comparison src/fftw-3.3.3/dft/simd/common/t3bv_32.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 07:39:19 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 t3bv_32 -include t3b.h -sign 1 */
29
30 /*
31 * This function contains 244 FP additions, 214 FP multiplications,
32 * (or, 146 additions, 116 multiplications, 98 fused multiply/add),
33 * 120 stack variables, 7 constants, and 64 memory accesses
34 */
35 #include "t3b.h"
36
37 static void t3bv_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 = ii;
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, T2F, T2N, T2H, T2z, T2P, T2L, T2C, T2M;
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, T3j, T28, TX, T3y, T3g, 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 = VZMUL(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 = VZMUL(T21, T22);
100 Te = VZMUL(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 = VZMUL(Tn, To);
113 TU = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
114 Ti = VZMUL(Tg, Th);
115 Ta = VZMUL(T8, T9);
116 TM = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
117 TR = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
118 Tm = VZMUL(Tk, Tl);
119 TJ = VZMUL(TH, TI);
120 }
121 {
122 V Tu, TE, Tw, TA;
123 {
124 V T3h, TO, T3i, 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 = VZMUL(TT, TU);
129 T2U = VADD(Te, Ti);
130 Tj = VSUB(Te, Ti);
131 T2R = VADD(T1, Ta);
132 Tb = VSUB(T1, Ta);
133 TN = VZMUL(TL, TM);
134 TS = VZMUL(TQ, TR);
135 T2V = VADD(Tm, Tp);
136 Tq = VSUB(Tm, Tp);
137 Tu = VZMUL(T4, Tt);
138 TD = LD(&(x[WS(rs, 26)]), ms, &(x[0]));
139 T3x = VSUB(T2R, T2S);
140 T2T = VADD(T2R, T2S);
141 T3h = VADD(TJ, TN);
142 TO = VSUB(TJ, TN);
143 T3i = VADD(TV, TS);
144 TW = VSUB(TS, TV);
145 T3K = VSUB(T2U, T2V);
146 T2W = VADD(T2U, T2V);
147 T25 = VSUB(Tj, Tq);
148 Tr = VADD(Tj, Tq);
149 TE = VZMUL(TC, TD);
150 }
151 Tw = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
152 T3z = VSUB(T3h, T3i);
153 T3j = VADD(T3h, T3i);
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, T3e, T1S, T1Q, T1I, T3f, 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 = VZMUL(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 = VZMUL(Tz, TA);
174 T1v = VZMUL(T1t, T1u);
175 T1y = VZMUL(T1w, T1x);
176 Ty = VSUB(Tu, Tx);
177 T3e = VADD(Tu, Tx);
178 T1S = VZMUL(Tf, T1R);
179 T1Q = VZMUL(T1O, T1P);
180 T1I = VZMUL(T7, T1H);
181 T3f = VADD(TB, TE);
182 TF = VSUB(TB, TE);
183 T1K = VZMUL(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(T3e, T3f);
193 T3g = VADD(T3e, T3f);
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 = VZMUL(T3, T1A);
199 T1F = VZMUL(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 = VZMUL(T2, T10);
218 T14 = VZMUL(T12, T13);
219 T1o = VZMUL(T1m, T1n);
220 T1l = VZMUL(T5, T1k);
221 T1e = VZMUL(T1c, T1d);
222 T3G = VSUB(T39, T38);
223 T3a = VADD(T38, T39);
224 T1U = VSUB(T1L, T1G);
225 T1M = VADD(T1G, T1L);
226 T1g = VZMUL(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 = VZMUL(TP, T16);
242 T1a = VZMUL(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(T3g, T3j);
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 = VSUB(T3p, T3q);
260 T3v = VADD(T3p, T3q);
261 {
262 V T1r, T2t, T1j, T2s, T3S, T3Y, T3R, T3V;
263 {
264 V T3B, T3T, T3M, T3W, T3U, T3P, T3X, T3I, T3l, T3c, T3w, T3u;
265 {
266 V T3L, T3A, T33, T3D, T1i, T1q, T3O, T3H;
267 T3L = VSUB(T3y, T3z);
268 T3A = VADD(T3y, T3z);
269 T33 = VADD(T31, T32);
270 T3D = VSUB(T31, T32);
271 T1i = VADD(T1b, T1h);
272 T1q = VSUB(T1b, T1h);
273 T3O = VFMA(LDK(KP414213562), T3F, T3G);
274 T3H = VFNMS(LDK(KP414213562), T3G, T3F);
275 T3B = VFMA(LDK(KP707106781), T3A, T3x);
276 T3T = VFNMS(LDK(KP707106781), T3A, T3x);
277 T3M = VFMA(LDK(KP707106781), T3L, T3K);
278 T3W = VFNMS(LDK(KP707106781), T3L, T3K);
279 {
280 V T3E, T3N, T3s, T34;
281 T3E = VFNMS(LDK(KP414213562), T3D, T3C);
282 T3N = VFMA(LDK(KP414213562), T3C, T3D);
283 T3s = VADD(T30, T33);
284 T34 = VSUB(T30, T33);
285 T1r = VFMA(LDK(KP707106781), T1q, T1p);
286 T2t = VFNMS(LDK(KP707106781), T1q, T1p);
287 T1j = VFMA(LDK(KP707106781), T1i, T15);
288 T2s = VFNMS(LDK(KP707106781), T1i, T15);
289 T3U = VADD(T3N, T3O);
290 T3P = VSUB(T3N, T3O);
291 T3X = VSUB(T3E, T3H);
292 T3I = VADD(T3E, T3H);
293 T3l = VSUB(T34, T3b);
294 T3c = VADD(T34, T3b);
295 T3w = VADD(T3s, T3t);
296 T3u = VSUB(T3s, T3t);
297 }
298 }
299 {
300 V T40, T3Z, T3Q, T3J;
301 T3S = VFMA(LDK(KP923879532), T3P, T3M);
302 T3Q = VFNMS(LDK(KP923879532), T3P, T3M);
303 T40 = VFNMS(LDK(KP923879532), T3X, T3W);
304 T3Y = VFMA(LDK(KP923879532), T3X, T3W);
305 T3R = VFMA(LDK(KP923879532), T3I, T3B);
306 T3J = VFNMS(LDK(KP923879532), T3I, T3B);
307 {
308 V T3o, T3m, T3n, T3d;
309 T3o = VFMA(LDK(KP707106781), T3l, T3k);
310 T3m = VFNMS(LDK(KP707106781), T3l, T3k);
311 T3n = VFMA(LDK(KP707106781), T3c, T2X);
312 T3d = VFNMS(LDK(KP707106781), T3c, T2X);
313 ST(&(x[WS(rs, 16)]), VSUB(T3v, T3w), ms, &(x[0]));
314 ST(&(x[0]), VADD(T3v, T3w), ms, &(x[0]));
315 ST(&(x[WS(rs, 8)]), VFMAI(T3u, T3r), ms, &(x[0]));
316 ST(&(x[WS(rs, 24)]), VFNMSI(T3u, T3r), ms, &(x[0]));
317 T3Z = VFMA(LDK(KP923879532), T3U, T3T);
318 T3V = VFNMS(LDK(KP923879532), T3U, T3T);
319 ST(&(x[WS(rs, 18)]), VFMAI(T3Q, T3J), ms, &(x[0]));
320 ST(&(x[WS(rs, 14)]), VFNMSI(T3Q, T3J), ms, &(x[0]));
321 ST(&(x[WS(rs, 28)]), VFNMSI(T3o, T3n), ms, &(x[0]));
322 ST(&(x[WS(rs, 4)]), VFMAI(T3o, T3n), ms, &(x[0]));
323 ST(&(x[WS(rs, 20)]), VFMAI(T3m, T3d), ms, &(x[0]));
324 ST(&(x[WS(rs, 12)]), VFNMSI(T3m, T3d), ms, &(x[0]));
325 }
326 ST(&(x[WS(rs, 26)]), VFMAI(T40, T3Z), ms, &(x[0]));
327 ST(&(x[WS(rs, 6)]), VFNMSI(T40, T3Z), ms, &(x[0]));
328 }
329 }
330 {
331 V T2p, T1s, T1W, T2h, TZ, T2i, T2d, T26, T29, T2q;
332 {
333 V Ts, TY, T2b, T2c;
334 T2p = VFNMS(LDK(KP707106781), Tr, Tb);
335 Ts = VFMA(LDK(KP707106781), Tr, Tb);
336 TY = VADD(TG, TX);
337 T2B = VSUB(TG, TX);
338 T1s = VFNMS(LDK(KP198912367), T1r, T1j);
339 T2b = VFMA(LDK(KP198912367), T1j, T1r);
340 T2c = VFMA(LDK(KP198912367), T1N, T1V);
341 T1W = VFNMS(LDK(KP198912367), T1V, T1N);
342 ST(&(x[WS(rs, 2)]), VFMAI(T3S, T3R), ms, &(x[0]));
343 ST(&(x[WS(rs, 30)]), VFNMSI(T3S, T3R), ms, &(x[0]));
344 ST(&(x[WS(rs, 22)]), VFNMSI(T3Y, T3V), ms, &(x[0]));
345 ST(&(x[WS(rs, 10)]), VFMAI(T3Y, T3V), ms, &(x[0]));
346 T2h = VFNMS(LDK(KP923879532), TY, Ts);
347 TZ = VFMA(LDK(KP923879532), TY, Ts);
348 T2i = VADD(T2b, T2c);
349 T2d = VSUB(T2b, T2c);
350 T2A = VFNMS(LDK(KP707106781), T25, T24);
351 T26 = VFMA(LDK(KP707106781), T25, T24);
352 T29 = VSUB(T27, T28);
353 T2q = VADD(T27, T28);
354 }
355 {
356 V T2J, T2r, T2K, T2y;
357 {
358 V T2u, T2D, T2j, T2n, T2l, T1X, T2k, T2a, T2E, T2x;
359 T2u = VFMA(LDK(KP668178637), T2t, T2s);
360 T2D = VFNMS(LDK(KP668178637), T2s, T2t);
361 T2j = VFNMS(LDK(KP980785280), T2i, T2h);
362 T2n = VFMA(LDK(KP980785280), T2i, T2h);
363 T2l = VSUB(T1s, T1W);
364 T1X = VADD(T1s, T1W);
365 T2k = VFNMS(LDK(KP923879532), T29, T26);
366 T2a = VFMA(LDK(KP923879532), T29, T26);
367 T2J = VFNMS(LDK(KP923879532), T2q, T2p);
368 T2r = VFMA(LDK(KP923879532), T2q, T2p);
369 T2E = VFNMS(LDK(KP668178637), T2v, T2w);
370 T2x = VFMA(LDK(KP668178637), T2w, T2v);
371 {
372 V T1Y, T2f, T2o, T2m, T2e, T2g;
373 T1Y = VFNMS(LDK(KP980785280), T1X, TZ);
374 T2f = VFMA(LDK(KP980785280), T1X, TZ);
375 T2o = VFNMS(LDK(KP980785280), T2l, T2k);
376 T2m = VFMA(LDK(KP980785280), T2l, T2k);
377 T2e = VFNMS(LDK(KP980785280), T2d, T2a);
378 T2g = VFMA(LDK(KP980785280), T2d, T2a);
379 T2F = VSUB(T2D, T2E);
380 T2K = VADD(T2D, T2E);
381 T2N = VSUB(T2u, T2x);
382 T2y = VADD(T2u, T2x);
383 ST(&(x[WS(rs, 23)]), VFNMSI(T2m, T2j), ms, &(x[WS(rs, 1)]));
384 ST(&(x[WS(rs, 9)]), VFMAI(T2m, T2j), ms, &(x[WS(rs, 1)]));
385 ST(&(x[WS(rs, 25)]), VFMAI(T2o, T2n), ms, &(x[WS(rs, 1)]));
386 ST(&(x[WS(rs, 7)]), VFNMSI(T2o, T2n), ms, &(x[WS(rs, 1)]));
387 ST(&(x[WS(rs, 1)]), VFMAI(T2g, T2f), ms, &(x[WS(rs, 1)]));
388 ST(&(x[WS(rs, 31)]), VFNMSI(T2g, T2f), ms, &(x[WS(rs, 1)]));
389 ST(&(x[WS(rs, 17)]), VFMAI(T2e, T1Y), ms, &(x[WS(rs, 1)]));
390 ST(&(x[WS(rs, 15)]), VFNMSI(T2e, T1Y), ms, &(x[WS(rs, 1)]));
391 }
392 }
393 T2H = VFMA(LDK(KP831469612), T2y, T2r);
394 T2z = VFNMS(LDK(KP831469612), T2y, T2r);
395 T2P = VFNMS(LDK(KP831469612), T2K, T2J);
396 T2L = VFMA(LDK(KP831469612), T2K, T2J);
397 }
398 }
399 }
400 }
401 }
402 }
403 T2C = VFNMS(LDK(KP923879532), T2B, T2A);
404 T2M = VFMA(LDK(KP923879532), T2B, T2A);
405 {
406 V T2Q, T2O, T2G, T2I;
407 T2Q = VFMA(LDK(KP831469612), T2N, T2M);
408 T2O = VFNMS(LDK(KP831469612), T2N, T2M);
409 T2G = VFNMS(LDK(KP831469612), T2F, T2C);
410 T2I = VFMA(LDK(KP831469612), T2F, T2C);
411 ST(&(x[WS(rs, 21)]), VFMAI(T2O, T2L), ms, &(x[WS(rs, 1)]));
412 ST(&(x[WS(rs, 11)]), VFNMSI(T2O, T2L), ms, &(x[WS(rs, 1)]));
413 ST(&(x[WS(rs, 27)]), VFNMSI(T2Q, T2P), ms, &(x[WS(rs, 1)]));
414 ST(&(x[WS(rs, 5)]), VFMAI(T2Q, T2P), ms, &(x[WS(rs, 1)]));
415 ST(&(x[WS(rs, 29)]), VFMAI(T2I, T2H), ms, &(x[WS(rs, 1)]));
416 ST(&(x[WS(rs, 3)]), VFNMSI(T2I, T2H), ms, &(x[WS(rs, 1)]));
417 ST(&(x[WS(rs, 13)]), VFMAI(T2G, T2z), ms, &(x[WS(rs, 1)]));
418 ST(&(x[WS(rs, 19)]), VFNMSI(T2G, T2z), ms, &(x[WS(rs, 1)]));
419 }
420 }
421 }
422 VLEAVE();
423 }
424
425 static const tw_instr twinstr[] = {
426 VTW(0, 1),
427 VTW(0, 3),
428 VTW(0, 9),
429 VTW(0, 27),
430 {TW_NEXT, VL, 0}
431 };
432
433 static const ct_desc desc = { 32, XSIMD_STRING("t3bv_32"), twinstr, &GENUS, {146, 116, 98, 0}, 0, 0, 0 };
434
435 void XSIMD(codelet_t3bv_32) (planner *p) {
436 X(kdft_dit_register) (p, t3bv_32, &desc);
437 }
438 #else /* HAVE_FMA */
439
440 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 32 -name t3bv_32 -include t3b.h -sign 1 */
441
442 /*
443 * This function contains 244 FP additions, 158 FP multiplications,
444 * (or, 228 additions, 142 multiplications, 16 fused multiply/add),
445 * 90 stack variables, 7 constants, and 64 memory accesses
446 */
447 #include "t3b.h"
448
449 static void t3bv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
450 {
451 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
452 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
453 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
454 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
455 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
456 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
457 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
458 {
459 INT m;
460 R *x;
461 x = ii;
462 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)) {
463 V T2, T5, T3, T4, Tc, T1v, TH, Tz, Tn, T6, TS, Tf, TK, T7, T8;
464 V Tv, T1I, T25, Tg, Tk, T1N, T1Q, TC, T16, T12, T1w, TL, TP, TT, T1m;
465 V T1f;
466 T2 = LDW(&(W[0]));
467 T5 = LDW(&(W[TWVL * 4]));
468 T3 = LDW(&(W[TWVL * 2]));
469 T4 = VZMULJ(T2, T3);
470 Tc = VZMUL(T2, T3);
471 T1v = VZMULJ(T2, T5);
472 TH = VZMULJ(T3, T5);
473 Tz = VZMUL(T2, T5);
474 Tn = VZMUL(T3, T5);
475 T6 = VZMUL(T4, T5);
476 TS = VZMUL(Tc, T5);
477 Tf = VZMULJ(T4, T5);
478 TK = VZMULJ(Tc, T5);
479 T7 = LDW(&(W[TWVL * 6]));
480 T8 = VZMULJ(T6, T7);
481 Tv = VZMULJ(T5, T7);
482 T1I = VZMULJ(Tc, T7);
483 T25 = VZMULJ(T3, T7);
484 Tg = VZMULJ(Tf, T7);
485 Tk = VZMUL(T2, T7);
486 T1N = VZMUL(Tc, T7);
487 T1Q = VZMULJ(Tn, T7);
488 TC = VZMULJ(T2, T7);
489 T16 = VZMUL(T4, T7);
490 T12 = VZMULJ(TH, T7);
491 T1w = VZMULJ(T1v, T7);
492 TL = VZMULJ(TK, T7);
493 TP = VZMUL(T3, T7);
494 TT = VZMULJ(TS, T7);
495 T1m = VZMULJ(Tz, T7);
496 T1f = VZMULJ(T4, T7);
497 {
498 V Tb, T28, T3k, T3M, Tr, T22, T3f, T3N, TX, T20, T3b, T3J, TG, T1Z, T38;
499 V T3I, T1M, T2v, T33, T3F, T1V, T2w, T30, T3E, T1j, T2s, T2W, T3C, T1r, T2t;
500 V T2T, T3B;
501 {
502 V T1, T27, Ta, T24, T26, T9, T23, T3i, T3j;
503 T1 = LD(&(x[0]), ms, &(x[0]));
504 T26 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
505 T27 = VZMUL(T25, T26);
506 T9 = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
507 Ta = VZMUL(T8, T9);
508 T23 = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
509 T24 = VZMUL(T1v, T23);
510 Tb = VSUB(T1, Ta);
511 T28 = VSUB(T24, T27);
512 T3i = VADD(T1, Ta);
513 T3j = VADD(T24, T27);
514 T3k = VSUB(T3i, T3j);
515 T3M = VADD(T3i, T3j);
516 }
517 {
518 V Te, Tp, Ti, Tm;
519 {
520 V Td, To, Th, Tl;
521 Td = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
522 Te = VZMUL(Tc, Td);
523 To = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
524 Tp = VZMUL(Tn, To);
525 Th = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
526 Ti = VZMUL(Tg, Th);
527 Tl = LD(&(x[WS(rs, 28)]), ms, &(x[0]));
528 Tm = VZMUL(Tk, Tl);
529 }
530 {
531 V Tj, Tq, T3d, T3e;
532 Tj = VSUB(Te, Ti);
533 Tq = VSUB(Tm, Tp);
534 Tr = VMUL(LDK(KP707106781), VADD(Tj, Tq));
535 T22 = VMUL(LDK(KP707106781), VSUB(Tj, Tq));
536 T3d = VADD(Te, Ti);
537 T3e = VADD(Tm, Tp);
538 T3f = VSUB(T3d, T3e);
539 T3N = VADD(T3d, T3e);
540 }
541 }
542 {
543 V TJ, TV, TN, TR;
544 {
545 V TI, TU, TM, TQ;
546 TI = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
547 TJ = VZMUL(TH, TI);
548 TU = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
549 TV = VZMUL(TT, TU);
550 TM = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
551 TN = VZMUL(TL, TM);
552 TQ = LD(&(x[WS(rs, 30)]), ms, &(x[0]));
553 TR = VZMUL(TP, TQ);
554 }
555 {
556 V TO, TW, T39, T3a;
557 TO = VSUB(TJ, TN);
558 TW = VSUB(TR, TV);
559 TX = VFNMS(LDK(KP382683432), TW, VMUL(LDK(KP923879532), TO));
560 T20 = VFMA(LDK(KP923879532), TW, VMUL(LDK(KP382683432), TO));
561 T39 = VADD(TR, TV);
562 T3a = VADD(TJ, TN);
563 T3b = VSUB(T39, T3a);
564 T3J = VADD(T39, T3a);
565 }
566 }
567 {
568 V Tu, TE, Tx, TB;
569 {
570 V Tt, TD, Tw, TA;
571 Tt = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
572 Tu = VZMUL(T4, Tt);
573 TD = LD(&(x[WS(rs, 26)]), ms, &(x[0]));
574 TE = VZMUL(TC, TD);
575 Tw = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
576 Tx = VZMUL(Tv, Tw);
577 TA = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
578 TB = VZMUL(Tz, TA);
579 }
580 {
581 V Ty, TF, T36, T37;
582 Ty = VSUB(Tu, Tx);
583 TF = VSUB(TB, TE);
584 TG = VFMA(LDK(KP382683432), Ty, VMUL(LDK(KP923879532), TF));
585 T1Z = VFNMS(LDK(KP382683432), TF, VMUL(LDK(KP923879532), Ty));
586 T36 = VADD(Tu, Tx);
587 T37 = VADD(TB, TE);
588 T38 = VSUB(T36, T37);
589 T3I = VADD(T36, T37);
590 }
591 }
592 {
593 V T1H, T1K, T1S, T1P, T1B, T1D, T1E, T1u, T1y, T1z;
594 {
595 V T1G, T1J, T1R, T1O;
596 T1G = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
597 T1H = VZMUL(Tf, T1G);
598 T1J = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
599 T1K = VZMUL(T1I, T1J);
600 T1R = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
601 T1S = VZMUL(T1Q, T1R);
602 T1O = LD(&(x[WS(rs, 31)]), ms, &(x[WS(rs, 1)]));
603 T1P = VZMUL(T1N, T1O);
604 {
605 V T1A, T1C, T1t, T1x;
606 T1A = LD(&(x[WS(rs, 27)]), ms, &(x[WS(rs, 1)]));
607 T1B = VZMUL(T7, T1A);
608 T1C = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
609 T1D = VZMUL(T6, T1C);
610 T1E = VSUB(T1B, T1D);
611 T1t = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
612 T1u = VZMUL(T3, T1t);
613 T1x = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
614 T1y = VZMUL(T1w, T1x);
615 T1z = VSUB(T1u, T1y);
616 }
617 }
618 {
619 V T1F, T1L, T31, T32;
620 T1F = VMUL(LDK(KP707106781), VSUB(T1z, T1E));
621 T1L = VSUB(T1H, T1K);
622 T1M = VSUB(T1F, T1L);
623 T2v = VADD(T1L, T1F);
624 T31 = VADD(T1u, T1y);
625 T32 = VADD(T1B, T1D);
626 T33 = VSUB(T31, T32);
627 T3F = VADD(T31, T32);
628 }
629 {
630 V T1T, T1U, T2Y, T2Z;
631 T1T = VSUB(T1P, T1S);
632 T1U = VMUL(LDK(KP707106781), VADD(T1z, T1E));
633 T1V = VSUB(T1T, T1U);
634 T2w = VADD(T1T, T1U);
635 T2Y = VADD(T1P, T1S);
636 T2Z = VADD(T1H, T1K);
637 T30 = VSUB(T2Y, T2Z);
638 T3E = VADD(T2Y, T2Z);
639 }
640 }
641 {
642 V T1e, T1h, T1o, T1l, T18, T1a, T1b, T11, T14, T15;
643 {
644 V T1d, T1g, T1n, T1k;
645 T1d = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
646 T1e = VZMUL(T5, T1d);
647 T1g = LD(&(x[WS(rs, 25)]), ms, &(x[WS(rs, 1)]));
648 T1h = VZMUL(T1f, T1g);
649 T1n = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
650 T1o = VZMUL(T1m, T1n);
651 T1k = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
652 T1l = VZMUL(T2, T1k);
653 {
654 V T17, T19, T10, T13;
655 T17 = LD(&(x[WS(rs, 29)]), ms, &(x[WS(rs, 1)]));
656 T18 = VZMUL(T16, T17);
657 T19 = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
658 T1a = VZMUL(TS, T19);
659 T1b = VSUB(T18, T1a);
660 T10 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
661 T11 = VZMUL(TK, T10);
662 T13 = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
663 T14 = VZMUL(T12, T13);
664 T15 = VSUB(T11, T14);
665 }
666 }
667 {
668 V T1c, T1i, T2U, T2V;
669 T1c = VMUL(LDK(KP707106781), VSUB(T15, T1b));
670 T1i = VSUB(T1e, T1h);
671 T1j = VSUB(T1c, T1i);
672 T2s = VADD(T1i, T1c);
673 T2U = VADD(T11, T14);
674 T2V = VADD(T18, T1a);
675 T2W = VSUB(T2U, T2V);
676 T3C = VADD(T2U, T2V);
677 }
678 {
679 V T1p, T1q, T2R, T2S;
680 T1p = VSUB(T1l, T1o);
681 T1q = VMUL(LDK(KP707106781), VADD(T15, T1b));
682 T1r = VSUB(T1p, T1q);
683 T2t = VADD(T1p, T1q);
684 T2R = VADD(T1l, T1o);
685 T2S = VADD(T1e, T1h);
686 T2T = VSUB(T2R, T2S);
687 T3B = VADD(T2R, T2S);
688 }
689 }
690 {
691 V T3V, T3Z, T3Y, T40;
692 {
693 V T3T, T3U, T3W, T3X;
694 T3T = VADD(T3M, T3N);
695 T3U = VADD(T3I, T3J);
696 T3V = VSUB(T3T, T3U);
697 T3Z = VADD(T3T, T3U);
698 T3W = VADD(T3B, T3C);
699 T3X = VADD(T3E, T3F);
700 T3Y = VBYI(VSUB(T3W, T3X));
701 T40 = VADD(T3W, T3X);
702 }
703 ST(&(x[WS(rs, 24)]), VSUB(T3V, T3Y), ms, &(x[0]));
704 ST(&(x[0]), VADD(T3Z, T40), ms, &(x[0]));
705 ST(&(x[WS(rs, 8)]), VADD(T3V, T3Y), ms, &(x[0]));
706 ST(&(x[WS(rs, 16)]), VSUB(T3Z, T40), ms, &(x[0]));
707 }
708 {
709 V T3K, T3O, T3H, T3P, T3D, T3G;
710 T3K = VSUB(T3I, T3J);
711 T3O = VSUB(T3M, T3N);
712 T3D = VSUB(T3B, T3C);
713 T3G = VSUB(T3E, T3F);
714 T3H = VMUL(LDK(KP707106781), VSUB(T3D, T3G));
715 T3P = VMUL(LDK(KP707106781), VADD(T3D, T3G));
716 {
717 V T3L, T3Q, T3R, T3S;
718 T3L = VBYI(VSUB(T3H, T3K));
719 T3Q = VSUB(T3O, T3P);
720 ST(&(x[WS(rs, 12)]), VADD(T3L, T3Q), ms, &(x[0]));
721 ST(&(x[WS(rs, 20)]), VSUB(T3Q, T3L), ms, &(x[0]));
722 T3R = VBYI(VADD(T3K, T3H));
723 T3S = VADD(T3O, T3P);
724 ST(&(x[WS(rs, 4)]), VADD(T3R, T3S), ms, &(x[0]));
725 ST(&(x[WS(rs, 28)]), VSUB(T3S, T3R), ms, &(x[0]));
726 }
727 }
728 {
729 V T3g, T3w, T3m, T3t, T35, T3u, T3p, T3x, T3c, T3l;
730 T3c = VMUL(LDK(KP707106781), VSUB(T38, T3b));
731 T3g = VSUB(T3c, T3f);
732 T3w = VADD(T3f, T3c);
733 T3l = VMUL(LDK(KP707106781), VADD(T38, T3b));
734 T3m = VSUB(T3k, T3l);
735 T3t = VADD(T3k, T3l);
736 {
737 V T2X, T34, T3n, T3o;
738 T2X = VFNMS(LDK(KP382683432), T2W, VMUL(LDK(KP923879532), T2T));
739 T34 = VFMA(LDK(KP923879532), T30, VMUL(LDK(KP382683432), T33));
740 T35 = VSUB(T2X, T34);
741 T3u = VADD(T2X, T34);
742 T3n = VFMA(LDK(KP382683432), T2T, VMUL(LDK(KP923879532), T2W));
743 T3o = VFNMS(LDK(KP382683432), T30, VMUL(LDK(KP923879532), T33));
744 T3p = VSUB(T3n, T3o);
745 T3x = VADD(T3n, T3o);
746 }
747 {
748 V T3h, T3q, T3z, T3A;
749 T3h = VBYI(VSUB(T35, T3g));
750 T3q = VSUB(T3m, T3p);
751 ST(&(x[WS(rs, 10)]), VADD(T3h, T3q), ms, &(x[0]));
752 ST(&(x[WS(rs, 22)]), VSUB(T3q, T3h), ms, &(x[0]));
753 T3z = VSUB(T3t, T3u);
754 T3A = VBYI(VSUB(T3x, T3w));
755 ST(&(x[WS(rs, 18)]), VSUB(T3z, T3A), ms, &(x[0]));
756 ST(&(x[WS(rs, 14)]), VADD(T3z, T3A), ms, &(x[0]));
757 }
758 {
759 V T3r, T3s, T3v, T3y;
760 T3r = VBYI(VADD(T3g, T35));
761 T3s = VADD(T3m, T3p);
762 ST(&(x[WS(rs, 6)]), VADD(T3r, T3s), ms, &(x[0]));
763 ST(&(x[WS(rs, 26)]), VSUB(T3s, T3r), ms, &(x[0]));
764 T3v = VADD(T3t, T3u);
765 T3y = VBYI(VADD(T3w, T3x));
766 ST(&(x[WS(rs, 30)]), VSUB(T3v, T3y), ms, &(x[0]));
767 ST(&(x[WS(rs, 2)]), VADD(T3v, T3y), ms, &(x[0]));
768 }
769 }
770 {
771 V TZ, T2k, T2d, T2l, T1X, T2h, T2a, T2i;
772 {
773 V Ts, TY, T2b, T2c;
774 Ts = VSUB(Tb, Tr);
775 TY = VSUB(TG, TX);
776 TZ = VSUB(Ts, TY);
777 T2k = VADD(Ts, TY);
778 T2b = VFNMS(LDK(KP555570233), T1j, VMUL(LDK(KP831469612), T1r));
779 T2c = VFMA(LDK(KP555570233), T1M, VMUL(LDK(KP831469612), T1V));
780 T2d = VSUB(T2b, T2c);
781 T2l = VADD(T2b, T2c);
782 }
783 {
784 V T1s, T1W, T21, T29;
785 T1s = VFMA(LDK(KP831469612), T1j, VMUL(LDK(KP555570233), T1r));
786 T1W = VFNMS(LDK(KP555570233), T1V, VMUL(LDK(KP831469612), T1M));
787 T1X = VSUB(T1s, T1W);
788 T2h = VADD(T1s, T1W);
789 T21 = VSUB(T1Z, T20);
790 T29 = VSUB(T22, T28);
791 T2a = VSUB(T21, T29);
792 T2i = VADD(T29, T21);
793 }
794 {
795 V T1Y, T2e, T2n, T2o;
796 T1Y = VADD(TZ, T1X);
797 T2e = VBYI(VADD(T2a, T2d));
798 ST(&(x[WS(rs, 27)]), VSUB(T1Y, T2e), ms, &(x[WS(rs, 1)]));
799 ST(&(x[WS(rs, 5)]), VADD(T1Y, T2e), ms, &(x[WS(rs, 1)]));
800 T2n = VBYI(VADD(T2i, T2h));
801 T2o = VADD(T2k, T2l);
802 ST(&(x[WS(rs, 3)]), VADD(T2n, T2o), ms, &(x[WS(rs, 1)]));
803 ST(&(x[WS(rs, 29)]), VSUB(T2o, T2n), ms, &(x[WS(rs, 1)]));
804 }
805 {
806 V T2f, T2g, T2j, T2m;
807 T2f = VSUB(TZ, T1X);
808 T2g = VBYI(VSUB(T2d, T2a));
809 ST(&(x[WS(rs, 21)]), VSUB(T2f, T2g), ms, &(x[WS(rs, 1)]));
810 ST(&(x[WS(rs, 11)]), VADD(T2f, T2g), ms, &(x[WS(rs, 1)]));
811 T2j = VBYI(VSUB(T2h, T2i));
812 T2m = VSUB(T2k, T2l);
813 ST(&(x[WS(rs, 13)]), VADD(T2j, T2m), ms, &(x[WS(rs, 1)]));
814 ST(&(x[WS(rs, 19)]), VSUB(T2m, T2j), ms, &(x[WS(rs, 1)]));
815 }
816 }
817 {
818 V T2r, T2M, T2F, T2N, T2y, T2J, T2C, T2K;
819 {
820 V T2p, T2q, T2D, T2E;
821 T2p = VADD(Tb, Tr);
822 T2q = VADD(T1Z, T20);
823 T2r = VSUB(T2p, T2q);
824 T2M = VADD(T2p, T2q);
825 T2D = VFNMS(LDK(KP195090322), T2s, VMUL(LDK(KP980785280), T2t));
826 T2E = VFMA(LDK(KP195090322), T2v, VMUL(LDK(KP980785280), T2w));
827 T2F = VSUB(T2D, T2E);
828 T2N = VADD(T2D, T2E);
829 }
830 {
831 V T2u, T2x, T2A, T2B;
832 T2u = VFMA(LDK(KP980785280), T2s, VMUL(LDK(KP195090322), T2t));
833 T2x = VFNMS(LDK(KP195090322), T2w, VMUL(LDK(KP980785280), T2v));
834 T2y = VSUB(T2u, T2x);
835 T2J = VADD(T2u, T2x);
836 T2A = VADD(TG, TX);
837 T2B = VADD(T28, T22);
838 T2C = VSUB(T2A, T2B);
839 T2K = VADD(T2B, T2A);
840 }
841 {
842 V T2z, T2G, T2P, T2Q;
843 T2z = VADD(T2r, T2y);
844 T2G = VBYI(VADD(T2C, T2F));
845 ST(&(x[WS(rs, 25)]), VSUB(T2z, T2G), ms, &(x[WS(rs, 1)]));
846 ST(&(x[WS(rs, 7)]), VADD(T2z, T2G), ms, &(x[WS(rs, 1)]));
847 T2P = VBYI(VADD(T2K, T2J));
848 T2Q = VADD(T2M, T2N);
849 ST(&(x[WS(rs, 1)]), VADD(T2P, T2Q), ms, &(x[WS(rs, 1)]));
850 ST(&(x[WS(rs, 31)]), VSUB(T2Q, T2P), ms, &(x[WS(rs, 1)]));
851 }
852 {
853 V T2H, T2I, T2L, T2O;
854 T2H = VSUB(T2r, T2y);
855 T2I = VBYI(VSUB(T2F, T2C));
856 ST(&(x[WS(rs, 23)]), VSUB(T2H, T2I), ms, &(x[WS(rs, 1)]));
857 ST(&(x[WS(rs, 9)]), VADD(T2H, T2I), ms, &(x[WS(rs, 1)]));
858 T2L = VBYI(VSUB(T2J, T2K));
859 T2O = VSUB(T2M, T2N);
860 ST(&(x[WS(rs, 15)]), VADD(T2L, T2O), ms, &(x[WS(rs, 1)]));
861 ST(&(x[WS(rs, 17)]), VSUB(T2O, T2L), ms, &(x[WS(rs, 1)]));
862 }
863 }
864 }
865 }
866 }
867 VLEAVE();
868 }
869
870 static const tw_instr twinstr[] = {
871 VTW(0, 1),
872 VTW(0, 3),
873 VTW(0, 9),
874 VTW(0, 27),
875 {TW_NEXT, VL, 0}
876 };
877
878 static const ct_desc desc = { 32, XSIMD_STRING("t3bv_32"), twinstr, &GENUS, {228, 142, 16, 0}, 0, 0, 0 };
879
880 void XSIMD(codelet_t3bv_32) (planner *p) {
881 X(kdft_dit_register) (p, t3bv_32, &desc);
882 }
883 #endif /* HAVE_FMA */