Mercurial > hg > sv-dependency-builds

comparison src/fftw-3.3.3/dft/scalar/codelets/n1_14.c @ 10:37bf6b4a2645

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW3
author Chris Cannam
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
comparison
equal deleted inserted replaced
9:c0fb53affa76 10:37bf6b4a2645
1 /*
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Sun Nov 25 07:35:43 EST 2012 */
23
24 #include "codelet-dft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_notw.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 14 -name n1_14 -include n.h */
29
30 /*
31 * This function contains 148 FP additions, 84 FP multiplications,
32 * (or, 64 additions, 0 multiplications, 84 fused multiply/add),
33 * 80 stack variables, 6 constants, and 56 memory accesses
34 */
35 #include "n.h"
36
37 static void n1_14(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
40 DK(KP801937735, +0.801937735804838252472204639014890102331838324);
41 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
42 DK(KP554958132, +0.554958132087371191422194871006410481067288862);
43 DK(KP692021471, +0.692021471630095869627814897002069140197260599);
44 DK(KP356895867, +0.356895867892209443894399510021300583399127187);
45 {
46 INT i;
47 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(56, is), MAKE_VOLATILE_STRIDE(56, os)) {
48 E Tp, T1L, T24, T1W, T1X, T28, T2a, T1Y, T29, T2b;
49 {
50 E T3, T1x, T1b, To, T1i, T1M, Ts, Ta, T1k, Tv, Th, T1j, T1K, Ty, TZ;
51 E T14, Tz, T1Z, T27, T2c, T1d, TI, T23, T1G, T1D, TW, T1e, T22, T1A, TP;
52 E T1c, T1n, T1s, T1f, T1P;
53 {
54 E T1, T2, T19, T1a;
55 T1 = ri[0];
56 T2 = ri[WS(is, 7)];
57 T19 = ii[0];
58 T1a = ii[WS(is, 7)];
59 {
60 E Tq, T6, Tr, T9, Te, Tx, Tn, Tw, Tk, Tf, Tb, Tc;
61 {
62 E Tl, Tm, Ti, Tj;
63 {
64 E T4, T5, T7, T8;
65 T4 = ri[WS(is, 2)];
66 Tp = T1 + T2;
67 T3 = T1 - T2;
68 T1x = T19 + T1a;
69 T1b = T19 - T1a;
70 T5 = ri[WS(is, 9)];
71 T7 = ri[WS(is, 12)];
72 T8 = ri[WS(is, 5)];
73 Tl = ri[WS(is, 8)];
74 Tq = T4 + T5;
75 T6 = T4 - T5;
76 Tr = T7 + T8;
77 T9 = T7 - T8;
78 Tm = ri[WS(is, 1)];
79 }
80 Ti = ri[WS(is, 6)];
81 Tj = ri[WS(is, 13)];
82 Te = ri[WS(is, 10)];
83 Tx = Tl + Tm;
84 Tn = Tl - Tm;
85 Tw = Ti + Tj;
86 Tk = Ti - Tj;
87 Tf = ri[WS(is, 3)];
88 Tb = ri[WS(is, 4)];
89 Tc = ri[WS(is, 11)];
90 }
91 {
92 E Tu, Tg, Tt, Td;
93 To = Tk + Tn;
94 T1i = Tn - Tk;
95 Tu = Te + Tf;
96 Tg = Te - Tf;
97 Tt = Tb + Tc;
98 Td = Tb - Tc;
99 T1M = Tr - Tq;
100 Ts = Tq + Tr;
101 Ta = T6 + T9;
102 T1k = T9 - T6;
103 T1L = Tt - Tu;
104 Tv = Tt + Tu;
105 Th = Td + Tg;
106 T1j = Tg - Td;
107 T1K = Tw - Tx;
108 Ty = Tw + Tx;
109 TZ = FNMS(KP356895867, Ta, To);
110 T14 = FNMS(KP356895867, To, Th);
111 Tz = FNMS(KP356895867, Th, Ta);
112 T1Z = FNMS(KP356895867, Ts, Ty);
113 }
114 }
115 {
116 E T1B, TE, T1C, TH, T1F, TV, TJ, T1E, TS, T1z, TO, TK, T1y, TL;
117 {
118 E TF, TG, TT, TU, TC, TD;
119 TC = ii[WS(is, 4)];
120 TD = ii[WS(is, 11)];
121 T27 = FNMS(KP356895867, Tv, Ts);
122 T2c = FNMS(KP356895867, Ty, Tv);
123 TF = ii[WS(is, 10)];
124 T1B = TC + TD;
125 TE = TC - TD;
126 TG = ii[WS(is, 3)];
127 TT = ii[WS(is, 8)];
128 TU = ii[WS(is, 1)];
129 {
130 E TQ, TR, TM, TN;
131 TQ = ii[WS(is, 6)];
132 T1C = TF + TG;
133 TH = TF - TG;
134 T1F = TT + TU;
135 TV = TT - TU;
136 TR = ii[WS(is, 13)];
137 TM = ii[WS(is, 12)];
138 TN = ii[WS(is, 5)];
139 TJ = ii[WS(is, 2)];
140 T1E = TQ + TR;
141 TS = TQ - TR;
142 T1z = TM + TN;
143 TO = TM - TN;
144 TK = ii[WS(is, 9)];
145 }
146 }
147 T1d = TE + TH;
148 TI = TE - TH;
149 T23 = T1F - T1E;
150 T1G = T1E + T1F;
151 T1D = T1B + T1C;
152 T24 = T1C - T1B;
153 T1y = TJ + TK;
154 TL = TJ - TK;
155 TW = TS - TV;
156 T1e = TS + TV;
157 T22 = T1y - T1z;
158 T1A = T1y + T1z;
159 TP = TL - TO;
160 T1c = TL + TO;
161 T1n = FNMS(KP356895867, T1c, T1e);
162 T1s = FNMS(KP356895867, T1d, T1c);
163 T1f = FNMS(KP356895867, T1e, T1d);
164 T1P = FNMS(KP356895867, T1A, T1G);
165 }
166 }
167 {
168 E T1U, T1H, T11, T12, T1o, T1q;
169 ro[WS(os, 7)] = T3 + Ta + Th + To;
170 io[WS(os, 7)] = T1b + T1c + T1d + T1e;
171 T1U = FNMS(KP356895867, T1D, T1A);
172 T1H = FNMS(KP356895867, T1G, T1D);
173 ro[0] = Tp + Ts + Tv + Ty;
174 io[0] = T1x + T1A + T1D + T1G;
175 {
176 E TB, TY, T1u, T1w, T10;
177 {
178 E TA, TX, T1t, T1v;
179 TA = FNMS(KP692021471, Tz, To);
180 TX = FMA(KP554958132, TW, TP);
181 T1t = FNMS(KP692021471, T1s, T1e);
182 T1v = FMA(KP554958132, T1i, T1k);
183 TB = FNMS(KP900968867, TA, T3);
184 TY = FMA(KP801937735, TX, TI);
185 T1u = FNMS(KP900968867, T1t, T1b);
186 T1w = FMA(KP801937735, T1v, T1j);
187 }
188 T10 = FNMS(KP692021471, TZ, Th);
189 ro[WS(os, 1)] = FMA(KP974927912, TY, TB);
190 ro[WS(os, 13)] = FNMS(KP974927912, TY, TB);
191 io[WS(os, 13)] = FNMS(KP974927912, T1w, T1u);
192 io[WS(os, 1)] = FMA(KP974927912, T1w, T1u);
193 T11 = FNMS(KP900968867, T10, T3);
194 T12 = FMA(KP554958132, TI, TW);
195 T1o = FNMS(KP692021471, T1n, T1d);
196 T1q = FMA(KP554958132, T1j, T1i);
197 }
198 {
199 E T1J, T1N, T2d, T2f;
200 {
201 E T16, T17, T1g, T1l;
202 {
203 E T13, T1p, T1r, T15;
204 T15 = FNMS(KP692021471, T14, Ta);
205 T13 = FNMS(KP801937735, T12, TP);
206 T1p = FNMS(KP900968867, T1o, T1b);
207 T1r = FNMS(KP801937735, T1q, T1k);
208 T16 = FNMS(KP900968867, T15, T3);
209 ro[WS(os, 9)] = FMA(KP974927912, T13, T11);
210 ro[WS(os, 5)] = FNMS(KP974927912, T13, T11);
211 io[WS(os, 9)] = FMA(KP974927912, T1r, T1p);
212 io[WS(os, 5)] = FNMS(KP974927912, T1r, T1p);
213 T17 = FNMS(KP554958132, TP, TI);
214 }
215 T1g = FNMS(KP692021471, T1f, T1c);
216 T1l = FNMS(KP554958132, T1k, T1j);
217 {
218 E T18, T1h, T1m, T1I;
219 T1I = FNMS(KP692021471, T1H, T1A);
220 T18 = FNMS(KP801937735, T17, TW);
221 T1h = FNMS(KP900968867, T1g, T1b);
222 T1m = FNMS(KP801937735, T1l, T1i);
223 T1J = FNMS(KP900968867, T1I, T1x);
224 ro[WS(os, 3)] = FMA(KP974927912, T18, T16);
225 ro[WS(os, 11)] = FNMS(KP974927912, T18, T16);
226 io[WS(os, 11)] = FNMS(KP974927912, T1m, T1h);
227 io[WS(os, 3)] = FMA(KP974927912, T1m, T1h);
228 T1N = FMA(KP554958132, T1M, T1L);
229 }
230 T2d = FNMS(KP692021471, T2c, Ts);
231 T2f = FMA(KP554958132, T22, T24);
232 }
233 {
234 E T1R, T1S, T20, T25;
235 {
236 E T1O, T2e, T2g, T1Q;
237 T1Q = FNMS(KP692021471, T1P, T1D);
238 T1O = FNMS(KP801937735, T1N, T1K);
239 T2e = FNMS(KP900968867, T2d, Tp);
240 T2g = FNMS(KP801937735, T2f, T23);
241 T1R = FNMS(KP900968867, T1Q, T1x);
242 io[WS(os, 10)] = FNMS(KP974927912, T1O, T1J);
243 io[WS(os, 4)] = FMA(KP974927912, T1O, T1J);
244 ro[WS(os, 4)] = FMA(KP974927912, T2g, T2e);
245 ro[WS(os, 10)] = FNMS(KP974927912, T2g, T2e);
246 T1S = FMA(KP554958132, T1L, T1K);
247 }
248 T20 = FNMS(KP692021471, T1Z, Tv);
249 T25 = FMA(KP554958132, T24, T23);
250 {
251 E T1T, T21, T26, T1V;
252 T1V = FNMS(KP692021471, T1U, T1G);
253 T1T = FMA(KP801937735, T1S, T1M);
254 T21 = FNMS(KP900968867, T20, Tp);
255 T26 = FMA(KP801937735, T25, T22);
256 T1W = FNMS(KP900968867, T1V, T1x);
257 io[WS(os, 12)] = FNMS(KP974927912, T1T, T1R);
258 io[WS(os, 2)] = FMA(KP974927912, T1T, T1R);
259 ro[WS(os, 2)] = FMA(KP974927912, T26, T21);
260 ro[WS(os, 12)] = FNMS(KP974927912, T26, T21);
261 T1X = FNMS(KP554958132, T1K, T1M);
262 }
263 T28 = FNMS(KP692021471, T27, Ty);
264 T2a = FNMS(KP554958132, T23, T22);
265 }
266 }
267 }
268 }
269 T1Y = FNMS(KP801937735, T1X, T1L);
270 T29 = FNMS(KP900968867, T28, Tp);
271 T2b = FNMS(KP801937735, T2a, T24);
272 io[WS(os, 8)] = FNMS(KP974927912, T1Y, T1W);
273 io[WS(os, 6)] = FMA(KP974927912, T1Y, T1W);
274 ro[WS(os, 6)] = FMA(KP974927912, T2b, T29);
275 ro[WS(os, 8)] = FNMS(KP974927912, T2b, T29);
276 }
277 }
278 }
279
280 static const kdft_desc desc = { 14, "n1_14", {64, 0, 84, 0}, &GENUS, 0, 0, 0, 0 };
281
282 void X(codelet_n1_14) (planner *p) {
283 X(kdft_register) (p, n1_14, &desc);
284 }
285
286 #else /* HAVE_FMA */
287
288 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 14 -name n1_14 -include n.h */
289
290 /*
291 * This function contains 148 FP additions, 72 FP multiplications,
292 * (or, 100 additions, 24 multiplications, 48 fused multiply/add),
293 * 43 stack variables, 6 constants, and 56 memory accesses
294 */
295 #include "n.h"
296
297 static void n1_14(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
298 {
299 DK(KP222520933, +0.222520933956314404288902564496794759466355569);
300 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
301 DK(KP623489801, +0.623489801858733530525004884004239810632274731);
302 DK(KP433883739, +0.433883739117558120475768332848358754609990728);
303 DK(KP781831482, +0.781831482468029808708444526674057750232334519);
304 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
305 {
306 INT i;
307 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(56, is), MAKE_VOLATILE_STRIDE(56, os)) {
308 E T3, Tp, T16, T1f, Ta, T1q, Ts, T10, TG, T1z, T19, T1i, Th, T1s, Tv;
309 E T12, TU, T1B, T17, T1o, To, T1r, Ty, T11, TN, T1A, T18, T1l;
310 {
311 E T1, T2, T14, T15;
312 T1 = ri[0];
313 T2 = ri[WS(is, 7)];
314 T3 = T1 - T2;
315 Tp = T1 + T2;
316 T14 = ii[0];
317 T15 = ii[WS(is, 7)];
318 T16 = T14 - T15;
319 T1f = T14 + T15;
320 }
321 {
322 E T6, Tq, T9, Tr;
323 {
324 E T4, T5, T7, T8;
325 T4 = ri[WS(is, 2)];
326 T5 = ri[WS(is, 9)];
327 T6 = T4 - T5;
328 Tq = T4 + T5;
329 T7 = ri[WS(is, 12)];
330 T8 = ri[WS(is, 5)];
331 T9 = T7 - T8;
332 Tr = T7 + T8;
333 }
334 Ta = T6 + T9;
335 T1q = Tr - Tq;
336 Ts = Tq + Tr;
337 T10 = T9 - T6;
338 }
339 {
340 E TC, T1g, TF, T1h;
341 {
342 E TA, TB, TD, TE;
343 TA = ii[WS(is, 2)];
344 TB = ii[WS(is, 9)];
345 TC = TA - TB;
346 T1g = TA + TB;
347 TD = ii[WS(is, 12)];
348 TE = ii[WS(is, 5)];
349 TF = TD - TE;
350 T1h = TD + TE;
351 }
352 TG = TC - TF;
353 T1z = T1g - T1h;
354 T19 = TC + TF;
355 T1i = T1g + T1h;
356 }
357 {
358 E Td, Tt, Tg, Tu;
359 {
360 E Tb, Tc, Te, Tf;
361 Tb = ri[WS(is, 4)];
362 Tc = ri[WS(is, 11)];
363 Td = Tb - Tc;
364 Tt = Tb + Tc;
365 Te = ri[WS(is, 10)];
366 Tf = ri[WS(is, 3)];
367 Tg = Te - Tf;
368 Tu = Te + Tf;
369 }
370 Th = Td + Tg;
371 T1s = Tt - Tu;
372 Tv = Tt + Tu;
373 T12 = Tg - Td;
374 }
375 {
376 E TQ, T1m, TT, T1n;
377 {
378 E TO, TP, TR, TS;
379 TO = ii[WS(is, 4)];
380 TP = ii[WS(is, 11)];
381 TQ = TO - TP;
382 T1m = TO + TP;
383 TR = ii[WS(is, 10)];
384 TS = ii[WS(is, 3)];
385 TT = TR - TS;
386 T1n = TR + TS;
387 }
388 TU = TQ - TT;
389 T1B = T1n - T1m;
390 T17 = TQ + TT;
391 T1o = T1m + T1n;
392 }
393 {
394 E Tk, Tw, Tn, Tx;
395 {
396 E Ti, Tj, Tl, Tm;
397 Ti = ri[WS(is, 6)];
398 Tj = ri[WS(is, 13)];
399 Tk = Ti - Tj;
400 Tw = Ti + Tj;
401 Tl = ri[WS(is, 8)];
402 Tm = ri[WS(is, 1)];
403 Tn = Tl - Tm;
404 Tx = Tl + Tm;
405 }
406 To = Tk + Tn;
407 T1r = Tw - Tx;
408 Ty = Tw + Tx;
409 T11 = Tn - Tk;
410 }
411 {
412 E TJ, T1j, TM, T1k;
413 {
414 E TH, TI, TK, TL;
415 TH = ii[WS(is, 6)];
416 TI = ii[WS(is, 13)];
417 TJ = TH - TI;
418 T1j = TH + TI;
419 TK = ii[WS(is, 8)];
420 TL = ii[WS(is, 1)];
421 TM = TK - TL;
422 T1k = TK + TL;
423 }
424 TN = TJ - TM;
425 T1A = T1k - T1j;
426 T18 = TJ + TM;
427 T1l = T1j + T1k;
428 }
429 ro[WS(os, 7)] = T3 + Ta + Th + To;
430 io[WS(os, 7)] = T16 + T19 + T17 + T18;
431 ro[0] = Tp + Ts + Tv + Ty;
432 io[0] = T1f + T1i + T1o + T1l;
433 {
434 E TV, Tz, T1e, T1d;
435 TV = FNMS(KP781831482, TN, KP974927912 * TG) - (KP433883739 * TU);
436 Tz = FMA(KP623489801, To, T3) + FNMA(KP900968867, Th, KP222520933 * Ta);
437 ro[WS(os, 5)] = Tz - TV;
438 ro[WS(os, 9)] = Tz + TV;
439 T1e = FNMS(KP781831482, T11, KP974927912 * T10) - (KP433883739 * T12);
440 T1d = FMA(KP623489801, T18, T16) + FNMA(KP900968867, T17, KP222520933 * T19);
441 io[WS(os, 5)] = T1d - T1e;
442 io[WS(os, 9)] = T1e + T1d;
443 }
444 {
445 E TX, TW, T1b, T1c;
446 TX = FMA(KP781831482, TG, KP974927912 * TU) + (KP433883739 * TN);
447 TW = FMA(KP623489801, Ta, T3) + FNMA(KP900968867, To, KP222520933 * Th);
448 ro[WS(os, 13)] = TW - TX;
449 ro[WS(os, 1)] = TW + TX;
450 T1b = FMA(KP781831482, T10, KP974927912 * T12) + (KP433883739 * T11);
451 T1c = FMA(KP623489801, T19, T16) + FNMA(KP900968867, T18, KP222520933 * T17);
452 io[WS(os, 1)] = T1b + T1c;
453 io[WS(os, 13)] = T1c - T1b;
454 }
455 {
456 E TZ, TY, T13, T1a;
457 TZ = FMA(KP433883739, TG, KP974927912 * TN) - (KP781831482 * TU);
458 TY = FMA(KP623489801, Th, T3) + FNMA(KP222520933, To, KP900968867 * Ta);
459 ro[WS(os, 11)] = TY - TZ;
460 ro[WS(os, 3)] = TY + TZ;
461 T13 = FMA(KP433883739, T10, KP974927912 * T11) - (KP781831482 * T12);
462 T1a = FMA(KP623489801, T17, T16) + FNMA(KP222520933, T18, KP900968867 * T19);
463 io[WS(os, 3)] = T13 + T1a;
464 io[WS(os, 11)] = T1a - T13;
465 }
466 {
467 E T1t, T1p, T1C, T1y;
468 T1t = FNMS(KP433883739, T1r, KP781831482 * T1q) - (KP974927912 * T1s);
469 T1p = FMA(KP623489801, T1i, T1f) + FNMA(KP900968867, T1l, KP222520933 * T1o);
470 io[WS(os, 6)] = T1p - T1t;
471 io[WS(os, 8)] = T1t + T1p;
472 T1C = FNMS(KP433883739, T1A, KP781831482 * T1z) - (KP974927912 * T1B);
473 T1y = FMA(KP623489801, Ts, Tp) + FNMA(KP900968867, Ty, KP222520933 * Tv);
474 ro[WS(os, 6)] = T1y - T1C;
475 ro[WS(os, 8)] = T1y + T1C;
476 }
477 {
478 E T1v, T1u, T1E, T1D;
479 T1v = FMA(KP433883739, T1q, KP781831482 * T1s) - (KP974927912 * T1r);
480 T1u = FMA(KP623489801, T1o, T1f) + FNMA(KP222520933, T1l, KP900968867 * T1i);
481 io[WS(os, 4)] = T1u - T1v;
482 io[WS(os, 10)] = T1v + T1u;
483 T1E = FMA(KP433883739, T1z, KP781831482 * T1B) - (KP974927912 * T1A);
484 T1D = FMA(KP623489801, Tv, Tp) + FNMA(KP222520933, Ty, KP900968867 * Ts);
485 ro[WS(os, 4)] = T1D - T1E;
486 ro[WS(os, 10)] = T1D + T1E;
487 }
488 {
489 E T1w, T1x, T1G, T1F;
490 T1w = FMA(KP974927912, T1q, KP433883739 * T1s) + (KP781831482 * T1r);
491 T1x = FMA(KP623489801, T1l, T1f) + FNMA(KP900968867, T1o, KP222520933 * T1i);
492 io[WS(os, 2)] = T1w + T1x;
493 io[WS(os, 12)] = T1x - T1w;
494 T1G = FMA(KP974927912, T1z, KP433883739 * T1B) + (KP781831482 * T1A);
495 T1F = FMA(KP623489801, Ty, Tp) + FNMA(KP900968867, Tv, KP222520933 * Ts);
496 ro[WS(os, 12)] = T1F - T1G;
497 ro[WS(os, 2)] = T1F + T1G;
498 }
499 }
500 }
501 }
502
503 static const kdft_desc desc = { 14, "n1_14", {100, 24, 48, 0}, &GENUS, 0, 0, 0, 0 };
504
505 void X(codelet_n1_14) (planner *p) {
506 X(kdft_register) (p, n1_14, &desc);
507 }
508
509 #endif /* HAVE_FMA */