Mercurial > hg > sv-dependency-builds

comparison src/fftw-3.3.3/dft/scalar/codelets/t1_7.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 07:35:48 EST 2012 */
23
24 #include "codelet-dft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 7 -name t1_7 -include t.h */
29
30 /*
31 * This function contains 72 FP additions, 66 FP multiplications,
32 * (or, 18 additions, 12 multiplications, 54 fused multiply/add),
33 * 66 stack variables, 6 constants, and 28 memory accesses
34 */
35 #include "t.h"
36
37 static void t1_7(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
38 {
39 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
40 DK(KP801937735, +0.801937735804838252472204639014890102331838324);
41 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
42 DK(KP692021471, +0.692021471630095869627814897002069140197260599);
43 DK(KP554958132, +0.554958132087371191422194871006410481067288862);
44 DK(KP356895867, +0.356895867892209443894399510021300583399127187);
45 {
46 INT m;
47 for (m = mb, W = W + (mb * 12); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 12, MAKE_VOLATILE_STRIDE(14, rs)) {
48 E T1c, T19, T1i, T18, T16, T1q, T1t, T1r, T1u, T1s;
49 {
50 E T1, TR, T1h, Te, Tt, Tw, T1a, TM, T1g, Tr, Tu, TS, Tz, TC, Ty;
51 E Tv, TB;
52 T1 = ri[0];
53 T1c = ii[0];
54 {
55 E T9, Tc, TP, Ta, Tb, TO, T7;
56 {
57 E T3, T6, T8, TN, T4, T2, T5;
58 T3 = ri[WS(rs, 1)];
59 T6 = ii[WS(rs, 1)];
60 T2 = W[0];
61 T9 = ri[WS(rs, 6)];
62 Tc = ii[WS(rs, 6)];
63 T8 = W[10];
64 TN = T2 * T6;
65 T4 = T2 * T3;
66 T5 = W[1];
67 TP = T8 * Tc;
68 Ta = T8 * T9;
69 Tb = W[11];
70 TO = FNMS(T5, T3, TN);
71 T7 = FMA(T5, T6, T4);
72 }
73 {
74 E Tg, Tj, Th, TI, Tm, Tp, Tl, Ti, To, TQ, Td, Tf;
75 Tg = ri[WS(rs, 2)];
76 TQ = FNMS(Tb, T9, TP);
77 Td = FMA(Tb, Tc, Ta);
78 Tj = ii[WS(rs, 2)];
79 Tf = W[2];
80 T19 = TO + TQ;
81 TR = TO - TQ;
82 T1h = Td - T7;
83 Te = T7 + Td;
84 Th = Tf * Tg;
85 TI = Tf * Tj;
86 Tm = ri[WS(rs, 5)];
87 Tp = ii[WS(rs, 5)];
88 Tl = W[8];
89 Ti = W[3];
90 To = W[9];
91 {
92 E TJ, Tk, TL, Tq, TK, Tn, Ts;
93 Tt = ri[WS(rs, 3)];
94 TK = Tl * Tp;
95 Tn = Tl * Tm;
96 TJ = FNMS(Ti, Tg, TI);
97 Tk = FMA(Ti, Tj, Th);
98 TL = FNMS(To, Tm, TK);
99 Tq = FMA(To, Tp, Tn);
100 Tw = ii[WS(rs, 3)];
101 Ts = W[4];
102 T1a = TJ + TL;
103 TM = TJ - TL;
104 T1g = Tq - Tk;
105 Tr = Tk + Tq;
106 Tu = Ts * Tt;
107 TS = Ts * Tw;
108 }
109 Tz = ri[WS(rs, 4)];
110 TC = ii[WS(rs, 4)];
111 Ty = W[6];
112 Tv = W[5];
113 TB = W[7];
114 }
115 }
116 {
117 E TF, TT, Tx, TV, TD, T1d, TU, TA;
118 TF = FNMS(KP356895867, Tr, Te);
119 TU = Ty * TC;
120 TA = Ty * Tz;
121 TT = FNMS(Tv, Tt, TS);
122 Tx = FMA(Tv, Tw, Tu);
123 TV = FNMS(TB, Tz, TU);
124 TD = FMA(TB, TC, TA);
125 T1d = FNMS(KP356895867, T1a, T19);
126 {
127 E T1b, T15, T17, TW;
128 T17 = FNMS(KP554958132, TR, TM);
129 T1b = TT + TV;
130 TW = TT - TV;
131 {
132 E TE, T1l, T1e, T12;
133 T1i = TD - Tx;
134 TE = Tx + TD;
135 T1l = FNMS(KP356895867, T19, T1b);
136 T1e = FNMS(KP692021471, T1d, T1b);
137 ii[0] = T19 + T1a + T1b + T1c;
138 T12 = FMA(KP554958132, TM, TW);
139 {
140 E TX, T1o, T1j, T14;
141 TX = FMA(KP554958132, TW, TR);
142 T1o = FMA(KP554958132, T1g, T1i);
143 T1j = FMA(KP554958132, T1i, T1h);
144 T14 = FNMS(KP356895867, TE, Tr);
145 {
146 E TZ, TG, T1m, T1f;
147 TZ = FNMS(KP356895867, Te, TE);
148 TG = FNMS(KP692021471, TF, TE);
149 ri[0] = T1 + Te + Tr + TE;
150 T1m = FNMS(KP692021471, T1l, T1a);
151 T1f = FNMS(KP900968867, T1e, T1c);
152 {
153 E T13, TY, T1p, T1k;
154 T13 = FNMS(KP801937735, T12, TR);
155 TY = FMA(KP801937735, TX, TM);
156 T1p = FNMS(KP801937735, T1o, T1h);
157 T1k = FMA(KP801937735, T1j, T1g);
158 T15 = FNMS(KP692021471, T14, Te);
159 {
160 E T10, TH, T1n, T11;
161 T10 = FNMS(KP692021471, TZ, Tr);
162 TH = FNMS(KP900968867, TG, T1);
163 T1n = FNMS(KP900968867, T1m, T1c);
164 ii[WS(rs, 6)] = FNMS(KP974927912, T1k, T1f);
165 ii[WS(rs, 1)] = FMA(KP974927912, T1k, T1f);
166 T11 = FNMS(KP900968867, T10, T1);
167 ri[WS(rs, 1)] = FMA(KP974927912, TY, TH);
168 ri[WS(rs, 6)] = FNMS(KP974927912, TY, TH);
169 ii[WS(rs, 5)] = FNMS(KP974927912, T1p, T1n);
170 ii[WS(rs, 2)] = FMA(KP974927912, T1p, T1n);
171 ri[WS(rs, 2)] = FMA(KP974927912, T13, T11);
172 ri[WS(rs, 5)] = FNMS(KP974927912, T13, T11);
173 T18 = FNMS(KP801937735, T17, TW);
174 }
175 }
176 }
177 }
178 }
179 T16 = FNMS(KP900968867, T15, T1);
180 T1q = FNMS(KP356895867, T1b, T1a);
181 T1t = FNMS(KP554958132, T1h, T1g);
182 }
183 }
184 }
185 ri[WS(rs, 3)] = FMA(KP974927912, T18, T16);
186 ri[WS(rs, 4)] = FNMS(KP974927912, T18, T16);
187 T1r = FNMS(KP692021471, T1q, T19);
188 T1u = FNMS(KP801937735, T1t, T1i);
189 T1s = FNMS(KP900968867, T1r, T1c);
190 ii[WS(rs, 4)] = FNMS(KP974927912, T1u, T1s);
191 ii[WS(rs, 3)] = FMA(KP974927912, T1u, T1s);
192 }
193 }
194 }
195
196 static const tw_instr twinstr[] = {
197 {TW_FULL, 0, 7},
198 {TW_NEXT, 1, 0}
199 };
200
201 static const ct_desc desc = { 7, "t1_7", twinstr, &GENUS, {18, 12, 54, 0}, 0, 0, 0 };
202
203 void X(codelet_t1_7) (planner *p) {
204 X(kdft_dit_register) (p, t1_7, &desc);
205 }
206 #else /* HAVE_FMA */
207
208 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 7 -name t1_7 -include t.h */
209
210 /*
211 * This function contains 72 FP additions, 60 FP multiplications,
212 * (or, 36 additions, 24 multiplications, 36 fused multiply/add),
213 * 29 stack variables, 6 constants, and 28 memory accesses
214 */
215 #include "t.h"
216
217 static void t1_7(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
218 {
219 DK(KP222520933, +0.222520933956314404288902564496794759466355569);
220 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
221 DK(KP623489801, +0.623489801858733530525004884004239810632274731);
222 DK(KP433883739, +0.433883739117558120475768332848358754609990728);
223 DK(KP781831482, +0.781831482468029808708444526674057750232334519);
224 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
225 {
226 INT m;
227 for (m = mb, W = W + (mb * 12); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 12, MAKE_VOLATILE_STRIDE(14, rs)) {
228 E T1, TR, Tc, TS, TC, TO, Tn, TT, TI, TP, Ty, TU, TF, TQ;
229 T1 = ri[0];
230 TR = ii[0];
231 {
232 E T6, TA, Tb, TB;
233 {
234 E T3, T5, T2, T4;
235 T3 = ri[WS(rs, 1)];
236 T5 = ii[WS(rs, 1)];
237 T2 = W[0];
238 T4 = W[1];
239 T6 = FMA(T2, T3, T4 * T5);
240 TA = FNMS(T4, T3, T2 * T5);
241 }
242 {
243 E T8, Ta, T7, T9;
244 T8 = ri[WS(rs, 6)];
245 Ta = ii[WS(rs, 6)];
246 T7 = W[10];
247 T9 = W[11];
248 Tb = FMA(T7, T8, T9 * Ta);
249 TB = FNMS(T9, T8, T7 * Ta);
250 }
251 Tc = T6 + Tb;
252 TS = Tb - T6;
253 TC = TA - TB;
254 TO = TA + TB;
255 }
256 {
257 E Th, TG, Tm, TH;
258 {
259 E Te, Tg, Td, Tf;
260 Te = ri[WS(rs, 2)];
261 Tg = ii[WS(rs, 2)];
262 Td = W[2];
263 Tf = W[3];
264 Th = FMA(Td, Te, Tf * Tg);
265 TG = FNMS(Tf, Te, Td * Tg);
266 }
267 {
268 E Tj, Tl, Ti, Tk;
269 Tj = ri[WS(rs, 5)];
270 Tl = ii[WS(rs, 5)];
271 Ti = W[8];
272 Tk = W[9];
273 Tm = FMA(Ti, Tj, Tk * Tl);
274 TH = FNMS(Tk, Tj, Ti * Tl);
275 }
276 Tn = Th + Tm;
277 TT = Tm - Th;
278 TI = TG - TH;
279 TP = TG + TH;
280 }
281 {
282 E Ts, TD, Tx, TE;
283 {
284 E Tp, Tr, To, Tq;
285 Tp = ri[WS(rs, 3)];
286 Tr = ii[WS(rs, 3)];
287 To = W[4];
288 Tq = W[5];
289 Ts = FMA(To, Tp, Tq * Tr);
290 TD = FNMS(Tq, Tp, To * Tr);
291 }
292 {
293 E Tu, Tw, Tt, Tv;
294 Tu = ri[WS(rs, 4)];
295 Tw = ii[WS(rs, 4)];
296 Tt = W[6];
297 Tv = W[7];
298 Tx = FMA(Tt, Tu, Tv * Tw);
299 TE = FNMS(Tv, Tu, Tt * Tw);
300 }
301 Ty = Ts + Tx;
302 TU = Tx - Ts;
303 TF = TD - TE;
304 TQ = TD + TE;
305 }
306 ri[0] = T1 + Tc + Tn + Ty;
307 ii[0] = TO + TP + TQ + TR;
308 {
309 E TJ, Tz, TX, TY;
310 TJ = FNMS(KP781831482, TF, KP974927912 * TC) - (KP433883739 * TI);
311 Tz = FMA(KP623489801, Ty, T1) + FNMA(KP900968867, Tn, KP222520933 * Tc);
312 ri[WS(rs, 5)] = Tz - TJ;
313 ri[WS(rs, 2)] = Tz + TJ;
314 TX = FNMS(KP781831482, TU, KP974927912 * TS) - (KP433883739 * TT);
315 TY = FMA(KP623489801, TQ, TR) + FNMA(KP900968867, TP, KP222520933 * TO);
316 ii[WS(rs, 2)] = TX + TY;
317 ii[WS(rs, 5)] = TY - TX;
318 }
319 {
320 E TL, TK, TV, TW;
321 TL = FMA(KP781831482, TC, KP974927912 * TI) + (KP433883739 * TF);
322 TK = FMA(KP623489801, Tc, T1) + FNMA(KP900968867, Ty, KP222520933 * Tn);
323 ri[WS(rs, 6)] = TK - TL;
324 ri[WS(rs, 1)] = TK + TL;
325 TV = FMA(KP781831482, TS, KP974927912 * TT) + (KP433883739 * TU);
326 TW = FMA(KP623489801, TO, TR) + FNMA(KP900968867, TQ, KP222520933 * TP);
327 ii[WS(rs, 1)] = TV + TW;
328 ii[WS(rs, 6)] = TW - TV;
329 }
330 {
331 E TN, TM, TZ, T10;
332 TN = FMA(KP433883739, TC, KP974927912 * TF) - (KP781831482 * TI);
333 TM = FMA(KP623489801, Tn, T1) + FNMA(KP222520933, Ty, KP900968867 * Tc);
334 ri[WS(rs, 4)] = TM - TN;
335 ri[WS(rs, 3)] = TM + TN;
336 TZ = FMA(KP433883739, TS, KP974927912 * TU) - (KP781831482 * TT);
337 T10 = FMA(KP623489801, TP, TR) + FNMA(KP222520933, TQ, KP900968867 * TO);
338 ii[WS(rs, 3)] = TZ + T10;
339 ii[WS(rs, 4)] = T10 - TZ;
340 }
341 }
342 }
343 }
344
345 static const tw_instr twinstr[] = {
346 {TW_FULL, 0, 7},
347 {TW_NEXT, 1, 0}
348 };
349
350 static const ct_desc desc = { 7, "t1_7", twinstr, &GENUS, {36, 24, 36, 0}, 0, 0, 0 };
351
352 void X(codelet_t1_7) (planner *p) {
353 X(kdft_dit_register) (p, t1_7, &desc);
354 }
355 #endif /* HAVE_FMA */