Mercurial > hg > sv-dependency-builds

comparison src/fftw-3.3.8/dft/scalar/codelets/t1_7.c @ 167:bd3cc4d1df30

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
comparison
equal deleted inserted replaced
166:cbd6d7e562c7 167:bd3cc4d1df30
1 /*
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
3 * Copyright (c) 2003, 2007-14 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 Thu May 24 08:04:13 EDT 2018 */
23
24 #include "dft/codelet-dft.h"
25
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -compact -variables 4 -pipeline-latency 4 -n 7 -name t1_7 -include dft/scalar/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 * 37 stack variables, 6 constants, and 28 memory accesses
34 */
35 #include "dft/scalar/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(KP900968867, +0.900968867902419126236102319507445051165919162);
41 DK(KP801937735, +0.801937735804838252472204639014890102331838324);
42 DK(KP554958132, +0.554958132087371191422194871006410481067288862);
43 DK(KP692021471, +0.692021471630095869627814897002069140197260599);
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 T1, T1c, Te, T1h, TR, T19, Tr, T1g, TM, T1a, TE, T1i, TW, T1b;
49 T1 = ri[0];
50 T1c = ii[0];
51 {
52 E T3, T6, T4, TN, T9, Tc, Ta, TP, T2, T8;
53 T3 = ri[WS(rs, 1)];
54 T6 = ii[WS(rs, 1)];
55 T2 = W[0];
56 T4 = T2 * T3;
57 TN = T2 * T6;
58 T9 = ri[WS(rs, 6)];
59 Tc = ii[WS(rs, 6)];
60 T8 = W[10];
61 Ta = T8 * T9;
62 TP = T8 * Tc;
63 {
64 E T7, TO, Td, TQ, T5, Tb;
65 T5 = W[1];
66 T7 = FMA(T5, T6, T4);
67 TO = FNMS(T5, T3, TN);
68 Tb = W[11];
69 Td = FMA(Tb, Tc, Ta);
70 TQ = FNMS(Tb, T9, TP);
71 Te = T7 + Td;
72 T1h = Td - T7;
73 TR = TO - TQ;
74 T19 = TO + TQ;
75 }
76 }
77 {
78 E Tg, Tj, Th, TI, Tm, Tp, Tn, TK, Tf, Tl;
79 Tg = ri[WS(rs, 2)];
80 Tj = ii[WS(rs, 2)];
81 Tf = W[2];
82 Th = Tf * Tg;
83 TI = Tf * Tj;
84 Tm = ri[WS(rs, 5)];
85 Tp = ii[WS(rs, 5)];
86 Tl = W[8];
87 Tn = Tl * Tm;
88 TK = Tl * Tp;
89 {
90 E Tk, TJ, Tq, TL, Ti, To;
91 Ti = W[3];
92 Tk = FMA(Ti, Tj, Th);
93 TJ = FNMS(Ti, Tg, TI);
94 To = W[9];
95 Tq = FMA(To, Tp, Tn);
96 TL = FNMS(To, Tm, TK);
97 Tr = Tk + Tq;
98 T1g = Tq - Tk;
99 TM = TJ - TL;
100 T1a = TJ + TL;
101 }
102 }
103 {
104 E Tt, Tw, Tu, TS, Tz, TC, TA, TU, Ts, Ty;
105 Tt = ri[WS(rs, 3)];
106 Tw = ii[WS(rs, 3)];
107 Ts = W[4];
108 Tu = Ts * Tt;
109 TS = Ts * Tw;
110 Tz = ri[WS(rs, 4)];
111 TC = ii[WS(rs, 4)];
112 Ty = W[6];
113 TA = Ty * Tz;
114 TU = Ty * TC;
115 {
116 E Tx, TT, TD, TV, Tv, TB;
117 Tv = W[5];
118 Tx = FMA(Tv, Tw, Tu);
119 TT = FNMS(Tv, Tt, TS);
120 TB = W[7];
121 TD = FMA(TB, TC, TA);
122 TV = FNMS(TB, Tz, TU);
123 TE = Tx + TD;
124 T1i = TD - Tx;
125 TW = TT - TV;
126 T1b = TT + TV;
127 }
128 }
129 ri[0] = T1 + Te + Tr + TE;
130 ii[0] = T19 + T1a + T1b + T1c;
131 {
132 E TG, TY, TF, TX, TH;
133 TF = FNMS(KP356895867, Tr, Te);
134 TG = FNMS(KP692021471, TF, TE);
135 TX = FMA(KP554958132, TW, TR);
136 TY = FMA(KP801937735, TX, TM);
137 TH = FNMS(KP900968867, TG, T1);
138 ri[WS(rs, 6)] = FNMS(KP974927912, TY, TH);
139 ri[WS(rs, 1)] = FMA(KP974927912, TY, TH);
140 }
141 {
142 E T1e, T1k, T1d, T1j, T1f;
143 T1d = FNMS(KP356895867, T1a, T19);
144 T1e = FNMS(KP692021471, T1d, T1b);
145 T1j = FMA(KP554958132, T1i, T1h);
146 T1k = FMA(KP801937735, T1j, T1g);
147 T1f = FNMS(KP900968867, T1e, T1c);
148 ii[WS(rs, 1)] = FMA(KP974927912, T1k, T1f);
149 ii[WS(rs, 6)] = FNMS(KP974927912, T1k, T1f);
150 }
151 {
152 E T10, T13, TZ, T12, T11;
153 TZ = FNMS(KP356895867, Te, TE);
154 T10 = FNMS(KP692021471, TZ, Tr);
155 T12 = FMA(KP554958132, TM, TW);
156 T13 = FNMS(KP801937735, T12, TR);
157 T11 = FNMS(KP900968867, T10, T1);
158 ri[WS(rs, 5)] = FNMS(KP974927912, T13, T11);
159 ri[WS(rs, 2)] = FMA(KP974927912, T13, T11);
160 }
161 {
162 E T1m, T1p, T1l, T1o, T1n;
163 T1l = FNMS(KP356895867, T19, T1b);
164 T1m = FNMS(KP692021471, T1l, T1a);
165 T1o = FMA(KP554958132, T1g, T1i);
166 T1p = FNMS(KP801937735, T1o, T1h);
167 T1n = FNMS(KP900968867, T1m, T1c);
168 ii[WS(rs, 2)] = FMA(KP974927912, T1p, T1n);
169 ii[WS(rs, 5)] = FNMS(KP974927912, T1p, T1n);
170 }
171 {
172 E T15, T18, T14, T17, T16;
173 T14 = FNMS(KP356895867, TE, Tr);
174 T15 = FNMS(KP692021471, T14, Te);
175 T17 = FNMS(KP554958132, TR, TM);
176 T18 = FNMS(KP801937735, T17, TW);
177 T16 = FNMS(KP900968867, T15, T1);
178 ri[WS(rs, 4)] = FNMS(KP974927912, T18, T16);
179 ri[WS(rs, 3)] = FMA(KP974927912, T18, T16);
180 }
181 {
182 E T1r, T1u, T1q, T1t, T1s;
183 T1q = FNMS(KP356895867, T1b, T1a);
184 T1r = FNMS(KP692021471, T1q, T19);
185 T1t = FNMS(KP554958132, T1h, T1g);
186 T1u = FNMS(KP801937735, T1t, T1i);
187 T1s = FNMS(KP900968867, T1r, T1c);
188 ii[WS(rs, 3)] = FMA(KP974927912, T1u, T1s);
189 ii[WS(rs, 4)] = FNMS(KP974927912, T1u, T1s);
190 }
191 }
192 }
193 }
194
195 static const tw_instr twinstr[] = {
196 {TW_FULL, 0, 7},
197 {TW_NEXT, 1, 0}
198 };
199
200 static const ct_desc desc = { 7, "t1_7", twinstr, &GENUS, {18, 12, 54, 0}, 0, 0, 0 };
201
202 void X(codelet_t1_7) (planner *p) {
203 X(kdft_dit_register) (p, t1_7, &desc);
204 }
205 #else
206
207 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 7 -name t1_7 -include dft/scalar/t.h */
208
209 /*
210 * This function contains 72 FP additions, 60 FP multiplications,
211 * (or, 36 additions, 24 multiplications, 36 fused multiply/add),
212 * 29 stack variables, 6 constants, and 28 memory accesses
213 */
214 #include "dft/scalar/t.h"
215
216 static void t1_7(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
217 {
218 DK(KP222520933, +0.222520933956314404288902564496794759466355569);
219 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
220 DK(KP623489801, +0.623489801858733530525004884004239810632274731);
221 DK(KP433883739, +0.433883739117558120475768332848358754609990728);
222 DK(KP781831482, +0.781831482468029808708444526674057750232334519);
223 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
224 {
225 INT m;
226 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)) {
227 E T1, TR, Tc, TS, TC, TO, Tn, TT, TI, TP, Ty, TU, TF, TQ;
228 T1 = ri[0];
229 TR = ii[0];
230 {
231 E T6, TA, Tb, TB;
232 {
233 E T3, T5, T2, T4;
234 T3 = ri[WS(rs, 1)];
235 T5 = ii[WS(rs, 1)];
236 T2 = W[0];
237 T4 = W[1];
238 T6 = FMA(T2, T3, T4 * T5);
239 TA = FNMS(T4, T3, T2 * T5);
240 }
241 {
242 E T8, Ta, T7, T9;
243 T8 = ri[WS(rs, 6)];
244 Ta = ii[WS(rs, 6)];
245 T7 = W[10];
246 T9 = W[11];
247 Tb = FMA(T7, T8, T9 * Ta);
248 TB = FNMS(T9, T8, T7 * Ta);
249 }
250 Tc = T6 + Tb;
251 TS = Tb - T6;
252 TC = TA - TB;
253 TO = TA + TB;
254 }
255 {
256 E Th, TG, Tm, TH;
257 {
258 E Te, Tg, Td, Tf;
259 Te = ri[WS(rs, 2)];
260 Tg = ii[WS(rs, 2)];
261 Td = W[2];
262 Tf = W[3];
263 Th = FMA(Td, Te, Tf * Tg);
264 TG = FNMS(Tf, Te, Td * Tg);
265 }
266 {
267 E Tj, Tl, Ti, Tk;
268 Tj = ri[WS(rs, 5)];
269 Tl = ii[WS(rs, 5)];
270 Ti = W[8];
271 Tk = W[9];
272 Tm = FMA(Ti, Tj, Tk * Tl);
273 TH = FNMS(Tk, Tj, Ti * Tl);
274 }
275 Tn = Th + Tm;
276 TT = Tm - Th;
277 TI = TG - TH;
278 TP = TG + TH;
279 }
280 {
281 E Ts, TD, Tx, TE;
282 {
283 E Tp, Tr, To, Tq;
284 Tp = ri[WS(rs, 3)];
285 Tr = ii[WS(rs, 3)];
286 To = W[4];
287 Tq = W[5];
288 Ts = FMA(To, Tp, Tq * Tr);
289 TD = FNMS(Tq, Tp, To * Tr);
290 }
291 {
292 E Tu, Tw, Tt, Tv;
293 Tu = ri[WS(rs, 4)];
294 Tw = ii[WS(rs, 4)];
295 Tt = W[6];
296 Tv = W[7];
297 Tx = FMA(Tt, Tu, Tv * Tw);
298 TE = FNMS(Tv, Tu, Tt * Tw);
299 }
300 Ty = Ts + Tx;
301 TU = Tx - Ts;
302 TF = TD - TE;
303 TQ = TD + TE;
304 }
305 ri[0] = T1 + Tc + Tn + Ty;
306 ii[0] = TO + TP + TQ + TR;
307 {
308 E TJ, Tz, TX, TY;
309 TJ = FNMS(KP781831482, TF, KP974927912 * TC) - (KP433883739 * TI);
310 Tz = FMA(KP623489801, Ty, T1) + FNMA(KP900968867, Tn, KP222520933 * Tc);
311 ri[WS(rs, 5)] = Tz - TJ;
312 ri[WS(rs, 2)] = Tz + TJ;
313 TX = FNMS(KP781831482, TU, KP974927912 * TS) - (KP433883739 * TT);
314 TY = FMA(KP623489801, TQ, TR) + FNMA(KP900968867, TP, KP222520933 * TO);
315 ii[WS(rs, 2)] = TX + TY;
316 ii[WS(rs, 5)] = TY - TX;
317 }
318 {
319 E TL, TK, TV, TW;
320 TL = FMA(KP781831482, TC, KP974927912 * TI) + (KP433883739 * TF);
321 TK = FMA(KP623489801, Tc, T1) + FNMA(KP900968867, Ty, KP222520933 * Tn);
322 ri[WS(rs, 6)] = TK - TL;
323 ri[WS(rs, 1)] = TK + TL;
324 TV = FMA(KP781831482, TS, KP974927912 * TT) + (KP433883739 * TU);
325 TW = FMA(KP623489801, TO, TR) + FNMA(KP900968867, TQ, KP222520933 * TP);
326 ii[WS(rs, 1)] = TV + TW;
327 ii[WS(rs, 6)] = TW - TV;
328 }
329 {
330 E TN, TM, TZ, T10;
331 TN = FMA(KP433883739, TC, KP974927912 * TF) - (KP781831482 * TI);
332 TM = FMA(KP623489801, Tn, T1) + FNMA(KP222520933, Ty, KP900968867 * Tc);
333 ri[WS(rs, 4)] = TM - TN;
334 ri[WS(rs, 3)] = TM + TN;
335 TZ = FMA(KP433883739, TS, KP974927912 * TU) - (KP781831482 * TT);
336 T10 = FMA(KP623489801, TP, TR) + FNMA(KP222520933, TQ, KP900968867 * TO);
337 ii[WS(rs, 3)] = TZ + T10;
338 ii[WS(rs, 4)] = T10 - TZ;
339 }
340 }
341 }
342 }
343
344 static const tw_instr twinstr[] = {
345 {TW_FULL, 0, 7},
346 {TW_NEXT, 1, 0}
347 };
348
349 static const ct_desc desc = { 7, "t1_7", twinstr, &GENUS, {36, 24, 36, 0}, 0, 0, 0 };
350
351 void X(codelet_t1_7) (planner *p) {
352 X(kdft_dit_register) (p, t1_7, &desc);
353 }
354 #endif