comparison src/fftw-3.3.3/dft/simd/common/t3fv_32.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 07:38:50 EST 2012 */
23
24 #include "codelet-dft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_twiddle_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 32 -name t3fv_32 -include t3f.h */
29
30 /*
31 * This function contains 244 FP additions, 214 FP multiplications,
32 * (or, 146 additions, 116 multiplications, 98 fused multiply/add),
33 * 118 stack variables, 7 constants, and 64 memory accesses
34 */
35 #include "t3f.h"
36
37 static void t3fv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
38 {
39 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
40 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
41 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
42 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
44 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
45 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
46 {
47 INT m;
48 R *x;
49 x = ri;
50 for (m = mb, W = W + (mb * ((TWVL / VL) * 8)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(32, rs)) {
51 V T2B, T2A, T2u, T2x, T2r, T2F, T2L, T2P;
52 {
53 V T2, T5, T3, T7;
54 T2 = LDW(&(W[0]));
55 T5 = LDW(&(W[TWVL * 4]));
56 T3 = LDW(&(W[TWVL * 2]));
57 T7 = LDW(&(W[TWVL * 6]));
58 {
59 V T24, Tb, T3x, T2T, T3K, T2W, T25, Tr, T3z, T3g, T28, TX, T3y, T3j, T27;
60 V TG, T37, T3F, T3G, T3a, T2Y, T15, T1p, T2Z, T2w, T1V, T2v, T1N, T32, T1h;
61 V T17, T1a;
62 {
63 V T1, Tz, TT, T4, TC, Tv, T12, T1D, T1w, T18, T1t, T1O, TK, TP, T1c;
64 V T1m, Tf, T6, Te, TL, TQ, T2S, Tp, TU, Ti, Ta, TM, TR, Tm, TJ;
65 V T22, T9, T1Z;
66 T1 = LD(&(x[0]), ms, &(x[0]));
67 T22 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
68 T9 = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
69 T1Z = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
70 {
71 V Tn, TH, Tk, To, Th, Tg, T8, Tl, T20, T23, TI;
72 {
73 V Td, T1C, Tc, T21;
74 Td = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
75 Tz = VZMUL(T2, T5);
76 T1C = VZMULJ(T2, T5);
77 Tn = VZMUL(T3, T5);
78 TT = VZMULJ(T3, T5);
79 Tc = VZMUL(T2, T3);
80 T4 = VZMULJ(T2, T3);
81 TH = VZMUL(T3, T7);
82 T21 = VZMULJ(T3, T7);
83 Tk = VZMUL(T2, T7);
84 TC = VZMULJ(T2, T7);
85 Tv = VZMULJ(T5, T7);
86 T12 = VZMULJ(Tz, T7);
87 T20 = VZMULJ(T1C, T1Z);
88 T1D = VZMULJ(T1C, T7);
89 T1w = VZMULJ(Tn, T7);
90 T18 = VZMULJ(TT, T7);
91 T1t = VZMUL(Tc, T7);
92 T1O = VZMULJ(Tc, T7);
93 TK = VZMUL(Tc, T5);
94 TP = VZMULJ(Tc, T5);
95 T1c = VZMUL(T4, T7);
96 T1m = VZMULJ(T4, T7);
97 Tf = VZMULJ(T4, T5);
98 T6 = VZMUL(T4, T5);
99 T23 = VZMULJ(T21, T22);
100 Te = VZMULJ(Tc, Td);
101 }
102 TL = VZMULJ(TK, T7);
103 TQ = VZMULJ(TP, T7);
104 To = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
105 Th = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
106 Tg = VZMULJ(Tf, T7);
107 T8 = VZMULJ(T6, T7);
108 T2S = VADD(T20, T23);
109 T24 = VSUB(T20, T23);
110 Tl = LD(&(x[WS(rs, 28)]), ms, &(x[0]));
111 TI = LD(&(x[WS(rs, 30)]), ms, &(x[0]));
112 Tp = VZMULJ(Tn, To);
113 TU = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
114 Ti = VZMULJ(Tg, Th);
115 Ta = VZMULJ(T8, T9);
116 TM = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
117 TR = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
118 Tm = VZMULJ(Tk, Tl);
119 TJ = VZMULJ(TH, TI);
120 }
121 {
122 V Tu, TE, Tw, TA;
123 {
124 V T3e, TO, T3f, TW;
125 {
126 V TV, T2U, Tj, T2R, TN, TS, T2V, Tq, Tt, TD;
127 Tt = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
128 TV = VZMULJ(TT, TU);
129 T2U = VADD(Te, Ti);
130 Tj = VSUB(Te, Ti);
131 T2R = VADD(T1, Ta);
132 Tb = VSUB(T1, Ta);
133 TN = VZMULJ(TL, TM);
134 TS = VZMULJ(TQ, TR);
135 T2V = VADD(Tm, Tp);
136 Tq = VSUB(Tm, Tp);
137 Tu = VZMULJ(T4, Tt);
138 TD = LD(&(x[WS(rs, 26)]), ms, &(x[0]));
139 T3x = VSUB(T2R, T2S);
140 T2T = VADD(T2R, T2S);
141 T3e = VADD(TJ, TN);
142 TO = VSUB(TJ, TN);
143 T3f = VADD(TV, TS);
144 TW = VSUB(TS, TV);
145 T3K = VSUB(T2V, T2U);
146 T2W = VADD(T2U, T2V);
147 T25 = VSUB(Tq, Tj);
148 Tr = VADD(Tj, Tq);
149 TE = VZMULJ(TC, TD);
150 }
151 Tw = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
152 T3z = VSUB(T3e, T3f);
153 T3g = VADD(T3e, T3f);
154 T28 = VFMA(LDK(KP414213562), TO, TW);
155 TX = VFNMS(LDK(KP414213562), TW, TO);
156 TA = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
157 }
158 {
159 V T35, T1z, T1T, T36, T39, T1L, T1B, T1F;
160 {
161 V T1v, T1y, Ty, T3h, T1S, T1Q, T1I, T3i, TF, T1K, T1A, T1E;
162 {
163 V T1u, T1x, Tx, T1R;
164 T1u = LD(&(x[WS(rs, 31)]), ms, &(x[WS(rs, 1)]));
165 T1x = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
166 Tx = VZMULJ(Tv, Tw);
167 T1R = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
168 {
169 V T1P, T1H, T1J, TB;
170 T1P = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
171 T1H = LD(&(x[WS(rs, 27)]), ms, &(x[WS(rs, 1)]));
172 T1J = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
173 TB = VZMULJ(Tz, TA);
174 T1v = VZMULJ(T1t, T1u);
175 T1y = VZMULJ(T1w, T1x);
176 Ty = VSUB(Tu, Tx);
177 T3h = VADD(Tu, Tx);
178 T1S = VZMULJ(Tf, T1R);
179 T1Q = VZMULJ(T1O, T1P);
180 T1I = VZMULJ(T7, T1H);
181 T3i = VADD(TB, TE);
182 TF = VSUB(TB, TE);
183 T1K = VZMULJ(T6, T1J);
184 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
185 T1E = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
186 }
187 }
188 T35 = VADD(T1v, T1y);
189 T1z = VSUB(T1v, T1y);
190 T1T = VSUB(T1Q, T1S);
191 T36 = VADD(T1S, T1Q);
192 T3y = VSUB(T3h, T3i);
193 T3j = VADD(T3h, T3i);
194 T27 = VFMA(LDK(KP414213562), Ty, TF);
195 TG = VFNMS(LDK(KP414213562), TF, Ty);
196 T39 = VADD(T1I, T1K);
197 T1L = VSUB(T1I, T1K);
198 T1B = VZMULJ(T3, T1A);
199 T1F = VZMULJ(T1D, T1E);
200 }
201 {
202 V T11, T14, T1o, T1l, T1e, T1U, T1M, T1g, T16, T19;
203 {
204 V T10, T13, T1n, T1k;
205 T10 = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
206 T13 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
207 T1n = LD(&(x[WS(rs, 25)]), ms, &(x[WS(rs, 1)]));
208 T1k = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
209 {
210 V T1d, T1f, T1G, T38;
211 T1d = LD(&(x[WS(rs, 29)]), ms, &(x[WS(rs, 1)]));
212 T1f = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
213 T1G = VSUB(T1B, T1F);
214 T38 = VADD(T1B, T1F);
215 T37 = VADD(T35, T36);
216 T3F = VSUB(T35, T36);
217 T11 = VZMULJ(T2, T10);
218 T14 = VZMULJ(T12, T13);
219 T1o = VZMULJ(T1m, T1n);
220 T1l = VZMULJ(T5, T1k);
221 T1e = VZMULJ(T1c, T1d);
222 T3G = VSUB(T39, T38);
223 T3a = VADD(T38, T39);
224 T1U = VSUB(T1L, T1G);
225 T1M = VADD(T1G, T1L);
226 T1g = VZMULJ(TK, T1f);
227 }
228 T16 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
229 T19 = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
230 }
231 T2Y = VADD(T11, T14);
232 T15 = VSUB(T11, T14);
233 T1p = VSUB(T1l, T1o);
234 T2Z = VADD(T1l, T1o);
235 T2w = VFNMS(LDK(KP707106781), T1U, T1T);
236 T1V = VFMA(LDK(KP707106781), T1U, T1T);
237 T2v = VFNMS(LDK(KP707106781), T1M, T1z);
238 T1N = VFMA(LDK(KP707106781), T1M, T1z);
239 T32 = VADD(T1e, T1g);
240 T1h = VSUB(T1e, T1g);
241 T17 = VZMULJ(TP, T16);
242 T1a = VZMULJ(T18, T19);
243 }
244 }
245 }
246 }
247 {
248 V T2X, T3k, T3b, T3t, T1b, T31, T30, T3C, T3r, T3v, T3p, T3q;
249 T2X = VSUB(T2T, T2W);
250 T3p = VADD(T2T, T2W);
251 T3q = VADD(T3j, T3g);
252 T3k = VSUB(T3g, T3j);
253 T3b = VSUB(T37, T3a);
254 T3t = VADD(T37, T3a);
255 T1b = VSUB(T17, T1a);
256 T31 = VADD(T17, T1a);
257 T30 = VADD(T2Y, T2Z);
258 T3C = VSUB(T2Y, T2Z);
259 T3r = VADD(T3p, T3q);
260 T3v = VSUB(T3p, T3q);
261 {
262 V T3N, T3B, T3T, T3M, T3W, T3O, T2t, T1r, T2s, T1j, T3I, T3X, T3c, T3l, T3u;
263 V T3w;
264 {
265 V T3L, T3A, T33, T3D, T1i, T1q;
266 T3L = VSUB(T3z, T3y);
267 T3A = VADD(T3y, T3z);
268 T33 = VADD(T31, T32);
269 T3D = VSUB(T31, T32);
270 T1i = VADD(T1b, T1h);
271 T1q = VSUB(T1b, T1h);
272 {
273 V T3H, T3E, T34, T3s;
274 T3N = VFMA(LDK(KP414213562), T3F, T3G);
275 T3H = VFNMS(LDK(KP414213562), T3G, T3F);
276 T3B = VFMA(LDK(KP707106781), T3A, T3x);
277 T3T = VFNMS(LDK(KP707106781), T3A, T3x);
278 T3M = VFMA(LDK(KP707106781), T3L, T3K);
279 T3W = VFNMS(LDK(KP707106781), T3L, T3K);
280 T3O = VFMA(LDK(KP414213562), T3C, T3D);
281 T3E = VFNMS(LDK(KP414213562), T3D, T3C);
282 T34 = VSUB(T30, T33);
283 T3s = VADD(T30, T33);
284 T2t = VFNMS(LDK(KP707106781), T1q, T1p);
285 T1r = VFMA(LDK(KP707106781), T1q, T1p);
286 T2s = VFNMS(LDK(KP707106781), T1i, T15);
287 T1j = VFMA(LDK(KP707106781), T1i, T15);
288 T3I = VADD(T3E, T3H);
289 T3X = VSUB(T3H, T3E);
290 T3c = VADD(T34, T3b);
291 T3l = VSUB(T3b, T34);
292 T3u = VADD(T3s, T3t);
293 T3w = VSUB(T3t, T3s);
294 }
295 }
296 {
297 V T2p, Ts, TY, T1s, T2b, T2c, T1W, T26, T29, T2q, T3U, T3P, T2J, T2K;
298 T2p = VFNMS(LDK(KP707106781), Tr, Tb);
299 Ts = VFMA(LDK(KP707106781), Tr, Tb);
300 T3U = VADD(T3O, T3N);
301 T3P = VSUB(T3N, T3O);
302 {
303 V T3Y, T40, T3R, T3J;
304 T3Y = VFMA(LDK(KP923879532), T3X, T3W);
305 T40 = VFNMS(LDK(KP923879532), T3X, T3W);
306 T3R = VFMA(LDK(KP923879532), T3I, T3B);
307 T3J = VFNMS(LDK(KP923879532), T3I, T3B);
308 {
309 V T3o, T3m, T3n, T3d;
310 T3o = VFMA(LDK(KP707106781), T3l, T3k);
311 T3m = VFNMS(LDK(KP707106781), T3l, T3k);
312 T3n = VFMA(LDK(KP707106781), T3c, T2X);
313 T3d = VFNMS(LDK(KP707106781), T3c, T2X);
314 ST(&(x[WS(rs, 24)]), VFNMSI(T3w, T3v), ms, &(x[0]));
315 ST(&(x[WS(rs, 8)]), VFMAI(T3w, T3v), ms, &(x[0]));
316 ST(&(x[0]), VADD(T3r, T3u), ms, &(x[0]));
317 ST(&(x[WS(rs, 16)]), VSUB(T3r, T3u), ms, &(x[0]));
318 {
319 V T3V, T3Z, T3S, T3Q;
320 T3V = VFNMS(LDK(KP923879532), T3U, T3T);
321 T3Z = VFMA(LDK(KP923879532), T3U, T3T);
322 T3S = VFMA(LDK(KP923879532), T3P, T3M);
323 T3Q = VFNMS(LDK(KP923879532), T3P, T3M);
324 ST(&(x[WS(rs, 4)]), VFMAI(T3o, T3n), ms, &(x[0]));
325 ST(&(x[WS(rs, 28)]), VFNMSI(T3o, T3n), ms, &(x[0]));
326 ST(&(x[WS(rs, 20)]), VFMAI(T3m, T3d), ms, &(x[0]));
327 ST(&(x[WS(rs, 12)]), VFNMSI(T3m, T3d), ms, &(x[0]));
328 ST(&(x[WS(rs, 22)]), VFNMSI(T3Y, T3V), ms, &(x[0]));
329 ST(&(x[WS(rs, 10)]), VFMAI(T3Y, T3V), ms, &(x[0]));
330 ST(&(x[WS(rs, 26)]), VFMAI(T40, T3Z), ms, &(x[0]));
331 ST(&(x[WS(rs, 6)]), VFNMSI(T40, T3Z), ms, &(x[0]));
332 ST(&(x[WS(rs, 2)]), VFMAI(T3S, T3R), ms, &(x[0]));
333 ST(&(x[WS(rs, 30)]), VFNMSI(T3S, T3R), ms, &(x[0]));
334 ST(&(x[WS(rs, 18)]), VFMAI(T3Q, T3J), ms, &(x[0]));
335 ST(&(x[WS(rs, 14)]), VFNMSI(T3Q, T3J), ms, &(x[0]));
336 TY = VADD(TG, TX);
337 T2B = VSUB(TX, TG);
338 }
339 }
340 }
341 T1s = VFNMS(LDK(KP198912367), T1r, T1j);
342 T2b = VFMA(LDK(KP198912367), T1j, T1r);
343 T2c = VFMA(LDK(KP198912367), T1N, T1V);
344 T1W = VFNMS(LDK(KP198912367), T1V, T1N);
345 T2A = VFMA(LDK(KP707106781), T25, T24);
346 T26 = VFNMS(LDK(KP707106781), T25, T24);
347 T29 = VSUB(T27, T28);
348 T2q = VADD(T27, T28);
349 {
350 V T2j, T2n, T1Y, T2f, T2o, T2m, T2e, T2g;
351 {
352 V T2h, TZ, T2i, T2d, T2l, T1X, T2k, T2a, T2D, T2E;
353 T2h = VFNMS(LDK(KP923879532), TY, Ts);
354 TZ = VFMA(LDK(KP923879532), TY, Ts);
355 T2i = VADD(T2b, T2c);
356 T2d = VSUB(T2b, T2c);
357 T2l = VSUB(T1W, T1s);
358 T1X = VADD(T1s, T1W);
359 T2k = VFNMS(LDK(KP923879532), T29, T26);
360 T2a = VFMA(LDK(KP923879532), T29, T26);
361 T2u = VFMA(LDK(KP668178637), T2t, T2s);
362 T2D = VFNMS(LDK(KP668178637), T2s, T2t);
363 T2j = VFNMS(LDK(KP980785280), T2i, T2h);
364 T2n = VFMA(LDK(KP980785280), T2i, T2h);
365 T2E = VFNMS(LDK(KP668178637), T2v, T2w);
366 T2x = VFMA(LDK(KP668178637), T2w, T2v);
367 T1Y = VFNMS(LDK(KP980785280), T1X, TZ);
368 T2f = VFMA(LDK(KP980785280), T1X, TZ);
369 T2o = VFMA(LDK(KP980785280), T2l, T2k);
370 T2m = VFNMS(LDK(KP980785280), T2l, T2k);
371 T2e = VFNMS(LDK(KP980785280), T2d, T2a);
372 T2g = VFMA(LDK(KP980785280), T2d, T2a);
373 T2r = VFMA(LDK(KP923879532), T2q, T2p);
374 T2J = VFNMS(LDK(KP923879532), T2q, T2p);
375 T2K = VADD(T2D, T2E);
376 T2F = VSUB(T2D, T2E);
377 }
378 ST(&(x[WS(rs, 23)]), VFMAI(T2m, T2j), ms, &(x[WS(rs, 1)]));
379 ST(&(x[WS(rs, 9)]), VFNMSI(T2m, T2j), ms, &(x[WS(rs, 1)]));
380 ST(&(x[WS(rs, 25)]), VFNMSI(T2o, T2n), ms, &(x[WS(rs, 1)]));
381 ST(&(x[WS(rs, 7)]), VFMAI(T2o, T2n), ms, &(x[WS(rs, 1)]));
382 ST(&(x[WS(rs, 31)]), VFMAI(T2g, T2f), ms, &(x[WS(rs, 1)]));
383 ST(&(x[WS(rs, 1)]), VFNMSI(T2g, T2f), ms, &(x[WS(rs, 1)]));
384 ST(&(x[WS(rs, 15)]), VFMAI(T2e, T1Y), ms, &(x[WS(rs, 1)]));
385 ST(&(x[WS(rs, 17)]), VFNMSI(T2e, T1Y), ms, &(x[WS(rs, 1)]));
386 }
387 T2L = VFMA(LDK(KP831469612), T2K, T2J);
388 T2P = VFNMS(LDK(KP831469612), T2K, T2J);
389 }
390 }
391 }
392 }
393 }
394 {
395 V T2y, T2N, T2C, T2M;
396 T2y = VADD(T2u, T2x);
397 T2N = VSUB(T2x, T2u);
398 T2C = VFMA(LDK(KP923879532), T2B, T2A);
399 T2M = VFNMS(LDK(KP923879532), T2B, T2A);
400 {
401 V T2z, T2H, T2Q, T2O, T2G, T2I;
402 T2z = VFNMS(LDK(KP831469612), T2y, T2r);
403 T2H = VFMA(LDK(KP831469612), T2y, T2r);
404 T2Q = VFNMS(LDK(KP831469612), T2N, T2M);
405 T2O = VFMA(LDK(KP831469612), T2N, T2M);
406 T2G = VFNMS(LDK(KP831469612), T2F, T2C);
407 T2I = VFMA(LDK(KP831469612), T2F, T2C);
408 ST(&(x[WS(rs, 21)]), VFNMSI(T2O, T2L), ms, &(x[WS(rs, 1)]));
409 ST(&(x[WS(rs, 11)]), VFMAI(T2O, T2L), ms, &(x[WS(rs, 1)]));
410 ST(&(x[WS(rs, 27)]), VFMAI(T2Q, T2P), ms, &(x[WS(rs, 1)]));
411 ST(&(x[WS(rs, 5)]), VFNMSI(T2Q, T2P), ms, &(x[WS(rs, 1)]));
412 ST(&(x[WS(rs, 3)]), VFMAI(T2I, T2H), ms, &(x[WS(rs, 1)]));
413 ST(&(x[WS(rs, 29)]), VFNMSI(T2I, T2H), ms, &(x[WS(rs, 1)]));
414 ST(&(x[WS(rs, 19)]), VFMAI(T2G, T2z), ms, &(x[WS(rs, 1)]));
415 ST(&(x[WS(rs, 13)]), VFNMSI(T2G, T2z), ms, &(x[WS(rs, 1)]));
416 }
417 }
418 }
419 }
420 VLEAVE();
421 }
422
423 static const tw_instr twinstr[] = {
424 VTW(0, 1),
425 VTW(0, 3),
426 VTW(0, 9),
427 VTW(0, 27),
428 {TW_NEXT, VL, 0}
429 };
430
431 static const ct_desc desc = { 32, XSIMD_STRING("t3fv_32"), twinstr, &GENUS, {146, 116, 98, 0}, 0, 0, 0 };
432
433 void XSIMD(codelet_t3fv_32) (planner *p) {
434 X(kdft_dit_register) (p, t3fv_32, &desc);
435 }
436 #else /* HAVE_FMA */
437
438 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 32 -name t3fv_32 -include t3f.h */
439
440 /*
441 * This function contains 244 FP additions, 158 FP multiplications,
442 * (or, 228 additions, 142 multiplications, 16 fused multiply/add),
443 * 90 stack variables, 7 constants, and 64 memory accesses
444 */
445 #include "t3f.h"
446
447 static void t3fv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
448 {
449 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
450 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
451 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
452 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
453 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
454 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
455 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
456 {
457 INT m;
458 R *x;
459 x = ri;
460 for (m = mb, W = W + (mb * ((TWVL / VL) * 8)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(32, rs)) {
461 V T2, T5, T3, T4, Tc, T1C, TP, Tz, Tn, T6, TS, Tf, TK, T7, T8;
462 V Tv, T1w, T22, Tg, Tk, T1D, T1R, TC, T18, T12, T1t, TH, TL, TT, T1n;
463 V T1c;
464 T2 = LDW(&(W[0]));
465 T5 = LDW(&(W[TWVL * 4]));
466 T3 = LDW(&(W[TWVL * 2]));
467 T4 = VZMULJ(T2, T3);
468 Tc = VZMUL(T2, T3);
469 T1C = VZMULJ(T2, T5);
470 TP = VZMULJ(T3, T5);
471 Tz = VZMUL(T2, T5);
472 Tn = VZMUL(T3, T5);
473 T6 = VZMUL(T4, T5);
474 TS = VZMULJ(Tc, T5);
475 Tf = VZMULJ(T4, T5);
476 TK = VZMUL(Tc, T5);
477 T7 = LDW(&(W[TWVL * 6]));
478 T8 = VZMULJ(T6, T7);
479 Tv = VZMULJ(T5, T7);
480 T1w = VZMULJ(Tn, T7);
481 T22 = VZMULJ(T3, T7);
482 Tg = VZMULJ(Tf, T7);
483 Tk = VZMUL(T2, T7);
484 T1D = VZMULJ(T1C, T7);
485 T1R = VZMULJ(Tc, T7);
486 TC = VZMULJ(T2, T7);
487 T18 = VZMULJ(TP, T7);
488 T12 = VZMULJ(Tz, T7);
489 T1t = VZMUL(Tc, T7);
490 TH = VZMUL(T3, T7);
491 TL = VZMULJ(TK, T7);
492 TT = VZMULJ(TS, T7);
493 T1n = VZMULJ(T4, T7);
494 T1c = VZMUL(T4, T7);
495 {
496 V Tb, T25, T2T, T3x, Tr, T1Z, T2W, T3K, TX, T27, T3g, T3z, TG, T28, T3j;
497 V T3y, T1N, T2v, T3a, T3G, T1V, T2w, T37, T3F, T1j, T2s, T33, T3D, T1r, T2t;
498 V T30, T3C;
499 {
500 V T1, T24, Ta, T21, T23, T9, T20, T2R, T2S;
501 T1 = LD(&(x[0]), ms, &(x[0]));
502 T23 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
503 T24 = VZMULJ(T22, T23);
504 T9 = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
505 Ta = VZMULJ(T8, T9);
506 T20 = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
507 T21 = VZMULJ(T1C, T20);
508 Tb = VSUB(T1, Ta);
509 T25 = VSUB(T21, T24);
510 T2R = VADD(T1, Ta);
511 T2S = VADD(T21, T24);
512 T2T = VADD(T2R, T2S);
513 T3x = VSUB(T2R, T2S);
514 }
515 {
516 V Te, Tp, Ti, Tm;
517 {
518 V Td, To, Th, Tl;
519 Td = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
520 Te = VZMULJ(Tc, Td);
521 To = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
522 Tp = VZMULJ(Tn, To);
523 Th = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
524 Ti = VZMULJ(Tg, Th);
525 Tl = LD(&(x[WS(rs, 28)]), ms, &(x[0]));
526 Tm = VZMULJ(Tk, Tl);
527 }
528 {
529 V Tj, Tq, T2U, T2V;
530 Tj = VSUB(Te, Ti);
531 Tq = VSUB(Tm, Tp);
532 Tr = VMUL(LDK(KP707106781), VADD(Tj, Tq));
533 T1Z = VMUL(LDK(KP707106781), VSUB(Tq, Tj));
534 T2U = VADD(Te, Ti);
535 T2V = VADD(Tm, Tp);
536 T2W = VADD(T2U, T2V);
537 T3K = VSUB(T2V, T2U);
538 }
539 }
540 {
541 V TJ, TV, TN, TR;
542 {
543 V TI, TU, TM, TQ;
544 TI = LD(&(x[WS(rs, 30)]), ms, &(x[0]));
545 TJ = VZMULJ(TH, TI);
546 TU = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
547 TV = VZMULJ(TT, TU);
548 TM = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
549 TN = VZMULJ(TL, TM);
550 TQ = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
551 TR = VZMULJ(TP, TQ);
552 }
553 {
554 V TO, TW, T3e, T3f;
555 TO = VSUB(TJ, TN);
556 TW = VSUB(TR, TV);
557 TX = VFMA(LDK(KP923879532), TO, VMUL(LDK(KP382683432), TW));
558 T27 = VFNMS(LDK(KP923879532), TW, VMUL(LDK(KP382683432), TO));
559 T3e = VADD(TJ, TN);
560 T3f = VADD(TR, TV);
561 T3g = VADD(T3e, T3f);
562 T3z = VSUB(T3e, T3f);
563 }
564 }
565 {
566 V Tu, TE, Tx, TB;
567 {
568 V Tt, TD, Tw, TA;
569 Tt = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
570 Tu = VZMULJ(T4, Tt);
571 TD = LD(&(x[WS(rs, 26)]), ms, &(x[0]));
572 TE = VZMULJ(TC, TD);
573 Tw = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
574 Tx = VZMULJ(Tv, Tw);
575 TA = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
576 TB = VZMULJ(Tz, TA);
577 }
578 {
579 V Ty, TF, T3h, T3i;
580 Ty = VSUB(Tu, Tx);
581 TF = VSUB(TB, TE);
582 TG = VFNMS(LDK(KP382683432), TF, VMUL(LDK(KP923879532), Ty));
583 T28 = VFMA(LDK(KP382683432), Ty, VMUL(LDK(KP923879532), TF));
584 T3h = VADD(Tu, Tx);
585 T3i = VADD(TB, TE);
586 T3j = VADD(T3h, T3i);
587 T3y = VSUB(T3h, T3i);
588 }
589 }
590 {
591 V T1v, T1y, T1T, T1Q, T1I, T1K, T1L, T1B, T1F, T1G;
592 {
593 V T1u, T1x, T1S, T1P;
594 T1u = LD(&(x[WS(rs, 31)]), ms, &(x[WS(rs, 1)]));
595 T1v = VZMULJ(T1t, T1u);
596 T1x = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
597 T1y = VZMULJ(T1w, T1x);
598 T1S = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
599 T1T = VZMULJ(T1R, T1S);
600 T1P = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
601 T1Q = VZMULJ(Tf, T1P);
602 {
603 V T1H, T1J, T1A, T1E;
604 T1H = LD(&(x[WS(rs, 27)]), ms, &(x[WS(rs, 1)]));
605 T1I = VZMULJ(T7, T1H);
606 T1J = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
607 T1K = VZMULJ(T6, T1J);
608 T1L = VSUB(T1I, T1K);
609 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
610 T1B = VZMULJ(T3, T1A);
611 T1E = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
612 T1F = VZMULJ(T1D, T1E);
613 T1G = VSUB(T1B, T1F);
614 }
615 }
616 {
617 V T1z, T1M, T38, T39;
618 T1z = VSUB(T1v, T1y);
619 T1M = VMUL(LDK(KP707106781), VADD(T1G, T1L));
620 T1N = VADD(T1z, T1M);
621 T2v = VSUB(T1z, T1M);
622 T38 = VADD(T1B, T1F);
623 T39 = VADD(T1I, T1K);
624 T3a = VADD(T38, T39);
625 T3G = VSUB(T39, T38);
626 }
627 {
628 V T1O, T1U, T35, T36;
629 T1O = VMUL(LDK(KP707106781), VSUB(T1L, T1G));
630 T1U = VSUB(T1Q, T1T);
631 T1V = VSUB(T1O, T1U);
632 T2w = VADD(T1U, T1O);
633 T35 = VADD(T1v, T1y);
634 T36 = VADD(T1Q, T1T);
635 T37 = VADD(T35, T36);
636 T3F = VSUB(T35, T36);
637 }
638 }
639 {
640 V T11, T14, T1p, T1m, T1e, T1g, T1h, T17, T1a, T1b;
641 {
642 V T10, T13, T1o, T1l;
643 T10 = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
644 T11 = VZMULJ(T2, T10);
645 T13 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
646 T14 = VZMULJ(T12, T13);
647 T1o = LD(&(x[WS(rs, 25)]), ms, &(x[WS(rs, 1)]));
648 T1p = VZMULJ(T1n, T1o);
649 T1l = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
650 T1m = VZMULJ(T5, T1l);
651 {
652 V T1d, T1f, T16, T19;
653 T1d = LD(&(x[WS(rs, 29)]), ms, &(x[WS(rs, 1)]));
654 T1e = VZMULJ(T1c, T1d);
655 T1f = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
656 T1g = VZMULJ(TK, T1f);
657 T1h = VSUB(T1e, T1g);
658 T16 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
659 T17 = VZMULJ(TS, T16);
660 T19 = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
661 T1a = VZMULJ(T18, T19);
662 T1b = VSUB(T17, T1a);
663 }
664 }
665 {
666 V T15, T1i, T31, T32;
667 T15 = VSUB(T11, T14);
668 T1i = VMUL(LDK(KP707106781), VADD(T1b, T1h));
669 T1j = VADD(T15, T1i);
670 T2s = VSUB(T15, T1i);
671 T31 = VADD(T17, T1a);
672 T32 = VADD(T1e, T1g);
673 T33 = VADD(T31, T32);
674 T3D = VSUB(T32, T31);
675 }
676 {
677 V T1k, T1q, T2Y, T2Z;
678 T1k = VMUL(LDK(KP707106781), VSUB(T1h, T1b));
679 T1q = VSUB(T1m, T1p);
680 T1r = VSUB(T1k, T1q);
681 T2t = VADD(T1q, T1k);
682 T2Y = VADD(T11, T14);
683 T2Z = VADD(T1m, T1p);
684 T30 = VADD(T2Y, T2Z);
685 T3C = VSUB(T2Y, T2Z);
686 }
687 }
688 {
689 V T3r, T3v, T3u, T3w;
690 {
691 V T3p, T3q, T3s, T3t;
692 T3p = VADD(T2T, T2W);
693 T3q = VADD(T3j, T3g);
694 T3r = VADD(T3p, T3q);
695 T3v = VSUB(T3p, T3q);
696 T3s = VADD(T30, T33);
697 T3t = VADD(T37, T3a);
698 T3u = VADD(T3s, T3t);
699 T3w = VBYI(VSUB(T3t, T3s));
700 }
701 ST(&(x[WS(rs, 16)]), VSUB(T3r, T3u), ms, &(x[0]));
702 ST(&(x[WS(rs, 8)]), VADD(T3v, T3w), ms, &(x[0]));
703 ST(&(x[0]), VADD(T3r, T3u), ms, &(x[0]));
704 ST(&(x[WS(rs, 24)]), VSUB(T3v, T3w), ms, &(x[0]));
705 }
706 {
707 V T2X, T3k, T3c, T3l, T34, T3b;
708 T2X = VSUB(T2T, T2W);
709 T3k = VSUB(T3g, T3j);
710 T34 = VSUB(T30, T33);
711 T3b = VSUB(T37, T3a);
712 T3c = VMUL(LDK(KP707106781), VADD(T34, T3b));
713 T3l = VMUL(LDK(KP707106781), VSUB(T3b, T34));
714 {
715 V T3d, T3m, T3n, T3o;
716 T3d = VADD(T2X, T3c);
717 T3m = VBYI(VADD(T3k, T3l));
718 ST(&(x[WS(rs, 28)]), VSUB(T3d, T3m), ms, &(x[0]));
719 ST(&(x[WS(rs, 4)]), VADD(T3d, T3m), ms, &(x[0]));
720 T3n = VSUB(T2X, T3c);
721 T3o = VBYI(VSUB(T3l, T3k));
722 ST(&(x[WS(rs, 20)]), VSUB(T3n, T3o), ms, &(x[0]));
723 ST(&(x[WS(rs, 12)]), VADD(T3n, T3o), ms, &(x[0]));
724 }
725 }
726 {
727 V T3B, T3W, T3M, T3U, T3I, T3T, T3P, T3X, T3A, T3L;
728 T3A = VMUL(LDK(KP707106781), VADD(T3y, T3z));
729 T3B = VADD(T3x, T3A);
730 T3W = VSUB(T3x, T3A);
731 T3L = VMUL(LDK(KP707106781), VSUB(T3z, T3y));
732 T3M = VADD(T3K, T3L);
733 T3U = VSUB(T3L, T3K);
734 {
735 V T3E, T3H, T3N, T3O;
736 T3E = VFMA(LDK(KP923879532), T3C, VMUL(LDK(KP382683432), T3D));
737 T3H = VFNMS(LDK(KP382683432), T3G, VMUL(LDK(KP923879532), T3F));
738 T3I = VADD(T3E, T3H);
739 T3T = VSUB(T3H, T3E);
740 T3N = VFNMS(LDK(KP382683432), T3C, VMUL(LDK(KP923879532), T3D));
741 T3O = VFMA(LDK(KP382683432), T3F, VMUL(LDK(KP923879532), T3G));
742 T3P = VADD(T3N, T3O);
743 T3X = VSUB(T3O, T3N);
744 }
745 {
746 V T3J, T3Q, T3Z, T40;
747 T3J = VADD(T3B, T3I);
748 T3Q = VBYI(VADD(T3M, T3P));
749 ST(&(x[WS(rs, 30)]), VSUB(T3J, T3Q), ms, &(x[0]));
750 ST(&(x[WS(rs, 2)]), VADD(T3J, T3Q), ms, &(x[0]));
751 T3Z = VBYI(VADD(T3U, T3T));
752 T40 = VADD(T3W, T3X);
753 ST(&(x[WS(rs, 6)]), VADD(T3Z, T40), ms, &(x[0]));
754 ST(&(x[WS(rs, 26)]), VSUB(T40, T3Z), ms, &(x[0]));
755 }
756 {
757 V T3R, T3S, T3V, T3Y;
758 T3R = VSUB(T3B, T3I);
759 T3S = VBYI(VSUB(T3P, T3M));
760 ST(&(x[WS(rs, 18)]), VSUB(T3R, T3S), ms, &(x[0]));
761 ST(&(x[WS(rs, 14)]), VADD(T3R, T3S), ms, &(x[0]));
762 T3V = VBYI(VSUB(T3T, T3U));
763 T3Y = VSUB(T3W, T3X);
764 ST(&(x[WS(rs, 10)]), VADD(T3V, T3Y), ms, &(x[0]));
765 ST(&(x[WS(rs, 22)]), VSUB(T3Y, T3V), ms, &(x[0]));
766 }
767 }
768 {
769 V TZ, T2k, T2d, T2l, T1X, T2h, T2a, T2i;
770 {
771 V Ts, TY, T2b, T2c;
772 Ts = VADD(Tb, Tr);
773 TY = VADD(TG, TX);
774 TZ = VADD(Ts, TY);
775 T2k = VSUB(Ts, TY);
776 T2b = VFNMS(LDK(KP195090322), T1j, VMUL(LDK(KP980785280), T1r));
777 T2c = VFMA(LDK(KP195090322), T1N, VMUL(LDK(KP980785280), T1V));
778 T2d = VADD(T2b, T2c);
779 T2l = VSUB(T2c, T2b);
780 }
781 {
782 V T1s, T1W, T26, T29;
783 T1s = VFMA(LDK(KP980785280), T1j, VMUL(LDK(KP195090322), T1r));
784 T1W = VFNMS(LDK(KP195090322), T1V, VMUL(LDK(KP980785280), T1N));
785 T1X = VADD(T1s, T1W);
786 T2h = VSUB(T1W, T1s);
787 T26 = VSUB(T1Z, T25);
788 T29 = VSUB(T27, T28);
789 T2a = VADD(T26, T29);
790 T2i = VSUB(T29, T26);
791 }
792 {
793 V T1Y, T2e, T2n, T2o;
794 T1Y = VADD(TZ, T1X);
795 T2e = VBYI(VADD(T2a, T2d));
796 ST(&(x[WS(rs, 31)]), VSUB(T1Y, T2e), ms, &(x[WS(rs, 1)]));
797 ST(&(x[WS(rs, 1)]), VADD(T1Y, T2e), ms, &(x[WS(rs, 1)]));
798 T2n = VBYI(VADD(T2i, T2h));
799 T2o = VADD(T2k, T2l);
800 ST(&(x[WS(rs, 7)]), VADD(T2n, T2o), ms, &(x[WS(rs, 1)]));
801 ST(&(x[WS(rs, 25)]), VSUB(T2o, T2n), ms, &(x[WS(rs, 1)]));
802 }
803 {
804 V T2f, T2g, T2j, T2m;
805 T2f = VSUB(TZ, T1X);
806 T2g = VBYI(VSUB(T2d, T2a));
807 ST(&(x[WS(rs, 17)]), VSUB(T2f, T2g), ms, &(x[WS(rs, 1)]));
808 ST(&(x[WS(rs, 15)]), VADD(T2f, T2g), ms, &(x[WS(rs, 1)]));
809 T2j = VBYI(VSUB(T2h, T2i));
810 T2m = VSUB(T2k, T2l);
811 ST(&(x[WS(rs, 9)]), VADD(T2j, T2m), ms, &(x[WS(rs, 1)]));
812 ST(&(x[WS(rs, 23)]), VSUB(T2m, T2j), ms, &(x[WS(rs, 1)]));
813 }
814 }
815 {
816 V T2r, T2M, T2F, T2N, T2y, T2J, T2C, T2K;
817 {
818 V T2p, T2q, T2D, T2E;
819 T2p = VSUB(Tb, Tr);
820 T2q = VADD(T28, T27);
821 T2r = VADD(T2p, T2q);
822 T2M = VSUB(T2p, T2q);
823 T2D = VFNMS(LDK(KP555570233), T2s, VMUL(LDK(KP831469612), T2t));
824 T2E = VFMA(LDK(KP555570233), T2v, VMUL(LDK(KP831469612), T2w));
825 T2F = VADD(T2D, T2E);
826 T2N = VSUB(T2E, T2D);
827 }
828 {
829 V T2u, T2x, T2A, T2B;
830 T2u = VFMA(LDK(KP831469612), T2s, VMUL(LDK(KP555570233), T2t));
831 T2x = VFNMS(LDK(KP555570233), T2w, VMUL(LDK(KP831469612), T2v));
832 T2y = VADD(T2u, T2x);
833 T2J = VSUB(T2x, T2u);
834 T2A = VADD(T25, T1Z);
835 T2B = VSUB(TX, TG);
836 T2C = VADD(T2A, T2B);
837 T2K = VSUB(T2B, T2A);
838 }
839 {
840 V T2z, T2G, T2P, T2Q;
841 T2z = VADD(T2r, T2y);
842 T2G = VBYI(VADD(T2C, T2F));
843 ST(&(x[WS(rs, 29)]), VSUB(T2z, T2G), ms, &(x[WS(rs, 1)]));
844 ST(&(x[WS(rs, 3)]), VADD(T2z, T2G), ms, &(x[WS(rs, 1)]));
845 T2P = VBYI(VADD(T2K, T2J));
846 T2Q = VADD(T2M, T2N);
847 ST(&(x[WS(rs, 5)]), VADD(T2P, T2Q), ms, &(x[WS(rs, 1)]));
848 ST(&(x[WS(rs, 27)]), VSUB(T2Q, T2P), ms, &(x[WS(rs, 1)]));
849 }
850 {
851 V T2H, T2I, T2L, T2O;
852 T2H = VSUB(T2r, T2y);
853 T2I = VBYI(VSUB(T2F, T2C));
854 ST(&(x[WS(rs, 19)]), VSUB(T2H, T2I), ms, &(x[WS(rs, 1)]));
855 ST(&(x[WS(rs, 13)]), VADD(T2H, T2I), ms, &(x[WS(rs, 1)]));
856 T2L = VBYI(VSUB(T2J, T2K));
857 T2O = VSUB(T2M, T2N);
858 ST(&(x[WS(rs, 11)]), VADD(T2L, T2O), ms, &(x[WS(rs, 1)]));
859 ST(&(x[WS(rs, 21)]), VSUB(T2O, T2L), ms, &(x[WS(rs, 1)]));
860 }
861 }
862 }
863 }
864 }
865 VLEAVE();
866 }
867
868 static const tw_instr twinstr[] = {
869 VTW(0, 1),
870 VTW(0, 3),
871 VTW(0, 9),
872 VTW(0, 27),
873 {TW_NEXT, VL, 0}
874 };
875
876 static const ct_desc desc = { 32, XSIMD_STRING("t3fv_32"), twinstr, &GENUS, {228, 142, 16, 0}, 0, 0, 0 };
877
878 void XSIMD(codelet_t3fv_32) (planner *p) {
879 X(kdft_dit_register) (p, t3fv_32, &desc);
880 }
881 #endif /* HAVE_FMA */