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 07:40:41 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 -n 20 -dit -name hc2cf_20 -include hc2cf.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 * 97 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hc2cf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hc2cf_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 T4P, T4Y, T50, T4U, T4S, T4T, T4Z, T4V;
|
Chris@10
|
47 {
|
Chris@10
|
48 E T4N, T4r, T8, T2i, T4n, T2n, T4O, Tl, T2v, T3v, T3T, T4f, TN, T2b, T3F;
|
Chris@10
|
49 E T3p, T2R, T3z, T43, T4b, T27, T2f, T3J, T33, T2K, T3y, T40, T4c, T1G, T2e;
|
Chris@10
|
50 E T3I, T3a, T2C, T3w, T3W, T4e, T1e, T2c, T3G, T3i;
|
Chris@10
|
51 {
|
Chris@10
|
52 E T1, T4q, T3, T6, T2, T5;
|
Chris@10
|
53 T1 = Rp[0];
|
Chris@10
|
54 T4q = Rm[0];
|
Chris@10
|
55 T3 = Rp[WS(rs, 5)];
|
Chris@10
|
56 T6 = Rm[WS(rs, 5)];
|
Chris@10
|
57 T2 = W[18];
|
Chris@10
|
58 T5 = W[19];
|
Chris@10
|
59 {
|
Chris@10
|
60 E Ta, Td, Tg, T2j, Tb, Tj, Tf, Tc, Ti;
|
Chris@10
|
61 {
|
Chris@10
|
62 E T4o, T4, T9, T4p, T7;
|
Chris@10
|
63 Ta = Ip[WS(rs, 2)];
|
Chris@10
|
64 Td = Im[WS(rs, 2)];
|
Chris@10
|
65 T4o = T2 * T6;
|
Chris@10
|
66 T4 = T2 * T3;
|
Chris@10
|
67 T9 = W[8];
|
Chris@10
|
68 Tg = Ip[WS(rs, 7)];
|
Chris@10
|
69 T4p = FNMS(T5, T3, T4o);
|
Chris@10
|
70 T7 = FMA(T5, T6, T4);
|
Chris@10
|
71 T2j = T9 * Td;
|
Chris@10
|
72 Tb = T9 * Ta;
|
Chris@10
|
73 T4N = T4q - T4p;
|
Chris@10
|
74 T4r = T4p + T4q;
|
Chris@10
|
75 T8 = T1 + T7;
|
Chris@10
|
76 T2i = T1 - T7;
|
Chris@10
|
77 Tj = Im[WS(rs, 7)];
|
Chris@10
|
78 Tf = W[28];
|
Chris@10
|
79 }
|
Chris@10
|
80 Tc = W[9];
|
Chris@10
|
81 Ti = W[29];
|
Chris@10
|
82 {
|
Chris@10
|
83 E T3l, Ts, T2t, TL, TB, TE, TD, T3n, Ty, T2q, TC;
|
Chris@10
|
84 {
|
Chris@10
|
85 E TH, TK, TJ, T2s, TI;
|
Chris@10
|
86 {
|
Chris@10
|
87 E To, Tr, Tp, T3k, Tq, TG;
|
Chris@10
|
88 {
|
Chris@10
|
89 E T2k, Te, T2m, Tk, T2l, Th, Tn;
|
Chris@10
|
90 To = Rp[WS(rs, 2)];
|
Chris@10
|
91 T2l = Tf * Tj;
|
Chris@10
|
92 Th = Tf * Tg;
|
Chris@10
|
93 T2k = FNMS(Tc, Ta, T2j);
|
Chris@10
|
94 Te = FMA(Tc, Td, Tb);
|
Chris@10
|
95 T2m = FNMS(Ti, Tg, T2l);
|
Chris@10
|
96 Tk = FMA(Ti, Tj, Th);
|
Chris@10
|
97 Tr = Rm[WS(rs, 2)];
|
Chris@10
|
98 Tn = W[6];
|
Chris@10
|
99 T4n = T2k + T2m;
|
Chris@10
|
100 T2n = T2k - T2m;
|
Chris@10
|
101 T4O = Te - Tk;
|
Chris@10
|
102 Tl = Te + Tk;
|
Chris@10
|
103 Tp = Tn * To;
|
Chris@10
|
104 T3k = Tn * Tr;
|
Chris@10
|
105 }
|
Chris@10
|
106 Tq = W[7];
|
Chris@10
|
107 TH = Ip[WS(rs, 9)];
|
Chris@10
|
108 TK = Im[WS(rs, 9)];
|
Chris@10
|
109 TG = W[36];
|
Chris@10
|
110 T3l = FNMS(Tq, To, T3k);
|
Chris@10
|
111 Ts = FMA(Tq, Tr, Tp);
|
Chris@10
|
112 TJ = W[37];
|
Chris@10
|
113 T2s = TG * TK;
|
Chris@10
|
114 TI = TG * TH;
|
Chris@10
|
115 }
|
Chris@10
|
116 {
|
Chris@10
|
117 E Tu, Tx, Tt, Tw, T3m, Tv, TA;
|
Chris@10
|
118 Tu = Rp[WS(rs, 7)];
|
Chris@10
|
119 Tx = Rm[WS(rs, 7)];
|
Chris@10
|
120 T2t = FNMS(TJ, TH, T2s);
|
Chris@10
|
121 TL = FMA(TJ, TK, TI);
|
Chris@10
|
122 Tt = W[26];
|
Chris@10
|
123 Tw = W[27];
|
Chris@10
|
124 TB = Ip[WS(rs, 4)];
|
Chris@10
|
125 TE = Im[WS(rs, 4)];
|
Chris@10
|
126 T3m = Tt * Tx;
|
Chris@10
|
127 Tv = Tt * Tu;
|
Chris@10
|
128 TA = W[16];
|
Chris@10
|
129 TD = W[17];
|
Chris@10
|
130 T3n = FNMS(Tw, Tu, T3m);
|
Chris@10
|
131 Ty = FMA(Tw, Tx, Tv);
|
Chris@10
|
132 T2q = TA * TE;
|
Chris@10
|
133 TC = TA * TB;
|
Chris@10
|
134 }
|
Chris@10
|
135 }
|
Chris@10
|
136 {
|
Chris@10
|
137 E T3o, T3R, Tz, T2p, T2r, TF;
|
Chris@10
|
138 T3o = T3l - T3n;
|
Chris@10
|
139 T3R = T3l + T3n;
|
Chris@10
|
140 Tz = Ts + Ty;
|
Chris@10
|
141 T2p = Ts - Ty;
|
Chris@10
|
142 T2r = FNMS(TD, TB, T2q);
|
Chris@10
|
143 TF = FMA(TD, TE, TC);
|
Chris@10
|
144 {
|
Chris@10
|
145 E T3S, T2u, TM, T3j;
|
Chris@10
|
146 T3S = T2r + T2t;
|
Chris@10
|
147 T2u = T2r - T2t;
|
Chris@10
|
148 TM = TF + TL;
|
Chris@10
|
149 T3j = TL - TF;
|
Chris@10
|
150 T2v = T2p - T2u;
|
Chris@10
|
151 T3v = T2p + T2u;
|
Chris@10
|
152 T3T = T3R + T3S;
|
Chris@10
|
153 T4f = T3S - T3R;
|
Chris@10
|
154 TN = Tz - TM;
|
Chris@10
|
155 T2b = Tz + TM;
|
Chris@10
|
156 T3F = T3o + T3j;
|
Chris@10
|
157 T3p = T3j - T3o;
|
Chris@10
|
158 }
|
Chris@10
|
159 }
|
Chris@10
|
160 }
|
Chris@10
|
161 }
|
Chris@10
|
162 }
|
Chris@10
|
163 {
|
Chris@10
|
164 E T2Z, T1M, T2P, T25, T1V, T1Y, T1X, T31, T1S, T2M, T1W;
|
Chris@10
|
165 {
|
Chris@10
|
166 E T21, T24, T23, T2O, T22;
|
Chris@10
|
167 {
|
Chris@10
|
168 E T1I, T1L, T1H, T1K, T2Y, T1J, T20;
|
Chris@10
|
169 T1I = Rp[WS(rs, 6)];
|
Chris@10
|
170 T1L = Rm[WS(rs, 6)];
|
Chris@10
|
171 T1H = W[22];
|
Chris@10
|
172 T1K = W[23];
|
Chris@10
|
173 T21 = Ip[WS(rs, 3)];
|
Chris@10
|
174 T24 = Im[WS(rs, 3)];
|
Chris@10
|
175 T2Y = T1H * T1L;
|
Chris@10
|
176 T1J = T1H * T1I;
|
Chris@10
|
177 T20 = W[12];
|
Chris@10
|
178 T23 = W[13];
|
Chris@10
|
179 T2Z = FNMS(T1K, T1I, T2Y);
|
Chris@10
|
180 T1M = FMA(T1K, T1L, T1J);
|
Chris@10
|
181 T2O = T20 * T24;
|
Chris@10
|
182 T22 = T20 * T21;
|
Chris@10
|
183 }
|
Chris@10
|
184 {
|
Chris@10
|
185 E T1O, T1R, T1N, T1Q, T30, T1P, T1U;
|
Chris@10
|
186 T1O = Rp[WS(rs, 1)];
|
Chris@10
|
187 T1R = Rm[WS(rs, 1)];
|
Chris@10
|
188 T2P = FNMS(T23, T21, T2O);
|
Chris@10
|
189 T25 = FMA(T23, T24, T22);
|
Chris@10
|
190 T1N = W[2];
|
Chris@10
|
191 T1Q = W[3];
|
Chris@10
|
192 T1V = Ip[WS(rs, 8)];
|
Chris@10
|
193 T1Y = Im[WS(rs, 8)];
|
Chris@10
|
194 T30 = T1N * T1R;
|
Chris@10
|
195 T1P = T1N * T1O;
|
Chris@10
|
196 T1U = W[32];
|
Chris@10
|
197 T1X = W[33];
|
Chris@10
|
198 T31 = FNMS(T1Q, T1O, T30);
|
Chris@10
|
199 T1S = FMA(T1Q, T1R, T1P);
|
Chris@10
|
200 T2M = T1U * T1Y;
|
Chris@10
|
201 T1W = T1U * T1V;
|
Chris@10
|
202 }
|
Chris@10
|
203 }
|
Chris@10
|
204 {
|
Chris@10
|
205 E T32, T41, T1T, T2L, T2N, T1Z;
|
Chris@10
|
206 T32 = T2Z - T31;
|
Chris@10
|
207 T41 = T2Z + T31;
|
Chris@10
|
208 T1T = T1M + T1S;
|
Chris@10
|
209 T2L = T1M - T1S;
|
Chris@10
|
210 T2N = FNMS(T1X, T1V, T2M);
|
Chris@10
|
211 T1Z = FMA(T1X, T1Y, T1W);
|
Chris@10
|
212 {
|
Chris@10
|
213 E T42, T2Q, T26, T2X;
|
Chris@10
|
214 T42 = T2N + T2P;
|
Chris@10
|
215 T2Q = T2N - T2P;
|
Chris@10
|
216 T26 = T1Z + T25;
|
Chris@10
|
217 T2X = T25 - T1Z;
|
Chris@10
|
218 T2R = T2L - T2Q;
|
Chris@10
|
219 T3z = T2L + T2Q;
|
Chris@10
|
220 T43 = T41 + T42;
|
Chris@10
|
221 T4b = T42 - T41;
|
Chris@10
|
222 T27 = T1T - T26;
|
Chris@10
|
223 T2f = T1T + T26;
|
Chris@10
|
224 T3J = T32 + T2X;
|
Chris@10
|
225 T33 = T2X - T32;
|
Chris@10
|
226 }
|
Chris@10
|
227 }
|
Chris@10
|
228 }
|
Chris@10
|
229 {
|
Chris@10
|
230 E T36, T1l, T2I, T1E, T1u, T1x, T1w, T38, T1r, T2F, T1v;
|
Chris@10
|
231 {
|
Chris@10
|
232 E T1A, T1D, T1C, T2H, T1B;
|
Chris@10
|
233 {
|
Chris@10
|
234 E T1h, T1k, T1g, T1j, T35, T1i, T1z;
|
Chris@10
|
235 T1h = Rp[WS(rs, 4)];
|
Chris@10
|
236 T1k = Rm[WS(rs, 4)];
|
Chris@10
|
237 T1g = W[14];
|
Chris@10
|
238 T1j = W[15];
|
Chris@10
|
239 T1A = Ip[WS(rs, 1)];
|
Chris@10
|
240 T1D = Im[WS(rs, 1)];
|
Chris@10
|
241 T35 = T1g * T1k;
|
Chris@10
|
242 T1i = T1g * T1h;
|
Chris@10
|
243 T1z = W[4];
|
Chris@10
|
244 T1C = W[5];
|
Chris@10
|
245 T36 = FNMS(T1j, T1h, T35);
|
Chris@10
|
246 T1l = FMA(T1j, T1k, T1i);
|
Chris@10
|
247 T2H = T1z * T1D;
|
Chris@10
|
248 T1B = T1z * T1A;
|
Chris@10
|
249 }
|
Chris@10
|
250 {
|
Chris@10
|
251 E T1n, T1q, T1m, T1p, T37, T1o, T1t;
|
Chris@10
|
252 T1n = Rp[WS(rs, 9)];
|
Chris@10
|
253 T1q = Rm[WS(rs, 9)];
|
Chris@10
|
254 T2I = FNMS(T1C, T1A, T2H);
|
Chris@10
|
255 T1E = FMA(T1C, T1D, T1B);
|
Chris@10
|
256 T1m = W[34];
|
Chris@10
|
257 T1p = W[35];
|
Chris@10
|
258 T1u = Ip[WS(rs, 6)];
|
Chris@10
|
259 T1x = Im[WS(rs, 6)];
|
Chris@10
|
260 T37 = T1m * T1q;
|
Chris@10
|
261 T1o = T1m * T1n;
|
Chris@10
|
262 T1t = W[24];
|
Chris@10
|
263 T1w = W[25];
|
Chris@10
|
264 T38 = FNMS(T1p, T1n, T37);
|
Chris@10
|
265 T1r = FMA(T1p, T1q, T1o);
|
Chris@10
|
266 T2F = T1t * T1x;
|
Chris@10
|
267 T1v = T1t * T1u;
|
Chris@10
|
268 }
|
Chris@10
|
269 }
|
Chris@10
|
270 {
|
Chris@10
|
271 E T39, T3Y, T1s, T2E, T2G, T1y;
|
Chris@10
|
272 T39 = T36 - T38;
|
Chris@10
|
273 T3Y = T36 + T38;
|
Chris@10
|
274 T1s = T1l + T1r;
|
Chris@10
|
275 T2E = T1l - T1r;
|
Chris@10
|
276 T2G = FNMS(T1w, T1u, T2F);
|
Chris@10
|
277 T1y = FMA(T1w, T1x, T1v);
|
Chris@10
|
278 {
|
Chris@10
|
279 E T3Z, T2J, T1F, T34;
|
Chris@10
|
280 T3Z = T2G + T2I;
|
Chris@10
|
281 T2J = T2G - T2I;
|
Chris@10
|
282 T1F = T1y + T1E;
|
Chris@10
|
283 T34 = T1E - T1y;
|
Chris@10
|
284 T2K = T2E - T2J;
|
Chris@10
|
285 T3y = T2E + T2J;
|
Chris@10
|
286 T40 = T3Y + T3Z;
|
Chris@10
|
287 T4c = T3Z - T3Y;
|
Chris@10
|
288 T1G = T1s - T1F;
|
Chris@10
|
289 T2e = T1s + T1F;
|
Chris@10
|
290 T3I = T39 + T34;
|
Chris@10
|
291 T3a = T34 - T39;
|
Chris@10
|
292 }
|
Chris@10
|
293 }
|
Chris@10
|
294 }
|
Chris@10
|
295 {
|
Chris@10
|
296 E T3e, TT, T2A, T1c, T12, T15, T14, T3g, TZ, T2x, T13;
|
Chris@10
|
297 {
|
Chris@10
|
298 E T18, T1b, T1a, T2z, T19;
|
Chris@10
|
299 {
|
Chris@10
|
300 E TP, TS, TO, TR, T3d, TQ, T17;
|
Chris@10
|
301 TP = Rp[WS(rs, 8)];
|
Chris@10
|
302 TS = Rm[WS(rs, 8)];
|
Chris@10
|
303 TO = W[30];
|
Chris@10
|
304 TR = W[31];
|
Chris@10
|
305 T18 = Ip[WS(rs, 5)];
|
Chris@10
|
306 T1b = Im[WS(rs, 5)];
|
Chris@10
|
307 T3d = TO * TS;
|
Chris@10
|
308 TQ = TO * TP;
|
Chris@10
|
309 T17 = W[20];
|
Chris@10
|
310 T1a = W[21];
|
Chris@10
|
311 T3e = FNMS(TR, TP, T3d);
|
Chris@10
|
312 TT = FMA(TR, TS, TQ);
|
Chris@10
|
313 T2z = T17 * T1b;
|
Chris@10
|
314 T19 = T17 * T18;
|
Chris@10
|
315 }
|
Chris@10
|
316 {
|
Chris@10
|
317 E TV, TY, TU, TX, T3f, TW, T11;
|
Chris@10
|
318 TV = Rp[WS(rs, 3)];
|
Chris@10
|
319 TY = Rm[WS(rs, 3)];
|
Chris@10
|
320 T2A = FNMS(T1a, T18, T2z);
|
Chris@10
|
321 T1c = FMA(T1a, T1b, T19);
|
Chris@10
|
322 TU = W[10];
|
Chris@10
|
323 TX = W[11];
|
Chris@10
|
324 T12 = Ip[0];
|
Chris@10
|
325 T15 = Im[0];
|
Chris@10
|
326 T3f = TU * TY;
|
Chris@10
|
327 TW = TU * TV;
|
Chris@10
|
328 T11 = W[0];
|
Chris@10
|
329 T14 = W[1];
|
Chris@10
|
330 T3g = FNMS(TX, TV, T3f);
|
Chris@10
|
331 TZ = FMA(TX, TY, TW);
|
Chris@10
|
332 T2x = T11 * T15;
|
Chris@10
|
333 T13 = T11 * T12;
|
Chris@10
|
334 }
|
Chris@10
|
335 }
|
Chris@10
|
336 {
|
Chris@10
|
337 E T3h, T3U, T10, T2w, T2y, T16;
|
Chris@10
|
338 T3h = T3e - T3g;
|
Chris@10
|
339 T3U = T3e + T3g;
|
Chris@10
|
340 T10 = TT + TZ;
|
Chris@10
|
341 T2w = TT - TZ;
|
Chris@10
|
342 T2y = FNMS(T14, T12, T2x);
|
Chris@10
|
343 T16 = FMA(T14, T15, T13);
|
Chris@10
|
344 {
|
Chris@10
|
345 E T3V, T2B, T1d, T3c;
|
Chris@10
|
346 T3V = T2y + T2A;
|
Chris@10
|
347 T2B = T2y - T2A;
|
Chris@10
|
348 T1d = T16 + T1c;
|
Chris@10
|
349 T3c = T1c - T16;
|
Chris@10
|
350 T2C = T2w - T2B;
|
Chris@10
|
351 T3w = T2w + T2B;
|
Chris@10
|
352 T3W = T3U + T3V;
|
Chris@10
|
353 T4e = T3V - T3U;
|
Chris@10
|
354 T1e = T10 - T1d;
|
Chris@10
|
355 T2c = T10 + T1d;
|
Chris@10
|
356 T3G = T3h + T3c;
|
Chris@10
|
357 T3i = T3c - T3h;
|
Chris@10
|
358 }
|
Chris@10
|
359 }
|
Chris@10
|
360 }
|
Chris@10
|
361 {
|
Chris@10
|
362 E T4s, T4k, T4l, T45, T47, T3P, T4y, T4A, T3O;
|
Chris@10
|
363 {
|
Chris@10
|
364 E T4C, T4B, T2a, T4j, T4h, T4E, T4M, T4K, T4i, T4a;
|
Chris@10
|
365 {
|
Chris@10
|
366 E Tm, T1f, T4J, T4I, T28, T4d, T4g, T29, T49, T48;
|
Chris@10
|
367 T4C = T4c + T4b;
|
Chris@10
|
368 T4d = T4b - T4c;
|
Chris@10
|
369 T4g = T4e - T4f;
|
Chris@10
|
370 T4B = T4f + T4e;
|
Chris@10
|
371 T2a = T8 + Tl;
|
Chris@10
|
372 Tm = T8 - Tl;
|
Chris@10
|
373 T1f = TN + T1e;
|
Chris@10
|
374 T4J = T1e - TN;
|
Chris@10
|
375 T4I = T1G - T27;
|
Chris@10
|
376 T28 = T1G + T27;
|
Chris@10
|
377 T4j = FMA(KP618033988, T4d, T4g);
|
Chris@10
|
378 T4h = FNMS(KP618033988, T4g, T4d);
|
Chris@10
|
379 T29 = T1f + T28;
|
Chris@10
|
380 T49 = T1f - T28;
|
Chris@10
|
381 T4E = T4r - T4n;
|
Chris@10
|
382 T4s = T4n + T4r;
|
Chris@10
|
383 Rm[WS(rs, 9)] = Tm + T29;
|
Chris@10
|
384 T48 = FNMS(KP250000000, T29, Tm);
|
Chris@10
|
385 T4M = FNMS(KP618033988, T4I, T4J);
|
Chris@10
|
386 T4K = FMA(KP618033988, T4J, T4I);
|
Chris@10
|
387 T4i = FMA(KP559016994, T49, T48);
|
Chris@10
|
388 T4a = FNMS(KP559016994, T49, T48);
|
Chris@10
|
389 }
|
Chris@10
|
390 {
|
Chris@10
|
391 E T2d, T4w, T4x, T2g, T2h;
|
Chris@10
|
392 {
|
Chris@10
|
393 E T3X, T4G, T4F, T44, T4D, T4L, T4H;
|
Chris@10
|
394 T4k = T3T + T3W;
|
Chris@10
|
395 T3X = T3T - T3W;
|
Chris@10
|
396 T4G = T4C - T4B;
|
Chris@10
|
397 T4D = T4B + T4C;
|
Chris@10
|
398 Rm[WS(rs, 1)] = FMA(KP951056516, T4h, T4a);
|
Chris@10
|
399 Rp[WS(rs, 2)] = FNMS(KP951056516, T4h, T4a);
|
Chris@10
|
400 Rp[WS(rs, 6)] = FMA(KP951056516, T4j, T4i);
|
Chris@10
|
401 Rm[WS(rs, 5)] = FNMS(KP951056516, T4j, T4i);
|
Chris@10
|
402 Im[WS(rs, 9)] = T4D - T4E;
|
Chris@10
|
403 T4F = FMA(KP250000000, T4D, T4E);
|
Chris@10
|
404 T44 = T40 - T43;
|
Chris@10
|
405 T4l = T40 + T43;
|
Chris@10
|
406 T2d = T2b + T2c;
|
Chris@10
|
407 T4w = T2b - T2c;
|
Chris@10
|
408 T4L = FMA(KP559016994, T4G, T4F);
|
Chris@10
|
409 T4H = FNMS(KP559016994, T4G, T4F);
|
Chris@10
|
410 T45 = FMA(KP618033988, T44, T3X);
|
Chris@10
|
411 T47 = FNMS(KP618033988, T3X, T44);
|
Chris@10
|
412 Ip[WS(rs, 2)] = FMA(KP951056516, T4K, T4H);
|
Chris@10
|
413 Im[WS(rs, 1)] = FMS(KP951056516, T4K, T4H);
|
Chris@10
|
414 Ip[WS(rs, 6)] = FMA(KP951056516, T4M, T4L);
|
Chris@10
|
415 Im[WS(rs, 5)] = FMS(KP951056516, T4M, T4L);
|
Chris@10
|
416 T4x = T2f - T2e;
|
Chris@10
|
417 T2g = T2e + T2f;
|
Chris@10
|
418 }
|
Chris@10
|
419 T2h = T2d + T2g;
|
Chris@10
|
420 T3P = T2d - T2g;
|
Chris@10
|
421 T4y = FNMS(KP618033988, T4x, T4w);
|
Chris@10
|
422 T4A = FMA(KP618033988, T4w, T4x);
|
Chris@10
|
423 Rp[0] = T2a + T2h;
|
Chris@10
|
424 T3O = FNMS(KP250000000, T2h, T2a);
|
Chris@10
|
425 }
|
Chris@10
|
426 }
|
Chris@10
|
427 {
|
Chris@10
|
428 E T3u, T54, T5a, T5c, T56, T53;
|
Chris@10
|
429 {
|
Chris@10
|
430 E T52, T51, T3t, T3r, T2o, T58, T59, T2T, T2V, T4u, T4t, T2U, T3s, T2W;
|
Chris@10
|
431 {
|
Chris@10
|
432 E T3b, T3q, T46, T3Q, T4m;
|
Chris@10
|
433 T52 = T3a + T33;
|
Chris@10
|
434 T3b = T33 - T3a;
|
Chris@10
|
435 T3q = T3i - T3p;
|
Chris@10
|
436 T51 = T3p + T3i;
|
Chris@10
|
437 T46 = FNMS(KP559016994, T3P, T3O);
|
Chris@10
|
438 T3Q = FMA(KP559016994, T3P, T3O);
|
Chris@10
|
439 T4m = T4k + T4l;
|
Chris@10
|
440 T4u = T4k - T4l;
|
Chris@10
|
441 Rm[WS(rs, 3)] = FMA(KP951056516, T45, T3Q);
|
Chris@10
|
442 Rp[WS(rs, 4)] = FNMS(KP951056516, T45, T3Q);
|
Chris@10
|
443 Rp[WS(rs, 8)] = FMA(KP951056516, T47, T46);
|
Chris@10
|
444 Rm[WS(rs, 7)] = FNMS(KP951056516, T47, T46);
|
Chris@10
|
445 Ip[0] = T4m + T4s;
|
Chris@10
|
446 T4t = FNMS(KP250000000, T4m, T4s);
|
Chris@10
|
447 T3t = FMA(KP618033988, T3b, T3q);
|
Chris@10
|
448 T3r = FNMS(KP618033988, T3q, T3b);
|
Chris@10
|
449 }
|
Chris@10
|
450 T3u = T2i + T2n;
|
Chris@10
|
451 T2o = T2i - T2n;
|
Chris@10
|
452 {
|
Chris@10
|
453 E T4v, T4z, T2D, T2S;
|
Chris@10
|
454 T4v = FMA(KP559016994, T4u, T4t);
|
Chris@10
|
455 T4z = FNMS(KP559016994, T4u, T4t);
|
Chris@10
|
456 T2D = T2v + T2C;
|
Chris@10
|
457 T58 = T2v - T2C;
|
Chris@10
|
458 T59 = T2K - T2R;
|
Chris@10
|
459 T2S = T2K + T2R;
|
Chris@10
|
460 Ip[WS(rs, 4)] = FMA(KP951056516, T4y, T4v);
|
Chris@10
|
461 Im[WS(rs, 3)] = FMS(KP951056516, T4y, T4v);
|
Chris@10
|
462 Ip[WS(rs, 8)] = FMA(KP951056516, T4A, T4z);
|
Chris@10
|
463 Im[WS(rs, 7)] = FMS(KP951056516, T4A, T4z);
|
Chris@10
|
464 T2T = T2D + T2S;
|
Chris@10
|
465 T2V = T2D - T2S;
|
Chris@10
|
466 }
|
Chris@10
|
467 Rm[WS(rs, 4)] = T2o + T2T;
|
Chris@10
|
468 T2U = FNMS(KP250000000, T2T, T2o);
|
Chris@10
|
469 T54 = T4O + T4N;
|
Chris@10
|
470 T4P = T4N - T4O;
|
Chris@10
|
471 T5a = FMA(KP618033988, T59, T58);
|
Chris@10
|
472 T5c = FNMS(KP618033988, T58, T59);
|
Chris@10
|
473 T3s = FMA(KP559016994, T2V, T2U);
|
Chris@10
|
474 T2W = FNMS(KP559016994, T2V, T2U);
|
Chris@10
|
475 Rp[WS(rs, 7)] = FNMS(KP951056516, T3r, T2W);
|
Chris@10
|
476 Rp[WS(rs, 3)] = FMA(KP951056516, T3r, T2W);
|
Chris@10
|
477 Rm[0] = FNMS(KP951056516, T3t, T3s);
|
Chris@10
|
478 Rm[WS(rs, 8)] = FMA(KP951056516, T3t, T3s);
|
Chris@10
|
479 T56 = T51 - T52;
|
Chris@10
|
480 T53 = T51 + T52;
|
Chris@10
|
481 }
|
Chris@10
|
482 {
|
Chris@10
|
483 E T4Q, T4R, T3N, T3L, T4X, T4W, T3B, T3D, T3H, T3K, T55, T3C, T3M, T3E;
|
Chris@10
|
484 T4Q = T3F + T3G;
|
Chris@10
|
485 T3H = T3F - T3G;
|
Chris@10
|
486 T3K = T3I - T3J;
|
Chris@10
|
487 T4R = T3I + T3J;
|
Chris@10
|
488 Im[WS(rs, 4)] = T53 - T54;
|
Chris@10
|
489 T55 = FMA(KP250000000, T53, T54);
|
Chris@10
|
490 T3N = FNMS(KP618033988, T3H, T3K);
|
Chris@10
|
491 T3L = FMA(KP618033988, T3K, T3H);
|
Chris@10
|
492 {
|
Chris@10
|
493 E T57, T5b, T3x, T3A;
|
Chris@10
|
494 T57 = FNMS(KP559016994, T56, T55);
|
Chris@10
|
495 T5b = FMA(KP559016994, T56, T55);
|
Chris@10
|
496 T3x = T3v + T3w;
|
Chris@10
|
497 T4X = T3v - T3w;
|
Chris@10
|
498 T4W = T3y - T3z;
|
Chris@10
|
499 T3A = T3y + T3z;
|
Chris@10
|
500 Im[0] = -(FMA(KP951056516, T5a, T57));
|
Chris@10
|
501 Im[WS(rs, 8)] = FMS(KP951056516, T5a, T57);
|
Chris@10
|
502 Ip[WS(rs, 7)] = FMA(KP951056516, T5c, T5b);
|
Chris@10
|
503 Ip[WS(rs, 3)] = FNMS(KP951056516, T5c, T5b);
|
Chris@10
|
504 T3B = T3x + T3A;
|
Chris@10
|
505 T3D = T3x - T3A;
|
Chris@10
|
506 }
|
Chris@10
|
507 Rp[WS(rs, 5)] = T3u + T3B;
|
Chris@10
|
508 T3C = FNMS(KP250000000, T3B, T3u);
|
Chris@10
|
509 T4Y = FNMS(KP618033988, T4X, T4W);
|
Chris@10
|
510 T50 = FMA(KP618033988, T4W, T4X);
|
Chris@10
|
511 T3M = FNMS(KP559016994, T3D, T3C);
|
Chris@10
|
512 T3E = FMA(KP559016994, T3D, T3C);
|
Chris@10
|
513 Rp[WS(rs, 9)] = FNMS(KP951056516, T3L, T3E);
|
Chris@10
|
514 Rp[WS(rs, 1)] = FMA(KP951056516, T3L, T3E);
|
Chris@10
|
515 Rm[WS(rs, 2)] = FNMS(KP951056516, T3N, T3M);
|
Chris@10
|
516 Rm[WS(rs, 6)] = FMA(KP951056516, T3N, T3M);
|
Chris@10
|
517 T4U = T4Q - T4R;
|
Chris@10
|
518 T4S = T4Q + T4R;
|
Chris@10
|
519 }
|
Chris@10
|
520 }
|
Chris@10
|
521 }
|
Chris@10
|
522 }
|
Chris@10
|
523 Ip[WS(rs, 5)] = T4S + T4P;
|
Chris@10
|
524 T4T = FNMS(KP250000000, T4S, T4P);
|
Chris@10
|
525 T4Z = FMA(KP559016994, T4U, T4T);
|
Chris@10
|
526 T4V = FNMS(KP559016994, T4U, T4T);
|
Chris@10
|
527 Im[WS(rs, 2)] = -(FMA(KP951056516, T4Y, T4V));
|
Chris@10
|
528 Im[WS(rs, 6)] = FMS(KP951056516, T4Y, T4V);
|
Chris@10
|
529 Ip[WS(rs, 9)] = FMA(KP951056516, T50, T4Z);
|
Chris@10
|
530 Ip[WS(rs, 1)] = FNMS(KP951056516, T50, T4Z);
|
Chris@10
|
531 }
|
Chris@10
|
532 }
|
Chris@10
|
533 }
|
Chris@10
|
534
|
Chris@10
|
535 static const tw_instr twinstr[] = {
|
Chris@10
|
536 {TW_FULL, 1, 20},
|
Chris@10
|
537 {TW_NEXT, 1, 0}
|
Chris@10
|
538 };
|
Chris@10
|
539
|
Chris@10
|
540 static const hc2c_desc desc = { 20, "hc2cf_20", twinstr, &GENUS, {136, 38, 110, 0} };
|
Chris@10
|
541
|
Chris@10
|
542 void X(codelet_hc2cf_20) (planner *p) {
|
Chris@10
|
543 X(khc2c_register) (p, hc2cf_20, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
544 }
|
Chris@10
|
545 #else /* HAVE_FMA */
|
Chris@10
|
546
|
Chris@10
|
547 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hc2cf_20 -include hc2cf.h */
|
Chris@10
|
548
|
Chris@10
|
549 /*
|
Chris@10
|
550 * This function contains 246 FP additions, 124 FP multiplications,
|
Chris@10
|
551 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
|
Chris@10
|
552 * 85 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
553 */
|
Chris@10
|
554 #include "hc2cf.h"
|
Chris@10
|
555
|
Chris@10
|
556 static void hc2cf_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
557 {
|
Chris@10
|
558 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
559 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
560 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
561 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
562 {
|
Chris@10
|
563 INT m;
|
Chris@10
|
564 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
|
565 E Tj, T1R, T4j, T4s, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3p, T3s, T3K, T3A;
|
Chris@10
|
566 E T3B, T3Z, T1V, T1W, T1X, T23, T28, T4q, T2W, T2X, T4f, T33, T34, T35, T2G;
|
Chris@10
|
567 E T2L, T2M, TG, T13, T14, T3i, T3l, T3J, T3D, T3E, T40, T1S, T1T, T1U, T2e;
|
Chris@10
|
568 E T2j, T4p, T2T, T2U, T4e, T30, T31, T32, T2v, T2A, T2B;
|
Chris@10
|
569 {
|
Chris@10
|
570 E T1, T3O, T6, T3N, Tc, T2n, Th, T2o;
|
Chris@10
|
571 T1 = Rp[0];
|
Chris@10
|
572 T3O = Rm[0];
|
Chris@10
|
573 {
|
Chris@10
|
574 E T3, T5, T2, T4;
|
Chris@10
|
575 T3 = Rp[WS(rs, 5)];
|
Chris@10
|
576 T5 = Rm[WS(rs, 5)];
|
Chris@10
|
577 T2 = W[18];
|
Chris@10
|
578 T4 = W[19];
|
Chris@10
|
579 T6 = FMA(T2, T3, T4 * T5);
|
Chris@10
|
580 T3N = FNMS(T4, T3, T2 * T5);
|
Chris@10
|
581 }
|
Chris@10
|
582 {
|
Chris@10
|
583 E T9, Tb, T8, Ta;
|
Chris@10
|
584 T9 = Ip[WS(rs, 2)];
|
Chris@10
|
585 Tb = Im[WS(rs, 2)];
|
Chris@10
|
586 T8 = W[8];
|
Chris@10
|
587 Ta = W[9];
|
Chris@10
|
588 Tc = FMA(T8, T9, Ta * Tb);
|
Chris@10
|
589 T2n = FNMS(Ta, T9, T8 * Tb);
|
Chris@10
|
590 }
|
Chris@10
|
591 {
|
Chris@10
|
592 E Te, Tg, Td, Tf;
|
Chris@10
|
593 Te = Ip[WS(rs, 7)];
|
Chris@10
|
594 Tg = Im[WS(rs, 7)];
|
Chris@10
|
595 Td = W[28];
|
Chris@10
|
596 Tf = W[29];
|
Chris@10
|
597 Th = FMA(Td, Te, Tf * Tg);
|
Chris@10
|
598 T2o = FNMS(Tf, Te, Td * Tg);
|
Chris@10
|
599 }
|
Chris@10
|
600 {
|
Chris@10
|
601 E T7, Ti, T4h, T4i;
|
Chris@10
|
602 T7 = T1 + T6;
|
Chris@10
|
603 Ti = Tc + Th;
|
Chris@10
|
604 Tj = T7 - Ti;
|
Chris@10
|
605 T1R = T7 + Ti;
|
Chris@10
|
606 T4h = T3O - T3N;
|
Chris@10
|
607 T4i = Tc - Th;
|
Chris@10
|
608 T4j = T4h - T4i;
|
Chris@10
|
609 T4s = T4i + T4h;
|
Chris@10
|
610 }
|
Chris@10
|
611 {
|
Chris@10
|
612 E T2m, T2p, T3M, T3P;
|
Chris@10
|
613 T2m = T1 - T6;
|
Chris@10
|
614 T2p = T2n - T2o;
|
Chris@10
|
615 T2q = T2m - T2p;
|
Chris@10
|
616 T37 = T2m + T2p;
|
Chris@10
|
617 T3M = T2n + T2o;
|
Chris@10
|
618 T3P = T3N + T3O;
|
Chris@10
|
619 T3Q = T3M + T3P;
|
Chris@10
|
620 T42 = T3P - T3M;
|
Chris@10
|
621 }
|
Chris@10
|
622 }
|
Chris@10
|
623 {
|
Chris@10
|
624 E T1f, T3n, T21, T2C, T1N, T3r, T27, T2K, T1q, T3o, T22, T2F, T1C, T3q, T26;
|
Chris@10
|
625 E T2H;
|
Chris@10
|
626 {
|
Chris@10
|
627 E T19, T1Z, T1e, T20;
|
Chris@10
|
628 {
|
Chris@10
|
629 E T16, T18, T15, T17;
|
Chris@10
|
630 T16 = Rp[WS(rs, 4)];
|
Chris@10
|
631 T18 = Rm[WS(rs, 4)];
|
Chris@10
|
632 T15 = W[14];
|
Chris@10
|
633 T17 = W[15];
|
Chris@10
|
634 T19 = FMA(T15, T16, T17 * T18);
|
Chris@10
|
635 T1Z = FNMS(T17, T16, T15 * T18);
|
Chris@10
|
636 }
|
Chris@10
|
637 {
|
Chris@10
|
638 E T1b, T1d, T1a, T1c;
|
Chris@10
|
639 T1b = Rp[WS(rs, 9)];
|
Chris@10
|
640 T1d = Rm[WS(rs, 9)];
|
Chris@10
|
641 T1a = W[34];
|
Chris@10
|
642 T1c = W[35];
|
Chris@10
|
643 T1e = FMA(T1a, T1b, T1c * T1d);
|
Chris@10
|
644 T20 = FNMS(T1c, T1b, T1a * T1d);
|
Chris@10
|
645 }
|
Chris@10
|
646 T1f = T19 + T1e;
|
Chris@10
|
647 T3n = T1Z + T20;
|
Chris@10
|
648 T21 = T1Z - T20;
|
Chris@10
|
649 T2C = T19 - T1e;
|
Chris@10
|
650 }
|
Chris@10
|
651 {
|
Chris@10
|
652 E T1H, T2I, T1M, T2J;
|
Chris@10
|
653 {
|
Chris@10
|
654 E T1E, T1G, T1D, T1F;
|
Chris@10
|
655 T1E = Ip[WS(rs, 8)];
|
Chris@10
|
656 T1G = Im[WS(rs, 8)];
|
Chris@10
|
657 T1D = W[32];
|
Chris@10
|
658 T1F = W[33];
|
Chris@10
|
659 T1H = FMA(T1D, T1E, T1F * T1G);
|
Chris@10
|
660 T2I = FNMS(T1F, T1E, T1D * T1G);
|
Chris@10
|
661 }
|
Chris@10
|
662 {
|
Chris@10
|
663 E T1J, T1L, T1I, T1K;
|
Chris@10
|
664 T1J = Ip[WS(rs, 3)];
|
Chris@10
|
665 T1L = Im[WS(rs, 3)];
|
Chris@10
|
666 T1I = W[12];
|
Chris@10
|
667 T1K = W[13];
|
Chris@10
|
668 T1M = FMA(T1I, T1J, T1K * T1L);
|
Chris@10
|
669 T2J = FNMS(T1K, T1J, T1I * T1L);
|
Chris@10
|
670 }
|
Chris@10
|
671 T1N = T1H + T1M;
|
Chris@10
|
672 T3r = T2I + T2J;
|
Chris@10
|
673 T27 = T1H - T1M;
|
Chris@10
|
674 T2K = T2I - T2J;
|
Chris@10
|
675 }
|
Chris@10
|
676 {
|
Chris@10
|
677 E T1k, T2D, T1p, T2E;
|
Chris@10
|
678 {
|
Chris@10
|
679 E T1h, T1j, T1g, T1i;
|
Chris@10
|
680 T1h = Ip[WS(rs, 6)];
|
Chris@10
|
681 T1j = Im[WS(rs, 6)];
|
Chris@10
|
682 T1g = W[24];
|
Chris@10
|
683 T1i = W[25];
|
Chris@10
|
684 T1k = FMA(T1g, T1h, T1i * T1j);
|
Chris@10
|
685 T2D = FNMS(T1i, T1h, T1g * T1j);
|
Chris@10
|
686 }
|
Chris@10
|
687 {
|
Chris@10
|
688 E T1m, T1o, T1l, T1n;
|
Chris@10
|
689 T1m = Ip[WS(rs, 1)];
|
Chris@10
|
690 T1o = Im[WS(rs, 1)];
|
Chris@10
|
691 T1l = W[4];
|
Chris@10
|
692 T1n = W[5];
|
Chris@10
|
693 T1p = FMA(T1l, T1m, T1n * T1o);
|
Chris@10
|
694 T2E = FNMS(T1n, T1m, T1l * T1o);
|
Chris@10
|
695 }
|
Chris@10
|
696 T1q = T1k + T1p;
|
Chris@10
|
697 T3o = T2D + T2E;
|
Chris@10
|
698 T22 = T1k - T1p;
|
Chris@10
|
699 T2F = T2D - T2E;
|
Chris@10
|
700 }
|
Chris@10
|
701 {
|
Chris@10
|
702 E T1w, T24, T1B, T25;
|
Chris@10
|
703 {
|
Chris@10
|
704 E T1t, T1v, T1s, T1u;
|
Chris@10
|
705 T1t = Rp[WS(rs, 6)];
|
Chris@10
|
706 T1v = Rm[WS(rs, 6)];
|
Chris@10
|
707 T1s = W[22];
|
Chris@10
|
708 T1u = W[23];
|
Chris@10
|
709 T1w = FMA(T1s, T1t, T1u * T1v);
|
Chris@10
|
710 T24 = FNMS(T1u, T1t, T1s * T1v);
|
Chris@10
|
711 }
|
Chris@10
|
712 {
|
Chris@10
|
713 E T1y, T1A, T1x, T1z;
|
Chris@10
|
714 T1y = Rp[WS(rs, 1)];
|
Chris@10
|
715 T1A = Rm[WS(rs, 1)];
|
Chris@10
|
716 T1x = W[2];
|
Chris@10
|
717 T1z = W[3];
|
Chris@10
|
718 T1B = FMA(T1x, T1y, T1z * T1A);
|
Chris@10
|
719 T25 = FNMS(T1z, T1y, T1x * T1A);
|
Chris@10
|
720 }
|
Chris@10
|
721 T1C = T1w + T1B;
|
Chris@10
|
722 T3q = T24 + T25;
|
Chris@10
|
723 T26 = T24 - T25;
|
Chris@10
|
724 T2H = T1w - T1B;
|
Chris@10
|
725 }
|
Chris@10
|
726 T1r = T1f - T1q;
|
Chris@10
|
727 T1O = T1C - T1N;
|
Chris@10
|
728 T1P = T1r + T1O;
|
Chris@10
|
729 T3p = T3n + T3o;
|
Chris@10
|
730 T3s = T3q + T3r;
|
Chris@10
|
731 T3K = T3p + T3s;
|
Chris@10
|
732 T3A = T3n - T3o;
|
Chris@10
|
733 T3B = T3r - T3q;
|
Chris@10
|
734 T3Z = T3B - T3A;
|
Chris@10
|
735 T1V = T1f + T1q;
|
Chris@10
|
736 T1W = T1C + T1N;
|
Chris@10
|
737 T1X = T1V + T1W;
|
Chris@10
|
738 T23 = T21 + T22;
|
Chris@10
|
739 T28 = T26 + T27;
|
Chris@10
|
740 T4q = T23 + T28;
|
Chris@10
|
741 T2W = T21 - T22;
|
Chris@10
|
742 T2X = T26 - T27;
|
Chris@10
|
743 T4f = T2W + T2X;
|
Chris@10
|
744 T33 = T2C + T2F;
|
Chris@10
|
745 T34 = T2H + T2K;
|
Chris@10
|
746 T35 = T33 + T34;
|
Chris@10
|
747 T2G = T2C - T2F;
|
Chris@10
|
748 T2L = T2H - T2K;
|
Chris@10
|
749 T2M = T2G + T2L;
|
Chris@10
|
750 }
|
Chris@10
|
751 {
|
Chris@10
|
752 E Tu, T3g, T2c, T2r, T12, T3k, T2f, T2z, TF, T3h, T2d, T2u, TR, T3j, T2i;
|
Chris@10
|
753 E T2w;
|
Chris@10
|
754 {
|
Chris@10
|
755 E To, T2a, Tt, T2b;
|
Chris@10
|
756 {
|
Chris@10
|
757 E Tl, Tn, Tk, Tm;
|
Chris@10
|
758 Tl = Rp[WS(rs, 2)];
|
Chris@10
|
759 Tn = Rm[WS(rs, 2)];
|
Chris@10
|
760 Tk = W[6];
|
Chris@10
|
761 Tm = W[7];
|
Chris@10
|
762 To = FMA(Tk, Tl, Tm * Tn);
|
Chris@10
|
763 T2a = FNMS(Tm, Tl, Tk * Tn);
|
Chris@10
|
764 }
|
Chris@10
|
765 {
|
Chris@10
|
766 E Tq, Ts, Tp, Tr;
|
Chris@10
|
767 Tq = Rp[WS(rs, 7)];
|
Chris@10
|
768 Ts = Rm[WS(rs, 7)];
|
Chris@10
|
769 Tp = W[26];
|
Chris@10
|
770 Tr = W[27];
|
Chris@10
|
771 Tt = FMA(Tp, Tq, Tr * Ts);
|
Chris@10
|
772 T2b = FNMS(Tr, Tq, Tp * Ts);
|
Chris@10
|
773 }
|
Chris@10
|
774 Tu = To + Tt;
|
Chris@10
|
775 T3g = T2a + T2b;
|
Chris@10
|
776 T2c = T2a - T2b;
|
Chris@10
|
777 T2r = To - Tt;
|
Chris@10
|
778 }
|
Chris@10
|
779 {
|
Chris@10
|
780 E TW, T2x, T11, T2y;
|
Chris@10
|
781 {
|
Chris@10
|
782 E TT, TV, TS, TU;
|
Chris@10
|
783 TT = Ip[0];
|
Chris@10
|
784 TV = Im[0];
|
Chris@10
|
785 TS = W[0];
|
Chris@10
|
786 TU = W[1];
|
Chris@10
|
787 TW = FMA(TS, TT, TU * TV);
|
Chris@10
|
788 T2x = FNMS(TU, TT, TS * TV);
|
Chris@10
|
789 }
|
Chris@10
|
790 {
|
Chris@10
|
791 E TY, T10, TX, TZ;
|
Chris@10
|
792 TY = Ip[WS(rs, 5)];
|
Chris@10
|
793 T10 = Im[WS(rs, 5)];
|
Chris@10
|
794 TX = W[20];
|
Chris@10
|
795 TZ = W[21];
|
Chris@10
|
796 T11 = FMA(TX, TY, TZ * T10);
|
Chris@10
|
797 T2y = FNMS(TZ, TY, TX * T10);
|
Chris@10
|
798 }
|
Chris@10
|
799 T12 = TW + T11;
|
Chris@10
|
800 T3k = T2x + T2y;
|
Chris@10
|
801 T2f = T11 - TW;
|
Chris@10
|
802 T2z = T2x - T2y;
|
Chris@10
|
803 }
|
Chris@10
|
804 {
|
Chris@10
|
805 E Tz, T2s, TE, T2t;
|
Chris@10
|
806 {
|
Chris@10
|
807 E Tw, Ty, Tv, Tx;
|
Chris@10
|
808 Tw = Ip[WS(rs, 4)];
|
Chris@10
|
809 Ty = Im[WS(rs, 4)];
|
Chris@10
|
810 Tv = W[16];
|
Chris@10
|
811 Tx = W[17];
|
Chris@10
|
812 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@10
|
813 T2s = FNMS(Tx, Tw, Tv * Ty);
|
Chris@10
|
814 }
|
Chris@10
|
815 {
|
Chris@10
|
816 E TB, TD, TA, TC;
|
Chris@10
|
817 TB = Ip[WS(rs, 9)];
|
Chris@10
|
818 TD = Im[WS(rs, 9)];
|
Chris@10
|
819 TA = W[36];
|
Chris@10
|
820 TC = W[37];
|
Chris@10
|
821 TE = FMA(TA, TB, TC * TD);
|
Chris@10
|
822 T2t = FNMS(TC, TB, TA * TD);
|
Chris@10
|
823 }
|
Chris@10
|
824 TF = Tz + TE;
|
Chris@10
|
825 T3h = T2s + T2t;
|
Chris@10
|
826 T2d = Tz - TE;
|
Chris@10
|
827 T2u = T2s - T2t;
|
Chris@10
|
828 }
|
Chris@10
|
829 {
|
Chris@10
|
830 E TL, T2g, TQ, T2h;
|
Chris@10
|
831 {
|
Chris@10
|
832 E TI, TK, TH, TJ;
|
Chris@10
|
833 TI = Rp[WS(rs, 8)];
|
Chris@10
|
834 TK = Rm[WS(rs, 8)];
|
Chris@10
|
835 TH = W[30];
|
Chris@10
|
836 TJ = W[31];
|
Chris@10
|
837 TL = FMA(TH, TI, TJ * TK);
|
Chris@10
|
838 T2g = FNMS(TJ, TI, TH * TK);
|
Chris@10
|
839 }
|
Chris@10
|
840 {
|
Chris@10
|
841 E TN, TP, TM, TO;
|
Chris@10
|
842 TN = Rp[WS(rs, 3)];
|
Chris@10
|
843 TP = Rm[WS(rs, 3)];
|
Chris@10
|
844 TM = W[10];
|
Chris@10
|
845 TO = W[11];
|
Chris@10
|
846 TQ = FMA(TM, TN, TO * TP);
|
Chris@10
|
847 T2h = FNMS(TO, TN, TM * TP);
|
Chris@10
|
848 }
|
Chris@10
|
849 TR = TL + TQ;
|
Chris@10
|
850 T3j = T2g + T2h;
|
Chris@10
|
851 T2i = T2g - T2h;
|
Chris@10
|
852 T2w = TL - TQ;
|
Chris@10
|
853 }
|
Chris@10
|
854 TG = Tu - TF;
|
Chris@10
|
855 T13 = TR - T12;
|
Chris@10
|
856 T14 = TG + T13;
|
Chris@10
|
857 T3i = T3g + T3h;
|
Chris@10
|
858 T3l = T3j + T3k;
|
Chris@10
|
859 T3J = T3i + T3l;
|
Chris@10
|
860 T3D = T3g - T3h;
|
Chris@10
|
861 T3E = T3j - T3k;
|
Chris@10
|
862 T40 = T3D + T3E;
|
Chris@10
|
863 T1S = Tu + TF;
|
Chris@10
|
864 T1T = TR + T12;
|
Chris@10
|
865 T1U = T1S + T1T;
|
Chris@10
|
866 T2e = T2c + T2d;
|
Chris@10
|
867 T2j = T2f - T2i;
|
Chris@10
|
868 T4p = T2j - T2e;
|
Chris@10
|
869 T2T = T2c - T2d;
|
Chris@10
|
870 T2U = T2i + T2f;
|
Chris@10
|
871 T4e = T2T + T2U;
|
Chris@10
|
872 T30 = T2r + T2u;
|
Chris@10
|
873 T31 = T2w + T2z;
|
Chris@10
|
874 T32 = T30 + T31;
|
Chris@10
|
875 T2v = T2r - T2u;
|
Chris@10
|
876 T2A = T2w - T2z;
|
Chris@10
|
877 T2B = T2v + T2A;
|
Chris@10
|
878 }
|
Chris@10
|
879 {
|
Chris@10
|
880 E T3y, T1Q, T3x, T3G, T3I, T3C, T3F, T3H, T3z;
|
Chris@10
|
881 T3y = KP559016994 * (T14 - T1P);
|
Chris@10
|
882 T1Q = T14 + T1P;
|
Chris@10
|
883 T3x = FNMS(KP250000000, T1Q, Tj);
|
Chris@10
|
884 T3C = T3A + T3B;
|
Chris@10
|
885 T3F = T3D - T3E;
|
Chris@10
|
886 T3G = FNMS(KP587785252, T3F, KP951056516 * T3C);
|
Chris@10
|
887 T3I = FMA(KP951056516, T3F, KP587785252 * T3C);
|
Chris@10
|
888 Rm[WS(rs, 9)] = Tj + T1Q;
|
Chris@10
|
889 T3H = T3y + T3x;
|
Chris@10
|
890 Rm[WS(rs, 5)] = T3H - T3I;
|
Chris@10
|
891 Rp[WS(rs, 6)] = T3H + T3I;
|
Chris@10
|
892 T3z = T3x - T3y;
|
Chris@10
|
893 Rp[WS(rs, 2)] = T3z - T3G;
|
Chris@10
|
894 Rm[WS(rs, 1)] = T3z + T3G;
|
Chris@10
|
895 }
|
Chris@10
|
896 {
|
Chris@10
|
897 E T47, T41, T46, T45, T49, T43, T44, T4a, T48;
|
Chris@10
|
898 T47 = KP559016994 * (T40 + T3Z);
|
Chris@10
|
899 T41 = T3Z - T40;
|
Chris@10
|
900 T46 = FMA(KP250000000, T41, T42);
|
Chris@10
|
901 T43 = T13 - TG;
|
Chris@10
|
902 T44 = T1r - T1O;
|
Chris@10
|
903 T45 = FMA(KP587785252, T43, KP951056516 * T44);
|
Chris@10
|
904 T49 = FNMS(KP587785252, T44, KP951056516 * T43);
|
Chris@10
|
905 Im[WS(rs, 9)] = T41 - T42;
|
Chris@10
|
906 T4a = T47 + T46;
|
Chris@10
|
907 Im[WS(rs, 5)] = T49 - T4a;
|
Chris@10
|
908 Ip[WS(rs, 6)] = T49 + T4a;
|
Chris@10
|
909 T48 = T46 - T47;
|
Chris@10
|
910 Im[WS(rs, 1)] = T45 - T48;
|
Chris@10
|
911 Ip[WS(rs, 2)] = T45 + T48;
|
Chris@10
|
912 }
|
Chris@10
|
913 {
|
Chris@10
|
914 E T3d, T1Y, T3e, T3u, T3w, T3m, T3t, T3v, T3f;
|
Chris@10
|
915 T3d = KP559016994 * (T1U - T1X);
|
Chris@10
|
916 T1Y = T1U + T1X;
|
Chris@10
|
917 T3e = FNMS(KP250000000, T1Y, T1R);
|
Chris@10
|
918 T3m = T3i - T3l;
|
Chris@10
|
919 T3t = T3p - T3s;
|
Chris@10
|
920 T3u = FMA(KP951056516, T3m, KP587785252 * T3t);
|
Chris@10
|
921 T3w = FNMS(KP587785252, T3m, KP951056516 * T3t);
|
Chris@10
|
922 Rp[0] = T1R + T1Y;
|
Chris@10
|
923 T3v = T3e - T3d;
|
Chris@10
|
924 Rm[WS(rs, 7)] = T3v - T3w;
|
Chris@10
|
925 Rp[WS(rs, 8)] = T3v + T3w;
|
Chris@10
|
926 T3f = T3d + T3e;
|
Chris@10
|
927 Rp[WS(rs, 4)] = T3f - T3u;
|
Chris@10
|
928 Rm[WS(rs, 3)] = T3f + T3u;
|
Chris@10
|
929 }
|
Chris@10
|
930 {
|
Chris@10
|
931 E T3U, T3L, T3V, T3T, T3X, T3R, T3S, T3Y, T3W;
|
Chris@10
|
932 T3U = KP559016994 * (T3J - T3K);
|
Chris@10
|
933 T3L = T3J + T3K;
|
Chris@10
|
934 T3V = FNMS(KP250000000, T3L, T3Q);
|
Chris@10
|
935 T3R = T1S - T1T;
|
Chris@10
|
936 T3S = T1V - T1W;
|
Chris@10
|
937 T3T = FMA(KP951056516, T3R, KP587785252 * T3S);
|
Chris@10
|
938 T3X = FNMS(KP951056516, T3S, KP587785252 * T3R);
|
Chris@10
|
939 Ip[0] = T3L + T3Q;
|
Chris@10
|
940 T3Y = T3V - T3U;
|
Chris@10
|
941 Im[WS(rs, 7)] = T3X - T3Y;
|
Chris@10
|
942 Ip[WS(rs, 8)] = T3X + T3Y;
|
Chris@10
|
943 T3W = T3U + T3V;
|
Chris@10
|
944 Im[WS(rs, 3)] = T3T - T3W;
|
Chris@10
|
945 Ip[WS(rs, 4)] = T3T + T3W;
|
Chris@10
|
946 }
|
Chris@10
|
947 {
|
Chris@10
|
948 E T2P, T2N, T2O, T2l, T2R, T29, T2k, T2S, T2Q;
|
Chris@10
|
949 T2P = KP559016994 * (T2B - T2M);
|
Chris@10
|
950 T2N = T2B + T2M;
|
Chris@10
|
951 T2O = FNMS(KP250000000, T2N, T2q);
|
Chris@10
|
952 T29 = T23 - T28;
|
Chris@10
|
953 T2k = T2e + T2j;
|
Chris@10
|
954 T2l = FNMS(KP587785252, T2k, KP951056516 * T29);
|
Chris@10
|
955 T2R = FMA(KP951056516, T2k, KP587785252 * T29);
|
Chris@10
|
956 Rm[WS(rs, 4)] = T2q + T2N;
|
Chris@10
|
957 T2S = T2P + T2O;
|
Chris@10
|
958 Rm[WS(rs, 8)] = T2R + T2S;
|
Chris@10
|
959 Rm[0] = T2S - T2R;
|
Chris@10
|
960 T2Q = T2O - T2P;
|
Chris@10
|
961 Rp[WS(rs, 3)] = T2l + T2Q;
|
Chris@10
|
962 Rp[WS(rs, 7)] = T2Q - T2l;
|
Chris@10
|
963 }
|
Chris@10
|
964 {
|
Chris@10
|
965 E T4w, T4r, T4x, T4v, T4A, T4t, T4u, T4z, T4y;
|
Chris@10
|
966 T4w = KP559016994 * (T4p + T4q);
|
Chris@10
|
967 T4r = T4p - T4q;
|
Chris@10
|
968 T4x = FMA(KP250000000, T4r, T4s);
|
Chris@10
|
969 T4t = T2v - T2A;
|
Chris@10
|
970 T4u = T2G - T2L;
|
Chris@10
|
971 T4v = FMA(KP951056516, T4t, KP587785252 * T4u);
|
Chris@10
|
972 T4A = FNMS(KP587785252, T4t, KP951056516 * T4u);
|
Chris@10
|
973 Im[WS(rs, 4)] = T4r - T4s;
|
Chris@10
|
974 T4z = T4w + T4x;
|
Chris@10
|
975 Ip[WS(rs, 3)] = T4z - T4A;
|
Chris@10
|
976 Ip[WS(rs, 7)] = T4A + T4z;
|
Chris@10
|
977 T4y = T4w - T4x;
|
Chris@10
|
978 Im[WS(rs, 8)] = T4v + T4y;
|
Chris@10
|
979 Im[0] = T4y - T4v;
|
Chris@10
|
980 }
|
Chris@10
|
981 {
|
Chris@10
|
982 E T36, T38, T39, T2Z, T3b, T2V, T2Y, T3c, T3a;
|
Chris@10
|
983 T36 = KP559016994 * (T32 - T35);
|
Chris@10
|
984 T38 = T32 + T35;
|
Chris@10
|
985 T39 = FNMS(KP250000000, T38, T37);
|
Chris@10
|
986 T2V = T2T - T2U;
|
Chris@10
|
987 T2Y = T2W - T2X;
|
Chris@10
|
988 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
|
Chris@10
|
989 T3b = FNMS(KP587785252, T2V, KP951056516 * T2Y);
|
Chris@10
|
990 Rp[WS(rs, 5)] = T37 + T38;
|
Chris@10
|
991 T3c = T39 - T36;
|
Chris@10
|
992 Rm[WS(rs, 6)] = T3b + T3c;
|
Chris@10
|
993 Rm[WS(rs, 2)] = T3c - T3b;
|
Chris@10
|
994 T3a = T36 + T39;
|
Chris@10
|
995 Rp[WS(rs, 1)] = T2Z + T3a;
|
Chris@10
|
996 Rp[WS(rs, 9)] = T3a - T2Z;
|
Chris@10
|
997 }
|
Chris@10
|
998 {
|
Chris@10
|
999 E T4g, T4k, T4l, T4d, T4o, T4b, T4c, T4n, T4m;
|
Chris@10
|
1000 T4g = KP559016994 * (T4e - T4f);
|
Chris@10
|
1001 T4k = T4e + T4f;
|
Chris@10
|
1002 T4l = FNMS(KP250000000, T4k, T4j);
|
Chris@10
|
1003 T4b = T33 - T34;
|
Chris@10
|
1004 T4c = T30 - T31;
|
Chris@10
|
1005 T4d = FNMS(KP587785252, T4c, KP951056516 * T4b);
|
Chris@10
|
1006 T4o = FMA(KP951056516, T4c, KP587785252 * T4b);
|
Chris@10
|
1007 Ip[WS(rs, 5)] = T4k + T4j;
|
Chris@10
|
1008 T4n = T4g + T4l;
|
Chris@10
|
1009 Ip[WS(rs, 1)] = T4n - T4o;
|
Chris@10
|
1010 Ip[WS(rs, 9)] = T4o + T4n;
|
Chris@10
|
1011 T4m = T4g - T4l;
|
Chris@10
|
1012 Im[WS(rs, 6)] = T4d + T4m;
|
Chris@10
|
1013 Im[WS(rs, 2)] = T4m - T4d;
|
Chris@10
|
1014 }
|
Chris@10
|
1015 }
|
Chris@10
|
1016 }
|
Chris@10
|
1017 }
|
Chris@10
|
1018
|
Chris@10
|
1019 static const tw_instr twinstr[] = {
|
Chris@10
|
1020 {TW_FULL, 1, 20},
|
Chris@10
|
1021 {TW_NEXT, 1, 0}
|
Chris@10
|
1022 };
|
Chris@10
|
1023
|
Chris@10
|
1024 static const hc2c_desc desc = { 20, "hc2cf_20", twinstr, &GENUS, {184, 62, 62, 0} };
|
Chris@10
|
1025
|
Chris@10
|
1026 void X(codelet_hc2cf_20) (planner *p) {
|
Chris@10
|
1027 X(khc2c_register) (p, hc2cf_20, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
1028 }
|
Chris@10
|
1029 #endif /* HAVE_FMA */
|