Mercurial > hg > sv-dependency-builds

comparison src/fftw-3.3.8/rdft/scalar/r2cf/r2cf_13.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 08:06:26 EDT 2018 */
23
24 #include "rdft/codelet-rdft.h"
25
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 13 -name r2cf_13 -include rdft/scalar/r2cf.h */
29
30 /*
31 * This function contains 76 FP additions, 51 FP multiplications,
32 * (or, 31 additions, 6 multiplications, 45 fused multiply/add),
33 * 58 stack variables, 23 constants, and 26 memory accesses
34 */
35 #include "rdft/scalar/r2cf.h"
36
37 static void r2cf_13(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
38 {
39 DK(KP300462606, +0.300462606288665774426601772289207995520941381);
40 DK(KP516520780, +0.516520780623489722840901288569017135705033622);
41 DK(KP859542535, +0.859542535098774820163672132761689612766401925);
42 DK(KP581704778, +0.581704778510515730456870384989698884939833902);
43 DK(KP514918778, +0.514918778086315755491789696138117261566051239);
44 DK(KP769338817, +0.769338817572980603471413688209101117038278899);
45 DK(KP686558370, +0.686558370781754340655719594850823015421401653);
46 DK(KP226109445, +0.226109445035782405468510155372505010481906348);
47 DK(KP251768516, +0.251768516431883313623436926934233488546674281);
48 DK(KP503537032, +0.503537032863766627246873853868466977093348562);
49 DK(KP301479260, +0.301479260047709873958013540496673347309208464);
50 DK(KP083333333, +0.083333333333333333333333333333333333333333333);
51 DK(KP904176221, +0.904176221990848204433795481776887926501523162);
52 DK(KP575140729, +0.575140729474003121368385547455453388461001608);
53 DK(KP522026385, +0.522026385161275033714027226654165028300441940);
54 DK(KP957805992, +0.957805992594665126462521754605754580515587217);
55 DK(KP600477271, +0.600477271932665282925769253334763009352012849);
56 DK(KP853480001, +0.853480001859823990758994934970528322872359049);
57 DK(KP612264650, +0.612264650376756543746494474777125408779395514);
58 DK(KP038632954, +0.038632954644348171955506895830342264440241080);
59 DK(KP302775637, +0.302775637731994646559610633735247973125648287);
60 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
61 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
62 {
63 INT i;
64 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(52, rs), MAKE_VOLATILE_STRIDE(52, csr), MAKE_VOLATILE_STRIDE(52, csi)) {
65 E TN, TA, TD, TO, TR, TS, TZ, T12, Tu, Tx, Tj, Tw, TW, T13;
66 TN = R0[0];
67 {
68 E T3, TP, Th, TB, Tp, Te, TC, Tm, T6, Tr, T9, Ts, Ta, TQ, T1;
69 E T2;
70 T1 = R0[WS(rs, 4)];
71 T2 = R1[WS(rs, 2)];
72 T3 = T1 - T2;
73 TP = T1 + T2;
74 {
75 E Tn, Tf, Tg, To;
76 Tn = R0[WS(rs, 6)];
77 Tf = R0[WS(rs, 5)];
78 Tg = R0[WS(rs, 2)];
79 To = Tf + Tg;
80 Th = Tf - Tg;
81 TB = Tn + To;
82 Tp = FMS(KP500000000, To, Tn);
83 }
84 {
85 E Tk, Tc, Td, Tl;
86 Tk = R1[0];
87 Tc = R1[WS(rs, 4)];
88 Td = R1[WS(rs, 1)];
89 Tl = Td + Tc;
90 Te = Tc - Td;
91 TC = Tk + Tl;
92 Tm = FNMS(KP500000000, Tl, Tk);
93 }
94 {
95 E T4, T5, T7, T8;
96 T4 = R1[WS(rs, 5)];
97 T5 = R0[WS(rs, 3)];
98 T6 = T4 - T5;
99 Tr = T4 + T5;
100 T7 = R1[WS(rs, 3)];
101 T8 = R0[WS(rs, 1)];
102 T9 = T7 - T8;
103 Ts = T7 + T8;
104 }
105 Ta = T6 + T9;
106 TQ = Tr + Ts;
107 TA = T3 + Ta;
108 TD = TB - TC;
109 TO = TC + TB;
110 TR = TP + TQ;
111 TS = TO + TR;
112 {
113 E TX, TY, Tq, Tt;
114 TX = Tm - Tp;
115 TY = FNMS(KP500000000, TQ, TP);
116 TZ = TX + TY;
117 T12 = TX - TY;
118 Tq = Tm + Tp;
119 Tt = Tr - Ts;
120 Tu = FMA(KP866025403, Tt, Tq);
121 Tx = FNMS(KP866025403, Tt, Tq);
122 }
123 {
124 E Tb, Ti, TU, TV;
125 Tb = FNMS(KP500000000, Ta, T3);
126 Ti = Te + Th;
127 Tj = FMA(KP866025403, Ti, Tb);
128 Tw = FNMS(KP866025403, Ti, Tb);
129 TU = Th - Te;
130 TV = T6 - T9;
131 TW = TU + TV;
132 T13 = TU - TV;
133 }
134 }
135 Cr[0] = TN + TS;
136 {
137 E TE, TI, Tz, TK, TH, TM, TJ, TL;
138 TE = FMA(KP302775637, TD, TA);
139 TI = FNMS(KP302775637, TA, TD);
140 {
141 E Tv, Ty, TF, TG;
142 Tv = FMA(KP038632954, Tu, Tj);
143 Ty = FMA(KP612264650, Tx, Tw);
144 Tz = FNMS(KP853480001, Ty, Tv);
145 TK = FMA(KP853480001, Ty, Tv);
146 TF = FNMS(KP038632954, Tj, Tu);
147 TG = FNMS(KP612264650, Tw, Tx);
148 TH = FNMS(KP853480001, TG, TF);
149 TM = FMA(KP853480001, TG, TF);
150 }
151 Ci[WS(csi, 1)] = KP600477271 * (FMA(KP957805992, TE, Tz));
152 Ci[WS(csi, 5)] = -(KP600477271 * (FNMS(KP957805992, TI, TH)));
153 TJ = FMA(KP522026385, TH, TI);
154 Ci[WS(csi, 2)] = KP575140729 * (FNMS(KP904176221, TK, TJ));
155 Ci[WS(csi, 6)] = KP575140729 * (FMA(KP904176221, TK, TJ));
156 TL = FNMS(KP522026385, Tz, TE);
157 Ci[WS(csi, 3)] = KP575140729 * (FNMS(KP904176221, TM, TL));
158 Ci[WS(csi, 4)] = -(KP575140729 * (FMA(KP904176221, TM, TL)));
159 }
160 {
161 E T11, T17, T1c, T1e, T16, T18, TT, T10, T19, T1d;
162 TT = FNMS(KP083333333, TS, TN);
163 T10 = FMA(KP301479260, TZ, TW);
164 T11 = FMA(KP503537032, T10, TT);
165 T17 = FNMS(KP251768516, T10, TT);
166 {
167 E T1a, T1b, T14, T15;
168 T1a = FNMS(KP226109445, TW, TZ);
169 T1b = FMA(KP686558370, T12, T13);
170 T1c = FNMS(KP769338817, T1b, T1a);
171 T1e = FMA(KP769338817, T1b, T1a);
172 T14 = FNMS(KP514918778, T13, T12);
173 T15 = TO - TR;
174 T16 = FMA(KP581704778, T15, T14);
175 T18 = FNMS(KP859542535, T14, T15);
176 }
177 Cr[WS(csr, 5)] = FNMS(KP516520780, T16, T11);
178 Cr[WS(csr, 1)] = FMA(KP516520780, T16, T11);
179 T19 = FMA(KP300462606, T18, T17);
180 Cr[WS(csr, 4)] = FNMS(KP503537032, T1c, T19);
181 Cr[WS(csr, 3)] = FMA(KP503537032, T1c, T19);
182 T1d = FNMS(KP300462606, T18, T17);
183 Cr[WS(csr, 6)] = FNMS(KP503537032, T1e, T1d);
184 Cr[WS(csr, 2)] = FMA(KP503537032, T1e, T1d);
185 }
186 }
187 }
188 }
189
190 static const kr2c_desc desc = { 13, "r2cf_13", {31, 6, 45, 0}, &GENUS };
191
192 void X(codelet_r2cf_13) (planner *p) {
193 X(kr2c_register) (p, r2cf_13, &desc);
194 }
195
196 #else
197
198 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 13 -name r2cf_13 -include rdft/scalar/r2cf.h */
199
200 /*
201 * This function contains 76 FP additions, 34 FP multiplications,
202 * (or, 57 additions, 15 multiplications, 19 fused multiply/add),
203 * 55 stack variables, 20 constants, and 26 memory accesses
204 */
205 #include "rdft/scalar/r2cf.h"
206
207 static void r2cf_13(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
208 {
209 DK(KP083333333, +0.083333333333333333333333333333333333333333333);
210 DK(KP075902986, +0.075902986037193865983102897245103540356428373);
211 DK(KP251768516, +0.251768516431883313623436926934233488546674281);
212 DK(KP503537032, +0.503537032863766627246873853868466977093348562);
213 DK(KP113854479, +0.113854479055790798974654345867655310534642560);
214 DK(KP265966249, +0.265966249214837287587521063842185948798330267);
215 DK(KP387390585, +0.387390585467617292130675966426762851778775217);
216 DK(KP300462606, +0.300462606288665774426601772289207995520941381);
217 DK(KP132983124, +0.132983124607418643793760531921092974399165133);
218 DK(KP258260390, +0.258260390311744861420450644284508567852516811);
219 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
220 DK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
221 DK(KP300238635, +0.300238635966332641462884626667381504676006424);
222 DK(KP011599105, +0.011599105605768290721655456654083252189827041);
223 DK(KP156891391, +0.156891391051584611046832726756003269660212636);
224 DK(KP256247671, +0.256247671582936600958684654061725059144125175);
225 DK(KP174138601, +0.174138601152135905005660794929264742616964676);
226 DK(KP575140729, +0.575140729474003121368385547455453388461001608);
227 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
228 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
229 {
230 INT i;
231 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(52, rs), MAKE_VOLATILE_STRIDE(52, csr), MAKE_VOLATILE_STRIDE(52, csi)) {
232 E T13, Tb, Tm, TW, TX, T14, TU, T10, Tz, TB, Tu, TC, TR, T11;
233 T13 = R0[0];
234 {
235 E Te, TO, Ta, Tv, To, T5, Tw, Tp, Th, Tr, Tk, Ts, Tl, TP, Tc;
236 E Td;
237 Tc = R0[WS(rs, 4)];
238 Td = R1[WS(rs, 2)];
239 Te = Tc - Td;
240 TO = Tc + Td;
241 {
242 E T6, T7, T8, T9;
243 T6 = R1[0];
244 T7 = R1[WS(rs, 1)];
245 T8 = R1[WS(rs, 4)];
246 T9 = T7 + T8;
247 Ta = T6 + T9;
248 Tv = T7 - T8;
249 To = FNMS(KP500000000, T9, T6);
250 }
251 {
252 E T1, T2, T3, T4;
253 T1 = R0[WS(rs, 6)];
254 T2 = R0[WS(rs, 5)];
255 T3 = R0[WS(rs, 2)];
256 T4 = T2 + T3;
257 T5 = T1 + T4;
258 Tw = T2 - T3;
259 Tp = FNMS(KP500000000, T4, T1);
260 }
261 {
262 E Tf, Tg, Ti, Tj;
263 Tf = R1[WS(rs, 5)];
264 Tg = R0[WS(rs, 3)];
265 Th = Tf - Tg;
266 Tr = Tf + Tg;
267 Ti = R1[WS(rs, 3)];
268 Tj = R0[WS(rs, 1)];
269 Tk = Ti - Tj;
270 Ts = Ti + Tj;
271 }
272 Tl = Th + Tk;
273 TP = Tr + Ts;
274 Tb = T5 - Ta;
275 Tm = Te + Tl;
276 TW = Ta + T5;
277 TX = TO + TP;
278 T14 = TW + TX;
279 {
280 E TS, TT, Tx, Ty;
281 TS = Tv + Tw;
282 TT = Th - Tk;
283 TU = TS - TT;
284 T10 = TS + TT;
285 Tx = KP866025403 * (Tv - Tw);
286 Ty = FNMS(KP500000000, Tl, Te);
287 Tz = Tx + Ty;
288 TB = Ty - Tx;
289 }
290 {
291 E Tq, Tt, TN, TQ;
292 Tq = To - Tp;
293 Tt = KP866025403 * (Tr - Ts);
294 Tu = Tq - Tt;
295 TC = Tq + Tt;
296 TN = To + Tp;
297 TQ = FNMS(KP500000000, TP, TO);
298 TR = TN - TQ;
299 T11 = TN + TQ;
300 }
301 }
302 Cr[0] = T13 + T14;
303 {
304 E Tn, TG, TE, TF, TJ, TM, TK, TL;
305 Tn = FNMS(KP174138601, Tm, KP575140729 * Tb);
306 TG = FMA(KP174138601, Tb, KP575140729 * Tm);
307 {
308 E TA, TD, TH, TI;
309 TA = FNMS(KP156891391, Tz, KP256247671 * Tu);
310 TD = FNMS(KP300238635, TC, KP011599105 * TB);
311 TE = TA + TD;
312 TF = KP1_732050807 * (TD - TA);
313 TH = FMA(KP300238635, TB, KP011599105 * TC);
314 TI = FMA(KP256247671, Tz, KP156891391 * Tu);
315 TJ = TH - TI;
316 TM = KP1_732050807 * (TI + TH);
317 }
318 Ci[WS(csi, 5)] = FMA(KP2_000000000, TE, Tn);
319 Ci[WS(csi, 1)] = FMA(KP2_000000000, TJ, TG);
320 TK = TG - TJ;
321 Ci[WS(csi, 4)] = TF - TK;
322 Ci[WS(csi, 3)] = TF + TK;
323 TL = Tn - TE;
324 Ci[WS(csi, 2)] = TL - TM;
325 Ci[WS(csi, 6)] = TL + TM;
326 }
327 {
328 E TZ, T1b, T19, T1e, T16, T1a, TV, TY, T1c, T1d;
329 TV = FNMS(KP132983124, TU, KP258260390 * TR);
330 TY = KP300462606 * (TW - TX);
331 TZ = FMA(KP2_000000000, TV, TY);
332 T1b = TY - TV;
333 {
334 E T17, T18, T12, T15;
335 T17 = FMA(KP387390585, TU, KP265966249 * TR);
336 T18 = FNMS(KP503537032, T11, KP113854479 * T10);
337 T19 = T17 - T18;
338 T1e = T17 + T18;
339 T12 = FMA(KP251768516, T10, KP075902986 * T11);
340 T15 = FNMS(KP083333333, T14, T13);
341 T16 = FMA(KP2_000000000, T12, T15);
342 T1a = T15 - T12;
343 }
344 Cr[WS(csr, 1)] = TZ + T16;
345 Cr[WS(csr, 5)] = T16 - TZ;
346 T1c = T1a - T1b;
347 Cr[WS(csr, 2)] = T19 + T1c;
348 Cr[WS(csr, 6)] = T1c - T19;
349 T1d = T1b + T1a;
350 Cr[WS(csr, 3)] = T1d - T1e;
351 Cr[WS(csr, 4)] = T1e + T1d;
352 }
353 }
354 }
355 }
356
357 static const kr2c_desc desc = { 13, "r2cf_13", {57, 15, 19, 0}, &GENUS };
358
359 void X(codelet_r2cf_13) (planner *p) {
360 X(kr2c_register) (p, r2cf_13, &desc);
361 }
362
363 #endif