris@10
|
157 }
|
Chris@10
|
158 {
|
Chris@10
|
159 E Tf, Tg, TQ, TR;
|
Chris@10
|
160 Tf = Rm[WS(rs, 3)];
|
Chris@10
|
161 {
|
Chris@10
|
162 E T10, Tz, T21, T1t;
|
Chris@10
|
163 T10 = Tx - Ty;
|
Chris@10
|
164 Tz = Tx + Ty;
|
Chris@10
|
165 T21 = T1s - T1r;
|
Chris@10
|
166 T1t = T1r + T1s;
|
Chris@10
|
167 T14 = T10 - T13;
|
Chris@10
|
168 T4b = T10 + T13;
|
Chris@10
|
169 T3D = Tw - Tz;
|
Chris@10
|
170 TA = Tw + Tz;
|
Chris@10
|
171 T22 = T20 + T21;
|
Chris@10
|
172 T3K = T20 - T21;
|
Chris@10
|
173 T1u = T1q + T1t;
|
Chris@10
|
174 T4i = T1q - T1t;
|
Chris@10
|
175 Tg = Rp[WS(rs, 6)];
|
Chris@10
|
176 }
|
Chris@10
|
177 TQ = Im[WS(rs, 3)];
|
Chris@10
|
178 TR = Ip[WS(rs, 6)];
|
Chris@10
|
179 Ti = Rp[WS(rs, 1)];
|
Chris@10
|
180 T1f = Tf - Tg;
|
Chris@10
|
181 Th = Tf + Tg;
|
Chris@10
|
182 T1T = TR - TQ;
|
Chris@10
|
183 TS = TQ + TR;
|
Chris@10
|
184 Tj = Rm[WS(rs, 8)];
|
Chris@10
|
185 T1g = Ip[WS(rs, 1)];
|
Chris@10
|
186 T1h = Im[WS(rs, 8)];
|
Chris@10
|
187 }
|
Chris@10
|
188 }
|
Chris@10
|
189 }
|
Chris@10
|
190 }
|
Chris@10
|
191 {
|
Chris@10
|
192 E T1V, T3N, TB, T3B, Tm, T3E, T1F, T1G, T4t, T4j, T4m, T4s, T4c, T4y, T4z;
|
Chris@10
|
193 E T49, T3y, T7;
|
Chris@10
|
194 {
|
Chris@10
|
195 E TT, T48, T1j, T4l, T3A, Tl;
|
Chris@10
|
196 T3T = T25 - T26;
|
Chris@10
|
197 T27 = T25 + T26;
|
Chris@10
|
198 {
|
Chris@10
|
199 E TP, Tk, T1U, T1i;
|
Chris@10
|
200 TP = Ti - Tj;
|
Chris@10
|
201 Tk = Ti + Tj;
|
Chris@10
|
202 T1U = T1g - T1h;
|
Chris@10
|
203 T1i = T1g + T1h;
|
Chris@10
|
204 TT = TP - TS;
|
Chris@10
|
205 T48 = TP + TS;
|
Chris@10
|
206 T3A = Th - Tk;
|
Chris@10
|
207 Tl = Th + Tk;
|
Chris@10
|
208 T1V = T1T + T1U;
|
Chris@10
|
209 T3N = T1T - T1U;
|
Chris@10
|
210 T1j = T1f - T1i;
|
Chris@10
|
211 T4l = T1f + T1i;
|
Chris@10
|
212 T2o = Tt - TA;
|
Chris@10
|
213 TB = Tt + TA;
|
Chris@10
|
214 }
|
Chris@10
|
215 T41 = T3z - T3A;
|
Chris@10
|
216 T3B = T3z + T3A;
|
Chris@10
|
217 Tm = Te + Tl;
|
Chris@10
|
218 T2p = Te - Tl;
|
Chris@10
|
219 {
|
Chris@10
|
220 E T1L, T1M, T1k, T1v;
|
Chris@10
|
221 T40 = T3C - T3D;
|
Chris@10
|
222 T3E = T3C + T3D;
|
Chris@10
|
223 TU = TO + TT;
|
Chris@10
|
224 T1L = TO - TT;
|
Chris@10
|
225 T1M = TZ - T14;
|
Chris@10
|
226 T15 = TZ + T14;
|
Chris@10
|
227 T1F = T1e + T1j;
|
Chris@10
|
228 T1k = T1e - T1j;
|
Chris@10
|
229 T1v = T1p - T1u;
|
Chris@10
|
230 T1G = T1p + T1u;
|
Chris@10
|
231 T4t = T4h + T4i;
|
Chris@10
|
232 T4j = T4h - T4i;
|
Chris@10
|
233 T2Q = FNMS(KP618033988, T1L, T1M);
|
Chris@10
|
234 T1N = FMA(KP618033988, T1M, T1L);
|
Chris@10
|
235 T2L = FNMS(KP618033988, T1k, T1v);
|
Chris@10
|
236 T1w = FMA(KP618033988, T1v, T1k);
|
Chris@10
|
237 T4m = T4k - T4l;
|
Chris@10
|
238 T4s = T4k + T4l;
|
Chris@10
|
239 T4c = T4a - T4b;
|
Chris@10
|
240 T4y = T4a + T4b;
|
Chris@10
|
241 T4z = T47 + T48;
|
Chris@10
|
242 T49 = T47 - T48;
|
Chris@10
|
243 }
|
Chris@10
|
244 }
|
Chris@10
|
245 {
|
Chris@10
|
246 E T2g, T1W, T23, T2f;
|
Chris@10
|
247 T2g = T1S - T1V;
|
Chris@10
|
248 T1W = T1S + T1V;
|
Chris@10
|
249 T59 = FMA(KP618033988, T4j, T4m);
|
Chris@10
|
250 T4n = FNMS(KP618033988, T4m, T4j);
|
Chris@10
|
251 T5e = FMA(KP618033988, T4y, T4z);
|
Chris@10
|
252 T4A = FNMS(KP618033988, T4z, T4y);
|
Chris@10
|
253 T23 = T1Z + T22;
|
Chris@10
|
254 T2f = T1Z - T22;
|
Chris@10
|
255 {
|
Chris@10
|
256 E T3V, T3L, T3O, T3U;
|
Chris@10
|
257 T3V = T3J + T3K;
|
Chris@10
|
258 T3L = T3J - T3K;
|
Chris@10
|
259 T2m = T1W - T23;
|
Chris@10
|
260 T24 = T1W + T23;
|
Chris@10
|
261 T2Z = FMA(KP618033988, T2f, T2g);
|
Chris@10
|
262 T2h = FNMS(KP618033988, T2g, T2f);
|
Chris@10
|
263 T3O = T3M - T3N;
|
Chris@10
|
264 T3U = T3M + T3N;
|
Chris@10
|
265 T3y = T3 - T6;
|
Chris@10
|
266 T7 = T3 + T6;
|
Chris@10
|
267 T4J = FMA(KP618033988, T3L, T3O);
|
Chris@10
|
268 T3P = FNMS(KP618033988, T3O, T3L);
|
Chris@10
|
269 T3Y = T3U - T3V;
|
Chris@10
|
270 T3W = T3U + T3V;
|
Chris@10
|
271 }
|
Chris@10
|
272 }
|
Chris@10
|
273 {
|
Chris@10
|
274 E T46, TC, T3F, T4r, T4d, T4u;
|
Chris@10
|
275 TC = Tm + TB;
|
Chris@10
|
276 T2d = Tm - TB;
|
Chris@10
|
277 TJ = TF + TI;
|
Chris@10
|
278 T46 = TI - TF;
|
Chris@10
|
279 T3H = T3B - T3E;
|
Chris@10
|
280 T3F = T3B + T3E;
|
Chris@10
|
281 T2c = FNMS(KP250000000, TC, T7);
|
Chris@10
|
282 TD = T7 + TC;
|
Chris@10
|
283 T52 = T3y + T3F;
|
Chris@10
|
284 T3G = FNMS(KP250000000, T3F, T3y);
|
Chris@10
|
285 T4r = T1A + T1D;
|
Chris@10
|
286 T1E = T1A - T1D;
|
Chris@10
|
287 T4f = T49 - T4c;
|
Chris@10
|
288 T4d = T49 + T4c;
|
Chris@10
|
289 T5I = T46 + T4d;
|
Chris@10
|
290 T4e = FNMS(KP250000000, T4d, T46);
|
Chris@10
|
291 T4w = T4s - T4t;
|
Chris@10
|
292 T4u = T4s + T4t;
|
Chris@10
|
293 T5L = T4u + T4r;
|
Chris@10
|
294 T4v = FNMS(KP250000000, T4u, T4r);
|
Chris@10
|
295 T1J = T1F - T1G;
|
Chris@10
|
296 T1H = T1F + T1G;
|
Chris@10
|
297 }
|
Chris@10
|
298 }
|
Chris@10
|
299 }
|
Chris@10
|
300 {
|
Chris@10
|
301 E T38, T3b, T39, T3f, T36, T3e, T3a;
|
Chris@10
|
302 {
|
Chris@10
|
303 E T28, T3r, T3o, T3v, T3p, T2b, T2k, T35, T3l, T2H, T2r, T2j, T2z, T2D, T2G;
|
Chris@10
|
304 E T2X, T2F, T2T, T32, T3h, T3k, T31, T3d, T3j, T3t, T1x, T2u, T1O, T2x, T2v;
|
Chris@10
|
305 E T1y, T2B, T29, T2J, T2M, T2R, T2N, T2V;
|
Chris@10
|
306 {
|
Chris@10
|
307 E T2l, T1I, T18, T2q, T34, T17, T16, T3n;
|
Chris@10
|
308 T28 = T24 + T27;
|
Chris@10
|
309 T2l = FNMS(KP250000000, T24, T27);
|
Chris@10
|
310 T3r = T1H + T1E;
|
Chris@10
|
311 T1I = FNMS(KP250000000, T1H, T1E);
|
Chris@10
|
312 T18 = TU - T15;
|
Chris@10
|
313 T16 = TU + T15;
|
Chris@10
|
314 T3n = W[8];
|
Chris@10
|
315 T2q = FNMS(KP618033988, T2p, T2o);
|
Chris@10
|
316 T34 = FMA(KP618033988, T2o, T2p);
|
Chris@10
|
317 T17 = FNMS(KP250000000, T16, TJ);
|
Chris@10
|
318 T3o = TJ + T16;
|
Chris@10
|
319 T3v = T3n * T3r;
|
Chris@10
|
320 T3p = T3n * T3o;
|
Chris@10
|
321 {
|
Chris@10
|
322 E T2Y, T2E, T3i, T30;
|
Chris@10
|
323 {
|
Chris@10
|
324 E T2e, T33, T2n, T2i;
|
Chris@10
|
325 T2Y = FMA(KP559016994, T2d, T2c);
|
Chris@10
|
326 T2e = FNMS(KP559016994, T2d, T2c);
|
Chris@10
|
327 T2b = W[14];
|
Chris@10
|
328 T2k = W[15];
|
Chris@10
|
329 T33 = FMA(KP559016994, T2m, T2l);
|
Chris@10
|
330 T2n = FNMS(KP559016994, T2m, T2l);
|
Chris@10
|
331 T2E = FMA(KP951056516, T2h, T2e);
|
Chris@10
|
332 T2i = FNMS(KP951056516, T2h, T2e);
|
Chris@10
|
333 T35 = FMA(KP951056516, T34, T33);
|
Chris@10
|
334 T3l = FNMS(KP951056516, T34, T33);
|
Chris@10
|
335 T2H = FNMS(KP951056516, T2q, T2n);
|
Chris@10
|
336 T2r = FMA(KP951056516, T2q, T2n);
|
Chris@10
|
337 T2j = T2b * T2i;
|
Chris@10
|
338 T2z = T2k * T2i;
|
Chris@10
|
339 T2D = W[22];
|
Chris@10
|
340 T2G = W[23];
|
Chris@10
|
341 }
|
Chris@10
|
342 T2X = W[30];
|
Chris@10
|
343 T2F = T2D * T2E;
|
Chris@10
|
344 T2T = T2G * T2E;
|
Chris@10
|
345 T3i = FMA(KP951056516, T2Z, T2Y);
|
Chris@10
|
346 T30 = FNMS(KP951056516, T2Z, T2Y);
|
Chris@10
|
347 T32 = W[31];
|
Chris@10
|
348 T3h = W[6];
|
Chris@10
|
349 T3k = W[7];
|
Chris@10
|
350 T31 = T2X * T30;
|
Chris@10
|
351 T3d = T32 * T30;
|
Chris@10
|
352 T3j = T3h * T3i;
|
Chris@10
|
353 T3t = T3k * T3i;
|
Chris@10
|
354 }
|
Chris@10
|
355 {
|
Chris@10
|
356 E T2K, T2P, TE, T19, T1K, T2t, T37;
|
Chris@10
|
357 T2K = FNMS(KP559016994, T18, T17);
|
Chris@10
|
358 T19 = FMA(KP559016994, T18, T17);
|
Chris@10
|
359 T1K = FMA(KP559016994, T1J, T1I);
|
Chris@10
|
360 T2P = FNMS(KP559016994, T1J, T1I);
|
Chris@10
|
361 TE = W[0];
|
Chris@10
|
362 T2t = W[16];
|
Chris@10
|
363 T1x = FMA(KP951056516, T1w, T19);
|
Chris@10
|
364 T2u = FNMS(KP951056516, T1w, T19);
|
Chris@10
|
365 T1O = FNMS(KP951056516, T1N, T1K);
|
Chris@10
|
366 T2x = FMA(KP951056516, T1N, T1K);
|
Chris@10
|
367 T2v = T2t * T2u;
|
Chris@10
|
368 T1y = TE * T1x;
|
Chris@10
|
369 T2B = T2t * T2x;
|
Chris@10
|
370 T29 = TE * T1O;
|
Chris@10
|
371 T2J = W[24];
|
Chris@10
|
372 T37 = W[32];
|
Chris@10
|
373 T2M = FMA(KP951056516, T2L, T2K);
|
Chris@10
|
374 T38 = FNMS(KP951056516, T2L, T2K);
|
Chris@10
|
375 T2R = FNMS(KP951056516, T2Q, T2P);
|
Chris@10
|
376 T3b = FMA(KP951056516, T2Q, T2P);
|
Chris@10
|
377 T39 = T37 * T38;
|
Chris@10
|
378 T2N = T2J * T2M;
|
Chris@10
|
379 T3f = T37 * T3b;
|
Chris@10
|
380 }
|
Chris@10
|
381 }
|
Chris@10
|
382 T2V = T2J * T2R;
|
Chris@10
|
383 {
|
Chris@10
|
384 E T3m, T3u, T3q, T2a, T1P, T1z;
|
Chris@10
|
385 T1z = W[1];
|
Chris@10
|
386 T3m = FNMS(T3k, T3l, T3j);
|
Chris@10
|
387 T3u = FMA(T3h, T3l, T3t);
|
Chris@10
|
388 T3q = W[9];
|
Chris@10
|
389 T2a = FNMS(T1z, T1x, T29);
|
Chris@10
|
390 T1P = FMA(T1z, T1O, T1y);
|
Chris@10
|
391 {
|
Chris@10
|
392 E T2s, T2A, T2w, T3w, T3s;
|
Chris@10
|
393 T2s = FNMS(T2k, T2r, T2j);
|
Chris@10
|
394 T3w = FNMS(T3q, T3o, T3v);
|
Chris@10
|
395 T3s = FMA(T3q, T3r, T3p);
|
Chris@10
|
396 Im[0] = T2a - T28;
|
Chris@10
|
397 Ip[0] = T28 + T2a;
|
Chris@10
|
398 Rm[0] = TD + T1P;
|
Chris@10
|
399 Rp[0] = TD - T1P;
|
Chris@10
|
400 Im[WS(rs, 2)] = T3w - T3u;
|
Chris@10
|
401 Ip[WS(rs, 2)] = T3u + T3w;
|
Chris@10
|
402 Rm[WS(rs, 2)] = T3m + T3s;
|
Chris@10
|
403 Rp[WS(rs, 2)] = T3m - T3s;
|
Chris@10
|
404 T2A = FMA(T2b, T2r, T2z);
|
Chris@10
|
405 T2w = W[17];
|
Chris@10
|
406 {
|
Chris@10
|
407 E T2I, T2U, T2O, T2C, T2y, T2W, T2S;
|
Chris@10
|
408 T2I = FNMS(T2G, T2H, T2F);
|
Chris@10
|
409 T2U = FMA(T2D, T2H, T2T);
|
Chris@10
|
410 T2O = W[25];
|
Chris@10
|
411 T2C = FNMS(T2w, T2u, T2B);
|
Chris@10
|
412 T2y = FMA(T2w, T2x, T2v);
|
Chris@10
|
413 T36 = FNMS(T32, T35, T31);
|
Chris@10
|
414 T2W = FNMS(T2O, T2M, T2V);
|
Chris@10
|
415 T2S = FMA(T2O, T2R, T2N);
|
Chris@10
|
416 Im[WS(rs, 4)] = T2C - T2A;
|
Chris@10
|
417 Ip[WS(rs, 4)] = T2A + T2C;
|
Chris@10
|
418 Rm[WS(rs, 4)] = T2s + T2y;
|
Chris@10
|
419 Rp[WS(rs, 4)] = T2s - T2y;
|
Chris@10
|
420 Im[WS(rs, 6)] = T2W - T2U;
|
Chris@10
|
421 Ip[WS(rs, 6)] = T2U + T2W;
|
Chris@10
|
422 Rm[WS(rs, 6)] = T2I + T2S;
|
Chris@10
|
423 Rp[WS(rs, 6)] = T2I - T2S;
|
Chris@10
|
424 T3e = FMA(T2X, T35, T3d);
|
Chris@10
|
425 T3a = W[33];
|
Chris@10
|
426 }
|
Chris@10
|
427 }
|
Chris@10
|
428 }
|
Chris@10
|
429 }
|
Chris@10
|
430 {
|
Chris@10
|
431 E T55, T51, T54, T53, T5h, T5P, T5J, T3x, T4P, T5F, T5p, T43, T3R, T3S, T5l;
|
Chris@10
|
432 E T5o, T4D, T5n, T5x, T4H, T4M, T5B, T5E, T4L, T4X, T5D, T5N, T4S, T4o, T4V;
|
Chris@10
|
433 E T4B, T4T, T4p, T4Z, T4F, T57, T5a, T5f, T5b, T5j;
|
Chris@10
|
434 {
|
Chris@10
|
435 E T3X, T4O, T42, T3g, T3c, T5H;
|
Chris@10
|
436 T55 = T3W + T3T;
|
Chris@10
|
437 T3X = FNMS(KP250000000, T3W, T3T);
|
Chris@10
|
438 T51 = W[18];
|
Chris@10
|
439 T3g = FNMS(T3a, T38, T3f);
|
Chris@10
|
440 T3c = FMA(T3a, T3b, T39);
|
Chris@10
|
441 T54 = W[19];
|
Chris@10
|
442 T53 = T51 * T52;
|
Chris@10
|
443 Im[WS(rs, 8)] = T3g - T3e;
|
Chris@10
|
444 Ip[WS(rs, 8)] = T3e + T3g;
|
Chris@10
|
445 Rm[WS(rs, 8)] = T36 + T3c;
|
Chris@10
|
446 Rp[WS(rs, 8)] = T36 - T3c;
|
Chris@10
|
447 T5h = T54 * T52;
|
Chris@10
|
448 T5H = W[28];
|
Chris@10
|
449 T4O = FMA(KP618033988, T40, T41);
|
Chris@10
|
450 T42 = FNMS(KP618033988, T41, T40);
|
Chris@10
|
451 T5P = T5H * T5L;
|
Chris@10
|
452 T5J = T5H * T5I;
|
Chris@10
|
453 {
|
Chris@10
|
454 E T4I, T5m, T3Q, T3I, T3Z, T4N, T4K, T5C;
|
Chris@10
|
455 T3I = FNMS(KP559016994, T3H, T3G);
|
Chris@10
|
456 T4I = FMA(KP559016994, T3H, T3G);
|
Chris@10
|
457 T3Z = FNMS(KP559016994, T3Y, T3X);
|
Chris@10
|
458 T4N = FMA(KP559016994, T3Y, T3X);
|
Chris@10
|
459 T3x = W[2];
|
Chris@10
|
460 T5m = FNMS(KP951056516, T3P, T3I);
|
Chris@10
|
461 T3Q = FMA(KP951056516, T3P, T3I);
|
Chris@10
|
462 T4P = FMA(KP951056516, T4O, T4N);
|
Chris@10
|
463 T5F = FNMS(KP951056516, T4O, T4N);
|
Chris@10
|
464 T5p = FMA(KP951056516, T42, T3Z);
|
Chris@10
|
465 T43 = FNMS(KP951056516, T42, T3Z);
|
Chris@10
|
466 T3R = T3x * T3Q;
|
Chris@10
|
467 T3S = W[3];
|
Chris@10
|
468 T5l = W[34];
|
Chris@10
|
469 T5o = W[35];
|
Chris@10
|
470 T4D = T3S * T3Q;
|
Chris@10
|
471 T5n = T5l * T5m;
|
Chris@10
|
472 T5x = T5o * T5m;
|
Chris@10
|
473 T4K = FNMS(KP951056516, T4J, T4I);
|
Chris@10
|
474 T5C = FMA(KP951056516, T4J, T4I);
|
Chris@10
|
475 T4H = W[10];
|
Chris@10
|
476 T4M = W[11];
|
Chris@10
|
477 T5B = W[26];
|
Chris@10
|
478 T5E = W[27];
|
Chris@10
|
479 T4L = T4H * T4K;
|
Chris@10
|
480 T4X = T4M * T4K;
|
Chris@10
|
481 T5D = T5B * T5C;
|
Chris@10
|
482 T5N = T5E * T5C;
|
Chris@10
|
483 }
|
Chris@10
|
484 {
|
Chris@10
|
485 E T58, T5d, T45, T4g, T4x, T4R, T5r;
|
Chris@10
|
486 T4g = FNMS(KP559016994, T4f, T4e);
|
Chris@10
|
487 T58 = FMA(KP559016994, T4f, T4e);
|
Chris@10
|
488 T5d = FMA(KP559016994, T4w, T4v);
|
Chris@10
|
489 T4x = FNMS(KP559016994, T4w, T4v);
|
Chris@10
|
490 T45 = W[4];
|
Chris@10
|
491 T4R = W[12];
|
Chris@10
|
492 T4S = FNMS(KP951056516, T4n, T4g);
|
Chris@10
|
493 T4o = FMA(KP951056516, T4n, T4g);
|
Chris@10
|
494 T4V = FMA(KP951056516, T4A, T4x);
|
Chris@10
|
495 T4B = FNMS(KP951056516, T4A, T4x);
|
Chris@10
|
496 T4T = T4R * T4S;
|
Chris@10
|
497 T4p = T45 * T4o;
|
Chris@10
|
498 T4Z = T4R * T4V;
|
Chris@10
|
499 T4F = T45 * T4B;
|
Chris@10
|
500 T57 = W[20];
|
Chris@10
|
501 T5r = W[36];
|
Chris@10
|
502 T5s = FNMS(KP951056516, T59, T58);
|
Chris@10
|
503 T5a = FMA(KP951056516, T59, T58);
|
Chris@10
|
504 T5v = FMA(KP951056516, T5e, T5d);
|
Chris@10
|
505 T5f = FNMS(KP951056516, T5e, T5d);
|
Chris@10
|
506 T5t = T5r * T5s;
|
Chris@10
|
507 T5b = T57 * T5a;
|
Chris@10
|
508 T5z = T5r * T5v;
|
Chris@10
|
509 }
|
Chris@10
|
510 }
|
Chris@10
|
511 T5j = T57 * T5f;
|
Chris@10
|
512 {
|
Chris@10
|
513 E T44, T4E, T5G, T5O, T5K, T4G, T4C, T4q;
|
Chris@10
|
514 T44 = FNMS(T3S, T43, T3R);
|
Chris@10
|
515 T4E = FMA(T3x, T43, T4D);
|
Chris@10
|
516 T4q = W[5];
|
Chris@10
|
517 T5G = FNMS(T5E, T5F, T5D);
|
Chris@10
|
518 T5O = FMA(T5B, T5F, T5N);
|
Chris@10
|
519 T5K = W[29];
|
Chris@10
|
520 T4G = FNMS(T4q, T4o, T4F);
|
Chris@10
|
521 T4C = FMA(T4q, T4B, T4p);
|
Chris@10
|
522 {
|
Chris@10
|
523 E T4Q, T4Y, T4U, T5Q, T5M;
|
Chris@10
|
524 T4Q = FNMS(T4M, T4P, T4L);
|
Chris@10
|
525 T5Q = FNMS(T5K, T5I, T5P);
|
Chris@10
|
526 T5M = FMA(T5K, T5L, T5J);
|
Chris@10
|
527 Im[WS(rs, 1)] = T4G - T4E;
|
Chris@10
|
528 Ip[WS(rs, 1)] = T4E + T4G;
|
Chris@10
|
529 Rm[WS(rs, 1)] = T44 + T4C;
|
Chris@10
|
530 Rp[WS(rs, 1)] = T44 - T4C;
|
Chris@10
|
531 Im[WS(rs, 7)] = T5Q - T5O;
|
Chris@10
|
532 Ip[WS(rs, 7)] = T5O + T5Q;
|
Chris@10
|
533 Rm[WS(rs, 7)] = T5G + T5M;
|
Chris@10
|
534 Rp[WS(rs, 7)] = T5G - T5M;
|
Chris@10
|
535 T4Y = FMA(T4H, T4P, T4X);
|
Chris@10
|
536 T4U = W[13];
|
Chris@10
|
537 {
|
Chris@10
|
538 E T56, T5i, T5c, T50, T4W, T5k, T5g;
|
Chris@10
|
539 T56 = FNMS(T54, T55, T53);
|
Chris@10
|
540 T5i = FMA(T51, T55, T5h);
|
Chris@10
|
541 T5c = W[21];
|
Chris@10
|
542 T50 = FNMS(T4U, T4S, T4Z);
|
Chris@10
|
543 T4W = FMA(T4U, T4V, T4T);
|
Chris@10
|
544 T5q = FNMS(T5o, T5p, T5n);
|
Chris@10
|
545 T5k = FNMS(T5c, T5a, T5j);
|
Chris@10
|
546 T5g = FMA(T5c, T5f, T5b);
|
Chris@10
|
547 Im[WS(rs, 3)] = T50 - T4Y;
|
Chris@10
|
548 Ip[WS(rs, 3)] = T4Y + T50;
|
Chris@10
|
549 Rm[WS(rs, 3)] = T4Q + T4W;
|
Chris@10
|
550 Rp[WS(rs, 3)] = T4Q - T4W;
|
Chris@10
|
551 Im[WS(rs, 5)] = T5k - T5i;
|
Chris@10
|
552 Ip[WS(rs, 5)] = T5i + T5k;
|
Chris@10
|
553 Rm[WS(rs, 5)] = T56 + T5g;
|
Chris@10
|
554 Rp[WS(rs, 5)] = T56 - T5g;
|
Chris@10
|
555 T5y = FMA(T5l, T5p, T5x);
|
Chris@10
|
556 T5u = W[37];
|
Chris@10
|
557 }
|
Chris@10
|
558 }
|
Chris@10
|
559 }
|
Chris@10
|
560 }
|
Chris@10
|
561 }
|
Chris@10
|
562 }
|
Chris@10
|
563 T5A = FNMS(T5u, T5s, T5z);
|
Chris@10
|
564 T5w = FMA(T5u, T5v, T5t);
|
Chris@10
|
565 Im[WS(rs, 9)] = T5A - T5y;
|
Chris@10
|
566 Ip[WS(rs, 9)] = T5y + T5A;
|
Chris@10
|
567 Rm[WS(rs, 9)] = T5q + T5w;
|
Chris@10
|
568 Rp[WS(rs, 9)] = T5q - T5w;
|
Chris@10
|
569 }
|
Chris@10
|
570 }
|
Chris@10
|
571 }
|
Chris@10
|
572
|
Chris@10
|
573 static const tw_instr twinstr[] = {
|
Chris@10
|
574 {TW_FULL, 1, 20},
|
Chris@10
|
575 {TW_NEXT, 1, 0}
|
Chris@10
|
576 };
|
Chris@10
|
577
|
Chris@10
|
578 static const hc2c_desc desc = { 20, "hc2cbdft2_20", twinstr, &GENUS, {176, 38, 110, 0} };
|
Chris@10
|
579
|
Chris@10
|
580 void X(codelet_hc2cbdft2_20) (planner *p) {
|
Chris@10
|
581 X(khc2c_register) (p, hc2cbdft2_20, &desc, HC2C_VIA_DFT);
|
Chris@10
|
582 }
|
Chris@10
|
583 #else /* HAVE_FMA */
|
Chris@10
|
584
|
Chris@10
|
585 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 20 -dif -name hc2cbdft2_20 -include hc2cb.h */
|
Chris@10
|
586
|
Chris@10
|
587 /*
|
Chris@10
|
588 * This function contains 286 FP additions, 124 FP multiplications,
|
Chris@10
|
589 * (or, 224 additions, 62 multiplications, 62 fused multiply/add),
|
Chris@10
|
590 * 89 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
591 */
|
Chris@10
|
592 #include "hc2cb.h"
|
Chris@10
|
593
|
Chris@10
|
594 static void hc2cbdft2_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
595 {
|
Chris@10
|
596 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
597 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
598 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
599 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
600 {
|
Chris@10
|
601 INT m;
|
Chris@10
|
602 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
|
603 E T7, T3N, T4a, T16, T1G, T3g, T3D, T26, T1k, T3A, T3B, T1v, T2e, T48, T47;
|
Chris@10
|
604 E T2d, T1L, T43, T40, T1K, T2l, T3t, T2m, T3w, T3n, T3p, TC, T2b, T4d, T4f;
|
Chris@10
|
605 E T23, T2j, T1B, T1H, T3U, T3W, T3G, T3I, T11, T17;
|
Chris@10
|
606 {
|
Chris@10
|
607 E T3, T1C, T15, T24, T6, T12, T1F, T25;
|
Chris@10
|
608 {
|
Chris@10
|
609 E T1, T2, T13, T14;
|
Chris@10
|
610 T1 = Rp[0];
|
Chris@10
|
611 T2 = Rm[WS(rs, 9)];
|
Chris@10
|
612 T3 = T1 + T2;
|
Chris@10
|
613 T1C = T1 - T2;
|
Chris@10
|
614 T13 = Ip[0];
|
Chris@10
|
615 T14 = Im[WS(rs, 9)];
|
Chris@10
|
616 T15 = T13 + T14;
|
Chris@10
|
617 T24 = T13 - T14;
|
Chris@10
|
618 }
|
Chris@10
|
619 {
|
Chris@10
|
620 E T4, T5, T1D, T1E;
|
Chris@10
|
621 T4 = Rp[WS(rs, 5)];
|
Chris@10
|
622 T5 = Rm[WS(rs, 4)];
|
Chris@10
|
623 T6 = T4 + T5;
|
Chris@10
|
624 T12 = T4 - T5;
|
Chris@10
|
625 T1D = Ip[WS(rs, 5)];
|
Chris@10
|
626 T1E = Im[WS(rs, 4)];
|
Chris@10
|
627 T1F = T1D + T1E;
|
Chris@10
|
628 T25 = T1D - T1E;
|
Chris@10
|
629 }
|
Chris@10
|
630 T7 = T3 + T6;
|
Chris@10
|
631 T3N = T15 - T12;
|
Chris@10
|
632 T4a = T1C + T1F;
|
Chris@10
|
633 T16 = T12 + T15;
|
Chris@10
|
634 T1G = T1C - T1F;
|
Chris@10
|
635 T3g = T3 - T6;
|
Chris@10
|
636 T3D = T24 - T25;
|
Chris@10
|
637 T26 = T24 + T25;
|
Chris@10
|
638 }
|
Chris@10
|
639 {
|
Chris@10
|
640 E Te, T3O, T3Y, TJ, T1e, T3h, T3r, T1R, TA, T3S, T42, TZ, T1u, T3l, T3v;
|
Chris@10
|
641 E T21, Tl, T3P, T3Z, TO, T1j, T3i, T3s, T1U, Tt, T3R, T41, TU, T1p, T3k;
|
Chris@10
|
642 E T3u, T1Y;
|
Chris@10
|
643 {
|
Chris@10
|
644 E Ta, T1a, TI, T1P, Td, TF, T1d, T1Q;
|
Chris@10
|
645 {
|
Chris@10
|
646 E T8, T9, TG, TH;
|
Chris@10
|
647 T8 = Rp[WS(rs, 4)];
|
Chris@10
|
648 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
649 Ta = T8 + T9;
|
Chris@10
|
650 T1a = T8 - T9;
|
Chris@10
|
651 TG = Ip[WS(rs, 4)];
|
Chris@10
|
652 TH = Im[WS(rs, 5)];
|
Chris@10
|
653 TI = TG + TH;
|
Chris@10
|
654 T1P = TG - TH;
|
Chris@10
|
655 }
|
Chris@10
|
656 {
|
Chris@10
|
657 E Tb, Tc, T1b, T1c;
|
Chris@10
|
658 Tb = Rp[WS(rs, 9)];
|
Chris@10
|
659 Tc = Rm[0];
|
Chris@10
|
660 Td = Tb + Tc;
|
Chris@10
|
661 TF = Tb - Tc;
|
Chris@10
|
662 T1b = Ip[WS(rs, 9)];
|
Chris@10
|
663 T1c = Im[0];
|
Chris@10
|
664 T1d = T1b + T1c;
|
Chris@10
|
665 T1Q = T1b - T1c;
|
Chris@10
|
666 }
|
Chris@10
|
667 Te = Ta + Td;
|
Chris@10
|
668 T3O = TI - TF;
|
Chris@10
|
669 T3Y = T1a + T1d;
|
Chris@10
|
670 TJ = TF + TI;
|
Chris@10
|
671 T1e = T1a - T1d;
|
Chris@10
|
672 T3h = Ta - Td;
|
Chris@10
|
673 T3r = T1P - T1Q;
|
Chris@10
|
674 T1R = T1P + T1Q;
|
Chris@10
|
675 }
|
Chris@10
|
676 {
|
Chris@10
|
677 E Tw, T1q, TY, T1Z, Tz, TV, T1t, T20;
|
Chris@10
|
678 {
|
Chris@10
|
679 E Tu, Tv, TW, TX;
|
Chris@10
|
680 Tu = Rm[WS(rs, 7)];
|
Chris@10
|
681 Tv = Rp[WS(rs, 2)];
|
Chris@10
|
682 Tw = Tu + Tv;
|
Chris@10
|
683 T1q = Tu - Tv;
|
Chris@10
|
684 TW = Im[WS(rs, 7)];
|
Chris@10
|
685 TX = Ip[WS(rs, 2)];
|
Chris@10
|
686 TY = TW + TX;
|
Chris@10
|
687 T1Z = TX - TW;
|
Chris@10
|
688 }
|
Chris@10
|
689 {
|
Chris@10
|
690 E Tx, Ty, T1r, T1s;
|
Chris@10
|
691 Tx = Rm[WS(rs, 2)];
|
Chris@10
|
692 Ty = Rp[WS(rs, 7)];
|
Chris@10
|
693 Tz = Tx + Ty;
|
Chris@10
|
694 TV = Tx - Ty;
|
Chris@10
|
695 T1r = Im[WS(rs, 2)];
|
Chris@10
|
696 T1s = Ip[WS(rs, 7)];
|
Chris@10
|
697 T1t = T1r + T1s;
|
Chris@10
|
698 T20 = T1s - T1r;
|
Chris@10
|
699 }
|
Chris@10
|
700 TA = Tw + Tz;
|
Chris@10
|
701 T3S = TV + TY;
|
Chris@10
|
702 T42 = T1q - T1t;
|
Chris@10
|
703 TZ = TV - TY;
|
Chris@10
|
704 T1u = T1q + T1t;
|
Chris@10
|
705 T3l = Tw - Tz;
|
Chris@10
|
706 T3v = T1Z - T20;
|
Chris@10
|
707 T21 = T1Z + T20;
|
Chris@10
|
708 }
|
Chris@10
|
709 {
|
Chris@10
|
710 E Th, T1f, TN, T1S, Tk, TK, T1i, T1T;
|
Chris@10
|
711 {
|
Chris@10
|
712 E Tf, Tg, TL, TM;
|
Chris@10
|
713 Tf = Rm[WS(rs, 3)];
|
Chris@10
|
714 Tg = Rp[WS(rs, 6)];
|
Chris@10
|
715 Th = Tf + Tg;
|
Chris@10
|
716 T1f = Tf - Tg;
|
Chris@10
|
717 TL = Im[WS(rs, 3)];
|
Chris@10
|
718 TM = Ip[WS(rs, 6)];
|
Chris@10
|
719 TN = TL + TM;
|
Chris@10
|
720 T1S = TM - TL;
|
Chris@10
|
721 }
|
Chris@10
|
722 {
|
Chris@10
|
723 E Ti, Tj, T1g, T1h;
|
Chris@10
|
724 Ti = Rp[WS(rs, 1)];
|
Chris@10
|
725 Tj = Rm[WS(rs, 8)];
|
Chris@10
|
726 Tk = Ti + Tj;
|
Chris@10
|
727 TK = Ti - Tj;
|
Chris@10
|
728 T1g = Ip[WS(rs, 1)];
|
Chris@10
|
729 T1h = Im[WS(rs, 8)];
|
Chris@10
|
730 T1i = T1g + T1h;
|
Chris@10
|
731 T1T = T1g - T1h;
|
Chris@10
|
732 }
|
Chris@10
|
733 Tl = Th + Tk;
|
Chris@10
|
734 T3P = TK + TN;
|
Chris@10
|
735 T3Z = T1f + T1i;
|
Chris@10
|
736 TO = TK - TN;
|
Chris@10
|
737 T1j = T1f - T1i;
|
Chris@10
|
738 T3i = Th - Tk;
|
Chris@10
|
739 T3s = T1S - T1T;
|
Chris@10
|
740 T1U = T1S + T1T;
|
Chris@10
|
741 }
|
Chris@10
|
742 {
|
Chris@10
|
743 E Tp, T1l, TT, T1W, Ts, TQ, T1o, T1X;
|
Chris@10
|
744 {
|
Chris@10
|
745 E Tn, To, TR, TS;
|
Chris@10
|
746 Tn = Rp[WS(rs, 8)];
|
Chris@10
|
747 To = Rm[WS(rs, 1)];
|
Chris@10
|
748 Tp = Tn + To;
|
Chris@10
|
749 T1l = Tn - To;
|
Chris@10
|
750 TR = Ip[WS(rs, 8)];
|
Chris@10
|
751 TS = Im[WS(rs, 1)];
|
Chris@10
|
752 TT = TR + TS;
|
Chris@10
|
753 T1W = TR - TS;
|
Chris@10
|
754 }
|
Chris@10
|
755 {
|
Chris@10
|
756 E Tq, Tr, T1m, T1n;
|
Chris@10
|
757 Tq = Rm[WS(rs, 6)];
|
Chris@10
|
758 Tr = Rp[WS(rs, 3)];
|
Chris@10
|
759 Ts = Tq + Tr;
|
Chris@10
|
760 TQ = Tq - Tr;
|
Chris@10
|
761 T1m = Im[WS(rs, 6)];
|
Chris@10
|
762 T1n = Ip[WS(rs, 3)];
|
Chris@10
|
763 T1o = T1m + T1n;
|
Chris@10
|
764 T1X = T1n - T1m;
|
Chris@10
|
765 }
|
Chris@10
|
766 Tt = Tp + Ts;
|
Chris@10
|
767 T3R = TT - TQ;
|
Chris@10
|
768 T41 = T1l - T1o;
|
Chris@10
|
769 TU = TQ + TT;
|
Chris@10
|
770 T1p = T1l + T1o;
|
Chris@10
|
771 T3k = Tp - Ts;
|
Chris@10
|
772 T3u = T1W - T1X;
|
Chris@10
|
773 T1Y = T1W + T1X;
|
Chris@10
|
774 }
|
Chris@10
|
775 T1k = T1e - T1j;
|
Chris@10
|
776 T3A = T3h - T3i;
|
Chris@10
|
777 T3B = T3k - T3l;
|
Chris@10
|
778 T1v = T1p - T1u;
|
Chris@10
|
779 T2e = T1Y - T21;
|
Chris@10
|
780 T48 = T3R + T3S;
|
Chris@10
|
781 T47 = T3O + T3P;
|
Chris@10
|
782 T2d = T1R - T1U;
|
Chris@10
|
783 T1L = TU - TZ;
|
Chris@10
|
784 T43 = T41 - T42;
|
Chris@10
|
785 T40 = T3Y - T3Z;
|
Chris@10
|
786 T1K = TJ - TO;
|
Chris@10
|
787 T2l = Te - Tl;
|
Chris@10
|
788 T3t = T3r - T3s;
|
Chris@10
|
789 T2m = Tt - TA;
|
Chris@10
|
790 T3w = T3u - T3v;
|
Chris@10
|
791 {
|
Chris@10
|
792 E T3j, T3m, Tm, TB;
|
Chris@10
|
793 T3j = T3h + T3i;
|
Chris@10
|
794 T3m = T3k + T3l;
|
Chris@10
|
795 T3n = T3j + T3m;
|
Chris@10
|
796 T3p = KP559016994 * (T3j - T3m);
|
Chris@10
|
797 Tm = Te + Tl;
|
Chris@10
|
798 TB = Tt + TA;
|
Chris@10
|
799 TC = Tm + TB;
|
Chris@10
|
800 T2b = KP559016994 * (Tm - TB);
|
Chris@10
|
801 }
|
Chris@10
|
802 {
|
Chris@10
|
803 E T4b, T4c, T3Q, T3T;
|
Chris@10
|
804 T4b = T3Y + T3Z;
|
Chris@10
|
805 T4c = T41 + T42;
|
Chris@10
|
806 T4d = T4b + T4c;
|
Chris@10
|
807 T4f = KP559016994 * (T4b - T4c);
|
Chris@10
|
808 {
|
Chris@10
|
809 E T1V, T22, T1z, T1A;
|
Chris@10
|
810 T1V = T1R + T1U;
|
Chris@10
|
811 T22 = T1Y + T21;
|
Chris@10
|
812 T23 = T1V + T22;
|
Chris@10
|
813 T2j = KP559016994 * (T1V - T22);
|
Chris@10
|
814 T1z = T1e + T1j;
|
Chris@10
|
815 T1A = T1p + T1u;
|
Chris@10
|
816 T1B = KP559016994 * (T1z - T1A);
|
Chris@10
|
817 T1H = T1z + T1A;
|
Chris@10
|
818 }
|
Chris@10
|
819 T3Q = T3O - T3P;
|
Chris@10
|
820 T3T = T3R - T3S;
|
Chris@10
|
821 T3U = T3Q + T3T;
|
Chris@10
|
822 T3W = KP559016994 * (T3Q - T3T);
|
Chris@10
|
823 {
|
Chris@10
|
824 E T3E, T3F, TP, T10;
|
Chris@10
|
825 T3E = T3r + T3s;
|
Chris@10
|
826 T3F = T3u + T3v;
|
Chris@10
|
827 T3G = T3E + T3F;
|
Chris@10
|
828 T3I = KP559016994 * (T3E - T3F);
|
Chris@10
|
829 TP = TJ + TO;
|
Chris@10
|
830 T10 = TU + TZ;
|
Chris@10
|
831 T11 = KP559016994 * (TP - T10);
|
Chris@10
|
832 T17 = TP + T10;
|
Chris@10
|
833 }
|
Chris@10
|
834 }
|
Chris@10
|
835 }
|
Chris@10
|
836 {
|
Chris@10
|
837 E TD, T27, T3c, T3e, T2o, T36, T2A, T2U, T1N, T2Z, T2t, T2J, T1x, T2X, T2r;
|
Chris@10
|
838 E T2F, T2g, T34, T2y, T2Q;
|
Chris@10
|
839 TD = T7 + TC;
|
Chris@10
|
840 T27 = T23 + T26;
|
Chris@10
|
841 {
|
Chris@10
|
842 E T39, T3b, T38, T3a;
|
Chris@10
|
843 T39 = T16 + T17;
|
Chris@10
|
844 T3b = T1H + T1G;
|
Chris@10
|
845 T38 = W[8];
|
Chris@10
|
846 T3a = W[9];
|
Chris@10
|
847 T3c = FMA(T38, T39, T3a * T3b);
|
Chris@10
|
848 T3e = FNMS(T3a, T39, T38 * T3b);
|
Chris@10
|
849 }
|
Chris@10
|
850 {
|
Chris@10
|
851 E T2n, T2S, T2k, T2T, T2i;
|
Chris@10
|
852 T2n = FNMS(KP951056516, T2m, KP587785252 * T2l);
|
Chris@10
|
853 T2S = FMA(KP951056516, T2l, KP587785252 * T2m);
|
Chris@10
|
854 T2i = FNMS(KP250000000, T23, T26);
|
Chris@10
|
855 T2k = T2i - T2j;
|
Chris@10
|
856 T2T = T2j + T2i;
|
Chris@10
|
857 T2o = T2k - T2n;
|
Chris@10
|
858 T36 = T2T - T2S;
|
Chris@10
|
859 T2A = T2n + T2k;
|
Chris@10
|
860 T2U = T2S + T2T;
|
Chris@10
|
861 }
|
Chris@10
|
862 {
|
Chris@10
|
863 E T1M, T2H, T1J, T2I, T1I;
|
Chris@10
|
864 T1M = FMA(KP951056516, T1K, KP587785252 * T1L);
|
Chris@10
|
865 T2H = FNMS(KP951056516, T1L, KP587785252 * T1K);
|
Chris@10
|
866 T1I = FNMS(KP250000000, T1H, T1G);
|
Chris@10
|
867 T1J = T1B + T1I;
|
Chris@10
|
868 T2I = T1I - T1B;
|
Chris@10
|
869 T1N = T1J - T1M;
|
Chris@10
|
870 T2Z = T2I - T2H;
|
Chris@10
|
871 T2t = T1M + T1J;
|
Chris@10
|
872 T2J = T2H + T2I;
|
Chris@10
|
873 }
|
Chris@10
|
874 {
|
Chris@10
|
875 E T1w, T2E, T19, T2D, T18;
|
Chris@10
|
876 T1w = FMA(KP951056516, T1k, KP587785252 * T1v);
|
Chris@10
|
877 T2E = FNMS(KP951056516, T1v, KP587785252 * T1k);
|
Chris@10
|
878 T18 = FNMS(KP250000000, T17, T16);
|
Chris@10
|
879 T19 = T11 + T18;
|
Chris@10
|
880 T2D = T18 - T11;
|
Chris@10
|
881 T1x = T19 + T1w;
|
Chris@10
|
882 T2X = T2D + T2E;
|
Chris@10
|
883 T2r = T19 - T1w;
|
Chris@10
|
884 T2F = T2D - T2E;
|
Chris@10
|
885 }
|
Chris@10
|
886 {
|
Chris@10
|
887 E T2f, T2P, T2c, T2O, T2a;
|
Chris@10
|
888 T2f = FNMS(KP951056516, T2e, KP587785252 * T2d);
|
Chris@10
|
889 T2P = FMA(KP951056516, T2d, KP587785252 * T2e);
|
Chris@10
|
890 T2a = FNMS(KP250000000, TC, T7);
|
Chris@10
|
891 T2c = T2a - T2b;
|
Chris@10
|
892 T2O = T2b + T2a;
|
Chris@10
|
893 T2g = T2c + T2f;
|
Chris@10
|
894 T34 = T2O + T2P;
|
Chris@10
|
895 T2y = T2c - T2f;
|
Chris@10
|
896 T2Q = T2O - T2P;
|
Chris@10
|
897 }
|
Chris@10
|
898 {
|
Chris@10
|
899 E T1O, T28, TE, T1y;
|
Chris@10
|
900 TE = W[0];
|
Chris@10
|
901 T1y = W[1];
|
Chris@10
|
902 T1O = FMA(TE, T1x, T1y * T1N);
|
Chris@10
|
903 T28 = FNMS(T1y, T1x, TE * T1N);
|
Chris@10
|
904 Rp[0] = TD - T1O;
|
Chris@10
|
905 Ip[0] = T27 + T28;
|
Chris@10
|
906 Rm[0] = TD + T1O;
|
Chris@10
|
907 Im[0] = T28 - T27;
|
Chris@10
|
908 }
|
Chris@10
|
909 {
|
Chris@10
|
910 E T37, T3d, T33, T35;
|
Chris@10
|
911 T33 = W[6];
|
Chris@10
|
912 T35 = W[7];
|
Chris@10
|
913 T37 = FNMS(T35, T36, T33 * T34);
|
Chris@10
|
914 T3d = FMA(T35, T34, T33 * T36);
|
Chris@10
|
915 Rp[WS(rs, 2)] = T37 - T3c;
|
Chris@10
|
916 Ip[WS(rs, 2)] = T3d + T3e;
|
Chris@10
|
917 Rm[WS(rs, 2)] = T37 + T3c;
|
Chris@10
|
918 Im[WS(rs, 2)] = T3e - T3d;
|
Chris@10
|
919 }
|
Chris@10
|
920 {
|
Chris@10
|
921 E T2p, T2v, T2u, T2w;
|
Chris@10
|
922 {
|
Chris@10
|
923 E T29, T2h, T2q, T2s;
|
Chris@10
|
924 T29 = W[14];
|
Chris@10
|
925 T2h = W[15];
|
Chris@10
|
926 T2p = FNMS(T2h, T2o, T29 * T2g);
|
Chris@10
|
927 T2v = FMA(T2h, T2g, T29 * T2o);
|
Chris@10
|
928 T2q = W[16];
|
Chris@10
|
929 T2s = W[17];
|
Chris@10
|
930 T2u = FMA(T2q, T2r, T2s * T2t);
|
Chris@10
|
931 T2w = FNMS(T2s, T2r, T2q * T2t);
|
Chris@10
|
932 }
|
Chris@10
|
933 Rp[WS(rs, 4)] = T2p - T2u;
|
Chris@10
|
934 Ip[WS(rs, 4)] = T2v + T2w;
|
Chris@10
|
935 Rm[WS(rs, 4)] = T2p + T2u;
|
Chris@10
|
936 Im[WS(rs, 4)] = T2w - T2v;
|
Chris@10
|
937 }
|
Chris@10
|
938 {
|
Chris@10
|
939 E T2B, T2L, T2K, T2M;
|
Chris@10
|
940 {
|
Chris@10
|
941 E T2x, T2z, T2C, T2G;
|
Chris@10
|
942 T2x = W[22];
|
Chris@10
|
943 T2z = W[23];
|
Chris@10
|
944 T2B = FNMS(T2z, T2A, T2x * T2y);
|
Chris@10
|
945 T2L = FMA(T2z, T2y, T2x * T2A);
|
Chris@10
|
946 T2C = W[24];
|
Chris@10
|
947 T2G = W[25];
|
Chris@10
|
948 T2K = FMA(T2C, T2F, T2G * T2J);
|
Chris@10
|
949 T2M = FNMS(T2G, T2F, T2C * T2J);
|
Chris@10
|
950 }
|
Chris@10
|
951 Rp[WS(rs, 6)] = T2B - T2K;
|
Chris@10
|
952 Ip[WS(rs, 6)] = T2L + T2M;
|
Chris@10
|
953 Rm[WS(rs, 6)] = T2B + T2K;
|
Chris@10
|
954 Im[WS(rs, 6)] = T2M - T2L;
|
Chris@10
|
955 }
|
Chris@10
|
956 {
|
Chris@10
|
957 E T2V, T31, T30, T32;
|
Chris@10
|
958 {
|
Chris@10
|
959 E T2N, T2R, T2W, T2Y;
|
Chris@10
|
960 T2N = W[30];
|
Chris@10
|
961 T2R = W[31];
|
Chris@10
|
962 T2V = FNMS(T2R, T2U, T2N * T2Q);
|
Chris@10
|
963 T31 = FMA(T2R, T2Q, T2N * T2U);
|
Chris@10
|
964 T2W = W[32];
|
Chris@10
|
965 T2Y = W[33];
|
Chris@10
|
966 T30 = FMA(T2W, T2X, T2Y * T2Z);
|
Chris@10
|
967 T32 = FNMS(T2Y, T2X, T2W * T2Z);
|
Chris@10
|
968 }
|
Chris@10
|
969 Rp[WS(rs, 8)] = T2V - T30;
|
Chris@10
|
970 Ip[WS(rs, 8)] = T31 + T32;
|
Chris@10
|
971 Rm[WS(rs, 8)] = T2V + T30;
|
Chris@10
|
972 Im[WS(rs, 8)] = T32 - T31;
|
Chris@10
|
973 }
|
Chris@10
|
974 }
|
Chris@10
|
975 {
|
Chris@10
|
976 E T4F, T4P, T5c, T5e, T3y, T54, T4o, T4S, T4h, T4Z, T4x, T4N, T45, T4X, T4v;
|
Chris@10
|
977 E T4J, T3K, T56, T4s, T4U;
|
Chris@10
|
978 {
|
Chris@10
|
979 E T4C, T4E, T4B, T4D;
|
Chris@10
|
980 T4C = T3g + T3n;
|
Chris@10
|
981 T4E = T3G + T3D;
|
Chris@10
|
982 T4B = W[18];
|
Chris@10
|
983 T4D = W[19];
|
Chris@10
|
984 T4F = FNMS(T4D, T4E, T4B * T4C);
|
Chris@10
|
985 T4P = FMA(T4D, T4C, T4B * T4E);
|
Chris@10
|
986 }
|
Chris@10
|
987 {
|
Chris@10
|
988 E T59, T5b, T58, T5a;
|
Chris@10
|
989 T59 = T3N + T3U;
|
Chris@10
|
990 T5b = T4d + T4a;
|
Chris@10
|
991 T58 = W[28];
|
Chris@10
|
992 T5a = W[29];
|
Chris@10
|
993 T5c = FMA(T58, T59, T5a * T5b);
|
Chris@10
|
994 T5e = FNMS(T5a, T59, T58 * T5b);
|
Chris@10
|
995 }
|
Chris@10
|
996 {
|
Chris@10
|
997 E T3x, T4n, T3q, T4m, T3o;
|
Chris@10
|
998 T3x = FNMS(KP951056516, T3w, KP587785252 * T3t);
|
Chris@10
|
999 T4n = FMA(KP951056516, T3t, KP587785252 * T3w);
|
Chris@10
|
1000 T3o = FNMS(KP250000000, T3n, T3g);
|
Chris@10
|
1001 T3q = T3o - T3p;
|
Chris@10
|
1002 T4m = T3p + T3o;
|
Chris@10
|
1003 T3y = T3q - T3x;
|
Chris@10
|
1004 T54 = T4m + T4n;
|
Chris@10
|
1005 T4o = T4m - T4n;
|
Chris@10
|
1006 T4S = T3q + T3x;
|
Chris@10
|
1007 }
|
Chris@10
|
1008 {
|
Chris@10
|
1009 E T49, T4M, T4g, T4L, T4e;
|
Chris@10
|
1010 T49 = FNMS(KP951056516, T48, KP587785252 * T47);
|
Chris@10
|
1011 T4M = FMA(KP951056516, T47, KP587785252 * T48);
|
Chris@10
|
1012 T4e = FNMS(KP250000000, T4d, T4a);
|
Chris@10
|
1013 T4g = T4e - T4f;
|
Chris@10
|
1014 T4L = T4f + T4e;
|
Chris@10
|
1015 T4h = T49 + T4g;
|
Chris@10
|
1016 T4Z = T4M + T4L;
|
Chris@10
|
1017 T4x = T4g - T49;
|
Chris@10
|
1018 T4N = T4L - T4M;
|
Chris@10
|
1019 }
|
Chris@10
|
1020 {
|
Chris@10
|
1021 E T44, T4I, T3X, T4H, T3V;
|
Chris@10
|
1022 T44 = FNMS(KP951056516, T43, KP587785252 * T40);
|
Chris@10
|
1023 T4I = FMA(KP951056516, T40, KP587785252 * T43);
|
Chris@10
|
1024 T3V = FNMS(KP250000000, T3U, T3N);
|
Chris@10
|
1025 T3X = T3V - T3W;
|
Chris@10
|
1026 T4H = T3W + T3V;
|
Chris@10
|
1027 T45 = T3X - T44;
|
Chris@10
|
1028 T4X = T4H - T4I;
|
Chris@10
|
1029 T4v = T3X + T44;
|
Chris@10
|
1030 T4J = T4H + T4I;
|
Chris@10
|
1031 }
|
Chris@10
|
1032 {
|
Chris@10
|
1033 E T3C, T4q, T3J, T4r, T3H;
|
Chris@10
|
1034 T3C = FNMS(KP951056516, T3B, KP587785252 * T3A);
|
Chris@10
|
1035 T4q = FMA(KP951056516, T3A, KP587785252 * T3B);
|
Chris@10
|
1036 T3H = FNMS(KP250000000, T3G, T3D);
|
Chris@10
|
1037 T3J = T3H - T3I;
|
Chris@10
|
1038 T4r = T3I + T3H;
|
Chris@10
|
1039 T3K = T3C + T3J;
|
Chris@10
|
1040 T56 = T4r - T4q;
|
Chris@10
|
1041 T4s = T4q + T4r;
|
Chris@10
|
1042 T4U = T3J - T3C;
|
Chris@10
|
1043 }
|
Chris@10
|
1044 {
|
Chris@10
|
1045 E T4O, T4Q, T4G, T4K;
|
Chris@10
|
1046 T4G = W[20];
|
Chris@10
|
1047 T4K = W[21];
|
Chris@10
|
1048 T4O = FMA(T4G, T4J, T4K * T4N);
|
Chris@10
|
1049 T4Q = FNMS(T4K, T4J, T4G * T4N);
|
Chris@10
|
1050 Rp[WS(rs, 5)] = T4F - T4O;
|
Chris@10
|
1051 Ip[WS(rs, 5)] = T4P + T4Q;
|
Chris@10
|
1052 Rm[WS(rs, 5)] = T4F + T4O;
|
Chris@10
|
1053 Im[WS(rs, 5)] = T4Q - T4P;
|
Chris@10
|
1054 }
|
Chris@10
|
1055 {
|
Chris@10
|
1056 E T57, T5d, T53, T55;
|
Chris@10
|
1057 T53 = W[26];
|
Chris@10
|
1058 T55 = W[27];
|
Chris@10
|
1059 T57 = FNMS(T55, T56, T53 * T54);
|
Chris@10
|
1060 T5d = FMA(T55, T54, T53 * T56);
|
Chris@10
|
1061 Rp[WS(rs, 7)] = T57 - T5c;
|
Chris@10
|
1062 Ip[WS(rs, 7)] = T5d + T5e;
|
Chris@10
|
1063 Rm[WS(rs, 7)] = T57 + T5c;
|
Chris@10
|
1064 Im[WS(rs, 7)] = T5e - T5d;
|
Chris@10
|
1065 }
|
Chris@10
|
1066 {
|
Chris@10
|
1067 E T3L, T4j, T4i, T4k;
|
Chris@10
|
1068 {
|
Chris@10
|
1069 E T3f, T3z, T3M, T46;
|
Chris@10
|
1070 T3f = W[2];
|
Chris@10
|
1071 T3z = W[3];
|
Chris@10
|
1072 T3L = FNMS(T3z, T3K, T3f * T3y);
|
Chris@10
|
1073 T4j = FMA(T3z, T3y, T3f * T3K);
|
Chris@10
|
1074 T3M = W[4];
|
Chris@10
|
1075 T46 = W[5];
|
Chris@10
|
1076 T4i = FMA(T3M, T45, T46 * T4h);
|
Chris@10
|
1077 T4k = FNMS(T46, T45, T3M * T4h);
|
Chris@10
|
1078 }
|
Chris@10
|
1079 Rp[WS(rs, 1)] = T3L - T4i;
|
Chris@10
|
1080 Ip[WS(rs, 1)] = T4j + T4k;
|
Chris@10
|
1081 Rm[WS(rs, 1)] = T3L + T4i;
|
Chris@10
|
1082 Im[WS(rs, 1)] = T4k - T4j;
|
Chris@10
|
1083 }
|
Chris@10
|
1084 {
|
Chris@10
|
1085 E T4t, T4z, T4y, T4A;
|
Chris@10
|
1086 {
|
Chris@10
|
1087 E T4l, T4p, T4u, T4w;
|
Chris@10
|
1088 T4l = W[10];
|
Chris@10
|
1089 T4p = W[11];
|
Chris@10
|
1090 T4t = FNMS(T4p, T4s, T4l * T4o);
|
Chris@10
|
1091 T4z = FMA(T4p, T4o, T4l * T4s);
|
Chris@10
|
1092 T4u = W[12];
|
Chris@10
|
1093 T4w = W[13];
|
Chris@10
|
1094 T4y = FMA(T4u, T4v, T4w * T4x);
|
Chris@10
|
1095 T4A = FNMS(T4w, T4v, T4u * T4x);
|
Chris@10
|
1096 }
|
Chris@10
|
1097 Rp[WS(rs, 3)] = T4t - T4y;
|
Chris@10
|
1098 Ip[WS(rs, 3)] = T4z + T4A;
|
Chris@10
|
1099 Rm[WS(rs, 3)] = T4t + T4y;
|
Chris@10
|
1100 Im[WS(rs, 3)] = T4A - T4z;
|
Chris@10
|
1101 }
|
Chris@10
|
1102 {
|
Chris@10
|
1103 E T4V, T51, T50, T52;
|
Chris@10
|
1104 {
|
Chris@10
|
1105 E T4R, T4T, T4W, T4Y;
|
Chris@10
|
1106 T4R = W[34];
|
Chris@10
|
1107 T4T = W[35];
|
Chris@10
|
1108 T4V = FNMS(T4T, T4U, T4R * T4S);
|
Chris@10
|
1109 T51 = FMA(T4T, T4S, T4R * T4U);
|
Chris@10
|
1110 T4W = W[36];
|
Chris@10
|
1111 T4Y = W[37];
|
Chris@10
|
1112 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
Chris@10
|
1113 T52 = FNMS(T4Y, T4X, T4W * T4Z);
|
Chris@10
|
1114 }
|
Chris@10
|
1115 Rp[WS(rs, 9)] = T4V - T50;
|
Chris@10
|
1116 Ip[WS(rs, 9)] = T51 + T52;
|
Chris@10
|
1117 Rm[WS(rs, 9)] = T4V + T50;
|
Chris@10
|
1118 Im[WS(rs, 9)] = T52 - T51;
|
Chris@10
|
1119 }
|
Chris@10
|
1120 }
|
Chris@10
|
1121 }
|
Chris@10
|
1122 }
|
Chris@10
|
1123 }
|
Chris@10
|
1124
|
Chris@10
|
1125 static const tw_instr twinstr[] = {
|
Chris@10
|
1126 {TW_FULL, 1, 20},
|
Chris@10
|
1127 {TW_NEXT, 1, 0}
|
Chris@10
|
1128 };
|
Chris@10
|
1129
|
Chris@10
|
1130 static const hc2c_desc desc = { 20, "hc2cbdft2_20", twinstr, &GENUS, {224, 62, 62, 0} };
|
Chris@10
|
1131
|
Chris@10
|
1132 void X(codelet_hc2cbdft2_20) (planner *p) {
|
Chris@10
|
1133 X(khc2c_register) (p, hc2cbdft2_20, &desc, HC2C_VIA_DFT);
|
Chris@10
|
1134 }
|
Chris@10
|
1135 #endif /* HAVE_FMA */
|