Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/scalar/codelets/t1_15.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 07:35:50 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 15 -name t1_15 -include t.h */ | |
29 | |
30 /* | |
31 * This function contains 184 FP additions, 140 FP multiplications, | |
32 * (or, 72 additions, 28 multiplications, 112 fused multiply/add), | |
33 * 89 stack variables, 6 constants, and 60 memory accesses | |
34 */ | |
35 #include "t.h" | |
36 | |
37 static void t1_15(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
41 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
42 DK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
43 DK(KP866025403, +0.866025403784438646763723170752936183471402627); | |
44 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
45 { | |
46 INT m; | |
47 for (m = mb, W = W + (mb * 28); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) { | |
48 E T2d, T2O, T2Q, T2m, T2k, T2l, T2P, T2n; | |
49 { | |
50 E T1G, T3u, T3k, T3t, T1B, Tf, T37, T1y, T2V, T2M, T2a, T2i, T39, Tz, T2X; | |
51 E T2t, T1O, T2e, T3a, TT, T10, T2Y, T2z, T1V, T2f, T2C, T12, T15, T14, T21; | |
52 E T1c, T1Y, T13; | |
53 { | |
54 E T2I, T1k, T1m, T1p, T1o, T28, T1w, T25, T1n; | |
55 { | |
56 E T1, T3j, T9, Tc, Tb, T1D, T7, T1E, Ta, T1j, T1i, T1h; | |
57 T1 = ri[0]; | |
58 T3j = ii[0]; | |
59 { | |
60 E T3, T6, T2, T5, T1C, T4, T8; | |
61 T3 = ri[WS(rs, 5)]; | |
62 T6 = ii[WS(rs, 5)]; | |
63 T2 = W[8]; | |
64 T5 = W[9]; | |
65 T9 = ri[WS(rs, 10)]; | |
66 Tc = ii[WS(rs, 10)]; | |
67 T1C = T2 * T6; | |
68 T4 = T2 * T3; | |
69 T8 = W[18]; | |
70 Tb = W[19]; | |
71 T1D = FNMS(T5, T3, T1C); | |
72 T7 = FMA(T5, T6, T4); | |
73 T1E = T8 * Tc; | |
74 Ta = T8 * T9; | |
75 } | |
76 { | |
77 E T1g, T1F, Td, T1f, T3i, Te, T2H; | |
78 T1g = ri[WS(rs, 9)]; | |
79 T1j = ii[WS(rs, 9)]; | |
80 T1F = FNMS(Tb, T9, T1E); | |
81 Td = FMA(Tb, Tc, Ta); | |
82 T1f = W[16]; | |
83 T1i = W[17]; | |
84 T1G = T1D - T1F; | |
85 T3i = T1D + T1F; | |
86 T3u = Td - T7; | |
87 Te = T7 + Td; | |
88 T2H = T1f * T1j; | |
89 T1h = T1f * T1g; | |
90 T3k = T3i + T3j; | |
91 T3t = FNMS(KP500000000, T3i, T3j); | |
92 T1B = FNMS(KP500000000, Te, T1); | |
93 Tf = T1 + Te; | |
94 T2I = FNMS(T1i, T1g, T2H); | |
95 } | |
96 T1k = FMA(T1i, T1j, T1h); | |
97 { | |
98 E T1s, T1v, T1r, T1u, T27, T1t, T1l; | |
99 T1s = ri[WS(rs, 4)]; | |
100 T1v = ii[WS(rs, 4)]; | |
101 T1r = W[6]; | |
102 T1u = W[7]; | |
103 T1m = ri[WS(rs, 14)]; | |
104 T1p = ii[WS(rs, 14)]; | |
105 T27 = T1r * T1v; | |
106 T1t = T1r * T1s; | |
107 T1l = W[26]; | |
108 T1o = W[27]; | |
109 T28 = FNMS(T1u, T1s, T27); | |
110 T1w = FMA(T1u, T1v, T1t); | |
111 T25 = T1l * T1p; | |
112 T1n = T1l * T1m; | |
113 } | |
114 } | |
115 { | |
116 E Tl, T2p, Tn, Tq, Tp, T1M, Tx, T1J, To; | |
117 { | |
118 E Th, Tk, T26, T1q, Tg, Tj; | |
119 Th = ri[WS(rs, 3)]; | |
120 Tk = ii[WS(rs, 3)]; | |
121 T26 = FNMS(T1o, T1m, T25); | |
122 T1q = FMA(T1o, T1p, T1n); | |
123 Tg = W[4]; | |
124 Tj = W[5]; | |
125 { | |
126 E T29, T2J, T1x, T2L; | |
127 T29 = T26 - T28; | |
128 T2J = T26 + T28; | |
129 T1x = T1q + T1w; | |
130 T2L = T1w - T1q; | |
131 { | |
132 E T2o, Ti, T2K, T24; | |
133 T2o = Tg * Tk; | |
134 Ti = Tg * Th; | |
135 T2K = FNMS(KP500000000, T2J, T2I); | |
136 T37 = T2I + T2J; | |
137 T24 = FNMS(KP500000000, T1x, T1k); | |
138 T1y = T1k + T1x; | |
139 Tl = FMA(Tj, Tk, Ti); | |
140 T2V = FNMS(KP866025403, T2L, T2K); | |
141 T2M = FMA(KP866025403, T2L, T2K); | |
142 T2a = FNMS(KP866025403, T29, T24); | |
143 T2i = FMA(KP866025403, T29, T24); | |
144 T2p = FNMS(Tj, Th, T2o); | |
145 } | |
146 } | |
147 } | |
148 { | |
149 E Tt, Tw, Ts, Tv, T1L, Tu, Tm; | |
150 Tt = ri[WS(rs, 13)]; | |
151 Tw = ii[WS(rs, 13)]; | |
152 Ts = W[24]; | |
153 Tv = W[25]; | |
154 Tn = ri[WS(rs, 8)]; | |
155 Tq = ii[WS(rs, 8)]; | |
156 T1L = Ts * Tw; | |
157 Tu = Ts * Tt; | |
158 Tm = W[14]; | |
159 Tp = W[15]; | |
160 T1M = FNMS(Tv, Tt, T1L); | |
161 Tx = FMA(Tv, Tw, Tu); | |
162 T1J = Tm * Tq; | |
163 To = Tm * Tn; | |
164 } | |
165 { | |
166 E TF, T2v, TH, TK, TJ, T1T, TR, T1Q, TI; | |
167 { | |
168 E TB, TE, T1K, Tr, TA, TD; | |
169 TB = ri[WS(rs, 12)]; | |
170 TE = ii[WS(rs, 12)]; | |
171 T1K = FNMS(Tp, Tn, T1J); | |
172 Tr = FMA(Tp, Tq, To); | |
173 TA = W[22]; | |
174 TD = W[23]; | |
175 { | |
176 E T1N, T2q, Ty, T2s; | |
177 T1N = T1K - T1M; | |
178 T2q = T1K + T1M; | |
179 Ty = Tr + Tx; | |
180 T2s = Tx - Tr; | |
181 { | |
182 E T2u, TC, T2r, T1I; | |
183 T2u = TA * TE; | |
184 TC = TA * TB; | |
185 T2r = FNMS(KP500000000, T2q, T2p); | |
186 T39 = T2p + T2q; | |
187 T1I = FNMS(KP500000000, Ty, Tl); | |
188 Tz = Tl + Ty; | |
189 TF = FMA(TD, TE, TC); | |
190 T2X = FNMS(KP866025403, T2s, T2r); | |
191 T2t = FMA(KP866025403, T2s, T2r); | |
192 T1O = FNMS(KP866025403, T1N, T1I); | |
193 T2e = FMA(KP866025403, T1N, T1I); | |
194 T2v = FNMS(TD, TB, T2u); | |
195 } | |
196 } | |
197 } | |
198 { | |
199 E TN, TQ, TM, TP, T1S, TO, TG; | |
200 TN = ri[WS(rs, 7)]; | |
201 TQ = ii[WS(rs, 7)]; | |
202 TM = W[12]; | |
203 TP = W[13]; | |
204 TH = ri[WS(rs, 2)]; | |
205 TK = ii[WS(rs, 2)]; | |
206 T1S = TM * TQ; | |
207 TO = TM * TN; | |
208 TG = W[2]; | |
209 TJ = W[3]; | |
210 T1T = FNMS(TP, TN, T1S); | |
211 TR = FMA(TP, TQ, TO); | |
212 T1Q = TG * TK; | |
213 TI = TG * TH; | |
214 } | |
215 { | |
216 E TW, TZ, T1R, TL, TV, TY; | |
217 TW = ri[WS(rs, 6)]; | |
218 TZ = ii[WS(rs, 6)]; | |
219 T1R = FNMS(TJ, TH, T1Q); | |
220 TL = FMA(TJ, TK, TI); | |
221 TV = W[10]; | |
222 TY = W[11]; | |
223 { | |
224 E T1U, T2w, TS, T2y; | |
225 T1U = T1R - T1T; | |
226 T2w = T1R + T1T; | |
227 TS = TL + TR; | |
228 T2y = TR - TL; | |
229 { | |
230 E T2B, TX, T2x, T1P; | |
231 T2B = TV * TZ; | |
232 TX = TV * TW; | |
233 T2x = FNMS(KP500000000, T2w, T2v); | |
234 T3a = T2v + T2w; | |
235 T1P = FNMS(KP500000000, TS, TF); | |
236 TT = TF + TS; | |
237 T10 = FMA(TY, TZ, TX); | |
238 T2Y = FNMS(KP866025403, T2y, T2x); | |
239 T2z = FMA(KP866025403, T2y, T2x); | |
240 T1V = FNMS(KP866025403, T1U, T1P); | |
241 T2f = FMA(KP866025403, T1U, T1P); | |
242 T2C = FNMS(TY, TW, T2B); | |
243 } | |
244 } | |
245 } | |
246 { | |
247 E T18, T1b, T17, T1a, T20, T19, T11; | |
248 T18 = ri[WS(rs, 1)]; | |
249 T1b = ii[WS(rs, 1)]; | |
250 T17 = W[0]; | |
251 T1a = W[1]; | |
252 T12 = ri[WS(rs, 11)]; | |
253 T15 = ii[WS(rs, 11)]; | |
254 T20 = T17 * T1b; | |
255 T19 = T17 * T18; | |
256 T11 = W[20]; | |
257 T14 = W[21]; | |
258 T21 = FNMS(T1a, T18, T20); | |
259 T1c = FMA(T1a, T1b, T19); | |
260 T1Y = T11 * T15; | |
261 T13 = T11 * T12; | |
262 } | |
263 } | |
264 } | |
265 } | |
266 { | |
267 E T2G, T2h, T3J, T3I, T32, T30, T1H, T1W, T3P, T3O, T2b; | |
268 { | |
269 E T3f, T3b, T1Z, T16, T3p, TU; | |
270 T3f = T39 + T3a; | |
271 T3b = T39 - T3a; | |
272 T1Z = FNMS(T14, T12, T1Y); | |
273 T16 = FMA(T14, T15, T13); | |
274 T3p = Tz - TT; | |
275 TU = Tz + TT; | |
276 { | |
277 E T3g, T2U, T23, T3c, T3e, T3q, T3s, T1A, T34, T3r, T3n; | |
278 { | |
279 E T22, T1d, T2F, T2E, T36, T2D; | |
280 T22 = T1Z - T21; | |
281 T2D = T1Z + T21; | |
282 T1d = T16 + T1c; | |
283 T2F = T1c - T16; | |
284 T2E = FNMS(KP500000000, T2D, T2C); | |
285 T36 = T2C + T2D; | |
286 { | |
287 E T1e, T1X, T38, T1z, T3o; | |
288 T1e = T10 + T1d; | |
289 T1X = FNMS(KP500000000, T1d, T10); | |
290 T38 = T36 - T37; | |
291 T3g = T36 + T37; | |
292 T2G = FMA(KP866025403, T2F, T2E); | |
293 T2U = FNMS(KP866025403, T2F, T2E); | |
294 T1z = T1e + T1y; | |
295 T3o = T1e - T1y; | |
296 T2h = FMA(KP866025403, T22, T1X); | |
297 T23 = FNMS(KP866025403, T22, T1X); | |
298 T3c = FNMS(KP618033988, T3b, T38); | |
299 T3e = FMA(KP618033988, T38, T3b); | |
300 T3q = FNMS(KP618033988, T3p, T3o); | |
301 T3s = FMA(KP618033988, T3o, T3p); | |
302 T1A = TU + T1z; | |
303 T34 = TU - T1z; | |
304 } | |
305 } | |
306 { | |
307 E T2W, T33, T3m, T3h, T2Z, T3d, T35, T3l; | |
308 T3J = T2U + T2V; | |
309 T2W = T2U - T2V; | |
310 ri[0] = Tf + T1A; | |
311 T33 = FNMS(KP250000000, T1A, Tf); | |
312 T3m = T3f - T3g; | |
313 T3h = T3f + T3g; | |
314 T2Z = T2X - T2Y; | |
315 T3I = T2X + T2Y; | |
316 T3d = FMA(KP559016994, T34, T33); | |
317 T35 = FNMS(KP559016994, T34, T33); | |
318 ii[0] = T3h + T3k; | |
319 T3l = FNMS(KP250000000, T3h, T3k); | |
320 ri[WS(rs, 3)] = FMA(KP951056516, T3c, T35); | |
321 ri[WS(rs, 12)] = FNMS(KP951056516, T3c, T35); | |
322 ri[WS(rs, 6)] = FMA(KP951056516, T3e, T3d); | |
323 ri[WS(rs, 9)] = FNMS(KP951056516, T3e, T3d); | |
324 T3r = FMA(KP559016994, T3m, T3l); | |
325 T3n = FNMS(KP559016994, T3m, T3l); | |
326 T32 = FMA(KP618033988, T2W, T2Z); | |
327 T30 = FNMS(KP618033988, T2Z, T2W); | |
328 } | |
329 ii[WS(rs, 12)] = FMA(KP951056516, T3q, T3n); | |
330 ii[WS(rs, 3)] = FNMS(KP951056516, T3q, T3n); | |
331 ii[WS(rs, 9)] = FMA(KP951056516, T3s, T3r); | |
332 ii[WS(rs, 6)] = FNMS(KP951056516, T3s, T3r); | |
333 T2d = FMA(KP866025403, T1G, T1B); | |
334 T1H = FNMS(KP866025403, T1G, T1B); | |
335 T1W = T1O + T1V; | |
336 T3P = T1O - T1V; | |
337 T3O = T23 - T2a; | |
338 T2b = T23 + T2a; | |
339 } | |
340 } | |
341 { | |
342 E T3H, T3v, T2S, T3Q, T3S, T2R, T2c; | |
343 T3H = FNMS(KP866025403, T3u, T3t); | |
344 T3v = FMA(KP866025403, T3u, T3t); | |
345 T2c = T1W + T2b; | |
346 T2S = T1W - T2b; | |
347 T3Q = FNMS(KP618033988, T3P, T3O); | |
348 T3S = FMA(KP618033988, T3O, T3P); | |
349 ri[WS(rs, 5)] = T1H + T2c; | |
350 T2R = FNMS(KP250000000, T2c, T1H); | |
351 { | |
352 E T2g, T2j, T3G, T3E, T2A, T2N, T3y, T3A, T3M, T3L, T3z, T3F, T3B; | |
353 { | |
354 E T3C, T3D, T31, T2T, T3K; | |
355 T2g = T2e + T2f; | |
356 T3C = T2e - T2f; | |
357 T3D = T2h - T2i; | |
358 T2j = T2h + T2i; | |
359 T31 = FMA(KP559016994, T2S, T2R); | |
360 T2T = FNMS(KP559016994, T2S, T2R); | |
361 T3K = T3I + T3J; | |
362 T3M = T3I - T3J; | |
363 ri[WS(rs, 8)] = FMA(KP951056516, T30, T2T); | |
364 ri[WS(rs, 2)] = FNMS(KP951056516, T30, T2T); | |
365 ri[WS(rs, 11)] = FMA(KP951056516, T32, T31); | |
366 ri[WS(rs, 14)] = FNMS(KP951056516, T32, T31); | |
367 ii[WS(rs, 5)] = T3K + T3H; | |
368 T3L = FNMS(KP250000000, T3K, T3H); | |
369 T3G = FNMS(KP618033988, T3C, T3D); | |
370 T3E = FMA(KP618033988, T3D, T3C); | |
371 } | |
372 { | |
373 E T3N, T3R, T3w, T3x; | |
374 T3N = FNMS(KP559016994, T3M, T3L); | |
375 T3R = FMA(KP559016994, T3M, T3L); | |
376 T3w = T2t + T2z; | |
377 T2A = T2t - T2z; | |
378 T2N = T2G - T2M; | |
379 T3x = T2G + T2M; | |
380 ii[WS(rs, 8)] = FNMS(KP951056516, T3Q, T3N); | |
381 ii[WS(rs, 2)] = FMA(KP951056516, T3Q, T3N); | |
382 ii[WS(rs, 14)] = FMA(KP951056516, T3S, T3R); | |
383 ii[WS(rs, 11)] = FNMS(KP951056516, T3S, T3R); | |
384 T3y = T3w + T3x; | |
385 T3A = T3w - T3x; | |
386 } | |
387 ii[WS(rs, 10)] = T3y + T3v; | |
388 T3z = FNMS(KP250000000, T3y, T3v); | |
389 T2O = FMA(KP618033988, T2N, T2A); | |
390 T2Q = FNMS(KP618033988, T2A, T2N); | |
391 T3F = FNMS(KP559016994, T3A, T3z); | |
392 T3B = FMA(KP559016994, T3A, T3z); | |
393 ii[WS(rs, 4)] = FMA(KP951056516, T3E, T3B); | |
394 ii[WS(rs, 1)] = FNMS(KP951056516, T3E, T3B); | |
395 ii[WS(rs, 13)] = FNMS(KP951056516, T3G, T3F); | |
396 ii[WS(rs, 7)] = FMA(KP951056516, T3G, T3F); | |
397 T2m = T2g - T2j; | |
398 T2k = T2g + T2j; | |
399 } | |
400 } | |
401 } | |
402 } | |
403 ri[WS(rs, 10)] = T2d + T2k; | |
404 T2l = FNMS(KP250000000, T2k, T2d); | |
405 T2P = FNMS(KP559016994, T2m, T2l); | |
406 T2n = FMA(KP559016994, T2m, T2l); | |
407 ri[WS(rs, 1)] = FMA(KP951056516, T2O, T2n); | |
408 ri[WS(rs, 4)] = FNMS(KP951056516, T2O, T2n); | |
409 ri[WS(rs, 13)] = FMA(KP951056516, T2Q, T2P); | |
410 ri[WS(rs, 7)] = FNMS(KP951056516, T2Q, T2P); | |
411 } | |
412 } | |
413 } | |
414 | |
415 static const tw_instr twinstr[] = { | |
416 {TW_FULL, 0, 15}, | |
417 {TW_NEXT, 1, 0} | |
418 }; | |
419 | |
420 static const ct_desc desc = { 15, "t1_15", twinstr, &GENUS, {72, 28, 112, 0}, 0, 0, 0 }; | |
421 | |
422 void X(codelet_t1_15) (planner *p) { | |
423 X(kdft_dit_register) (p, t1_15, &desc); | |
424 } | |
425 #else /* HAVE_FMA */ | |
426 | |
427 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 15 -name t1_15 -include t.h */ | |
428 | |
429 /* | |
430 * This function contains 184 FP additions, 112 FP multiplications, | |
431 * (or, 128 additions, 56 multiplications, 56 fused multiply/add), | |
432 * 65 stack variables, 6 constants, and 60 memory accesses | |
433 */ | |
434 #include "t.h" | |
435 | |
436 static void t1_15(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
437 { | |
438 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
439 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
440 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
441 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
442 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
443 DK(KP866025403, +0.866025403784438646763723170752936183471402627); | |
444 { | |
445 INT m; | |
446 for (m = mb, W = W + (mb * 28); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) { | |
447 E T1q, T34, Td, T1n, T2S, T35, T13, T1k, T1l, T2E, T2F, T2O, T1H, T1T, T2k; | |
448 E T2t, T2f, T2s, T1M, T1U, Tu, TL, TM, T2H, T2I, T2N, T1w, T1Q, T29, T2w; | |
449 E T24, T2v, T1B, T1R; | |
450 { | |
451 E T1, T2R, T6, T1o, Tb, T1p, Tc, T2Q; | |
452 T1 = ri[0]; | |
453 T2R = ii[0]; | |
454 { | |
455 E T3, T5, T2, T4; | |
456 T3 = ri[WS(rs, 5)]; | |
457 T5 = ii[WS(rs, 5)]; | |
458 T2 = W[8]; | |
459 T4 = W[9]; | |
460 T6 = FMA(T2, T3, T4 * T5); | |
461 T1o = FNMS(T4, T3, T2 * T5); | |
462 } | |
463 { | |
464 E T8, Ta, T7, T9; | |
465 T8 = ri[WS(rs, 10)]; | |
466 Ta = ii[WS(rs, 10)]; | |
467 T7 = W[18]; | |
468 T9 = W[19]; | |
469 Tb = FMA(T7, T8, T9 * Ta); | |
470 T1p = FNMS(T9, T8, T7 * Ta); | |
471 } | |
472 T1q = KP866025403 * (T1o - T1p); | |
473 T34 = KP866025403 * (Tb - T6); | |
474 Tc = T6 + Tb; | |
475 Td = T1 + Tc; | |
476 T1n = FNMS(KP500000000, Tc, T1); | |
477 T2Q = T1o + T1p; | |
478 T2S = T2Q + T2R; | |
479 T35 = FNMS(KP500000000, T2Q, T2R); | |
480 } | |
481 { | |
482 E TR, T2c, T18, T2h, TW, T1E, T11, T1F, T12, T2d, T1d, T1J, T1i, T1K, T1j; | |
483 E T2i; | |
484 { | |
485 E TO, TQ, TN, TP; | |
486 TO = ri[WS(rs, 6)]; | |
487 TQ = ii[WS(rs, 6)]; | |
488 TN = W[10]; | |
489 TP = W[11]; | |
490 TR = FMA(TN, TO, TP * TQ); | |
491 T2c = FNMS(TP, TO, TN * TQ); | |
492 } | |
493 { | |
494 E T15, T17, T14, T16; | |
495 T15 = ri[WS(rs, 9)]; | |
496 T17 = ii[WS(rs, 9)]; | |
497 T14 = W[16]; | |
498 T16 = W[17]; | |
499 T18 = FMA(T14, T15, T16 * T17); | |
500 T2h = FNMS(T16, T15, T14 * T17); | |
501 } | |
502 { | |
503 E TT, TV, TS, TU; | |
504 TT = ri[WS(rs, 11)]; | |
505 TV = ii[WS(rs, 11)]; | |
506 TS = W[20]; | |
507 TU = W[21]; | |
508 TW = FMA(TS, TT, TU * TV); | |
509 T1E = FNMS(TU, TT, TS * TV); | |
510 } | |
511 { | |
512 E TY, T10, TX, TZ; | |
513 TY = ri[WS(rs, 1)]; | |
514 T10 = ii[WS(rs, 1)]; | |
515 TX = W[0]; | |
516 TZ = W[1]; | |
517 T11 = FMA(TX, TY, TZ * T10); | |
518 T1F = FNMS(TZ, TY, TX * T10); | |
519 } | |
520 T12 = TW + T11; | |
521 T2d = T1E + T1F; | |
522 { | |
523 E T1a, T1c, T19, T1b; | |
524 T1a = ri[WS(rs, 14)]; | |
525 T1c = ii[WS(rs, 14)]; | |
526 T19 = W[26]; | |
527 T1b = W[27]; | |
528 T1d = FMA(T19, T1a, T1b * T1c); | |
529 T1J = FNMS(T1b, T1a, T19 * T1c); | |
530 } | |
531 { | |
532 E T1f, T1h, T1e, T1g; | |
533 T1f = ri[WS(rs, 4)]; | |
534 T1h = ii[WS(rs, 4)]; | |
535 T1e = W[6]; | |
536 T1g = W[7]; | |
537 T1i = FMA(T1e, T1f, T1g * T1h); | |
538 T1K = FNMS(T1g, T1f, T1e * T1h); | |
539 } | |
540 T1j = T1d + T1i; | |
541 T2i = T1J + T1K; | |
542 { | |
543 E T1D, T1G, T2g, T2j; | |
544 T13 = TR + T12; | |
545 T1k = T18 + T1j; | |
546 T1l = T13 + T1k; | |
547 T2E = T2c + T2d; | |
548 T2F = T2h + T2i; | |
549 T2O = T2E + T2F; | |
550 T1D = FNMS(KP500000000, T12, TR); | |
551 T1G = KP866025403 * (T1E - T1F); | |
552 T1H = T1D - T1G; | |
553 T1T = T1D + T1G; | |
554 T2g = KP866025403 * (T1i - T1d); | |
555 T2j = FNMS(KP500000000, T2i, T2h); | |
556 T2k = T2g + T2j; | |
557 T2t = T2j - T2g; | |
558 { | |
559 E T2b, T2e, T1I, T1L; | |
560 T2b = KP866025403 * (T11 - TW); | |
561 T2e = FNMS(KP500000000, T2d, T2c); | |
562 T2f = T2b + T2e; | |
563 T2s = T2e - T2b; | |
564 T1I = FNMS(KP500000000, T1j, T18); | |
565 T1L = KP866025403 * (T1J - T1K); | |
566 T1M = T1I - T1L; | |
567 T1U = T1I + T1L; | |
568 } | |
569 } | |
570 } | |
571 { | |
572 E Ti, T21, Tz, T26, Tn, T1t, Ts, T1u, Tt, T22, TE, T1y, TJ, T1z, TK; | |
573 E T27; | |
574 { | |
575 E Tf, Th, Te, Tg; | |
576 Tf = ri[WS(rs, 3)]; | |
577 Th = ii[WS(rs, 3)]; | |
578 Te = W[4]; | |
579 Tg = W[5]; | |
580 Ti = FMA(Te, Tf, Tg * Th); | |
581 T21 = FNMS(Tg, Tf, Te * Th); | |
582 } | |
583 { | |
584 E Tw, Ty, Tv, Tx; | |
585 Tw = ri[WS(rs, 12)]; | |
586 Ty = ii[WS(rs, 12)]; | |
587 Tv = W[22]; | |
588 Tx = W[23]; | |
589 Tz = FMA(Tv, Tw, Tx * Ty); | |
590 T26 = FNMS(Tx, Tw, Tv * Ty); | |
591 } | |
592 { | |
593 E Tk, Tm, Tj, Tl; | |
594 Tk = ri[WS(rs, 8)]; | |
595 Tm = ii[WS(rs, 8)]; | |
596 Tj = W[14]; | |
597 Tl = W[15]; | |
598 Tn = FMA(Tj, Tk, Tl * Tm); | |
599 T1t = FNMS(Tl, Tk, Tj * Tm); | |
600 } | |
601 { | |
602 E Tp, Tr, To, Tq; | |
603 Tp = ri[WS(rs, 13)]; | |
604 Tr = ii[WS(rs, 13)]; | |
605 To = W[24]; | |
606 Tq = W[25]; | |
607 Ts = FMA(To, Tp, Tq * Tr); | |
608 T1u = FNMS(Tq, Tp, To * Tr); | |
609 } | |
610 Tt = Tn + Ts; | |
611 T22 = T1t + T1u; | |
612 { | |
613 E TB, TD, TA, TC; | |
614 TB = ri[WS(rs, 2)]; | |
615 TD = ii[WS(rs, 2)]; | |
616 TA = W[2]; | |
617 TC = W[3]; | |
618 TE = FMA(TA, TB, TC * TD); | |
619 T1y = FNMS(TC, TB, TA * TD); | |
620 } | |
621 { | |
622 E TG, TI, TF, TH; | |
623 TG = ri[WS(rs, 7)]; | |
624 TI = ii[WS(rs, 7)]; | |
625 TF = W[12]; | |
626 TH = W[13]; | |
627 TJ = FMA(TF, TG, TH * TI); | |
628 T1z = FNMS(TH, TG, TF * TI); | |
629 } | |
630 TK = TE + TJ; | |
631 T27 = T1y + T1z; | |
632 { | |
633 E T1s, T1v, T25, T28; | |
634 Tu = Ti + Tt; | |
635 TL = Tz + TK; | |
636 TM = Tu + TL; | |
637 T2H = T21 + T22; | |
638 T2I = T26 + T27; | |
639 T2N = T2H + T2I; | |
640 T1s = FNMS(KP500000000, Tt, Ti); | |
641 T1v = KP866025403 * (T1t - T1u); | |
642 T1w = T1s - T1v; | |
643 T1Q = T1s + T1v; | |
644 T25 = KP866025403 * (TJ - TE); | |
645 T28 = FNMS(KP500000000, T27, T26); | |
646 T29 = T25 + T28; | |
647 T2w = T28 - T25; | |
648 { | |
649 E T20, T23, T1x, T1A; | |
650 T20 = KP866025403 * (Ts - Tn); | |
651 T23 = FNMS(KP500000000, T22, T21); | |
652 T24 = T20 + T23; | |
653 T2v = T23 - T20; | |
654 T1x = FNMS(KP500000000, TK, Tz); | |
655 T1A = KP866025403 * (T1y - T1z); | |
656 T1B = T1x - T1A; | |
657 T1R = T1x + T1A; | |
658 } | |
659 } | |
660 } | |
661 { | |
662 E T2C, T1m, T2B, T2K, T2M, T2G, T2J, T2L, T2D; | |
663 T2C = KP559016994 * (TM - T1l); | |
664 T1m = TM + T1l; | |
665 T2B = FNMS(KP250000000, T1m, Td); | |
666 T2G = T2E - T2F; | |
667 T2J = T2H - T2I; | |
668 T2K = FNMS(KP587785252, T2J, KP951056516 * T2G); | |
669 T2M = FMA(KP951056516, T2J, KP587785252 * T2G); | |
670 ri[0] = Td + T1m; | |
671 T2L = T2C + T2B; | |
672 ri[WS(rs, 9)] = T2L - T2M; | |
673 ri[WS(rs, 6)] = T2L + T2M; | |
674 T2D = T2B - T2C; | |
675 ri[WS(rs, 12)] = T2D - T2K; | |
676 ri[WS(rs, 3)] = T2D + T2K; | |
677 } | |
678 { | |
679 E T2U, T2P, T2T, T2Y, T30, T2W, T2X, T2Z, T2V; | |
680 T2U = KP559016994 * (T2N - T2O); | |
681 T2P = T2N + T2O; | |
682 T2T = FNMS(KP250000000, T2P, T2S); | |
683 T2W = T13 - T1k; | |
684 T2X = Tu - TL; | |
685 T2Y = FNMS(KP587785252, T2X, KP951056516 * T2W); | |
686 T30 = FMA(KP951056516, T2X, KP587785252 * T2W); | |
687 ii[0] = T2P + T2S; | |
688 T2Z = T2U + T2T; | |
689 ii[WS(rs, 6)] = T2Z - T30; | |
690 ii[WS(rs, 9)] = T30 + T2Z; | |
691 T2V = T2T - T2U; | |
692 ii[WS(rs, 3)] = T2V - T2Y; | |
693 ii[WS(rs, 12)] = T2Y + T2V; | |
694 } | |
695 { | |
696 E T2y, T2A, T1r, T1O, T2p, T2q, T2z, T2r; | |
697 { | |
698 E T2u, T2x, T1C, T1N; | |
699 T2u = T2s - T2t; | |
700 T2x = T2v - T2w; | |
701 T2y = FNMS(KP587785252, T2x, KP951056516 * T2u); | |
702 T2A = FMA(KP951056516, T2x, KP587785252 * T2u); | |
703 T1r = T1n - T1q; | |
704 T1C = T1w + T1B; | |
705 T1N = T1H + T1M; | |
706 T1O = T1C + T1N; | |
707 T2p = FNMS(KP250000000, T1O, T1r); | |
708 T2q = KP559016994 * (T1C - T1N); | |
709 } | |
710 ri[WS(rs, 5)] = T1r + T1O; | |
711 T2z = T2q + T2p; | |
712 ri[WS(rs, 14)] = T2z - T2A; | |
713 ri[WS(rs, 11)] = T2z + T2A; | |
714 T2r = T2p - T2q; | |
715 ri[WS(rs, 2)] = T2r - T2y; | |
716 ri[WS(rs, 8)] = T2r + T2y; | |
717 } | |
718 { | |
719 E T3h, T3q, T3i, T3l, T3m, T3n, T3p, T3o; | |
720 { | |
721 E T3f, T3g, T3j, T3k; | |
722 T3f = T1H - T1M; | |
723 T3g = T1w - T1B; | |
724 T3h = FNMS(KP587785252, T3g, KP951056516 * T3f); | |
725 T3q = FMA(KP951056516, T3g, KP587785252 * T3f); | |
726 T3i = T35 - T34; | |
727 T3j = T2v + T2w; | |
728 T3k = T2s + T2t; | |
729 T3l = T3j + T3k; | |
730 T3m = FNMS(KP250000000, T3l, T3i); | |
731 T3n = KP559016994 * (T3j - T3k); | |
732 } | |
733 ii[WS(rs, 5)] = T3l + T3i; | |
734 T3p = T3n + T3m; | |
735 ii[WS(rs, 11)] = T3p - T3q; | |
736 ii[WS(rs, 14)] = T3q + T3p; | |
737 T3o = T3m - T3n; | |
738 ii[WS(rs, 2)] = T3h + T3o; | |
739 ii[WS(rs, 8)] = T3o - T3h; | |
740 } | |
741 { | |
742 E T3c, T3d, T36, T37, T33, T38, T3e, T39; | |
743 { | |
744 E T3a, T3b, T31, T32; | |
745 T3a = T1Q - T1R; | |
746 T3b = T1T - T1U; | |
747 T3c = FMA(KP951056516, T3a, KP587785252 * T3b); | |
748 T3d = FNMS(KP587785252, T3a, KP951056516 * T3b); | |
749 T36 = T34 + T35; | |
750 T31 = T24 + T29; | |
751 T32 = T2f + T2k; | |
752 T37 = T31 + T32; | |
753 T33 = KP559016994 * (T31 - T32); | |
754 T38 = FNMS(KP250000000, T37, T36); | |
755 } | |
756 ii[WS(rs, 10)] = T37 + T36; | |
757 T3e = T38 - T33; | |
758 ii[WS(rs, 7)] = T3d + T3e; | |
759 ii[WS(rs, 13)] = T3e - T3d; | |
760 T39 = T33 + T38; | |
761 ii[WS(rs, 1)] = T39 - T3c; | |
762 ii[WS(rs, 4)] = T3c + T39; | |
763 } | |
764 { | |
765 E T2m, T2o, T1P, T1W, T1X, T1Y, T2n, T1Z; | |
766 { | |
767 E T2a, T2l, T1S, T1V; | |
768 T2a = T24 - T29; | |
769 T2l = T2f - T2k; | |
770 T2m = FMA(KP951056516, T2a, KP587785252 * T2l); | |
771 T2o = FNMS(KP587785252, T2a, KP951056516 * T2l); | |
772 T1P = T1n + T1q; | |
773 T1S = T1Q + T1R; | |
774 T1V = T1T + T1U; | |
775 T1W = T1S + T1V; | |
776 T1X = KP559016994 * (T1S - T1V); | |
777 T1Y = FNMS(KP250000000, T1W, T1P); | |
778 } | |
779 ri[WS(rs, 10)] = T1P + T1W; | |
780 T2n = T1Y - T1X; | |
781 ri[WS(rs, 7)] = T2n - T2o; | |
782 ri[WS(rs, 13)] = T2n + T2o; | |
783 T1Z = T1X + T1Y; | |
784 ri[WS(rs, 4)] = T1Z - T2m; | |
785 ri[WS(rs, 1)] = T1Z + T2m; | |
786 } | |
787 } | |
788 } | |
789 } | |
790 | |
791 static const tw_instr twinstr[] = { | |
792 {TW_FULL, 0, 15}, | |
793 {TW_NEXT, 1, 0} | |
794 }; | |
795 | |
796 static const ct_desc desc = { 15, "t1_15", twinstr, &GENUS, {128, 56, 56, 0}, 0, 0, 0 }; | |
797 | |
798 void X(codelet_t1_15) (planner *p) { | |
799 X(kdft_dit_register) (p, t1_15, &desc); | |
800 } | |
801 #endif /* HAVE_FMA */ |