Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/scalar/codelets/n1_13.c @ 10:37bf6b4a2645
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 13 -name n1_13 -include n.h */ | |
29 | |
30 /* | |
31 * This function contains 176 FP additions, 114 FP multiplications, | |
32 * (or, 62 additions, 0 multiplications, 114 fused multiply/add), | |
33 * 87 stack variables, 25 constants, and 52 memory accesses | |
34 */ | |
35 #include "n.h" | |
36 | |
37 static void n1_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP875502302, +0.875502302409147941146295545768755143177842006); | |
40 DK(KP520028571, +0.520028571888864619117130500499232802493238139); | |
41 DK(KP575140729, +0.575140729474003121368385547455453388461001608); | |
42 DK(KP600477271, +0.600477271932665282925769253334763009352012849); | |
43 DK(KP300462606, +0.300462606288665774426601772289207995520941381); | |
44 DK(KP516520780, +0.516520780623489722840901288569017135705033622); | |
45 DK(KP968287244, +0.968287244361984016049539446938120421179794516); | |
46 DK(KP503537032, +0.503537032863766627246873853868466977093348562); | |
47 DK(KP251768516, +0.251768516431883313623436926934233488546674281); | |
48 DK(KP581704778, +0.581704778510515730456870384989698884939833902); | |
49 DK(KP859542535, +0.859542535098774820163672132761689612766401925); | |
50 DK(KP083333333, +0.083333333333333333333333333333333333333333333); | |
51 DK(KP957805992, +0.957805992594665126462521754605754580515587217); | |
52 DK(KP522026385, +0.522026385161275033714027226654165028300441940); | |
53 DK(KP853480001, +0.853480001859823990758994934970528322872359049); | |
54 DK(KP769338817, +0.769338817572980603471413688209101117038278899); | |
55 DK(KP612264650, +0.612264650376756543746494474777125408779395514); | |
56 DK(KP038632954, +0.038632954644348171955506895830342264440241080); | |
57 DK(KP302775637, +0.302775637731994646559610633735247973125648287); | |
58 DK(KP514918778, +0.514918778086315755491789696138117261566051239); | |
59 DK(KP686558370, +0.686558370781754340655719594850823015421401653); | |
60 DK(KP226109445, +0.226109445035782405468510155372505010481906348); | |
61 DK(KP301479260, +0.301479260047709873958013540496673347309208464); | |
62 DK(KP866025403, +0.866025403784438646763723170752936183471402627); | |
63 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
64 { | |
65 INT i; | |
66 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(52, is), MAKE_VOLATILE_STRIDE(52, os)) { | |
67 E T2B, T2H, T2I, T2G; | |
68 { | |
69 E T1, T1P, T2n, T2o, To, TH, T2h, T2k, TE, TB, TF, Tw, T2j, T2c, T1m; | |
70 E T1W, T1X, T1c, T19, T1j, T12, T1f, T21, T24, T27, T1U; | |
71 T1 = ri[0]; | |
72 T1P = ii[0]; | |
73 { | |
74 E T2b, Tv, Ts, T2a; | |
75 { | |
76 E T2d, Tf, Tq, Ty, Tb, Tr, T6, Tx, Ti, Tt, Tu, Tl; | |
77 { | |
78 E T7, T8, T9, Td, Te; | |
79 Td = ri[WS(is, 8)]; | |
80 Te = ri[WS(is, 5)]; | |
81 T7 = ri[WS(is, 12)]; | |
82 T8 = ri[WS(is, 10)]; | |
83 T9 = ri[WS(is, 4)]; | |
84 T2d = Td - Te; | |
85 Tf = Td + Te; | |
86 { | |
87 E T2, Ta, T3, T4; | |
88 T2 = ri[WS(is, 1)]; | |
89 Ta = T8 + T9; | |
90 Tq = T8 - T9; | |
91 T3 = ri[WS(is, 3)]; | |
92 T4 = ri[WS(is, 9)]; | |
93 { | |
94 E Tg, T5, Th, Tj, Tk; | |
95 Tg = ri[WS(is, 11)]; | |
96 Ty = FMS(KP500000000, Ta, T7); | |
97 Tb = T7 + Ta; | |
98 Tr = T4 - T3; | |
99 T5 = T3 + T4; | |
100 Th = ri[WS(is, 6)]; | |
101 Tj = ri[WS(is, 7)]; | |
102 Tk = ri[WS(is, 2)]; | |
103 T6 = T2 + T5; | |
104 Tx = FNMS(KP500000000, T5, T2); | |
105 Ti = Tg + Th; | |
106 Tt = Tg - Th; | |
107 Tu = Tj - Tk; | |
108 Tl = Tj + Tk; | |
109 } | |
110 } | |
111 } | |
112 { | |
113 E Tc, Tm, T2e, T2g; | |
114 Tc = T6 + Tb; | |
115 T2n = T6 - Tb; | |
116 T2b = Ti - Tl; | |
117 Tm = Ti + Tl; | |
118 T2e = Tt + Tu; | |
119 Tv = Tt - Tu; | |
120 Ts = Tq - Tr; | |
121 T2g = Tr + Tq; | |
122 { | |
123 E Tz, TA, Tn, T2f; | |
124 Tz = Tx - Ty; | |
125 T2a = Tx + Ty; | |
126 TA = FNMS(KP500000000, Tm, Tf); | |
127 Tn = Tf + Tm; | |
128 T2f = FNMS(KP500000000, T2e, T2d); | |
129 T2o = T2d + T2e; | |
130 To = Tc + Tn; | |
131 TH = Tc - Tn; | |
132 T2h = FMA(KP866025403, T2g, T2f); | |
133 T2k = FNMS(KP866025403, T2g, T2f); | |
134 TE = Tz - TA; | |
135 TB = Tz + TA; | |
136 } | |
137 } | |
138 } | |
139 { | |
140 E T1R, TM, T10, T18, T1l, TX, T1k, T15, TP, T1a, T1b, TS; | |
141 { | |
142 E T16, TY, TZ, TK, TL; | |
143 TK = ii[WS(is, 8)]; | |
144 TF = Ts - Tv; | |
145 Tw = Ts + Tv; | |
146 T2j = FNMS(KP866025403, T2b, T2a); | |
147 T2c = FMA(KP866025403, T2b, T2a); | |
148 TL = ii[WS(is, 5)]; | |
149 T16 = ii[WS(is, 12)]; | |
150 TY = ii[WS(is, 10)]; | |
151 TZ = ii[WS(is, 4)]; | |
152 T1R = TK + TL; | |
153 TM = TK - TL; | |
154 { | |
155 E T13, T17, TV, TW; | |
156 T13 = ii[WS(is, 1)]; | |
157 T17 = TY + TZ; | |
158 T10 = TY - TZ; | |
159 TV = ii[WS(is, 9)]; | |
160 TW = ii[WS(is, 3)]; | |
161 { | |
162 E TN, T14, TO, TQ, TR; | |
163 TN = ii[WS(is, 11)]; | |
164 T18 = FMS(KP500000000, T17, T16); | |
165 T1l = T16 + T17; | |
166 TX = TV - TW; | |
167 T14 = TW + TV; | |
168 TO = ii[WS(is, 6)]; | |
169 TQ = ii[WS(is, 7)]; | |
170 TR = ii[WS(is, 2)]; | |
171 T1k = T13 + T14; | |
172 T15 = FNMS(KP500000000, T14, T13); | |
173 TP = TN - TO; | |
174 T1a = TN + TO; | |
175 T1b = TQ + TR; | |
176 TS = TQ - TR; | |
177 } | |
178 } | |
179 } | |
180 { | |
181 E T1Q, T11, TT, T1S; | |
182 T1Q = T1k + T1l; | |
183 T1m = T1k - T1l; | |
184 T11 = TX + T10; | |
185 T1W = T10 - TX; | |
186 T1X = TP - TS; | |
187 TT = TP + TS; | |
188 T1S = T1a + T1b; | |
189 T1c = T1a - T1b; | |
190 { | |
191 E T1Z, TU, T1T, T20; | |
192 T19 = T15 + T18; | |
193 T1Z = T15 - T18; | |
194 T1j = TM + TT; | |
195 TU = FNMS(KP500000000, TT, TM); | |
196 T1T = T1R + T1S; | |
197 T20 = FNMS(KP500000000, T1S, T1R); | |
198 T12 = FMA(KP866025403, T11, TU); | |
199 T1f = FNMS(KP866025403, T11, TU); | |
200 T21 = T1Z + T20; | |
201 T24 = T1Z - T20; | |
202 T27 = T1Q - T1T; | |
203 T1U = T1Q + T1T; | |
204 } | |
205 } | |
206 } | |
207 } | |
208 { | |
209 E T1g, T1d, T25, T1Y; | |
210 ro[0] = T1 + To; | |
211 T1g = FNMS(KP866025403, T1c, T19); | |
212 T1d = FMA(KP866025403, T1c, T19); | |
213 T25 = T1W - T1X; | |
214 T1Y = T1W + T1X; | |
215 io[0] = T1P + T1U; | |
216 { | |
217 E T1C, T1B, T1F, T1K; | |
218 { | |
219 E TC, T1J, T1z, T1w, T1I, T1O, Tp, T1E, T1q, TI, T1o, T1s; | |
220 { | |
221 E TG, T1n, T1G, T1u, T1e, T1h, T1v, T1x, T1y, T1H, T1i; | |
222 TC = FMA(KP301479260, TB, Tw); | |
223 T1x = FNMS(KP226109445, Tw, TB); | |
224 T1y = FMA(KP686558370, TE, TF); | |
225 TG = FNMS(KP514918778, TF, TE); | |
226 T1n = FNMS(KP302775637, T1m, T1j); | |
227 T1G = FMA(KP302775637, T1j, T1m); | |
228 T1u = FNMS(KP038632954, T12, T1d); | |
229 T1e = FMA(KP038632954, T1d, T12); | |
230 T1h = FMA(KP612264650, T1g, T1f); | |
231 T1v = FNMS(KP612264650, T1f, T1g); | |
232 T1J = FMA(KP769338817, T1y, T1x); | |
233 T1z = FNMS(KP769338817, T1y, T1x); | |
234 T1H = FNMS(KP853480001, T1v, T1u); | |
235 T1w = FMA(KP853480001, T1v, T1u); | |
236 T1I = FNMS(KP522026385, T1H, T1G); | |
237 T1O = FMA(KP957805992, T1G, T1H); | |
238 Tp = FNMS(KP083333333, To, T1); | |
239 T1E = FMA(KP853480001, T1h, T1e); | |
240 T1i = FNMS(KP853480001, T1h, T1e); | |
241 T1q = FNMS(KP859542535, TG, TH); | |
242 TI = FMA(KP581704778, TH, TG); | |
243 T1o = FMA(KP957805992, T1n, T1i); | |
244 T1s = FNMS(KP522026385, T1i, T1n); | |
245 } | |
246 { | |
247 E T1A, T1D, T1t, T1L, T1M; | |
248 { | |
249 E T1p, TD, TJ, T1N, T1r; | |
250 T1p = FNMS(KP251768516, TC, Tp); | |
251 TD = FMA(KP503537032, TC, Tp); | |
252 T1C = FNMS(KP968287244, T1z, T1w); | |
253 T1A = FMA(KP968287244, T1z, T1w); | |
254 TJ = FMA(KP516520780, TI, TD); | |
255 T1N = FNMS(KP516520780, TI, TD); | |
256 T1D = FNMS(KP300462606, T1q, T1p); | |
257 T1r = FMA(KP300462606, T1q, T1p); | |
258 ro[WS(os, 8)] = FNMS(KP600477271, T1O, T1N); | |
259 ro[WS(os, 12)] = FMA(KP600477271, T1o, TJ); | |
260 ro[WS(os, 1)] = FNMS(KP600477271, T1o, TJ); | |
261 T1t = FNMS(KP575140729, T1s, T1r); | |
262 T1B = FMA(KP575140729, T1s, T1r); | |
263 ro[WS(os, 5)] = FMA(KP600477271, T1O, T1N); | |
264 } | |
265 T1L = FNMS(KP520028571, T1E, T1D); | |
266 T1F = FMA(KP520028571, T1E, T1D); | |
267 T1K = FMA(KP875502302, T1J, T1I); | |
268 T1M = FNMS(KP875502302, T1J, T1I); | |
269 ro[WS(os, 3)] = FMA(KP520028571, T1A, T1t); | |
270 ro[WS(os, 9)] = FNMS(KP520028571, T1A, T1t); | |
271 ro[WS(os, 6)] = FMA(KP575140729, T1M, T1L); | |
272 ro[WS(os, 11)] = FNMS(KP575140729, T1M, T1L); | |
273 } | |
274 } | |
275 { | |
276 E T22, T2F, T2N, T2K, T2w, T2A, T1V, T2C, T28, T2y, T2M, T2q; | |
277 { | |
278 E T26, T2v, T2p, T2i, T2s, T2t, T2l, T2D, T2E, T2u, T2m; | |
279 T2D = FNMS(KP226109445, T1Y, T21); | |
280 T22 = FMA(KP301479260, T21, T1Y); | |
281 ro[WS(os, 2)] = FMA(KP575140729, T1K, T1F); | |
282 ro[WS(os, 7)] = FNMS(KP575140729, T1K, T1F); | |
283 ro[WS(os, 4)] = FMA(KP520028571, T1C, T1B); | |
284 ro[WS(os, 10)] = FNMS(KP520028571, T1C, T1B); | |
285 T26 = FNMS(KP514918778, T25, T24); | |
286 T2E = FMA(KP686558370, T24, T25); | |
287 T2v = FNMS(KP302775637, T2n, T2o); | |
288 T2p = FMA(KP302775637, T2o, T2n); | |
289 T2i = FNMS(KP038632954, T2h, T2c); | |
290 T2s = FMA(KP038632954, T2c, T2h); | |
291 T2t = FMA(KP612264650, T2j, T2k); | |
292 T2l = FNMS(KP612264650, T2k, T2j); | |
293 T2F = FNMS(KP769338817, T2E, T2D); | |
294 T2N = FMA(KP769338817, T2E, T2D); | |
295 T2K = FMA(KP853480001, T2t, T2s); | |
296 T2u = FNMS(KP853480001, T2t, T2s); | |
297 T2w = FMA(KP957805992, T2v, T2u); | |
298 T2A = FNMS(KP522026385, T2u, T2v); | |
299 T1V = FNMS(KP083333333, T1U, T1P); | |
300 T2m = FNMS(KP853480001, T2l, T2i); | |
301 T2C = FMA(KP853480001, T2l, T2i); | |
302 T28 = FMA(KP581704778, T27, T26); | |
303 T2y = FNMS(KP859542535, T26, T27); | |
304 T2M = FNMS(KP522026385, T2m, T2p); | |
305 T2q = FMA(KP957805992, T2p, T2m); | |
306 } | |
307 { | |
308 E T2O, T2Q, T2z, T2P, T2L; | |
309 { | |
310 E T23, T2x, T2r, T29, T2J; | |
311 T23 = FMA(KP503537032, T22, T1V); | |
312 T2x = FNMS(KP251768516, T22, T1V); | |
313 T2O = FNMS(KP875502302, T2N, T2M); | |
314 T2Q = FMA(KP875502302, T2N, T2M); | |
315 T2r = FMA(KP516520780, T28, T23); | |
316 T29 = FNMS(KP516520780, T28, T23); | |
317 T2z = FMA(KP300462606, T2y, T2x); | |
318 T2J = FNMS(KP300462606, T2y, T2x); | |
319 io[WS(os, 12)] = FNMS(KP600477271, T2w, T2r); | |
320 io[WS(os, 1)] = FMA(KP600477271, T2w, T2r); | |
321 io[WS(os, 8)] = FMA(KP600477271, T2q, T29); | |
322 io[WS(os, 5)] = FNMS(KP600477271, T2q, T29); | |
323 T2P = FMA(KP520028571, T2K, T2J); | |
324 T2L = FNMS(KP520028571, T2K, T2J); | |
325 } | |
326 T2B = FMA(KP575140729, T2A, T2z); | |
327 T2H = FNMS(KP575140729, T2A, T2z); | |
328 io[WS(os, 11)] = FMA(KP575140729, T2Q, T2P); | |
329 io[WS(os, 6)] = FNMS(KP575140729, T2Q, T2P); | |
330 io[WS(os, 7)] = FMA(KP575140729, T2O, T2L); | |
331 io[WS(os, 2)] = FNMS(KP575140729, T2O, T2L); | |
332 T2I = FMA(KP968287244, T2F, T2C); | |
333 T2G = FNMS(KP968287244, T2F, T2C); | |
334 } | |
335 } | |
336 } | |
337 } | |
338 } | |
339 io[WS(os, 10)] = FMA(KP520028571, T2I, T2H); | |
340 io[WS(os, 4)] = FNMS(KP520028571, T2I, T2H); | |
341 io[WS(os, 9)] = FMA(KP520028571, T2G, T2B); | |
342 io[WS(os, 3)] = FNMS(KP520028571, T2G, T2B); | |
343 } | |
344 } | |
345 } | |
346 | |
347 static const kdft_desc desc = { 13, "n1_13", {62, 0, 114, 0}, &GENUS, 0, 0, 0, 0 }; | |
348 | |
349 void X(codelet_n1_13) (planner *p) { | |
350 X(kdft_register) (p, n1_13, &desc); | |
351 } | |
352 | |
353 #else /* HAVE_FMA */ | |
354 | |
355 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 13 -name n1_13 -include n.h */ | |
356 | |
357 /* | |
358 * This function contains 176 FP additions, 68 FP multiplications, | |
359 * (or, 138 additions, 30 multiplications, 38 fused multiply/add), | |
360 * 71 stack variables, 20 constants, and 52 memory accesses | |
361 */ | |
362 #include "n.h" | |
363 | |
364 static void n1_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
365 { | |
366 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
367 DK(KP083333333, +0.083333333333333333333333333333333333333333333); | |
368 DK(KP251768516, +0.251768516431883313623436926934233488546674281); | |
369 DK(KP075902986, +0.075902986037193865983102897245103540356428373); | |
370 DK(KP132983124, +0.132983124607418643793760531921092974399165133); | |
371 DK(KP258260390, +0.258260390311744861420450644284508567852516811); | |
372 DK(KP1_732050807, +1.732050807568877293527446341505872366942805254); | |
373 DK(KP300238635, +0.300238635966332641462884626667381504676006424); | |
374 DK(KP011599105, +0.011599105605768290721655456654083252189827041); | |
375 DK(KP156891391, +0.156891391051584611046832726756003269660212636); | |
376 DK(KP256247671, +0.256247671582936600958684654061725059144125175); | |
377 DK(KP174138601, +0.174138601152135905005660794929264742616964676); | |
378 DK(KP575140729, +0.575140729474003121368385547455453388461001608); | |
379 DK(KP503537032, +0.503537032863766627246873853868466977093348562); | |
380 DK(KP113854479, +0.113854479055790798974654345867655310534642560); | |
381 DK(KP265966249, +0.265966249214837287587521063842185948798330267); | |
382 DK(KP387390585, +0.387390585467617292130675966426762851778775217); | |
383 DK(KP866025403, +0.866025403784438646763723170752936183471402627); | |
384 DK(KP300462606, +0.300462606288665774426601772289207995520941381); | |
385 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
386 { | |
387 INT i; | |
388 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(52, is), MAKE_VOLATILE_STRIDE(52, os)) { | |
389 E T1, T1q, Tt, Tu, To, T22, T20, T24, TF, TH, TA, TI, T1X, T25, T2a; | |
390 E T2d, T18, T1n, T2k, T2n, T1l, T1r, T1f, T1o, T2h, T2m; | |
391 T1 = ri[0]; | |
392 T1q = ii[0]; | |
393 { | |
394 E Tf, Tp, Tb, TC, Tx, T6, TB, Tw, Ti, Tq, Tl, Tr, Tm, Ts, Td; | |
395 E Te, Tc, Tn; | |
396 Td = ri[WS(is, 8)]; | |
397 Te = ri[WS(is, 5)]; | |
398 Tf = Td + Te; | |
399 Tp = Td - Te; | |
400 { | |
401 E T7, T8, T9, Ta; | |
402 T7 = ri[WS(is, 12)]; | |
403 T8 = ri[WS(is, 10)]; | |
404 T9 = ri[WS(is, 4)]; | |
405 Ta = T8 + T9; | |
406 Tb = T7 + Ta; | |
407 TC = T8 - T9; | |
408 Tx = FNMS(KP500000000, Ta, T7); | |
409 } | |
410 { | |
411 E T2, T3, T4, T5; | |
412 T2 = ri[WS(is, 1)]; | |
413 T3 = ri[WS(is, 3)]; | |
414 T4 = ri[WS(is, 9)]; | |
415 T5 = T3 + T4; | |
416 T6 = T2 + T5; | |
417 TB = T3 - T4; | |
418 Tw = FNMS(KP500000000, T5, T2); | |
419 } | |
420 { | |
421 E Tg, Th, Tj, Tk; | |
422 Tg = ri[WS(is, 11)]; | |
423 Th = ri[WS(is, 6)]; | |
424 Ti = Tg + Th; | |
425 Tq = Tg - Th; | |
426 Tj = ri[WS(is, 7)]; | |
427 Tk = ri[WS(is, 2)]; | |
428 Tl = Tj + Tk; | |
429 Tr = Tj - Tk; | |
430 } | |
431 Tm = Ti + Tl; | |
432 Ts = Tq + Tr; | |
433 Tt = Tp + Ts; | |
434 Tu = T6 - Tb; | |
435 Tc = T6 + Tb; | |
436 Tn = Tf + Tm; | |
437 To = Tc + Tn; | |
438 T22 = KP300462606 * (Tc - Tn); | |
439 { | |
440 E T1Y, T1Z, TD, TE; | |
441 T1Y = TB + TC; | |
442 T1Z = Tq - Tr; | |
443 T20 = T1Y - T1Z; | |
444 T24 = T1Y + T1Z; | |
445 TD = KP866025403 * (TB - TC); | |
446 TE = FNMS(KP500000000, Ts, Tp); | |
447 TF = TD - TE; | |
448 TH = TD + TE; | |
449 } | |
450 { | |
451 E Ty, Tz, T1V, T1W; | |
452 Ty = Tw - Tx; | |
453 Tz = KP866025403 * (Ti - Tl); | |
454 TA = Ty + Tz; | |
455 TI = Ty - Tz; | |
456 T1V = Tw + Tx; | |
457 T1W = FNMS(KP500000000, Tm, Tf); | |
458 T1X = T1V - T1W; | |
459 T25 = T1V + T1W; | |
460 } | |
461 } | |
462 { | |
463 E TZ, T2b, TV, T1i, T1a, TQ, T1h, T19, T12, T1d, T15, T1c, T16, T2c, TX; | |
464 E TY, TW, T17; | |
465 TX = ii[WS(is, 8)]; | |
466 TY = ii[WS(is, 5)]; | |
467 TZ = TX + TY; | |
468 T2b = TX - TY; | |
469 { | |
470 E TR, TS, TT, TU; | |
471 TR = ii[WS(is, 12)]; | |
472 TS = ii[WS(is, 10)]; | |
473 TT = ii[WS(is, 4)]; | |
474 TU = TS + TT; | |
475 TV = FNMS(KP500000000, TU, TR); | |
476 T1i = TR + TU; | |
477 T1a = TS - TT; | |
478 } | |
479 { | |
480 E TM, TN, TO, TP; | |
481 TM = ii[WS(is, 1)]; | |
482 TN = ii[WS(is, 3)]; | |
483 TO = ii[WS(is, 9)]; | |
484 TP = TN + TO; | |
485 TQ = FNMS(KP500000000, TP, TM); | |
486 T1h = TM + TP; | |
487 T19 = TN - TO; | |
488 } | |
489 { | |
490 E T10, T11, T13, T14; | |
491 T10 = ii[WS(is, 11)]; | |
492 T11 = ii[WS(is, 6)]; | |
493 T12 = T10 + T11; | |
494 T1d = T10 - T11; | |
495 T13 = ii[WS(is, 7)]; | |
496 T14 = ii[WS(is, 2)]; | |
497 T15 = T13 + T14; | |
498 T1c = T13 - T14; | |
499 } | |
500 T16 = T12 + T15; | |
501 T2c = T1d + T1c; | |
502 T2a = T1h - T1i; | |
503 T2d = T2b + T2c; | |
504 TW = TQ + TV; | |
505 T17 = FNMS(KP500000000, T16, TZ); | |
506 T18 = TW - T17; | |
507 T1n = TW + T17; | |
508 { | |
509 E T2i, T2j, T1j, T1k; | |
510 T2i = TQ - TV; | |
511 T2j = KP866025403 * (T15 - T12); | |
512 T2k = T2i + T2j; | |
513 T2n = T2i - T2j; | |
514 T1j = T1h + T1i; | |
515 T1k = TZ + T16; | |
516 T1l = KP300462606 * (T1j - T1k); | |
517 T1r = T1j + T1k; | |
518 } | |
519 { | |
520 E T1b, T1e, T2f, T2g; | |
521 T1b = T19 + T1a; | |
522 T1e = T1c - T1d; | |
523 T1f = T1b + T1e; | |
524 T1o = T1e - T1b; | |
525 T2f = FNMS(KP500000000, T2c, T2b); | |
526 T2g = KP866025403 * (T1a - T19); | |
527 T2h = T2f - T2g; | |
528 T2m = T2g + T2f; | |
529 } | |
530 } | |
531 ro[0] = T1 + To; | |
532 io[0] = T1q + T1r; | |
533 { | |
534 E T1D, T1N, T1y, T1x, T1E, T1O, Tv, TK, T1J, T1Q, T1m, T1R, T1t, T1I, TG; | |
535 E TJ; | |
536 { | |
537 E T1B, T1C, T1v, T1w; | |
538 T1B = FMA(KP387390585, T1f, KP265966249 * T18); | |
539 T1C = FMA(KP113854479, T1o, KP503537032 * T1n); | |
540 T1D = T1B + T1C; | |
541 T1N = T1C - T1B; | |
542 T1y = FMA(KP575140729, Tu, KP174138601 * Tt); | |
543 T1v = FNMS(KP156891391, TH, KP256247671 * TI); | |
544 T1w = FMA(KP011599105, TF, KP300238635 * TA); | |
545 T1x = T1v - T1w; | |
546 T1E = T1y + T1x; | |
547 T1O = KP1_732050807 * (T1v + T1w); | |
548 } | |
549 Tv = FNMS(KP174138601, Tu, KP575140729 * Tt); | |
550 TG = FNMS(KP300238635, TF, KP011599105 * TA); | |
551 TJ = FMA(KP256247671, TH, KP156891391 * TI); | |
552 TK = TG - TJ; | |
553 T1J = KP1_732050807 * (TJ + TG); | |
554 T1Q = Tv - TK; | |
555 { | |
556 E T1g, T1H, T1p, T1s, T1G; | |
557 T1g = FNMS(KP132983124, T1f, KP258260390 * T18); | |
558 T1H = T1l - T1g; | |
559 T1p = FNMS(KP251768516, T1o, KP075902986 * T1n); | |
560 T1s = FNMS(KP083333333, T1r, T1q); | |
561 T1G = T1s - T1p; | |
562 T1m = FMA(KP2_000000000, T1g, T1l); | |
563 T1R = T1H + T1G; | |
564 T1t = FMA(KP2_000000000, T1p, T1s); | |
565 T1I = T1G - T1H; | |
566 } | |
567 { | |
568 E TL, T1u, T1P, T1S; | |
569 TL = FMA(KP2_000000000, TK, Tv); | |
570 T1u = T1m + T1t; | |
571 io[WS(os, 1)] = TL + T1u; | |
572 io[WS(os, 12)] = T1u - TL; | |
573 { | |
574 E T1z, T1A, T1T, T1U; | |
575 T1z = FMS(KP2_000000000, T1x, T1y); | |
576 T1A = T1t - T1m; | |
577 io[WS(os, 5)] = T1z + T1A; | |
578 io[WS(os, 8)] = T1A - T1z; | |
579 T1T = T1R - T1Q; | |
580 T1U = T1O + T1N; | |
581 io[WS(os, 4)] = T1T - T1U; | |
582 io[WS(os, 10)] = T1U + T1T; | |
583 } | |
584 T1P = T1N - T1O; | |
585 T1S = T1Q + T1R; | |
586 io[WS(os, 3)] = T1P + T1S; | |
587 io[WS(os, 9)] = T1S - T1P; | |
588 { | |
589 E T1L, T1M, T1F, T1K; | |
590 T1L = T1J + T1I; | |
591 T1M = T1E + T1D; | |
592 io[WS(os, 6)] = T1L - T1M; | |
593 io[WS(os, 11)] = T1M + T1L; | |
594 T1F = T1D - T1E; | |
595 T1K = T1I - T1J; | |
596 io[WS(os, 2)] = T1F + T1K; | |
597 io[WS(os, 7)] = T1K - T1F; | |
598 } | |
599 } | |
600 } | |
601 { | |
602 E T2y, T2I, T2J, T2K, T2B, T2L, T2e, T2p, T2u, T2G, T23, T2F, T28, T2t, T2l; | |
603 E T2o; | |
604 { | |
605 E T2w, T2x, T2z, T2A; | |
606 T2w = FMA(KP387390585, T20, KP265966249 * T1X); | |
607 T2x = FNMS(KP503537032, T25, KP113854479 * T24); | |
608 T2y = T2w + T2x; | |
609 T2I = T2w - T2x; | |
610 T2J = FMA(KP575140729, T2a, KP174138601 * T2d); | |
611 T2z = FNMS(KP300238635, T2n, KP011599105 * T2m); | |
612 T2A = FNMS(KP156891391, T2h, KP256247671 * T2k); | |
613 T2K = T2z + T2A; | |
614 T2B = KP1_732050807 * (T2z - T2A); | |
615 T2L = T2J + T2K; | |
616 } | |
617 T2e = FNMS(KP575140729, T2d, KP174138601 * T2a); | |
618 T2l = FMA(KP256247671, T2h, KP156891391 * T2k); | |
619 T2o = FMA(KP300238635, T2m, KP011599105 * T2n); | |
620 T2p = T2l - T2o; | |
621 T2u = T2e - T2p; | |
622 T2G = KP1_732050807 * (T2o + T2l); | |
623 { | |
624 E T21, T2r, T26, T27, T2s; | |
625 T21 = FNMS(KP132983124, T20, KP258260390 * T1X); | |
626 T2r = T22 - T21; | |
627 T26 = FMA(KP251768516, T24, KP075902986 * T25); | |
628 T27 = FNMS(KP083333333, To, T1); | |
629 T2s = T27 - T26; | |
630 T23 = FMA(KP2_000000000, T21, T22); | |
631 T2F = T2s - T2r; | |
632 T28 = FMA(KP2_000000000, T26, T27); | |
633 T2t = T2r + T2s; | |
634 } | |
635 { | |
636 E T29, T2q, T2N, T2O; | |
637 T29 = T23 + T28; | |
638 T2q = FMA(KP2_000000000, T2p, T2e); | |
639 ro[WS(os, 12)] = T29 - T2q; | |
640 ro[WS(os, 1)] = T29 + T2q; | |
641 { | |
642 E T2v, T2C, T2P, T2Q; | |
643 T2v = T2t - T2u; | |
644 T2C = T2y - T2B; | |
645 ro[WS(os, 10)] = T2v - T2C; | |
646 ro[WS(os, 4)] = T2v + T2C; | |
647 T2P = T28 - T23; | |
648 T2Q = FMS(KP2_000000000, T2K, T2J); | |
649 ro[WS(os, 5)] = T2P - T2Q; | |
650 ro[WS(os, 8)] = T2P + T2Q; | |
651 } | |
652 T2N = T2F - T2G; | |
653 T2O = T2L - T2I; | |
654 ro[WS(os, 11)] = T2N - T2O; | |
655 ro[WS(os, 6)] = T2N + T2O; | |
656 { | |
657 E T2H, T2M, T2D, T2E; | |
658 T2H = T2F + T2G; | |
659 T2M = T2I + T2L; | |
660 ro[WS(os, 7)] = T2H - T2M; | |
661 ro[WS(os, 2)] = T2H + T2M; | |
662 T2D = T2t + T2u; | |
663 T2E = T2y + T2B; | |
664 ro[WS(os, 3)] = T2D - T2E; | |
665 ro[WS(os, 9)] = T2D + T2E; | |
666 } | |
667 } | |
668 } | |
669 } | |
670 } | |
671 } | |
672 | |
673 static const kdft_desc desc = { 13, "n1_13", {138, 30, 38, 0}, &GENUS, 0, 0, 0, 0 }; | |
674 | |
675 void X(codelet_n1_13) (planner *p) { | |
676 X(kdft_register) (p, n1_13, &desc); | |
677 } | |
678 | |
679 #endif /* HAVE_FMA */ |