Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/scalar/codelets/n1_20.c @ 95:89f5e221ed7b
Add FFTW3
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Wed, 20 Mar 2013 15:35:50 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
94:d278df1123f9 | 95:89f5e221ed7b |
---|---|
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 07:35:46 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 20 -name n1_20 -include n.h */ | |
29 | |
30 /* | |
31 * This function contains 208 FP additions, 72 FP multiplications, | |
32 * (or, 136 additions, 0 multiplications, 72 fused multiply/add), | |
33 * 86 stack variables, 4 constants, and 80 memory accesses | |
34 */ | |
35 #include "n.h" | |
36 | |
37 static void n1_20(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
41 DK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
42 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
43 { | |
44 INT i; | |
45 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(80, is), MAKE_VOLATILE_STRIDE(80, os)) { | |
46 E T1Y, T1Z, T1W, T1V; | |
47 { | |
48 E T1d, TP, TD, T7, T3b, T2N, T2f, T1R, T2U, TB, T2P, T2A, T3d, T37, T3j; | |
49 E TJ, T2n, T1b, T1T, T1y, T2b, T2h, T1j, T2V, Tm, T2O, T2H, T3c, T34, T1e; | |
50 E T1f, T3i, TG, T2m, T10, T1S, T1J, T28, T2g; | |
51 { | |
52 E T4, T1N, T3, T2L, TN, T5, T1O, T1P, T1h, T1i; | |
53 { | |
54 E T1, T2, TL, TM; | |
55 T1 = ri[0]; | |
56 T2 = ri[WS(is, 10)]; | |
57 TL = ii[0]; | |
58 TM = ii[WS(is, 10)]; | |
59 T4 = ri[WS(is, 5)]; | |
60 T1N = T1 - T2; | |
61 T3 = T1 + T2; | |
62 T2L = TL + TM; | |
63 TN = TL - TM; | |
64 T5 = ri[WS(is, 15)]; | |
65 T1O = ii[WS(is, 5)]; | |
66 T1P = ii[WS(is, 15)]; | |
67 } | |
68 { | |
69 E T1o, Tp, T2u, T13, T14, Ts, T2v, T1r, Tx, T1t, Tw, T2x, T18, Ty, T1u; | |
70 E T1v; | |
71 { | |
72 E Tq, Tr, T1p, T1q; | |
73 { | |
74 E Tn, To, T11, T12; | |
75 Tn = ri[WS(is, 8)]; | |
76 { | |
77 E TO, T6, T2M, T1Q; | |
78 TO = T4 - T5; | |
79 T6 = T4 + T5; | |
80 T2M = T1O + T1P; | |
81 T1Q = T1O - T1P; | |
82 T1d = TO + TN; | |
83 TP = TN - TO; | |
84 TD = T3 + T6; | |
85 T7 = T3 - T6; | |
86 T3b = T2L + T2M; | |
87 T2N = T2L - T2M; | |
88 T2f = T1N + T1Q; | |
89 T1R = T1N - T1Q; | |
90 To = ri[WS(is, 18)]; | |
91 } | |
92 T11 = ii[WS(is, 8)]; | |
93 T12 = ii[WS(is, 18)]; | |
94 Tq = ri[WS(is, 13)]; | |
95 T1o = Tn - To; | |
96 Tp = Tn + To; | |
97 T2u = T11 + T12; | |
98 T13 = T11 - T12; | |
99 Tr = ri[WS(is, 3)]; | |
100 T1p = ii[WS(is, 13)]; | |
101 T1q = ii[WS(is, 3)]; | |
102 } | |
103 { | |
104 E Tu, Tv, T16, T17; | |
105 Tu = ri[WS(is, 12)]; | |
106 T14 = Tq - Tr; | |
107 Ts = Tq + Tr; | |
108 T2v = T1p + T1q; | |
109 T1r = T1p - T1q; | |
110 Tv = ri[WS(is, 2)]; | |
111 T16 = ii[WS(is, 12)]; | |
112 T17 = ii[WS(is, 2)]; | |
113 Tx = ri[WS(is, 17)]; | |
114 T1t = Tu - Tv; | |
115 Tw = Tu + Tv; | |
116 T2x = T16 + T17; | |
117 T18 = T16 - T17; | |
118 Ty = ri[WS(is, 7)]; | |
119 T1u = ii[WS(is, 17)]; | |
120 T1v = ii[WS(is, 7)]; | |
121 } | |
122 } | |
123 { | |
124 E TH, T19, T1w, TI; | |
125 { | |
126 E Tt, T2w, T35, TA, T2z, T36, Tz, T2y; | |
127 TH = Tp + Ts; | |
128 Tt = Tp - Ts; | |
129 T19 = Tx - Ty; | |
130 Tz = Tx + Ty; | |
131 T2y = T1u + T1v; | |
132 T1w = T1u - T1v; | |
133 T2w = T2u - T2v; | |
134 T35 = T2u + T2v; | |
135 TI = Tw + Tz; | |
136 TA = Tw - Tz; | |
137 T2z = T2x - T2y; | |
138 T36 = T2x + T2y; | |
139 T2U = Tt - TA; | |
140 TB = Tt + TA; | |
141 T2P = T2w + T2z; | |
142 T2A = T2w - T2z; | |
143 T3d = T35 + T36; | |
144 T37 = T35 - T36; | |
145 } | |
146 { | |
147 E T1s, T29, T1x, T2a, T15, T1a; | |
148 T15 = T13 - T14; | |
149 T1h = T14 + T13; | |
150 T1i = T19 + T18; | |
151 T1a = T18 - T19; | |
152 T1s = T1o - T1r; | |
153 T29 = T1o + T1r; | |
154 T3j = TH - TI; | |
155 TJ = TH + TI; | |
156 T1x = T1t - T1w; | |
157 T2a = T1t + T1w; | |
158 T2n = T15 - T1a; | |
159 T1b = T15 + T1a; | |
160 T1T = T1s + T1x; | |
161 T1y = T1s - T1x; | |
162 T2b = T29 - T2a; | |
163 T2h = T29 + T2a; | |
164 } | |
165 } | |
166 } | |
167 { | |
168 E Ta, T1z, T2B, TS, TT, Td, T2C, T1C, Ti, T1E, Th, T2E, TX, Tj, T1F; | |
169 E T1G; | |
170 { | |
171 E Tb, Tc, T1A, T1B; | |
172 { | |
173 E TQ, TR, T8, T9; | |
174 T8 = ri[WS(is, 4)]; | |
175 T9 = ri[WS(is, 14)]; | |
176 T1j = T1h + T1i; | |
177 T1Y = T1h - T1i; | |
178 TQ = ii[WS(is, 4)]; | |
179 TR = ii[WS(is, 14)]; | |
180 Ta = T8 + T9; | |
181 T1z = T8 - T9; | |
182 Tb = ri[WS(is, 9)]; | |
183 T2B = TQ + TR; | |
184 TS = TQ - TR; | |
185 Tc = ri[WS(is, 19)]; | |
186 T1A = ii[WS(is, 9)]; | |
187 T1B = ii[WS(is, 19)]; | |
188 } | |
189 { | |
190 E Tf, Tg, TV, TW; | |
191 Tf = ri[WS(is, 16)]; | |
192 TT = Tb - Tc; | |
193 Td = Tb + Tc; | |
194 T2C = T1A + T1B; | |
195 T1C = T1A - T1B; | |
196 Tg = ri[WS(is, 6)]; | |
197 TV = ii[WS(is, 16)]; | |
198 TW = ii[WS(is, 6)]; | |
199 Ti = ri[WS(is, 1)]; | |
200 T1E = Tf - Tg; | |
201 Th = Tf + Tg; | |
202 T2E = TV + TW; | |
203 TX = TV - TW; | |
204 Tj = ri[WS(is, 11)]; | |
205 T1F = ii[WS(is, 1)]; | |
206 T1G = ii[WS(is, 11)]; | |
207 } | |
208 } | |
209 { | |
210 E TE, TY, T1H, TF; | |
211 { | |
212 E Te, T2D, T32, Tl, T2G, T33, Tk, T2F; | |
213 TE = Ta + Td; | |
214 Te = Ta - Td; | |
215 TY = Ti - Tj; | |
216 Tk = Ti + Tj; | |
217 T2F = T1F + T1G; | |
218 T1H = T1F - T1G; | |
219 T2D = T2B - T2C; | |
220 T32 = T2B + T2C; | |
221 TF = Th + Tk; | |
222 Tl = Th - Tk; | |
223 T2G = T2E - T2F; | |
224 T33 = T2E + T2F; | |
225 T2V = Te - Tl; | |
226 Tm = Te + Tl; | |
227 T2O = T2D + T2G; | |
228 T2H = T2D - T2G; | |
229 T3c = T32 + T33; | |
230 T34 = T32 - T33; | |
231 } | |
232 { | |
233 E T1D, T26, T1I, T27, TU, TZ; | |
234 TU = TS - TT; | |
235 T1e = TT + TS; | |
236 T1f = TY + TX; | |
237 TZ = TX - TY; | |
238 T1D = T1z - T1C; | |
239 T26 = T1z + T1C; | |
240 T3i = TE - TF; | |
241 TG = TE + TF; | |
242 T1I = T1E - T1H; | |
243 T27 = T1E + T1H; | |
244 T2m = TU - TZ; | |
245 T10 = TU + TZ; | |
246 T1S = T1D + T1I; | |
247 T1J = T1D - T1I; | |
248 T28 = T26 - T27; | |
249 T2g = T26 + T27; | |
250 } | |
251 } | |
252 } | |
253 } | |
254 { | |
255 E T1g, T3g, T3f, T2S, T2R, T2k, T2j; | |
256 { | |
257 E T2s, T2r, TC, T2Q; | |
258 T2s = Tm - TB; | |
259 TC = Tm + TB; | |
260 T1g = T1e + T1f; | |
261 T1Z = T1e - T1f; | |
262 T2r = FNMS(KP250000000, TC, T7); | |
263 ro[WS(os, 10)] = T7 + TC; | |
264 T2Q = T2O + T2P; | |
265 T2S = T2O - T2P; | |
266 { | |
267 E T2K, T2I, T2t, T2J; | |
268 T2K = FMA(KP618033988, T2A, T2H); | |
269 T2I = FNMS(KP618033988, T2H, T2A); | |
270 T2t = FNMS(KP559016994, T2s, T2r); | |
271 T2J = FMA(KP559016994, T2s, T2r); | |
272 ro[WS(os, 18)] = FMA(KP951056516, T2I, T2t); | |
273 ro[WS(os, 2)] = FNMS(KP951056516, T2I, T2t); | |
274 ro[WS(os, 6)] = FMA(KP951056516, T2K, T2J); | |
275 ro[WS(os, 14)] = FNMS(KP951056516, T2K, T2J); | |
276 T2R = FNMS(KP250000000, T2Q, T2N); | |
277 } | |
278 io[WS(os, 10)] = T2N + T2Q; | |
279 } | |
280 { | |
281 E T30, T2Z, TK, T3e; | |
282 TK = TG + TJ; | |
283 T30 = TG - TJ; | |
284 { | |
285 E T2T, T2X, T2Y, T2W; | |
286 T2T = FNMS(KP559016994, T2S, T2R); | |
287 T2X = FMA(KP559016994, T2S, T2R); | |
288 T2Y = FMA(KP618033988, T2U, T2V); | |
289 T2W = FNMS(KP618033988, T2V, T2U); | |
290 io[WS(os, 14)] = FMA(KP951056516, T2Y, T2X); | |
291 io[WS(os, 6)] = FNMS(KP951056516, T2Y, T2X); | |
292 io[WS(os, 18)] = FNMS(KP951056516, T2W, T2T); | |
293 io[WS(os, 2)] = FMA(KP951056516, T2W, T2T); | |
294 T2Z = FNMS(KP250000000, TK, TD); | |
295 } | |
296 ro[0] = TD + TK; | |
297 T3e = T3c + T3d; | |
298 T3g = T3c - T3d; | |
299 { | |
300 E T31, T39, T3a, T38; | |
301 T31 = FMA(KP559016994, T30, T2Z); | |
302 T39 = FNMS(KP559016994, T30, T2Z); | |
303 T3a = FNMS(KP618033988, T34, T37); | |
304 T38 = FMA(KP618033988, T37, T34); | |
305 ro[WS(os, 8)] = FMA(KP951056516, T3a, T39); | |
306 ro[WS(os, 12)] = FNMS(KP951056516, T3a, T39); | |
307 ro[WS(os, 16)] = FMA(KP951056516, T38, T31); | |
308 ro[WS(os, 4)] = FNMS(KP951056516, T38, T31); | |
309 T3f = FNMS(KP250000000, T3e, T3b); | |
310 } | |
311 io[0] = T3b + T3e; | |
312 } | |
313 { | |
314 E T24, T23, T1c, T2i; | |
315 T1c = T10 + T1b; | |
316 T24 = T10 - T1b; | |
317 { | |
318 E T3h, T3l, T3m, T3k; | |
319 T3h = FMA(KP559016994, T3g, T3f); | |
320 T3l = FNMS(KP559016994, T3g, T3f); | |
321 T3m = FNMS(KP618033988, T3i, T3j); | |
322 T3k = FMA(KP618033988, T3j, T3i); | |
323 io[WS(os, 12)] = FMA(KP951056516, T3m, T3l); | |
324 io[WS(os, 8)] = FNMS(KP951056516, T3m, T3l); | |
325 io[WS(os, 16)] = FNMS(KP951056516, T3k, T3h); | |
326 io[WS(os, 4)] = FMA(KP951056516, T3k, T3h); | |
327 T23 = FNMS(KP250000000, T1c, TP); | |
328 } | |
329 io[WS(os, 5)] = TP + T1c; | |
330 T2i = T2g + T2h; | |
331 T2k = T2g - T2h; | |
332 { | |
333 E T25, T2d, T2e, T2c; | |
334 T25 = FMA(KP559016994, T24, T23); | |
335 T2d = FNMS(KP559016994, T24, T23); | |
336 T2e = FNMS(KP618033988, T28, T2b); | |
337 T2c = FMA(KP618033988, T2b, T28); | |
338 io[WS(os, 17)] = FMA(KP951056516, T2e, T2d); | |
339 io[WS(os, 13)] = FNMS(KP951056516, T2e, T2d); | |
340 io[WS(os, 9)] = FMA(KP951056516, T2c, T25); | |
341 io[WS(os, 1)] = FNMS(KP951056516, T2c, T25); | |
342 T2j = FNMS(KP250000000, T2i, T2f); | |
343 } | |
344 ro[WS(os, 5)] = T2f + T2i; | |
345 } | |
346 { | |
347 E T1m, T1l, T1k, T1U; | |
348 T1k = T1g + T1j; | |
349 T1m = T1g - T1j; | |
350 { | |
351 E T2l, T2p, T2q, T2o; | |
352 T2l = FMA(KP559016994, T2k, T2j); | |
353 T2p = FNMS(KP559016994, T2k, T2j); | |
354 T2q = FNMS(KP618033988, T2m, T2n); | |
355 T2o = FMA(KP618033988, T2n, T2m); | |
356 ro[WS(os, 17)] = FNMS(KP951056516, T2q, T2p); | |
357 ro[WS(os, 13)] = FMA(KP951056516, T2q, T2p); | |
358 ro[WS(os, 9)] = FNMS(KP951056516, T2o, T2l); | |
359 ro[WS(os, 1)] = FMA(KP951056516, T2o, T2l); | |
360 T1l = FNMS(KP250000000, T1k, T1d); | |
361 } | |
362 io[WS(os, 15)] = T1d + T1k; | |
363 T1U = T1S + T1T; | |
364 T1W = T1S - T1T; | |
365 { | |
366 E T1n, T1L, T1M, T1K; | |
367 T1n = FNMS(KP559016994, T1m, T1l); | |
368 T1L = FMA(KP559016994, T1m, T1l); | |
369 T1M = FMA(KP618033988, T1y, T1J); | |
370 T1K = FNMS(KP618033988, T1J, T1y); | |
371 io[WS(os, 19)] = FMA(KP951056516, T1M, T1L); | |
372 io[WS(os, 11)] = FNMS(KP951056516, T1M, T1L); | |
373 io[WS(os, 7)] = FMA(KP951056516, T1K, T1n); | |
374 io[WS(os, 3)] = FNMS(KP951056516, T1K, T1n); | |
375 T1V = FNMS(KP250000000, T1U, T1R); | |
376 } | |
377 ro[WS(os, 15)] = T1R + T1U; | |
378 } | |
379 } | |
380 } | |
381 { | |
382 E T21, T1X, T20, T22; | |
383 T21 = FMA(KP559016994, T1W, T1V); | |
384 T1X = FNMS(KP559016994, T1W, T1V); | |
385 T20 = FNMS(KP618033988, T1Z, T1Y); | |
386 T22 = FMA(KP618033988, T1Y, T1Z); | |
387 ro[WS(os, 19)] = FNMS(KP951056516, T22, T21); | |
388 ro[WS(os, 11)] = FMA(KP951056516, T22, T21); | |
389 ro[WS(os, 7)] = FNMS(KP951056516, T20, T1X); | |
390 ro[WS(os, 3)] = FMA(KP951056516, T20, T1X); | |
391 } | |
392 } | |
393 } | |
394 } | |
395 | |
396 static const kdft_desc desc = { 20, "n1_20", {136, 0, 72, 0}, &GENUS, 0, 0, 0, 0 }; | |
397 | |
398 void X(codelet_n1_20) (planner *p) { | |
399 X(kdft_register) (p, n1_20, &desc); | |
400 } | |
401 | |
402 #else /* HAVE_FMA */ | |
403 | |
404 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 20 -name n1_20 -include n.h */ | |
405 | |
406 /* | |
407 * This function contains 208 FP additions, 48 FP multiplications, | |
408 * (or, 184 additions, 24 multiplications, 24 fused multiply/add), | |
409 * 81 stack variables, 4 constants, and 80 memory accesses | |
410 */ | |
411 #include "n.h" | |
412 | |
413 static void n1_20(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
414 { | |
415 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
416 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
417 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
418 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
419 { | |
420 INT i; | |
421 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(80, is), MAKE_VOLATILE_STRIDE(80, os)) { | |
422 E T7, T2Q, T3h, TD, TP, T1U, T2l, T1d, Tt, TA, TB, T2w, T2z, T2S, T35; | |
423 E T36, T3f, TH, TI, TJ, T15, T1a, T1b, T1s, T1x, T1W, T29, T2a, T2j, T1h; | |
424 E T1i, T1j, Te, Tl, Tm, T2D, T2G, T2R, T32, T33, T3e, TE, TF, TG, TU; | |
425 E TZ, T10, T1D, T1I, T1V, T26, T27, T2i, T1e, T1f, T1g; | |
426 { | |
427 E T3, T1Q, TN, T2O, T6, TO, T1T, T2P; | |
428 { | |
429 E T1, T2, TL, TM; | |
430 T1 = ri[0]; | |
431 T2 = ri[WS(is, 10)]; | |
432 T3 = T1 + T2; | |
433 T1Q = T1 - T2; | |
434 TL = ii[0]; | |
435 TM = ii[WS(is, 10)]; | |
436 TN = TL - TM; | |
437 T2O = TL + TM; | |
438 } | |
439 { | |
440 E T4, T5, T1R, T1S; | |
441 T4 = ri[WS(is, 5)]; | |
442 T5 = ri[WS(is, 15)]; | |
443 T6 = T4 + T5; | |
444 TO = T4 - T5; | |
445 T1R = ii[WS(is, 5)]; | |
446 T1S = ii[WS(is, 15)]; | |
447 T1T = T1R - T1S; | |
448 T2P = T1R + T1S; | |
449 } | |
450 T7 = T3 - T6; | |
451 T2Q = T2O - T2P; | |
452 T3h = T2O + T2P; | |
453 TD = T3 + T6; | |
454 TP = TN - TO; | |
455 T1U = T1Q - T1T; | |
456 T2l = T1Q + T1T; | |
457 T1d = TO + TN; | |
458 } | |
459 { | |
460 E Tp, T1o, T13, T2u, Ts, T14, T1r, T2v, Tw, T1t, T18, T2x, Tz, T19, T1w; | |
461 E T2y; | |
462 { | |
463 E Tn, To, T11, T12; | |
464 Tn = ri[WS(is, 8)]; | |
465 To = ri[WS(is, 18)]; | |
466 Tp = Tn + To; | |
467 T1o = Tn - To; | |
468 T11 = ii[WS(is, 8)]; | |
469 T12 = ii[WS(is, 18)]; | |
470 T13 = T11 - T12; | |
471 T2u = T11 + T12; | |
472 } | |
473 { | |
474 E Tq, Tr, T1p, T1q; | |
475 Tq = ri[WS(is, 13)]; | |
476 Tr = ri[WS(is, 3)]; | |
477 Ts = Tq + Tr; | |
478 T14 = Tq - Tr; | |
479 T1p = ii[WS(is, 13)]; | |
480 T1q = ii[WS(is, 3)]; | |
481 T1r = T1p - T1q; | |
482 T2v = T1p + T1q; | |
483 } | |
484 { | |
485 E Tu, Tv, T16, T17; | |
486 Tu = ri[WS(is, 12)]; | |
487 Tv = ri[WS(is, 2)]; | |
488 Tw = Tu + Tv; | |
489 T1t = Tu - Tv; | |
490 T16 = ii[WS(is, 12)]; | |
491 T17 = ii[WS(is, 2)]; | |
492 T18 = T16 - T17; | |
493 T2x = T16 + T17; | |
494 } | |
495 { | |
496 E Tx, Ty, T1u, T1v; | |
497 Tx = ri[WS(is, 17)]; | |
498 Ty = ri[WS(is, 7)]; | |
499 Tz = Tx + Ty; | |
500 T19 = Tx - Ty; | |
501 T1u = ii[WS(is, 17)]; | |
502 T1v = ii[WS(is, 7)]; | |
503 T1w = T1u - T1v; | |
504 T2y = T1u + T1v; | |
505 } | |
506 Tt = Tp - Ts; | |
507 TA = Tw - Tz; | |
508 TB = Tt + TA; | |
509 T2w = T2u - T2v; | |
510 T2z = T2x - T2y; | |
511 T2S = T2w + T2z; | |
512 T35 = T2u + T2v; | |
513 T36 = T2x + T2y; | |
514 T3f = T35 + T36; | |
515 TH = Tp + Ts; | |
516 TI = Tw + Tz; | |
517 TJ = TH + TI; | |
518 T15 = T13 - T14; | |
519 T1a = T18 - T19; | |
520 T1b = T15 + T1a; | |
521 T1s = T1o - T1r; | |
522 T1x = T1t - T1w; | |
523 T1W = T1s + T1x; | |
524 T29 = T1o + T1r; | |
525 T2a = T1t + T1w; | |
526 T2j = T29 + T2a; | |
527 T1h = T14 + T13; | |
528 T1i = T19 + T18; | |
529 T1j = T1h + T1i; | |
530 } | |
531 { | |
532 E Ta, T1z, TS, T2B, Td, TT, T1C, T2C, Th, T1E, TX, T2E, Tk, TY, T1H; | |
533 E T2F; | |
534 { | |
535 E T8, T9, TQ, TR; | |
536 T8 = ri[WS(is, 4)]; | |
537 T9 = ri[WS(is, 14)]; | |
538 Ta = T8 + T9; | |
539 T1z = T8 - T9; | |
540 TQ = ii[WS(is, 4)]; | |
541 TR = ii[WS(is, 14)]; | |
542 TS = TQ - TR; | |
543 T2B = TQ + TR; | |
544 } | |
545 { | |
546 E Tb, Tc, T1A, T1B; | |
547 Tb = ri[WS(is, 9)]; | |
548 Tc = ri[WS(is, 19)]; | |
549 Td = Tb + Tc; | |
550 TT = Tb - Tc; | |
551 T1A = ii[WS(is, 9)]; | |
552 T1B = ii[WS(is, 19)]; | |
553 T1C = T1A - T1B; | |
554 T2C = T1A + T1B; | |
555 } | |
556 { | |
557 E Tf, Tg, TV, TW; | |
558 Tf = ri[WS(is, 16)]; | |
559 Tg = ri[WS(is, 6)]; | |
560 Th = Tf + Tg; | |
561 T1E = Tf - Tg; | |
562 TV = ii[WS(is, 16)]; | |
563 TW = ii[WS(is, 6)]; | |
564 TX = TV - TW; | |
565 T2E = TV + TW; | |
566 } | |
567 { | |
568 E Ti, Tj, T1F, T1G; | |
569 Ti = ri[WS(is, 1)]; | |
570 Tj = ri[WS(is, 11)]; | |
571 Tk = Ti + Tj; | |
572 TY = Ti - Tj; | |
573 T1F = ii[WS(is, 1)]; | |
574 T1G = ii[WS(is, 11)]; | |
575 T1H = T1F - T1G; | |
576 T2F = T1F + T1G; | |
577 } | |
578 Te = Ta - Td; | |
579 Tl = Th - Tk; | |
580 Tm = Te + Tl; | |
581 T2D = T2B - T2C; | |
582 T2G = T2E - T2F; | |
583 T2R = T2D + T2G; | |
584 T32 = T2B + T2C; | |
585 T33 = T2E + T2F; | |
586 T3e = T32 + T33; | |
587 TE = Ta + Td; | |
588 TF = Th + Tk; | |
589 TG = TE + TF; | |
590 TU = TS - TT; | |
591 TZ = TX - TY; | |
592 T10 = TU + TZ; | |
593 T1D = T1z - T1C; | |
594 T1I = T1E - T1H; | |
595 T1V = T1D + T1I; | |
596 T26 = T1z + T1C; | |
597 T27 = T1E + T1H; | |
598 T2i = T26 + T27; | |
599 T1e = TT + TS; | |
600 T1f = TY + TX; | |
601 T1g = T1e + T1f; | |
602 } | |
603 { | |
604 E T2s, TC, T2r, T2I, T2K, T2A, T2H, T2J, T2t; | |
605 T2s = KP559016994 * (Tm - TB); | |
606 TC = Tm + TB; | |
607 T2r = FNMS(KP250000000, TC, T7); | |
608 T2A = T2w - T2z; | |
609 T2H = T2D - T2G; | |
610 T2I = FNMS(KP587785252, T2H, KP951056516 * T2A); | |
611 T2K = FMA(KP951056516, T2H, KP587785252 * T2A); | |
612 ro[WS(os, 10)] = T7 + TC; | |
613 T2J = T2s + T2r; | |
614 ro[WS(os, 14)] = T2J - T2K; | |
615 ro[WS(os, 6)] = T2J + T2K; | |
616 T2t = T2r - T2s; | |
617 ro[WS(os, 2)] = T2t - T2I; | |
618 ro[WS(os, 18)] = T2t + T2I; | |
619 } | |
620 { | |
621 E T2V, T2T, T2U, T2N, T2Y, T2L, T2M, T2X, T2W; | |
622 T2V = KP559016994 * (T2R - T2S); | |
623 T2T = T2R + T2S; | |
624 T2U = FNMS(KP250000000, T2T, T2Q); | |
625 T2L = Tt - TA; | |
626 T2M = Te - Tl; | |
627 T2N = FNMS(KP587785252, T2M, KP951056516 * T2L); | |
628 T2Y = FMA(KP951056516, T2M, KP587785252 * T2L); | |
629 io[WS(os, 10)] = T2Q + T2T; | |
630 T2X = T2V + T2U; | |
631 io[WS(os, 6)] = T2X - T2Y; | |
632 io[WS(os, 14)] = T2Y + T2X; | |
633 T2W = T2U - T2V; | |
634 io[WS(os, 2)] = T2N + T2W; | |
635 io[WS(os, 18)] = T2W - T2N; | |
636 } | |
637 { | |
638 E T2Z, TK, T30, T38, T3a, T34, T37, T39, T31; | |
639 T2Z = KP559016994 * (TG - TJ); | |
640 TK = TG + TJ; | |
641 T30 = FNMS(KP250000000, TK, TD); | |
642 T34 = T32 - T33; | |
643 T37 = T35 - T36; | |
644 T38 = FMA(KP951056516, T34, KP587785252 * T37); | |
645 T3a = FNMS(KP587785252, T34, KP951056516 * T37); | |
646 ro[0] = TD + TK; | |
647 T39 = T30 - T2Z; | |
648 ro[WS(os, 12)] = T39 - T3a; | |
649 ro[WS(os, 8)] = T39 + T3a; | |
650 T31 = T2Z + T30; | |
651 ro[WS(os, 4)] = T31 - T38; | |
652 ro[WS(os, 16)] = T31 + T38; | |
653 } | |
654 { | |
655 E T3g, T3i, T3j, T3d, T3m, T3b, T3c, T3l, T3k; | |
656 T3g = KP559016994 * (T3e - T3f); | |
657 T3i = T3e + T3f; | |
658 T3j = FNMS(KP250000000, T3i, T3h); | |
659 T3b = TE - TF; | |
660 T3c = TH - TI; | |
661 T3d = FMA(KP951056516, T3b, KP587785252 * T3c); | |
662 T3m = FNMS(KP587785252, T3b, KP951056516 * T3c); | |
663 io[0] = T3h + T3i; | |
664 T3l = T3j - T3g; | |
665 io[WS(os, 8)] = T3l - T3m; | |
666 io[WS(os, 12)] = T3m + T3l; | |
667 T3k = T3g + T3j; | |
668 io[WS(os, 4)] = T3d + T3k; | |
669 io[WS(os, 16)] = T3k - T3d; | |
670 } | |
671 { | |
672 E T23, T1c, T24, T2c, T2e, T28, T2b, T2d, T25; | |
673 T23 = KP559016994 * (T10 - T1b); | |
674 T1c = T10 + T1b; | |
675 T24 = FNMS(KP250000000, T1c, TP); | |
676 T28 = T26 - T27; | |
677 T2b = T29 - T2a; | |
678 T2c = FMA(KP951056516, T28, KP587785252 * T2b); | |
679 T2e = FNMS(KP587785252, T28, KP951056516 * T2b); | |
680 io[WS(os, 5)] = TP + T1c; | |
681 T2d = T24 - T23; | |
682 io[WS(os, 13)] = T2d - T2e; | |
683 io[WS(os, 17)] = T2d + T2e; | |
684 T25 = T23 + T24; | |
685 io[WS(os, 1)] = T25 - T2c; | |
686 io[WS(os, 9)] = T25 + T2c; | |
687 } | |
688 { | |
689 E T2k, T2m, T2n, T2h, T2p, T2f, T2g, T2q, T2o; | |
690 T2k = KP559016994 * (T2i - T2j); | |
691 T2m = T2i + T2j; | |
692 T2n = FNMS(KP250000000, T2m, T2l); | |
693 T2f = TU - TZ; | |
694 T2g = T15 - T1a; | |
695 T2h = FMA(KP951056516, T2f, KP587785252 * T2g); | |
696 T2p = FNMS(KP587785252, T2f, KP951056516 * T2g); | |
697 ro[WS(os, 5)] = T2l + T2m; | |
698 T2q = T2n - T2k; | |
699 ro[WS(os, 13)] = T2p + T2q; | |
700 ro[WS(os, 17)] = T2q - T2p; | |
701 T2o = T2k + T2n; | |
702 ro[WS(os, 1)] = T2h + T2o; | |
703 ro[WS(os, 9)] = T2o - T2h; | |
704 } | |
705 { | |
706 E T1m, T1k, T1l, T1K, T1M, T1y, T1J, T1L, T1n; | |
707 T1m = KP559016994 * (T1g - T1j); | |
708 T1k = T1g + T1j; | |
709 T1l = FNMS(KP250000000, T1k, T1d); | |
710 T1y = T1s - T1x; | |
711 T1J = T1D - T1I; | |
712 T1K = FNMS(KP587785252, T1J, KP951056516 * T1y); | |
713 T1M = FMA(KP951056516, T1J, KP587785252 * T1y); | |
714 io[WS(os, 15)] = T1d + T1k; | |
715 T1L = T1m + T1l; | |
716 io[WS(os, 11)] = T1L - T1M; | |
717 io[WS(os, 19)] = T1L + T1M; | |
718 T1n = T1l - T1m; | |
719 io[WS(os, 3)] = T1n - T1K; | |
720 io[WS(os, 7)] = T1n + T1K; | |
721 } | |
722 { | |
723 E T1Z, T1X, T1Y, T1P, T21, T1N, T1O, T22, T20; | |
724 T1Z = KP559016994 * (T1V - T1W); | |
725 T1X = T1V + T1W; | |
726 T1Y = FNMS(KP250000000, T1X, T1U); | |
727 T1N = T1h - T1i; | |
728 T1O = T1e - T1f; | |
729 T1P = FNMS(KP587785252, T1O, KP951056516 * T1N); | |
730 T21 = FMA(KP951056516, T1O, KP587785252 * T1N); | |
731 ro[WS(os, 15)] = T1U + T1X; | |
732 T22 = T1Z + T1Y; | |
733 ro[WS(os, 11)] = T21 + T22; | |
734 ro[WS(os, 19)] = T22 - T21; | |
735 T20 = T1Y - T1Z; | |
736 ro[WS(os, 3)] = T1P + T20; | |
737 ro[WS(os, 7)] = T20 - T1P; | |
738 } | |
739 } | |
740 } | |
741 } | |
742 | |
743 static const kdft_desc desc = { 20, "n1_20", {184, 24, 24, 0}, &GENUS, 0, 0, 0, 0 }; | |
744 | |
745 void X(codelet_n1_20) (planner *p) { | |
746 X(kdft_register) (p, n1_20, &desc); | |
747 } | |
748 | |
749 #endif /* HAVE_FMA */ |