comparison src/fftw-3.3.3/dft/scalar/codelets/n1_11.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 11 -name n1_11 -include n.h */
29
30 /*
31 * This function contains 140 FP additions, 110 FP multiplications,
32 * (or, 30 additions, 0 multiplications, 110 fused multiply/add),
33 * 84 stack variables, 10 constants, and 44 memory accesses
34 */
35 #include "n.h"
36
37 static void n1_11(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39 DK(KP989821441, +0.989821441880932732376092037776718787376519372);
40 DK(KP959492973, +0.959492973614497389890368057066327699062454848);
41 DK(KP918985947, +0.918985947228994779780736114132655398124909697);
42 DK(KP876768831, +0.876768831002589333891339807079336796764054852);
43 DK(KP830830026, +0.830830026003772851058548298459246407048009821);
44 DK(KP778434453, +0.778434453334651800608337670740821884709317477);
45 DK(KP715370323, +0.715370323453429719112414662767260662417897278);
46 DK(KP634356270, +0.634356270682424498893150776899916060542806975);
47 DK(KP342584725, +0.342584725681637509502641509861112333758894680);
48 DK(KP521108558, +0.521108558113202722944698153526659300680427422);
49 {
50 INT i;
51 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(44, is), MAKE_VOLATILE_STRIDE(44, os)) {
52 E T1, TA, T1p, T1y, T19, T1d, T1a, T1e;
53 {
54 E T1f, T1u, T4, T1q, Tg, T1t, T7, T1s, Ta, Td, T1r, TP, T1X, T26, Ti;
55 E TG, T1O, T1w, TY, T1F, T17, To, T1i, T1k, T1h, Tr, T1j, Tu, T1g, Tx;
56 E T21, TU, TL, TC, T1S, T1J, T1m, T12, T1z, T1b;
57 T1 = ri[0];
58 T1f = ii[0];
59 {
60 E T1E, T16, Tb, Tc, Tv, Tw;
61 {
62 E T2, T3, Te, Tf;
63 T2 = ri[WS(is, 1)];
64 T3 = ri[WS(is, 10)];
65 Te = ri[WS(is, 5)];
66 Tf = ri[WS(is, 6)];
67 {
68 E T5, T6, T8, T9;
69 T5 = ri[WS(is, 2)];
70 T1u = T3 - T2;
71 T4 = T2 + T3;
72 T1q = Tf - Te;
73 Tg = Te + Tf;
74 T6 = ri[WS(is, 9)];
75 T8 = ri[WS(is, 3)];
76 T9 = ri[WS(is, 8)];
77 Tb = ri[WS(is, 4)];
78 T1t = T6 - T5;
79 T7 = T5 + T6;
80 T1s = T9 - T8;
81 Ta = T8 + T9;
82 Tc = ri[WS(is, 7)];
83 }
84 }
85 {
86 E T25, Th, T1W, TO;
87 T25 = FMA(KP521108558, T1q, T1u);
88 T1W = FMA(KP521108558, T1s, T1q);
89 TO = FNMS(KP342584725, T4, Ta);
90 Th = FNMS(KP342584725, Ta, T7);
91 Td = Tb + Tc;
92 T1r = Tc - Tb;
93 TP = FNMS(KP634356270, TO, Tg);
94 T1X = FNMS(KP715370323, T1W, T1t);
95 T26 = FMA(KP715370323, T25, T1r);
96 {
97 E TF, T1N, T1v, TX;
98 TF = FNMS(KP342584725, Td, T4);
99 Ti = FNMS(KP634356270, Th, Td);
100 T1N = FNMS(KP521108558, T1t, T1r);
101 T1v = FNMS(KP521108558, T1u, T1t);
102 TG = FNMS(KP634356270, TF, T7);
103 TX = FNMS(KP342584725, T7, Tg);
104 T1O = FMA(KP715370323, T1N, T1q);
105 T1w = FNMS(KP715370323, T1v, T1s);
106 T1E = FMA(KP521108558, T1r, T1s);
107 TY = FNMS(KP634356270, TX, T4);
108 T16 = FNMS(KP342584725, Tg, Td);
109 }
110 }
111 {
112 E Ty, Tz, Tm, Tn;
113 Tm = ii[WS(is, 3)];
114 T1F = FMA(KP715370323, T1E, T1u);
115 Tn = ii[WS(is, 8)];
116 T17 = FNMS(KP634356270, T16, Ta);
117 Ty = ii[WS(is, 5)];
118 Tz = ii[WS(is, 6)];
119 To = Tm - Tn;
120 T1i = Tm + Tn;
121 {
122 E Tp, Tq, Ts, Tt;
123 Tp = ii[WS(is, 2)];
124 T1k = Ty + Tz;
125 TA = Ty - Tz;
126 Tq = ii[WS(is, 9)];
127 Ts = ii[WS(is, 4)];
128 Tt = ii[WS(is, 7)];
129 Tv = ii[WS(is, 1)];
130 T1h = Tp + Tq;
131 Tr = Tp - Tq;
132 T1j = Ts + Tt;
133 Tu = Ts - Tt;
134 Tw = ii[WS(is, 10)];
135 }
136 }
137 {
138 E TB, T1R, T20, TK, TT, T1I, T1l;
139 T20 = FNMS(KP342584725, T1i, T1h);
140 TK = FMA(KP521108558, To, TA);
141 TT = FNMS(KP521108558, Tr, Tu);
142 T1g = Tv + Tw;
143 Tx = Tv - Tw;
144 T21 = FNMS(KP634356270, T20, T1j);
145 TU = FMA(KP715370323, TT, TA);
146 TL = FNMS(KP715370323, TK, Tr);
147 TB = FMA(KP521108558, TA, Tx);
148 T1R = FNMS(KP342584725, T1j, T1g);
149 T1I = FNMS(KP342584725, T1g, T1i);
150 T1l = FNMS(KP342584725, T1k, T1j);
151 TC = FMA(KP715370323, TB, Tu);
152 T1S = FNMS(KP634356270, T1R, T1h);
153 T1J = FNMS(KP634356270, T1I, T1k);
154 T1m = FNMS(KP634356270, T1l, T1i);
155 T12 = FMA(KP521108558, Tu, To);
156 T1z = FNMS(KP342584725, T1h, T1k);
157 T1b = FNMS(KP521108558, Tx, Tr);
158 }
159 }
160 {
161 E T13, T1A, T1c, T1Z, T1V, TH, TM, Tj, TD;
162 ro[0] = T1 + T4 + T7 + Ta + Td + Tg;
163 T13 = FMA(KP715370323, T12, Tx);
164 T1A = FNMS(KP634356270, T1z, T1g);
165 T1c = FNMS(KP715370323, T1b, To);
166 io[0] = T1f + T1g + T1h + T1i + T1j + T1k;
167 Tj = FNMS(KP778434453, Ti, T4);
168 TD = FMA(KP830830026, TC, Tr);
169 {
170 E TE, T23, T28, Tl, Tk, T22, T27;
171 T22 = FNMS(KP778434453, T21, T1g);
172 T27 = FMA(KP830830026, T26, T1t);
173 Tk = FNMS(KP876768831, Tj, Tg);
174 TE = FMA(KP918985947, TD, To);
175 T23 = FNMS(KP876768831, T22, T1k);
176 T28 = FMA(KP918985947, T27, T1s);
177 Tl = FNMS(KP959492973, Tk, T1);
178 {
179 E T1U, T1T, T24, T1Y;
180 T1T = FNMS(KP778434453, T1S, T1k);
181 T24 = FNMS(KP959492973, T23, T1f);
182 T1Y = FMA(KP830830026, T1X, T1u);
183 ro[WS(os, 1)] = FMA(KP989821441, TE, Tl);
184 ro[WS(os, 10)] = FNMS(KP989821441, TE, Tl);
185 T1U = FNMS(KP876768831, T1T, T1i);
186 io[WS(os, 10)] = FNMS(KP989821441, T28, T24);
187 io[WS(os, 1)] = FMA(KP989821441, T28, T24);
188 T1Z = FNMS(KP918985947, T1Y, T1r);
189 T1V = FNMS(KP959492973, T1U, T1f);
190 }
191 TH = FNMS(KP778434453, TG, Tg);
192 TM = FMA(KP830830026, TL, Tx);
193 }
194 {
195 E T1M, TZ, T14, T1Q;
196 {
197 E TN, TR, TV, TJ, TI, TQ, T1P;
198 TQ = FNMS(KP778434453, TP, Td);
199 io[WS(os, 9)] = FMA(KP989821441, T1Z, T1V);
200 io[WS(os, 2)] = FNMS(KP989821441, T1Z, T1V);
201 TI = FNMS(KP876768831, TH, Ta);
202 TN = FNMS(KP918985947, TM, Tu);
203 TR = FNMS(KP876768831, TQ, T7);
204 TV = FNMS(KP830830026, TU, To);
205 TJ = FNMS(KP959492973, TI, T1);
206 {
207 E T1L, TS, TW, T1K;
208 T1K = FNMS(KP778434453, T1J, T1j);
209 TS = FNMS(KP959492973, TR, T1);
210 TW = FNMS(KP918985947, TV, Tx);
211 ro[WS(os, 9)] = FMA(KP989821441, TN, TJ);
212 ro[WS(os, 2)] = FNMS(KP989821441, TN, TJ);
213 T1L = FNMS(KP876768831, T1K, T1h);
214 ro[WS(os, 3)] = FMA(KP989821441, TW, TS);
215 ro[WS(os, 8)] = FNMS(KP989821441, TW, TS);
216 T1P = FNMS(KP830830026, T1O, T1s);
217 T1M = FNMS(KP959492973, T1L, T1f);
218 }
219 TZ = FNMS(KP778434453, TY, Ta);
220 T14 = FNMS(KP830830026, T13, TA);
221 T1Q = FNMS(KP918985947, T1P, T1u);
222 }
223 {
224 E T15, T11, T1C, T1G, T1B, T10;
225 T1B = FNMS(KP778434453, T1A, T1i);
226 T10 = FNMS(KP876768831, TZ, Td);
227 T15 = FMA(KP918985947, T14, Tr);
228 io[WS(os, 8)] = FNMS(KP989821441, T1Q, T1M);
229 io[WS(os, 3)] = FMA(KP989821441, T1Q, T1M);
230 T11 = FNMS(KP959492973, T10, T1);
231 T1C = FNMS(KP876768831, T1B, T1j);
232 T1G = FNMS(KP830830026, T1F, T1q);
233 {
234 E T1D, T1H, T1o, T1x, T1n, T18;
235 T1n = FNMS(KP778434453, T1m, T1h);
236 ro[WS(os, 7)] = FMA(KP989821441, T15, T11);
237 ro[WS(os, 4)] = FNMS(KP989821441, T15, T11);
238 T1D = FNMS(KP959492973, T1C, T1f);
239 T1H = FMA(KP918985947, T1G, T1t);
240 T1o = FNMS(KP876768831, T1n, T1g);
241 T1x = FNMS(KP830830026, T1w, T1r);
242 T18 = FNMS(KP778434453, T17, T7);
243 io[WS(os, 7)] = FMA(KP989821441, T1H, T1D);
244 io[WS(os, 4)] = FNMS(KP989821441, T1H, T1D);
245 T1p = FNMS(KP959492973, T1o, T1f);
246 T1y = FNMS(KP918985947, T1x, T1q);
247 T19 = FNMS(KP876768831, T18, T4);
248 T1d = FNMS(KP830830026, T1c, Tu);
249 }
250 }
251 }
252 }
253 }
254 io[WS(os, 6)] = FNMS(KP989821441, T1y, T1p);
255 io[WS(os, 5)] = FMA(KP989821441, T1y, T1p);
256 T1a = FNMS(KP959492973, T19, T1);
257 T1e = FNMS(KP918985947, T1d, TA);
258 ro[WS(os, 5)] = FMA(KP989821441, T1e, T1a);
259 ro[WS(os, 6)] = FNMS(KP989821441, T1e, T1a);
260 }
261 }
262 }
263
264 static const kdft_desc desc = { 11, "n1_11", {30, 0, 110, 0}, &GENUS, 0, 0, 0, 0 };
265
266 void X(codelet_n1_11) (planner *p) {
267 X(kdft_register) (p, n1_11, &desc);
268 }
269
270 #else /* HAVE_FMA */
271
272 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 11 -name n1_11 -include n.h */
273
274 /*
275 * This function contains 140 FP additions, 100 FP multiplications,
276 * (or, 60 additions, 20 multiplications, 80 fused multiply/add),
277 * 41 stack variables, 10 constants, and 44 memory accesses
278 */
279 #include "n.h"
280
281 static void n1_11(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
282 {
283 DK(KP654860733, +0.654860733945285064056925072466293553183791199);
284 DK(KP142314838, +0.142314838273285140443792668616369668791051361);
285 DK(KP959492973, +0.959492973614497389890368057066327699062454848);
286 DK(KP415415013, +0.415415013001886425529274149229623203524004910);
287 DK(KP841253532, +0.841253532831181168861811648919367717513292498);
288 DK(KP989821441, +0.989821441880932732376092037776718787376519372);
289 DK(KP909631995, +0.909631995354518371411715383079028460060241051);
290 DK(KP281732556, +0.281732556841429697711417915346616899035777899);
291 DK(KP540640817, +0.540640817455597582107635954318691695431770608);
292 DK(KP755749574, +0.755749574354258283774035843972344420179717445);
293 {
294 INT i;
295 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(44, is), MAKE_VOLATILE_STRIDE(44, os)) {
296 E T1, TM, T4, TG, Tk, TR, Tw, TN, T7, TK, Ta, TH, Tn, TQ, Td;
297 E TJ, Tq, TO, Tt, TP, Tg, TI;
298 {
299 E T2, T3, Ti, Tj;
300 T1 = ri[0];
301 TM = ii[0];
302 T2 = ri[WS(is, 1)];
303 T3 = ri[WS(is, 10)];
304 T4 = T2 + T3;
305 TG = T3 - T2;
306 Ti = ii[WS(is, 1)];
307 Tj = ii[WS(is, 10)];
308 Tk = Ti - Tj;
309 TR = Ti + Tj;
310 {
311 E Tu, Tv, T5, T6;
312 Tu = ii[WS(is, 2)];
313 Tv = ii[WS(is, 9)];
314 Tw = Tu - Tv;
315 TN = Tu + Tv;
316 T5 = ri[WS(is, 2)];
317 T6 = ri[WS(is, 9)];
318 T7 = T5 + T6;
319 TK = T6 - T5;
320 }
321 }
322 {
323 E T8, T9, To, Tp;
324 T8 = ri[WS(is, 3)];
325 T9 = ri[WS(is, 8)];
326 Ta = T8 + T9;
327 TH = T9 - T8;
328 {
329 E Tl, Tm, Tb, Tc;
330 Tl = ii[WS(is, 3)];
331 Tm = ii[WS(is, 8)];
332 Tn = Tl - Tm;
333 TQ = Tl + Tm;
334 Tb = ri[WS(is, 4)];
335 Tc = ri[WS(is, 7)];
336 Td = Tb + Tc;
337 TJ = Tc - Tb;
338 }
339 To = ii[WS(is, 4)];
340 Tp = ii[WS(is, 7)];
341 Tq = To - Tp;
342 TO = To + Tp;
343 {
344 E Tr, Ts, Te, Tf;
345 Tr = ii[WS(is, 5)];
346 Ts = ii[WS(is, 6)];
347 Tt = Tr - Ts;
348 TP = Tr + Ts;
349 Te = ri[WS(is, 5)];
350 Tf = ri[WS(is, 6)];
351 Tg = Te + Tf;
352 TI = Tf - Te;
353 }
354 }
355 {
356 E Tx, Th, TZ, T10;
357 ro[0] = T1 + T4 + T7 + Ta + Td + Tg;
358 io[0] = TM + TR + TN + TQ + TO + TP;
359 Tx = FMA(KP755749574, Tk, KP540640817 * Tn) + FNMS(KP909631995, Tt, KP281732556 * Tq) - (KP989821441 * Tw);
360 Th = FMA(KP841253532, Ta, T1) + FNMS(KP959492973, Td, KP415415013 * Tg) + FNMA(KP142314838, T7, KP654860733 * T4);
361 ro[WS(os, 7)] = Th - Tx;
362 ro[WS(os, 4)] = Th + Tx;
363 TZ = FMA(KP755749574, TG, KP540640817 * TH) + FNMS(KP909631995, TI, KP281732556 * TJ) - (KP989821441 * TK);
364 T10 = FMA(KP841253532, TQ, TM) + FNMS(KP959492973, TO, KP415415013 * TP) + FNMA(KP142314838, TN, KP654860733 * TR);
365 io[WS(os, 4)] = TZ + T10;
366 io[WS(os, 7)] = T10 - TZ;
367 {
368 E TX, TY, Tz, Ty;
369 TX = FMA(KP909631995, TG, KP755749574 * TK) + FNMA(KP540640817, TI, KP989821441 * TJ) - (KP281732556 * TH);
370 TY = FMA(KP415415013, TR, TM) + FNMS(KP142314838, TO, KP841253532 * TP) + FNMA(KP959492973, TQ, KP654860733 * TN);
371 io[WS(os, 2)] = TX + TY;
372 io[WS(os, 9)] = TY - TX;
373 Tz = FMA(KP909631995, Tk, KP755749574 * Tw) + FNMA(KP540640817, Tt, KP989821441 * Tq) - (KP281732556 * Tn);
374 Ty = FMA(KP415415013, T4, T1) + FNMS(KP142314838, Td, KP841253532 * Tg) + FNMA(KP959492973, Ta, KP654860733 * T7);
375 ro[WS(os, 9)] = Ty - Tz;
376 ro[WS(os, 2)] = Ty + Tz;
377 }
378 }
379 {
380 E TB, TA, TT, TU;
381 TB = FMA(KP540640817, Tk, KP909631995 * Tw) + FMA(KP989821441, Tn, KP755749574 * Tq) + (KP281732556 * Tt);
382 TA = FMA(KP841253532, T4, T1) + FNMS(KP959492973, Tg, KP415415013 * T7) + FNMA(KP654860733, Td, KP142314838 * Ta);
383 ro[WS(os, 10)] = TA - TB;
384 ro[WS(os, 1)] = TA + TB;
385 {
386 E TV, TW, TD, TC;
387 TV = FMA(KP540640817, TG, KP909631995 * TK) + FMA(KP989821441, TH, KP755749574 * TJ) + (KP281732556 * TI);
388 TW = FMA(KP841253532, TR, TM) + FNMS(KP959492973, TP, KP415415013 * TN) + FNMA(KP654860733, TO, KP142314838 * TQ);
389 io[WS(os, 1)] = TV + TW;
390 io[WS(os, 10)] = TW - TV;
391 TD = FMA(KP989821441, Tk, KP540640817 * Tq) + FNMS(KP909631995, Tn, KP755749574 * Tt) - (KP281732556 * Tw);
392 TC = FMA(KP415415013, Ta, T1) + FNMS(KP654860733, Tg, KP841253532 * Td) + FNMA(KP959492973, T7, KP142314838 * T4);
393 ro[WS(os, 8)] = TC - TD;
394 ro[WS(os, 3)] = TC + TD;
395 }
396 TT = FMA(KP989821441, TG, KP540640817 * TJ) + FNMS(KP909631995, TH, KP755749574 * TI) - (KP281732556 * TK);
397 TU = FMA(KP415415013, TQ, TM) + FNMS(KP654860733, TP, KP841253532 * TO) + FNMA(KP959492973, TN, KP142314838 * TR);
398 io[WS(os, 3)] = TT + TU;
399 io[WS(os, 8)] = TU - TT;
400 {
401 E TL, TS, TF, TE;
402 TL = FMA(KP281732556, TG, KP755749574 * TH) + FNMS(KP909631995, TJ, KP989821441 * TI) - (KP540640817 * TK);
403 TS = FMA(KP841253532, TN, TM) + FNMS(KP142314838, TP, KP415415013 * TO) + FNMA(KP654860733, TQ, KP959492973 * TR);
404 io[WS(os, 5)] = TL + TS;
405 io[WS(os, 6)] = TS - TL;
406 TF = FMA(KP281732556, Tk, KP755749574 * Tn) + FNMS(KP909631995, Tq, KP989821441 * Tt) - (KP540640817 * Tw);
407 TE = FMA(KP841253532, T7, T1) + FNMS(KP142314838, Tg, KP415415013 * Td) + FNMA(KP654860733, Ta, KP959492973 * T4);
408 ro[WS(os, 6)] = TE - TF;
409 ro[WS(os, 5)] = TE + TF;
410 }
411 }
412 }
413 }
414 }
415
416 static const kdft_desc desc = { 11, "n1_11", {60, 20, 80, 0}, &GENUS, 0, 0, 0, 0 };
417
418 void X(codelet_n1_11) (planner *p) {
419 X(kdft_register) (p, n1_11, &desc);
420 }
421
422 #endif /* HAVE_FMA */