Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:41:55 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_hc2c.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 20 -dif -name hc2cb_20 -include hc2cb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 246 FP additions, 148 FP multiplications,
|
Chris@10
|
32 * (or, 136 additions, 38 multiplications, 110 fused multiply/add),
|
Chris@10
|
33 * 112 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 hc2cb_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(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
42 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
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 T1T, T1Q, T1P;
|
Chris@10
|
47 {
|
Chris@10
|
48 E T3z, T4z, TE, T7, T2W, T4e, T2l, T1t, T33, T3H, T3G, T3a, T1i, T2g, T13;
|
Chris@10
|
49 E T4H, T4G, T2d, T1B, T4u, T4B, T4A, T4r, T1A, T2s, T3l, T2t, T3s, T2o, T2q;
|
Chris@10
|
50 E T1w, T1y, TC, T29, T3E, T3C, T4n, T4l, TN, TL;
|
Chris@10
|
51 {
|
Chris@10
|
52 E T4, T2U, T3, T3x, T1p, T5, T1q, T1r;
|
Chris@10
|
53 {
|
Chris@10
|
54 E T1, T2, T1n, T1o;
|
Chris@10
|
55 T1 = Rp[0];
|
Chris@10
|
56 T2 = Rm[WS(rs, 9)];
|
Chris@10
|
57 T1n = Ip[0];
|
Chris@10
|
58 T1o = Im[WS(rs, 9)];
|
Chris@10
|
59 T4 = Rp[WS(rs, 5)];
|
Chris@10
|
60 T2U = T1 - T2;
|
Chris@10
|
61 T3 = T1 + T2;
|
Chris@10
|
62 T3x = T1n + T1o;
|
Chris@10
|
63 T1p = T1n - T1o;
|
Chris@10
|
64 T5 = Rm[WS(rs, 4)];
|
Chris@10
|
65 T1q = Ip[WS(rs, 5)];
|
Chris@10
|
66 T1r = Im[WS(rs, 4)];
|
Chris@10
|
67 }
|
Chris@10
|
68 {
|
Chris@10
|
69 E T3o, T4p, TF, Te, T2Z, T4f, T2b, T1a, T3k, T4t, TJ, TA, T39, T4j, T2f;
|
Chris@10
|
70 E T12, T3r, T4q, TG, Tl, T32, T4g, T2c, T1h, Tq, T34, Tp, T3f, TR, Tr;
|
Chris@10
|
71 E TS, TT;
|
Chris@10
|
72 {
|
Chris@10
|
73 E Tx, T37, Tw, T3j, TY, Ty, TZ, T10;
|
Chris@10
|
74 {
|
Chris@10
|
75 E Tb, T2X, Ta, T3m, T16, Tc, T17, T18;
|
Chris@10
|
76 {
|
Chris@10
|
77 E T8, T9, T14, T15;
|
Chris@10
|
78 T8 = Rp[WS(rs, 4)];
|
Chris@10
|
79 {
|
Chris@10
|
80 E T3y, T6, T2V, T1s;
|
Chris@10
|
81 T3y = T4 - T5;
|
Chris@10
|
82 T6 = T4 + T5;
|
Chris@10
|
83 T2V = T1q + T1r;
|
Chris@10
|
84 T1s = T1q - T1r;
|
Chris@10
|
85 T3z = T3x - T3y;
|
Chris@10
|
86 T4z = T3y + T3x;
|
Chris@10
|
87 TE = T3 - T6;
|
Chris@10
|
88 T7 = T3 + T6;
|
Chris@10
|
89 T2W = T2U + T2V;
|
Chris@10
|
90 T4e = T2U - T2V;
|
Chris@10
|
91 T2l = T1p + T1s;
|
Chris@10
|
92 T1t = T1p - T1s;
|
Chris@10
|
93 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
94 }
|
Chris@10
|
95 T14 = Ip[WS(rs, 4)];
|
Chris@10
|
96 T15 = Im[WS(rs, 5)];
|
Chris@10
|
97 Tb = Rp[WS(rs, 9)];
|
Chris@10
|
98 T2X = T8 - T9;
|
Chris@10
|
99 Ta = T8 + T9;
|
Chris@10
|
100 T3m = T14 + T15;
|
Chris@10
|
101 T16 = T14 - T15;
|
Chris@10
|
102 Tc = Rm[0];
|
Chris@10
|
103 T17 = Ip[WS(rs, 9)];
|
Chris@10
|
104 T18 = Im[0];
|
Chris@10
|
105 }
|
Chris@10
|
106 {
|
Chris@10
|
107 E Tu, Tv, TW, TX;
|
Chris@10
|
108 Tu = Rm[WS(rs, 7)];
|
Chris@10
|
109 {
|
Chris@10
|
110 E T3n, Td, T2Y, T19;
|
Chris@10
|
111 T3n = Tb - Tc;
|
Chris@10
|
112 Td = Tb + Tc;
|
Chris@10
|
113 T2Y = T17 + T18;
|
Chris@10
|
114 T19 = T17 - T18;
|
Chris@10
|
115 T3o = T3m - T3n;
|
Chris@10
|
116 T4p = T3n + T3m;
|
Chris@10
|
117 TF = Ta - Td;
|
Chris@10
|
118 Te = Ta + Td;
|
Chris@10
|
119 T2Z = T2X + T2Y;
|
Chris@10
|
120 T4f = T2X - T2Y;
|
Chris@10
|
121 T2b = T16 + T19;
|
Chris@10
|
122 T1a = T16 - T19;
|
Chris@10
|
123 Tv = Rp[WS(rs, 2)];
|
Chris@10
|
124 }
|
Chris@10
|
125 TW = Ip[WS(rs, 2)];
|
Chris@10
|
126 TX = Im[WS(rs, 7)];
|
Chris@10
|
127 Tx = Rm[WS(rs, 2)];
|
Chris@10
|
128 T37 = Tu - Tv;
|
Chris@10
|
129 Tw = Tu + Tv;
|
Chris@10
|
130 T3j = TW + TX;
|
Chris@10
|
131 TY = TW - TX;
|
Chris@10
|
132 Ty = Rp[WS(rs, 7)];
|
Chris@10
|
133 TZ = Ip[WS(rs, 7)];
|
Chris@10
|
134 T10 = Im[WS(rs, 2)];
|
Chris@10
|
135 }
|
Chris@10
|
136 }
|
Chris@10
|
137 {
|
Chris@10
|
138 E Ti, T30, Th, T3q, T1d, Tj, T1e, T1f;
|
Chris@10
|
139 {
|
Chris@10
|
140 E Tf, Tg, T1b, T1c;
|
Chris@10
|
141 Tf = Rm[WS(rs, 3)];
|
Chris@10
|
142 {
|
Chris@10
|
143 E T3i, Tz, T38, T11;
|
Chris@10
|
144 T3i = Tx - Ty;
|
Chris@10
|
145 Tz = Tx + Ty;
|
Chris@10
|
146 T38 = TZ + T10;
|
Chris@10
|
147 T11 = TZ - T10;
|
Chris@10
|
148 T3k = T3i + T3j;
|
Chris@10
|
149 T4t = T3i - T3j;
|
Chris@10
|
150 TJ = Tw - Tz;
|
Chris@10
|
151 TA = Tw + Tz;
|
Chris@10
|
152 T39 = T37 - T38;
|
Chris@10
|
153 T4j = T37 + T38;
|
Chris@10
|
154 T2f = TY + T11;
|
Chris@10
|
155 T12 = TY - T11;
|
Chris@10
|
156 Tg = Rp[WS(rs, 6)];
|
Chris@10
|
157 }
|
Chris@10
|
158 T1b = Ip[WS(rs, 6)];
|
Chris@10
|
159 T1c = Im[WS(rs, 3)];
|
Chris@10
|
160 Ti = Rp[WS(rs, 1)];
|
Chris@10
|
161 T30 = Tf - Tg;
|
Chris@10
|
162 Th = Tf + Tg;
|
Chris@10
|
163 T3q = T1b + T1c;
|
Chris@10
|
164 T1d = T1b - T1c;
|
Chris@10
|
165 Tj = Rm[WS(rs, 8)];
|
Chris@10
|
166 T1e = Ip[WS(rs, 1)];
|
Chris@10
|
167 T1f = Im[WS(rs, 8)];
|
Chris@10
|
168 }
|
Chris@10
|
169 {
|
Chris@10
|
170 E Tn, To, TP, TQ;
|
Chris@10
|
171 Tn = Rp[WS(rs, 8)];
|
Chris@10
|
172 {
|
Chris@10
|
173 E T3p, Tk, T31, T1g;
|
Chris@10
|
174 T3p = Ti - Tj;
|
Chris@10
|
175 Tk = Ti + Tj;
|
Chris@10
|
176 T31 = T1e + T1f;
|
Chris@10
|
177 T1g = T1e - T1f;
|
Chris@10
|
178 T3r = T3p + T3q;
|
Chris@10
|
179 T4q = T3p - T3q;
|
Chris@10
|
180 TG = Th - Tk;
|
Chris@10
|
181 Tl = Th + Tk;
|
Chris@10
|
182 T32 = T30 + T31;
|
Chris@10
|
183 T4g = T30 - T31;
|
Chris@10
|
184 T2c = T1d + T1g;
|
Chris@10
|
185 T1h = T1d - T1g;
|
Chris@10
|
186 To = Rm[WS(rs, 1)];
|
Chris@10
|
187 }
|
Chris@10
|
188 TP = Ip[WS(rs, 8)];
|
Chris@10
|
189 TQ = Im[WS(rs, 1)];
|
Chris@10
|
190 Tq = Rm[WS(rs, 6)];
|
Chris@10
|
191 T34 = Tn - To;
|
Chris@10
|
192 Tp = Tn + To;
|
Chris@10
|
193 T3f = TP + TQ;
|
Chris@10
|
194 TR = TP - TQ;
|
Chris@10
|
195 Tr = Rp[WS(rs, 3)];
|
Chris@10
|
196 TS = Ip[WS(rs, 3)];
|
Chris@10
|
197 TT = Im[WS(rs, 6)];
|
Chris@10
|
198 }
|
Chris@10
|
199 }
|
Chris@10
|
200 }
|
Chris@10
|
201 {
|
Chris@10
|
202 E T3h, Tt, T1u, T2n, T1v, T4k, T4h, T2m, TH, TK, T4s, TI;
|
Chris@10
|
203 T33 = T2Z + T32;
|
Chris@10
|
204 T3H = T2Z - T32;
|
Chris@10
|
205 {
|
Chris@10
|
206 E T3g, Ts, T35, TU;
|
Chris@10
|
207 T3g = Tq - Tr;
|
Chris@10
|
208 Ts = Tq + Tr;
|
Chris@10
|
209 T35 = TS + TT;
|
Chris@10
|
210 TU = TS - TT;
|
Chris@10
|
211 T3h = T3f - T3g;
|
Chris@10
|
212 T4s = T3g + T3f;
|
Chris@10
|
213 TI = Tp - Ts;
|
Chris@10
|
214 Tt = Tp + Ts;
|
Chris@10
|
215 {
|
Chris@10
|
216 E T36, T4i, T2e, TV;
|
Chris@10
|
217 T36 = T34 - T35;
|
Chris@10
|
218 T4i = T34 + T35;
|
Chris@10
|
219 T2e = TR + TU;
|
Chris@10
|
220 TV = TR - TU;
|
Chris@10
|
221 T3G = T36 - T39;
|
Chris@10
|
222 T3a = T36 + T39;
|
Chris@10
|
223 T1u = T1a + T1h;
|
Chris@10
|
224 T1i = T1a - T1h;
|
Chris@10
|
225 T2g = T2e - T2f;
|
Chris@10
|
226 T2n = T2e + T2f;
|
Chris@10
|
227 T1v = TV + T12;
|
Chris@10
|
228 T13 = TV - T12;
|
Chris@10
|
229 T4H = T4i - T4j;
|
Chris@10
|
230 T4k = T4i + T4j;
|
Chris@10
|
231 }
|
Chris@10
|
232 }
|
Chris@10
|
233 T4h = T4f + T4g;
|
Chris@10
|
234 T4G = T4f - T4g;
|
Chris@10
|
235 T2d = T2b - T2c;
|
Chris@10
|
236 T2m = T2b + T2c;
|
Chris@10
|
237 TH = TF + TG;
|
Chris@10
|
238 T1B = TF - TG;
|
Chris@10
|
239 T4u = T4s - T4t;
|
Chris@10
|
240 T4B = T4s + T4t;
|
Chris@10
|
241 T4A = T4p + T4q;
|
Chris@10
|
242 T4r = T4p - T4q;
|
Chris@10
|
243 T1A = TI - TJ;
|
Chris@10
|
244 TK = TI + TJ;
|
Chris@10
|
245 {
|
Chris@10
|
246 E Tm, T3B, TB, T3A;
|
Chris@10
|
247 Tm = Te + Tl;
|
Chris@10
|
248 T2s = Te - Tl;
|
Chris@10
|
249 T3l = T3h + T3k;
|
Chris@10
|
250 T3B = T3h - T3k;
|
Chris@10
|
251 TB = Tt + TA;
|
Chris@10
|
252 T2t = Tt - TA;
|
Chris@10
|
253 T3s = T3o + T3r;
|
Chris@10
|
254 T3A = T3o - T3r;
|
Chris@10
|
255 T2o = T2m + T2n;
|
Chris@10
|
256 T2q = T2m - T2n;
|
Chris@10
|
257 T1w = T1u + T1v;
|
Chris@10
|
258 T1y = T1u - T1v;
|
Chris@10
|
259 TC = Tm + TB;
|
Chris@10
|
260 T29 = Tm - TB;
|
Chris@10
|
261 T3E = T3A - T3B;
|
Chris@10
|
262 T3C = T3A + T3B;
|
Chris@10
|
263 T4n = T4h - T4k;
|
Chris@10
|
264 T4l = T4h + T4k;
|
Chris@10
|
265 TN = TH - TK;
|
Chris@10
|
266 TL = TH + TK;
|
Chris@10
|
267 }
|
Chris@10
|
268 }
|
Chris@10
|
269 }
|
Chris@10
|
270 }
|
Chris@10
|
271 {
|
Chris@10
|
272 E T3d, T3b, T4E, T1x, TM, T4m, T58, T5b, T4D, T5a, T5c, T59, T4C;
|
Chris@10
|
273 Rp[0] = T7 + TC;
|
Chris@10
|
274 T3d = T33 - T3a;
|
Chris@10
|
275 T3b = T33 + T3a;
|
Chris@10
|
276 T4E = T4A - T4B;
|
Chris@10
|
277 T4C = T4A + T4B;
|
Chris@10
|
278 Rm[0] = T2l + T2o;
|
Chris@10
|
279 {
|
Chris@10
|
280 E T25, T22, T21, T24, T23, T26, T57;
|
Chris@10
|
281 T1x = FNMS(KP250000000, T1w, T1t);
|
Chris@10
|
282 T25 = T1t + T1w;
|
Chris@10
|
283 T22 = TE + TL;
|
Chris@10
|
284 TM = FNMS(KP250000000, TL, TE);
|
Chris@10
|
285 T21 = W[18];
|
Chris@10
|
286 T24 = W[19];
|
Chris@10
|
287 T4m = FNMS(KP250000000, T4l, T4e);
|
Chris@10
|
288 T58 = T4e + T4l;
|
Chris@10
|
289 T5b = T4z + T4C;
|
Chris@10
|
290 T4D = FNMS(KP250000000, T4C, T4z);
|
Chris@10
|
291 T23 = T21 * T22;
|
Chris@10
|
292 T26 = T24 * T22;
|
Chris@10
|
293 T57 = W[8];
|
Chris@10
|
294 T5a = W[9];
|
Chris@10
|
295 Rp[WS(rs, 5)] = FNMS(T24, T25, T23);
|
Chris@10
|
296 Rm[WS(rs, 5)] = FMA(T21, T25, T26);
|
Chris@10
|
297 T5c = T57 * T5b;
|
Chris@10
|
298 T59 = T57 * T58;
|
Chris@10
|
299 }
|
Chris@10
|
300 {
|
Chris@10
|
301 E T3U, T3Z, T3W, T40, T3V;
|
Chris@10
|
302 {
|
Chris@10
|
303 E T3c, T48, T4b, T3D, T47, T4a;
|
Chris@10
|
304 T3c = FNMS(KP250000000, T3b, T2W);
|
Chris@10
|
305 T48 = T2W + T3b;
|
Chris@10
|
306 T4b = T3z + T3C;
|
Chris@10
|
307 T3D = FNMS(KP250000000, T3C, T3z);
|
Chris@10
|
308 Im[WS(rs, 2)] = FMA(T5a, T58, T5c);
|
Chris@10
|
309 Ip[WS(rs, 2)] = FNMS(T5a, T5b, T59);
|
Chris@10
|
310 T47 = W[28];
|
Chris@10
|
311 T4a = W[29];
|
Chris@10
|
312 {
|
Chris@10
|
313 E T3I, T3Y, T42, T3u, T3M, T3X, T3F;
|
Chris@10
|
314 {
|
Chris@10
|
315 E T3T, T3t, T4c, T49, T3e, T3S;
|
Chris@10
|
316 T3T = FMA(KP618033988, T3l, T3s);
|
Chris@10
|
317 T3t = FNMS(KP618033988, T3s, T3l);
|
Chris@10
|
318 T4c = T47 * T4b;
|
Chris@10
|
319 T49 = T47 * T48;
|
Chris@10
|
320 T3I = FNMS(KP618033988, T3H, T3G);
|
Chris@10
|
321 T3Y = FMA(KP618033988, T3G, T3H);
|
Chris@10
|
322 Im[WS(rs, 7)] = FMA(T4a, T48, T4c);
|
Chris@10
|
323 Ip[WS(rs, 7)] = FNMS(T4a, T4b, T49);
|
Chris@10
|
324 T3e = FNMS(KP559016994, T3d, T3c);
|
Chris@10
|
325 T3S = FMA(KP559016994, T3d, T3c);
|
Chris@10
|
326 T42 = FMA(KP951056516, T3T, T3S);
|
Chris@10
|
327 T3U = FNMS(KP951056516, T3T, T3S);
|
Chris@10
|
328 T3u = FNMS(KP951056516, T3t, T3e);
|
Chris@10
|
329 T3M = FMA(KP951056516, T3t, T3e);
|
Chris@10
|
330 T3X = FMA(KP559016994, T3E, T3D);
|
Chris@10
|
331 T3F = FNMS(KP559016994, T3E, T3D);
|
Chris@10
|
332 }
|
Chris@10
|
333 {
|
Chris@10
|
334 E T3P, T45, T44, T46, T43;
|
Chris@10
|
335 {
|
Chris@10
|
336 E T3w, T3J, T3v, T3K, T2T, T41;
|
Chris@10
|
337 T2T = W[4];
|
Chris@10
|
338 T3w = W[5];
|
Chris@10
|
339 T3J = FMA(KP951056516, T3I, T3F);
|
Chris@10
|
340 T3P = FNMS(KP951056516, T3I, T3F);
|
Chris@10
|
341 T45 = FNMS(KP951056516, T3Y, T3X);
|
Chris@10
|
342 T3Z = FMA(KP951056516, T3Y, T3X);
|
Chris@10
|
343 T3v = T2T * T3u;
|
Chris@10
|
344 T3K = T2T * T3J;
|
Chris@10
|
345 T41 = W[36];
|
Chris@10
|
346 T44 = W[37];
|
Chris@10
|
347 Ip[WS(rs, 1)] = FNMS(T3w, T3J, T3v);
|
Chris@10
|
348 Im[WS(rs, 1)] = FMA(T3w, T3u, T3K);
|
Chris@10
|
349 T46 = T41 * T45;
|
Chris@10
|
350 T43 = T41 * T42;
|
Chris@10
|
351 }
|
Chris@10
|
352 {
|
Chris@10
|
353 E T3O, T3Q, T3N, T3L, T3R;
|
Chris@10
|
354 T3L = W[12];
|
Chris@10
|
355 T3O = W[13];
|
Chris@10
|
356 Im[WS(rs, 9)] = FMA(T44, T42, T46);
|
Chris@10
|
357 Ip[WS(rs, 9)] = FNMS(T44, T45, T43);
|
Chris@10
|
358 T3Q = T3L * T3P;
|
Chris@10
|
359 T3N = T3L * T3M;
|
Chris@10
|
360 T3R = W[20];
|
Chris@10
|
361 T3W = W[21];
|
Chris@10
|
362 Im[WS(rs, 3)] = FMA(T3O, T3M, T3Q);
|
Chris@10
|
363 Ip[WS(rs, 3)] = FNMS(T3O, T3P, T3N);
|
Chris@10
|
364 T40 = T3R * T3Z;
|
Chris@10
|
365 T3V = T3R * T3U;
|
Chris@10
|
366 }
|
Chris@10
|
367 }
|
Chris@10
|
368 }
|
Chris@10
|
369 }
|
Chris@10
|
370 {
|
Chris@10
|
371 E T4U, T4Z, T4W, T50, T4V, T2L, T2I, T2H;
|
Chris@10
|
372 {
|
Chris@10
|
373 E T4T, T4v, T4I, T4Y, T4o, T4S;
|
Chris@10
|
374 T4T = FNMS(KP618033988, T4r, T4u);
|
Chris@10
|
375 T4v = FMA(KP618033988, T4u, T4r);
|
Chris@10
|
376 Im[WS(rs, 5)] = FMA(T3W, T3U, T40);
|
Chris@10
|
377 Ip[WS(rs, 5)] = FNMS(T3W, T3Z, T3V);
|
Chris@10
|
378 T4I = FMA(KP618033988, T4H, T4G);
|
Chris@10
|
379 T4Y = FNMS(KP618033988, T4G, T4H);
|
Chris@10
|
380 T4o = FMA(KP559016994, T4n, T4m);
|
Chris@10
|
381 T4S = FNMS(KP559016994, T4n, T4m);
|
Chris@10
|
382 {
|
Chris@10
|
383 E T52, T4M, T55, T4P, T54, T56, T53;
|
Chris@10
|
384 {
|
Chris@10
|
385 E T4d, T4w, T4J, T4x, T4y, T4X, T4F, T51, T4K;
|
Chris@10
|
386 T4d = W[0];
|
Chris@10
|
387 T4X = FNMS(KP559016994, T4E, T4D);
|
Chris@10
|
388 T4F = FMA(KP559016994, T4E, T4D);
|
Chris@10
|
389 T4U = FNMS(KP951056516, T4T, T4S);
|
Chris@10
|
390 T52 = FMA(KP951056516, T4T, T4S);
|
Chris@10
|
391 T4M = FMA(KP951056516, T4v, T4o);
|
Chris@10
|
392 T4w = FNMS(KP951056516, T4v, T4o);
|
Chris@10
|
393 T4Z = FMA(KP951056516, T4Y, T4X);
|
Chris@10
|
394 T55 = FNMS(KP951056516, T4Y, T4X);
|
Chris@10
|
395 T4P = FNMS(KP951056516, T4I, T4F);
|
Chris@10
|
396 T4J = FMA(KP951056516, T4I, T4F);
|
Chris@10
|
397 T4x = T4d * T4w;
|
Chris@10
|
398 T4y = W[1];
|
Chris@10
|
399 T51 = W[32];
|
Chris@10
|
400 T4K = T4d * T4J;
|
Chris@10
|
401 T54 = W[33];
|
Chris@10
|
402 Ip[0] = FNMS(T4y, T4J, T4x);
|
Chris@10
|
403 T56 = T51 * T55;
|
Chris@10
|
404 T53 = T51 * T52;
|
Chris@10
|
405 Im[0] = FMA(T4y, T4w, T4K);
|
Chris@10
|
406 }
|
Chris@10
|
407 {
|
Chris@10
|
408 E T4O, T4Q, T4N, T4L, T4R;
|
Chris@10
|
409 T4L = W[16];
|
Chris@10
|
410 Im[WS(rs, 8)] = FMA(T54, T52, T56);
|
Chris@10
|
411 Ip[WS(rs, 8)] = FNMS(T54, T55, T53);
|
Chris@10
|
412 T4O = W[17];
|
Chris@10
|
413 T4Q = T4L * T4P;
|
Chris@10
|
414 T4N = T4L * T4M;
|
Chris@10
|
415 T4R = W[24];
|
Chris@10
|
416 T4W = W[25];
|
Chris@10
|
417 Im[WS(rs, 4)] = FMA(T4O, T4M, T4Q);
|
Chris@10
|
418 Ip[WS(rs, 4)] = FNMS(T4O, T4P, T4N);
|
Chris@10
|
419 T50 = T4R * T4Z;
|
Chris@10
|
420 T4V = T4R * T4U;
|
Chris@10
|
421 }
|
Chris@10
|
422 }
|
Chris@10
|
423 }
|
Chris@10
|
424 {
|
Chris@10
|
425 E T2K, T2u, T2F, T2h, T28, T2J, T2r, T2p;
|
Chris@10
|
426 T2K = FNMS(KP618033988, T2s, T2t);
|
Chris@10
|
427 T2u = FMA(KP618033988, T2t, T2s);
|
Chris@10
|
428 Im[WS(rs, 6)] = FMA(T4W, T4U, T50);
|
Chris@10
|
429 Ip[WS(rs, 6)] = FNMS(T4W, T4Z, T4V);
|
Chris@10
|
430 T2p = FNMS(KP250000000, T2o, T2l);
|
Chris@10
|
431 T2F = FNMS(KP618033988, T2d, T2g);
|
Chris@10
|
432 T2h = FMA(KP618033988, T2g, T2d);
|
Chris@10
|
433 T28 = FNMS(KP250000000, TC, T7);
|
Chris@10
|
434 T2J = FNMS(KP559016994, T2q, T2p);
|
Chris@10
|
435 T2r = FMA(KP559016994, T2q, T2p);
|
Chris@10
|
436 {
|
Chris@10
|
437 E T2B, T2G, T2y, T2R, T2Q, T2P, T2A, T2x;
|
Chris@10
|
438 {
|
Chris@10
|
439 E T2k, T2v, T27, T2O, T2i, T2a, T2E;
|
Chris@10
|
440 T2k = W[7];
|
Chris@10
|
441 T2a = FMA(KP559016994, T29, T28);
|
Chris@10
|
442 T2E = FNMS(KP559016994, T29, T28);
|
Chris@10
|
443 T2B = FMA(KP951056516, T2u, T2r);
|
Chris@10
|
444 T2v = FNMS(KP951056516, T2u, T2r);
|
Chris@10
|
445 T27 = W[6];
|
Chris@10
|
446 T2O = FMA(KP951056516, T2F, T2E);
|
Chris@10
|
447 T2G = FNMS(KP951056516, T2F, T2E);
|
Chris@10
|
448 T2i = FMA(KP951056516, T2h, T2a);
|
Chris@10
|
449 T2y = FNMS(KP951056516, T2h, T2a);
|
Chris@10
|
450 {
|
Chris@10
|
451 E T2N, T2j, T2w, T2S;
|
Chris@10
|
452 T2L = FMA(KP951056516, T2K, T2J);
|
Chris@10
|
453 T2R = FNMS(KP951056516, T2K, T2J);
|
Chris@10
|
454 T2Q = W[23];
|
Chris@10
|
455 T2N = W[22];
|
Chris@10
|
456 T2j = T27 * T2i;
|
Chris@10
|
457 T2w = T2k * T2i;
|
Chris@10
|
458 T2S = T2Q * T2O;
|
Chris@10
|
459 T2P = T2N * T2O;
|
Chris@10
|
460 Rp[WS(rs, 2)] = FNMS(T2k, T2v, T2j);
|
Chris@10
|
461 Rm[WS(rs, 2)] = FMA(T27, T2v, T2w);
|
Chris@10
|
462 Rm[WS(rs, 6)] = FMA(T2N, T2R, T2S);
|
Chris@10
|
463 }
|
Chris@10
|
464 }
|
Chris@10
|
465 Rp[WS(rs, 6)] = FNMS(T2Q, T2R, T2P);
|
Chris@10
|
466 T2A = W[31];
|
Chris@10
|
467 T2x = W[30];
|
Chris@10
|
468 {
|
Chris@10
|
469 E T2D, T2M, T2C, T2z;
|
Chris@10
|
470 T2I = W[15];
|
Chris@10
|
471 T2C = T2A * T2y;
|
Chris@10
|
472 T2z = T2x * T2y;
|
Chris@10
|
473 T2D = W[14];
|
Chris@10
|
474 T2M = T2I * T2G;
|
Chris@10
|
475 Rm[WS(rs, 8)] = FMA(T2x, T2B, T2C);
|
Chris@10
|
476 Rp[WS(rs, 8)] = FNMS(T2A, T2B, T2z);
|
Chris@10
|
477 T2H = T2D * T2G;
|
Chris@10
|
478 Rm[WS(rs, 4)] = FMA(T2D, T2L, T2M);
|
Chris@10
|
479 }
|
Chris@10
|
480 }
|
Chris@10
|
481 }
|
Chris@10
|
482 {
|
Chris@10
|
483 E T1S, T1C, T1j, T1N, T1z, T1R;
|
Chris@10
|
484 T1S = FMA(KP618033988, T1A, T1B);
|
Chris@10
|
485 T1C = FNMS(KP618033988, T1B, T1A);
|
Chris@10
|
486 Rp[WS(rs, 4)] = FNMS(T2I, T2L, T2H);
|
Chris@10
|
487 T1j = FNMS(KP618033988, T1i, T13);
|
Chris@10
|
488 T1N = FMA(KP618033988, T13, T1i);
|
Chris@10
|
489 T1z = FNMS(KP559016994, T1y, T1x);
|
Chris@10
|
490 T1R = FMA(KP559016994, T1y, T1x);
|
Chris@10
|
491 {
|
Chris@10
|
492 E T1J, T1O, T1G, T1Z, T1Y, T1X, T1I, T1F;
|
Chris@10
|
493 {
|
Chris@10
|
494 E T1m, T1D, TD, T1W, T1k, T1M, TO;
|
Chris@10
|
495 T1m = W[3];
|
Chris@10
|
496 T1M = FMA(KP559016994, TN, TM);
|
Chris@10
|
497 TO = FNMS(KP559016994, TN, TM);
|
Chris@10
|
498 T1D = FNMS(KP951056516, T1C, T1z);
|
Chris@10
|
499 T1J = FMA(KP951056516, T1C, T1z);
|
Chris@10
|
500 TD = W[2];
|
Chris@10
|
501 T1O = FNMS(KP951056516, T1N, T1M);
|
Chris@10
|
502 T1W = FMA(KP951056516, T1N, T1M);
|
Chris@10
|
503 T1G = FNMS(KP951056516, T1j, TO);
|
Chris@10
|
504 T1k = FMA(KP951056516, T1j, TO);
|
Chris@10
|
505 {
|
Chris@10
|
506 E T1V, T1l, T1E, T20;
|
Chris@10
|
507 T1Z = FNMS(KP951056516, T1S, T1R);
|
Chris@10
|
508 T1T = FMA(KP951056516, T1S, T1R);
|
Chris@10
|
509 T1Y = W[27];
|
Chris@10
|
510 T1V = W[26];
|
Chris@10
|
511 T1l = TD * T1k;
|
Chris@10
|
512 T1E = T1m * T1k;
|
Chris@10
|
513 T20 = T1Y * T1W;
|
Chris@10
|
514 T1X = T1V * T1W;
|
Chris@10
|
515 Rp[WS(rs, 1)] = FNMS(T1m, T1D, T1l);
|
Chris@10
|
516 Rm[WS(rs, 1)] = FMA(TD, T1D, T1E);
|
Chris@10
|
517 Rm[WS(rs, 7)] = FMA(T1V, T1Z, T20);
|
Chris@10
|
518 }
|
Chris@10
|
519 }
|
Chris@10
|
520 Rp[WS(rs, 7)] = FNMS(T1Y, T1Z, T1X);
|
Chris@10
|
521 T1I = W[35];
|
Chris@10
|
522 T1F = W[34];
|
Chris@10
|
523 {
|
Chris@10
|
524 E T1L, T1U, T1K, T1H;
|
Chris@10
|
525 T1Q = W[11];
|
Chris@10
|
526 T1K = T1I * T1G;
|
Chris@10
|
527 T1H = T1F * T1G;
|
Chris@10
|
528 T1L = W[10];
|
Chris@10
|
529 T1U = T1Q * T1O;
|
Chris@10
|
530 Rm[WS(rs, 9)] = FMA(T1F, T1J, T1K);
|
Chris@10
|
531 Rp[WS(rs, 9)] = FNMS(T1I, T1J, T1H);
|
Chris@10
|
532 T1P = T1L * T1O;
|
Chris@10
|
533 Rm[WS(rs, 3)] = FMA(T1L, T1T, T1U);
|
Chris@10
|
534 }
|
Chris@10
|
535 }
|
Chris@10
|
536 }
|
Chris@10
|
537 }
|
Chris@10
|
538 }
|
Chris@10
|
539 }
|
Chris@10
|
540 }
|
Chris@10
|
541 Rp[WS(rs, 3)] = FNMS(T1Q, T1T, T1P);
|
Chris@10
|
542 }
|
Chris@10
|
543 }
|
Chris@10
|
544 }
|
Chris@10
|
545
|
Chris@10
|
546 static const tw_instr twinstr[] = {
|
Chris@10
|
547 {TW_FULL, 1, 20},
|
Chris@10
|
548 {TW_NEXT, 1, 0}
|
Chris@10
|
549 };
|
Chris@10
|
550
|
Chris@10
|
551 static const hc2c_desc desc = { 20, "hc2cb_20", twinstr, &GENUS, {136, 38, 110, 0} };
|
Chris@10
|
552
|
Chris@10
|
553 void X(codelet_hc2cb_20) (planner *p) {
|
Chris@10
|
554 X(khc2c_register) (p, hc2cb_20, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
555 }
|
Chris@10
|
556 #else /* HAVE_FMA */
|
Chris@10
|
557
|
Chris@10
|
558 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 20 -dif -name hc2cb_20 -include hc2cb.h */
|
Chris@10
|
559
|
Chris@10
|
560 /*
|
Chris@10
|
561 * This function contains 246 FP additions, 124 FP multiplications,
|
Chris@10
|
562 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
|
Chris@10
|
563 * 97 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
564 */
|
Chris@10
|
565 #include "hc2cb.h"
|
Chris@10
|
566
|
Chris@10
|
567 static void hc2cb_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
568 {
|
Chris@10
|
569 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
570 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
571 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
572 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
573 {
|
Chris@10
|
574 INT m;
|
Chris@10
|
575 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
|
576 E T7, T3T, T49, TE, T1v, T2T, T3g, T2d, T13, T3n, T3o, T1i, T26, T4e, T4d;
|
Chris@10
|
577 E T23, T1n, T42, T3Z, T1m, T2h, T2I, T2i, T2P, T30, T37, T38, Tm, TB, TC;
|
Chris@10
|
578 E T46, T47, T4a, T2a, T2b, T2e, T1w, T1x, T1y, T3O, T3R, T3U, T3h, T3i, T3j;
|
Chris@10
|
579 E TH, TK, TL;
|
Chris@10
|
580 {
|
Chris@10
|
581 E T3, T2R, T1r, T3e, T6, T3f, T1u, T2S;
|
Chris@10
|
582 {
|
Chris@10
|
583 E T1, T2, T1p, T1q;
|
Chris@10
|
584 T1 = Rp[0];
|
Chris@10
|
585 T2 = Rm[WS(rs, 9)];
|
Chris@10
|
586 T3 = T1 + T2;
|
Chris@10
|
587 T2R = T1 - T2;
|
Chris@10
|
588 T1p = Ip[0];
|
Chris@10
|
589 T1q = Im[WS(rs, 9)];
|
Chris@10
|
590 T1r = T1p - T1q;
|
Chris@10
|
591 T3e = T1p + T1q;
|
Chris@10
|
592 }
|
Chris@10
|
593 {
|
Chris@10
|
594 E T4, T5, T1s, T1t;
|
Chris@10
|
595 T4 = Rp[WS(rs, 5)];
|
Chris@10
|
596 T5 = Rm[WS(rs, 4)];
|
Chris@10
|
597 T6 = T4 + T5;
|
Chris@10
|
598 T3f = T4 - T5;
|
Chris@10
|
599 T1s = Ip[WS(rs, 5)];
|
Chris@10
|
600 T1t = Im[WS(rs, 4)];
|
Chris@10
|
601 T1u = T1s - T1t;
|
Chris@10
|
602 T2S = T1s + T1t;
|
Chris@10
|
603 }
|
Chris@10
|
604 T7 = T3 + T6;
|
Chris@10
|
605 T3T = T2R - T2S;
|
Chris@10
|
606 T49 = T3f + T3e;
|
Chris@10
|
607 TE = T3 - T6;
|
Chris@10
|
608 T1v = T1r - T1u;
|
Chris@10
|
609 T2T = T2R + T2S;
|
Chris@10
|
610 T3g = T3e - T3f;
|
Chris@10
|
611 T2d = T1r + T1u;
|
Chris@10
|
612 }
|
Chris@10
|
613 {
|
Chris@10
|
614 E Te, T3M, T3X, TF, TV, T2E, T2W, T21, TA, T3Q, T41, TJ, T1h, T2O, T36;
|
Chris@10
|
615 E T25, Tl, T3N, T3Y, TG, T12, T2H, T2Z, T22, Tt, T3P, T40, TI, T1a, T2L;
|
Chris@10
|
616 E T33, T24;
|
Chris@10
|
617 {
|
Chris@10
|
618 E Ta, T2U, TR, T2C, Td, T2D, TU, T2V;
|
Chris@10
|
619 {
|
Chris@10
|
620 E T8, T9, TP, TQ;
|
Chris@10
|
621 T8 = Rp[WS(rs, 4)];
|
Chris@10
|
622 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
623 Ta = T8 + T9;
|
Chris@10
|
624 T2U = T8 - T9;
|
Chris@10
|
625 TP = Ip[WS(rs, 4)];
|
Chris@10
|
626 TQ = Im[WS(rs, 5)];
|
Chris@10
|
627 TR = TP - TQ;
|
Chris@10
|
628 T2C = TP + TQ;
|
Chris@10
|
629 }
|
Chris@10
|
630 {
|
Chris@10
|
631 E Tb, Tc, TS, TT;
|
Chris@10
|
632 Tb = Rp[WS(rs, 9)];
|
Chris@10
|
633 Tc = Rm[0];
|
Chris@10
|
634 Td = Tb + Tc;
|
Chris@10
|
635 T2D = Tb - Tc;
|
Chris@10
|
636 TS = Ip[WS(rs, 9)];
|
Chris@10
|
637 TT = Im[0];
|
Chris@10
|
638 TU = TS - TT;
|
Chris@10
|
639 T2V = TS + TT;
|
Chris@10
|
640 }
|
Chris@10
|
641 Te = Ta + Td;
|
Chris@10
|
642 T3M = T2U - T2V;
|
Chris@10
|
643 T3X = T2D + T2C;
|
Chris@10
|
644 TF = Ta - Td;
|
Chris@10
|
645 TV = TR - TU;
|
Chris@10
|
646 T2E = T2C - T2D;
|
Chris@10
|
647 T2W = T2U + T2V;
|
Chris@10
|
648 T21 = TR + TU;
|
Chris@10
|
649 }
|
Chris@10
|
650 {
|
Chris@10
|
651 E Tw, T34, T1d, T2N, Tz, T2M, T1g, T35;
|
Chris@10
|
652 {
|
Chris@10
|
653 E Tu, Tv, T1b, T1c;
|
Chris@10
|
654 Tu = Rm[WS(rs, 7)];
|
Chris@10
|
655 Tv = Rp[WS(rs, 2)];
|
Chris@10
|
656 Tw = Tu + Tv;
|
Chris@10
|
657 T34 = Tu - Tv;
|
Chris@10
|
658 T1b = Ip[WS(rs, 2)];
|
Chris@10
|
659 T1c = Im[WS(rs, 7)];
|
Chris@10
|
660 T1d = T1b - T1c;
|
Chris@10
|
661 T2N = T1b + T1c;
|
Chris@10
|
662 }
|
Chris@10
|
663 {
|
Chris@10
|
664 E Tx, Ty, T1e, T1f;
|
Chris@10
|
665 Tx = Rm[WS(rs, 2)];
|
Chris@10
|
666 Ty = Rp[WS(rs, 7)];
|
Chris@10
|
667 Tz = Tx + Ty;
|
Chris@10
|
668 T2M = Tx - Ty;
|
Chris@10
|
669 T1e = Ip[WS(rs, 7)];
|
Chris@10
|
670 T1f = Im[WS(rs, 2)];
|
Chris@10
|
671 T1g = T1e - T1f;
|
Chris@10
|
672 T35 = T1e + T1f;
|
Chris@10
|
673 }
|
Chris@10
|
674 TA = Tw + Tz;
|
Chris@10
|
675 T3Q = T34 + T35;
|
Chris@10
|
676 T41 = T2M - T2N;
|
Chris@10
|
677 TJ = Tw - Tz;
|
Chris@10
|
678 T1h = T1d - T1g;
|
Chris@10
|
679 T2O = T2M + T2N;
|
Chris@10
|
680 T36 = T34 - T35;
|
Chris@10
|
681 T25 = T1d + T1g;
|
Chris@10
|
682 }
|
Chris@10
|
683 {
|
Chris@10
|
684 E Th, T2X, TY, T2G, Tk, T2F, T11, T2Y;
|
Chris@10
|
685 {
|
Chris@10
|
686 E Tf, Tg, TW, TX;
|
Chris@10
|
687 Tf = Rm[WS(rs, 3)];
|
Chris@10
|
688 Tg = Rp[WS(rs, 6)];
|
Chris@10
|
689 Th = Tf + Tg;
|
Chris@10
|
690 T2X = Tf - Tg;
|
Chris@10
|
691 TW = Ip[WS(rs, 6)];
|
Chris@10
|
692 TX = Im[WS(rs, 3)];
|
Chris@10
|
693 TY = TW - TX;
|
Chris@10
|
694 T2G = TW + TX;
|
Chris@10
|
695 }
|
Chris@10
|
696 {
|
Chris@10
|
697 E Ti, Tj, TZ, T10;
|
Chris@10
|
698 Ti = Rp[WS(rs, 1)];
|
Chris@10
|
699 Tj = Rm[WS(rs, 8)];
|
Chris@10
|
700 Tk = Ti + Tj;
|
Chris@10
|
701 T2F = Ti - Tj;
|
Chris@10
|
702 TZ = Ip[WS(rs, 1)];
|
Chris@10
|
703 T10 = Im[WS(rs, 8)];
|
Chris@10
|
704 T11 = TZ - T10;
|
Chris@10
|
705 T2Y = TZ + T10;
|
Chris@10
|
706 }
|
Chris@10
|
707 Tl = Th + Tk;
|
Chris@10
|
708 T3N = T2X - T2Y;
|
Chris@10
|
709 T3Y = T2F - T2G;
|
Chris@10
|
710 TG = Th - Tk;
|
Chris@10
|
711 T12 = TY - T11;
|
Chris@10
|
712 T2H = T2F + T2G;
|
Chris@10
|
713 T2Z = T2X + T2Y;
|
Chris@10
|
714 T22 = TY + T11;
|
Chris@10
|
715 }
|
Chris@10
|
716 {
|
Chris@10
|
717 E Tp, T31, T16, T2J, Ts, T2K, T19, T32;
|
Chris@10
|
718 {
|
Chris@10
|
719 E Tn, To, T14, T15;
|
Chris@10
|
720 Tn = Rp[WS(rs, 8)];
|
Chris@10
|
721 To = Rm[WS(rs, 1)];
|
Chris@10
|
722 Tp = Tn + To;
|
Chris@10
|
723 T31 = Tn - To;
|
Chris@10
|
724 T14 = Ip[WS(rs, 8)];
|
Chris@10
|
725 T15 = Im[WS(rs, 1)];
|
Chris@10
|
726 T16 = T14 - T15;
|
Chris@10
|
727 T2J = T14 + T15;
|
Chris@10
|
728 }
|
Chris@10
|
729 {
|
Chris@10
|
730 E Tq, Tr, T17, T18;
|
Chris@10
|
731 Tq = Rm[WS(rs, 6)];
|
Chris@10
|
732 Tr = Rp[WS(rs, 3)];
|
Chris@10
|
733 Ts = Tq + Tr;
|
Chris@10
|
734 T2K = Tq - Tr;
|
Chris@10
|
735 T17 = Ip[WS(rs, 3)];
|
Chris@10
|
736 T18 = Im[WS(rs, 6)];
|
Chris@10
|
737 T19 = T17 - T18;
|
Chris@10
|
738 T32 = T17 + T18;
|
Chris@10
|
739 }
|
Chris@10
|
740 Tt = Tp + Ts;
|
Chris@10
|
741 T3P = T31 + T32;
|
Chris@10
|
742 T40 = T2K + T2J;
|
Chris@10
|
743 TI = Tp - Ts;
|
Chris@10
|
744 T1a = T16 - T19;
|
Chris@10
|
745 T2L = T2J - T2K;
|
Chris@10
|
746 T33 = T31 - T32;
|
Chris@10
|
747 T24 = T16 + T19;
|
Chris@10
|
748 }
|
Chris@10
|
749 T13 = TV - T12;
|
Chris@10
|
750 T3n = T2W - T2Z;
|
Chris@10
|
751 T3o = T33 - T36;
|
Chris@10
|
752 T1i = T1a - T1h;
|
Chris@10
|
753 T26 = T24 - T25;
|
Chris@10
|
754 T4e = T3P - T3Q;
|
Chris@10
|
755 T4d = T3M - T3N;
|
Chris@10
|
756 T23 = T21 - T22;
|
Chris@10
|
757 T1n = TI - TJ;
|
Chris@10
|
758 T42 = T40 - T41;
|
Chris@10
|
759 T3Z = T3X - T3Y;
|
Chris@10
|
760 T1m = TF - TG;
|
Chris@10
|
761 T2h = Te - Tl;
|
Chris@10
|
762 T2I = T2E + T2H;
|
Chris@10
|
763 T2i = Tt - TA;
|
Chris@10
|
764 T2P = T2L + T2O;
|
Chris@10
|
765 T30 = T2W + T2Z;
|
Chris@10
|
766 T37 = T33 + T36;
|
Chris@10
|
767 T38 = T30 + T37;
|
Chris@10
|
768 Tm = Te + Tl;
|
Chris@10
|
769 TB = Tt + TA;
|
Chris@10
|
770 TC = Tm + TB;
|
Chris@10
|
771 T46 = T3X + T3Y;
|
Chris@10
|
772 T47 = T40 + T41;
|
Chris@10
|
773 T4a = T46 + T47;
|
Chris@10
|
774 T2a = T21 + T22;
|
Chris@10
|
775 T2b = T24 + T25;
|
Chris@10
|
776 T2e = T2a + T2b;
|
Chris@10
|
777 T1w = TV + T12;
|
Chris@10
|
778 T1x = T1a + T1h;
|
Chris@10
|
779 T1y = T1w + T1x;
|
Chris@10
|
780 T3O = T3M + T3N;
|
Chris@10
|
781 T3R = T3P + T3Q;
|
Chris@10
|
782 T3U = T3O + T3R;
|
Chris@10
|
783 T3h = T2E - T2H;
|
Chris@10
|
784 T3i = T2L - T2O;
|
Chris@10
|
785 T3j = T3h + T3i;
|
Chris@10
|
786 TH = TF + TG;
|
Chris@10
|
787 TK = TI + TJ;
|
Chris@10
|
788 TL = TH + TK;
|
Chris@10
|
789 }
|
Chris@10
|
790 Rp[0] = T7 + TC;
|
Chris@10
|
791 Rm[0] = T2d + T2e;
|
Chris@10
|
792 {
|
Chris@10
|
793 E T1U, T1W, T1T, T1V;
|
Chris@10
|
794 T1U = TE + TL;
|
Chris@10
|
795 T1W = T1v + T1y;
|
Chris@10
|
796 T1T = W[18];
|
Chris@10
|
797 T1V = W[19];
|
Chris@10
|
798 Rp[WS(rs, 5)] = FNMS(T1V, T1W, T1T * T1U);
|
Chris@10
|
799 Rm[WS(rs, 5)] = FMA(T1V, T1U, T1T * T1W);
|
Chris@10
|
800 }
|
Chris@10
|
801 {
|
Chris@10
|
802 E T4y, T4A, T4x, T4z;
|
Chris@10
|
803 T4y = T3T + T3U;
|
Chris@10
|
804 T4A = T49 + T4a;
|
Chris@10
|
805 T4x = W[8];
|
Chris@10
|
806 T4z = W[9];
|
Chris@10
|
807 Ip[WS(rs, 2)] = FNMS(T4z, T4A, T4x * T4y);
|
Chris@10
|
808 Im[WS(rs, 2)] = FMA(T4x, T4A, T4z * T4y);
|
Chris@10
|
809 }
|
Chris@10
|
810 {
|
Chris@10
|
811 E T3I, T3K, T3H, T3J;
|
Chris@10
|
812 T3I = T2T + T38;
|
Chris@10
|
813 T3K = T3g + T3j;
|
Chris@10
|
814 T3H = W[28];
|
Chris@10
|
815 T3J = W[29];
|
Chris@10
|
816 Ip[WS(rs, 7)] = FNMS(T3J, T3K, T3H * T3I);
|
Chris@10
|
817 Im[WS(rs, 7)] = FMA(T3H, T3K, T3J * T3I);
|
Chris@10
|
818 }
|
Chris@10
|
819 {
|
Chris@10
|
820 E T27, T2j, T2v, T2r, T2g, T2u, T20, T2q;
|
Chris@10
|
821 T27 = FMA(KP951056516, T23, KP587785252 * T26);
|
Chris@10
|
822 T2j = FMA(KP951056516, T2h, KP587785252 * T2i);
|
Chris@10
|
823 T2v = FNMS(KP951056516, T2i, KP587785252 * T2h);
|
Chris@10
|
824 T2r = FNMS(KP951056516, T26, KP587785252 * T23);
|
Chris@10
|
825 {
|
Chris@10
|
826 E T2c, T2f, T1Y, T1Z;
|
Chris@10
|
827 T2c = KP559016994 * (T2a - T2b);
|
Chris@10
|
828 T2f = FNMS(KP250000000, T2e, T2d);
|
Chris@10
|
829 T2g = T2c + T2f;
|
Chris@10
|
830 T2u = T2f - T2c;
|
Chris@10
|
831 T1Y = KP559016994 * (Tm - TB);
|
Chris@10
|
832 T1Z = FNMS(KP250000000, TC, T7);
|
Chris@10
|
833 T20 = T1Y + T1Z;
|
Chris@10
|
834 T2q = T1Z - T1Y;
|
Chris@10
|
835 }
|
Chris@10
|
836 {
|
Chris@10
|
837 E T28, T2k, T1X, T29;
|
Chris@10
|
838 T28 = T20 + T27;
|
Chris@10
|
839 T2k = T2g - T2j;
|
Chris@10
|
840 T1X = W[6];
|
Chris@10
|
841 T29 = W[7];
|
Chris@10
|
842 Rp[WS(rs, 2)] = FNMS(T29, T2k, T1X * T28);
|
Chris@10
|
843 Rm[WS(rs, 2)] = FMA(T29, T28, T1X * T2k);
|
Chris@10
|
844 }
|
Chris@10
|
845 {
|
Chris@10
|
846 E T2y, T2A, T2x, T2z;
|
Chris@10
|
847 T2y = T2q - T2r;
|
Chris@10
|
848 T2A = T2v + T2u;
|
Chris@10
|
849 T2x = W[22];
|
Chris@10
|
850 T2z = W[23];
|
Chris@10
|
851 Rp[WS(rs, 6)] = FNMS(T2z, T2A, T2x * T2y);
|
Chris@10
|
852 Rm[WS(rs, 6)] = FMA(T2z, T2y, T2x * T2A);
|
Chris@10
|
853 }
|
Chris@10
|
854 {
|
Chris@10
|
855 E T2m, T2o, T2l, T2n;
|
Chris@10
|
856 T2m = T20 - T27;
|
Chris@10
|
857 T2o = T2j + T2g;
|
Chris@10
|
858 T2l = W[30];
|
Chris@10
|
859 T2n = W[31];
|
Chris@10
|
860 Rp[WS(rs, 8)] = FNMS(T2n, T2o, T2l * T2m);
|
Chris@10
|
861 Rm[WS(rs, 8)] = FMA(T2n, T2m, T2l * T2o);
|
Chris@10
|
862 }
|
Chris@10
|
863 {
|
Chris@10
|
864 E T2s, T2w, T2p, T2t;
|
Chris@10
|
865 T2s = T2q + T2r;
|
Chris@10
|
866 T2w = T2u - T2v;
|
Chris@10
|
867 T2p = W[14];
|
Chris@10
|
868 T2t = W[15];
|
Chris@10
|
869 Rp[WS(rs, 4)] = FNMS(T2t, T2w, T2p * T2s);
|
Chris@10
|
870 Rm[WS(rs, 4)] = FMA(T2t, T2s, T2p * T2w);
|
Chris@10
|
871 }
|
Chris@10
|
872 }
|
Chris@10
|
873 {
|
Chris@10
|
874 E T43, T4f, T4r, T4m, T4c, T4q, T3W, T4n;
|
Chris@10
|
875 T43 = FMA(KP951056516, T3Z, KP587785252 * T42);
|
Chris@10
|
876 T4f = FMA(KP951056516, T4d, KP587785252 * T4e);
|
Chris@10
|
877 T4r = FNMS(KP951056516, T4e, KP587785252 * T4d);
|
Chris@10
|
878 T4m = FNMS(KP951056516, T42, KP587785252 * T3Z);
|
Chris@10
|
879 {
|
Chris@10
|
880 E T48, T4b, T3S, T3V;
|
Chris@10
|
881 T48 = KP559016994 * (T46 - T47);
|
Chris@10
|
882 T4b = FNMS(KP250000000, T4a, T49);
|
Chris@10
|
883 T4c = T48 + T4b;
|
Chris@10
|
884 T4q = T4b - T48;
|
Chris@10
|
885 T3S = KP559016994 * (T3O - T3R);
|
Chris@10
|
886 T3V = FNMS(KP250000000, T3U, T3T);
|
Chris@10
|
887 T3W = T3S + T3V;
|
Chris@10
|
888 T4n = T3V - T3S;
|
Chris@10
|
889 }
|
Chris@10
|
890 {
|
Chris@10
|
891 E T44, T4g, T3L, T45;
|
Chris@10
|
892 T44 = T3W - T43;
|
Chris@10
|
893 T4g = T4c + T4f;
|
Chris@10
|
894 T3L = W[0];
|
Chris@10
|
895 T45 = W[1];
|
Chris@10
|
896 Ip[0] = FNMS(T45, T4g, T3L * T44);
|
Chris@10
|
897 Im[0] = FMA(T3L, T4g, T45 * T44);
|
Chris@10
|
898 }
|
Chris@10
|
899 {
|
Chris@10
|
900 E T4u, T4w, T4t, T4v;
|
Chris@10
|
901 T4u = T4n - T4m;
|
Chris@10
|
902 T4w = T4q + T4r;
|
Chris@10
|
903 T4t = W[32];
|
Chris@10
|
904 T4v = W[33];
|
Chris@10
|
905 Ip[WS(rs, 8)] = FNMS(T4v, T4w, T4t * T4u);
|
Chris@10
|
906 Im[WS(rs, 8)] = FMA(T4t, T4w, T4v * T4u);
|
Chris@10
|
907 }
|
Chris@10
|
908 {
|
Chris@10
|
909 E T4i, T4k, T4h, T4j;
|
Chris@10
|
910 T4i = T43 + T3W;
|
Chris@10
|
911 T4k = T4c - T4f;
|
Chris@10
|
912 T4h = W[16];
|
Chris@10
|
913 T4j = W[17];
|
Chris@10
|
914 Ip[WS(rs, 4)] = FNMS(T4j, T4k, T4h * T4i);
|
Chris@10
|
915 Im[WS(rs, 4)] = FMA(T4h, T4k, T4j * T4i);
|
Chris@10
|
916 }
|
Chris@10
|
917 {
|
Chris@10
|
918 E T4o, T4s, T4l, T4p;
|
Chris@10
|
919 T4o = T4m + T4n;
|
Chris@10
|
920 T4s = T4q - T4r;
|
Chris@10
|
921 T4l = W[24];
|
Chris@10
|
922 T4p = W[25];
|
Chris@10
|
923 Ip[WS(rs, 6)] = FNMS(T4p, T4s, T4l * T4o);
|
Chris@10
|
924 Im[WS(rs, 6)] = FMA(T4l, T4s, T4p * T4o);
|
Chris@10
|
925 }
|
Chris@10
|
926 }
|
Chris@10
|
927 {
|
Chris@10
|
928 E T1j, T1o, T1M, T1J, T1B, T1N, TO, T1I;
|
Chris@10
|
929 T1j = FNMS(KP951056516, T1i, KP587785252 * T13);
|
Chris@10
|
930 T1o = FNMS(KP951056516, T1n, KP587785252 * T1m);
|
Chris@10
|
931 T1M = FMA(KP951056516, T1m, KP587785252 * T1n);
|
Chris@10
|
932 T1J = FMA(KP951056516, T13, KP587785252 * T1i);
|
Chris@10
|
933 {
|
Chris@10
|
934 E T1z, T1A, TM, TN;
|
Chris@10
|
935 T1z = FNMS(KP250000000, T1y, T1v);
|
Chris@10
|
936 T1A = KP559016994 * (T1w - T1x);
|
Chris@10
|
937 T1B = T1z - T1A;
|
Chris@10
|
938 T1N = T1A + T1z;
|
Chris@10
|
939 TM = FNMS(KP250000000, TL, TE);
|
Chris@10
|
940 TN = KP559016994 * (TH - TK);
|
Chris@10
|
941 TO = TM - TN;
|
Chris@10
|
942 T1I = TN + TM;
|
Chris@10
|
943 }
|
Chris@10
|
944 {
|
Chris@10
|
945 E T1k, T1C, TD, T1l;
|
Chris@10
|
946 T1k = TO - T1j;
|
Chris@10
|
947 T1C = T1o + T1B;
|
Chris@10
|
948 TD = W[2];
|
Chris@10
|
949 T1l = W[3];
|
Chris@10
|
950 Rp[WS(rs, 1)] = FNMS(T1l, T1C, TD * T1k);
|
Chris@10
|
951 Rm[WS(rs, 1)] = FMA(T1l, T1k, TD * T1C);
|
Chris@10
|
952 }
|
Chris@10
|
953 {
|
Chris@10
|
954 E T1Q, T1S, T1P, T1R;
|
Chris@10
|
955 T1Q = T1I + T1J;
|
Chris@10
|
956 T1S = T1N - T1M;
|
Chris@10
|
957 T1P = W[26];
|
Chris@10
|
958 T1R = W[27];
|
Chris@10
|
959 Rp[WS(rs, 7)] = FNMS(T1R, T1S, T1P * T1Q);
|
Chris@10
|
960 Rm[WS(rs, 7)] = FMA(T1R, T1Q, T1P * T1S);
|
Chris@10
|
961 }
|
Chris@10
|
962 {
|
Chris@10
|
963 E T1E, T1G, T1D, T1F;
|
Chris@10
|
964 T1E = TO + T1j;
|
Chris@10
|
965 T1G = T1B - T1o;
|
Chris@10
|
966 T1D = W[34];
|
Chris@10
|
967 T1F = W[35];
|
Chris@10
|
968 Rp[WS(rs, 9)] = FNMS(T1F, T1G, T1D * T1E);
|
Chris@10
|
969 Rm[WS(rs, 9)] = FMA(T1F, T1E, T1D * T1G);
|
Chris@10
|
970 }
|
Chris@10
|
971 {
|
Chris@10
|
972 E T1K, T1O, T1H, T1L;
|
Chris@10
|
973 T1K = T1I - T1J;
|
Chris@10
|
974 T1O = T1M + T1N;
|
Chris@10
|
975 T1H = W[10];
|
Chris@10
|
976 T1L = W[11];
|
Chris@10
|
977 Rp[WS(rs, 3)] = FNMS(T1L, T1O, T1H * T1K);
|
Chris@10
|
978 Rm[WS(rs, 3)] = FMA(T1L, T1K, T1H * T1O);
|
Chris@10
|
979 }
|
Chris@10
|
980 }
|
Chris@10
|
981 {
|
Chris@10
|
982 E T2Q, T3p, T3B, T3x, T3m, T3A, T3b, T3w;
|
Chris@10
|
983 T2Q = FNMS(KP951056516, T2P, KP587785252 * T2I);
|
Chris@10
|
984 T3p = FNMS(KP951056516, T3o, KP587785252 * T3n);
|
Chris@10
|
985 T3B = FMA(KP951056516, T3n, KP587785252 * T3o);
|
Chris@10
|
986 T3x = FMA(KP951056516, T2I, KP587785252 * T2P);
|
Chris@10
|
987 {
|
Chris@10
|
988 E T3k, T3l, T39, T3a;
|
Chris@10
|
989 T3k = FNMS(KP250000000, T3j, T3g);
|
Chris@10
|
990 T3l = KP559016994 * (T3h - T3i);
|
Chris@10
|
991 T3m = T3k - T3l;
|
Chris@10
|
992 T3A = T3l + T3k;
|
Chris@10
|
993 T39 = FNMS(KP250000000, T38, T2T);
|
Chris@10
|
994 T3a = KP559016994 * (T30 - T37);
|
Chris@10
|
995 T3b = T39 - T3a;
|
Chris@10
|
996 T3w = T3a + T39;
|
Chris@10
|
997 }
|
Chris@10
|
998 {
|
Chris@10
|
999 E T3c, T3q, T2B, T3d;
|
Chris@10
|
1000 T3c = T2Q + T3b;
|
Chris@10
|
1001 T3q = T3m - T3p;
|
Chris@10
|
1002 T2B = W[4];
|
Chris@10
|
1003 T3d = W[5];
|
Chris@10
|
1004 Ip[WS(rs, 1)] = FNMS(T3d, T3q, T2B * T3c);
|
Chris@10
|
1005 Im[WS(rs, 1)] = FMA(T2B, T3q, T3d * T3c);
|
Chris@10
|
1006 }
|
Chris@10
|
1007 {
|
Chris@10
|
1008 E T3E, T3G, T3D, T3F;
|
Chris@10
|
1009 T3E = T3x + T3w;
|
Chris@10
|
1010 T3G = T3A - T3B;
|
Chris@10
|
1011 T3D = W[36];
|
Chris@10
|
1012 T3F = W[37];
|
Chris@10
|
1013 Ip[WS(rs, 9)] = FNMS(T3F, T3G, T3D * T3E);
|
Chris@10
|
1014 Im[WS(rs, 9)] = FMA(T3D, T3G, T3F * T3E);
|
Chris@10
|
1015 }
|
Chris@10
|
1016 {
|
Chris@10
|
1017 E T3s, T3u, T3r, T3t;
|
Chris@10
|
1018 T3s = T3b - T2Q;
|
Chris@10
|
1019 T3u = T3m + T3p;
|
Chris@10
|
1020 T3r = W[12];
|
Chris@10
|
1021 T3t = W[13];
|
Chris@10
|
1022 Ip[WS(rs, 3)] = FNMS(T3t, T3u, T3r * T3s);
|
Chris@10
|
1023 Im[WS(rs, 3)] = FMA(T3r, T3u, T3t * T3s);
|
Chris@10
|
1024 }
|
Chris@10
|
1025 {
|
Chris@10
|
1026 E T3y, T3C, T3v, T3z;
|
Chris@10
|
1027 T3y = T3w - T3x;
|
Chris@10
|
1028 T3C = T3A + T3B;
|
Chris@10
|
1029 T3v = W[20];
|
Chris@10
|
1030 T3z = W[21];
|
Chris@10
|
1031 Ip[WS(rs, 5)] = FNMS(T3z, T3C, T3v * T3y);
|
Chris@10
|
1032 Im[WS(rs, 5)] = FMA(T3v, T3C, T3z * T3y);
|
Chris@10
|
1033 }
|
Chris@10
|
1034 }
|
Chris@10
|
1035 }
|
Chris@10
|
1036 }
|
Chris@10
|
1037 }
|
Chris@10
|
1038
|
Chris@10
|
1039 static const tw_instr twinstr[] = {
|
Chris@10
|
1040 {TW_FULL, 1, 20},
|
Chris@10
|
1041 {TW_NEXT, 1, 0}
|
Chris@10
|
1042 };
|
Chris@10
|
1043
|
Chris@10
|
1044 static const hc2c_desc desc = { 20, "hc2cb_20", twinstr, &GENUS, {184, 62, 62, 0} };
|
Chris@10
|
1045
|
Chris@10
|
1046 void X(codelet_hc2cb_20) (planner *p) {
|
Chris@10
|
1047 X(khc2c_register) (p, hc2cb_20, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
1048 }
|
Chris@10
|
1049 #endif /* HAVE_FMA */
|