Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/simd/common/n2fv_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 07:37:24 EST 2012 */ | |
23 | |
24 #include "codelet-dft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2fv_32 -with-ostride 2 -include n2f.h -store-multiple 2 */ | |
29 | |
30 /* | |
31 * This function contains 186 FP additions, 98 FP multiplications, | |
32 * (or, 88 additions, 0 multiplications, 98 fused multiply/add), | |
33 * 120 stack variables, 7 constants, and 80 memory accesses | |
34 */ | |
35 #include "n2f.h" | |
36 | |
37 static void n2fv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
38 { | |
39 DVK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
40 DVK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
41 DVK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
44 DVK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
46 { | |
47 INT i; | |
48 const R *xi; | |
49 R *xo; | |
50 xi = ri; | |
51 xo = ro; | |
52 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(64, is), MAKE_VOLATILE_STRIDE(64, os)) { | |
53 V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c, T1h, Tr, T3d; | |
54 V T3e, T3f, T3g, T1a, T1k, TI, T1b, T1L, T1P, T1I, T1G, T1O, T1Q, T1H, T1z; | |
55 V T1c, TZ; | |
56 { | |
57 V T2x, T1T, T2K, T1W, T1p, Tb, T1A, T16, Tu, TF, T2N, T2H, T2b, T2t, TY; | |
58 V T1w, TT, T1v, T20, T2C, Tj, Te, T2h, To, T2f, T23, T2D, TB, TG, Th; | |
59 V T2i, Tk; | |
60 { | |
61 V TL, TW, TP, TQ, T2F, T27, T28, TO; | |
62 { | |
63 V T1, T2, T12, T13, T4, T5, T7, T8; | |
64 T1 = LD(&(xi[0]), ivs, &(xi[0])); | |
65 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); | |
66 T12 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); | |
67 T13 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); | |
68 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); | |
69 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); | |
70 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); | |
71 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); | |
72 { | |
73 V TM, T25, T26, TN; | |
74 { | |
75 V TJ, T3, T14, T1U, T6, T1V, T9, TK, TU, TV, T1R, T1S, Ta, T15; | |
76 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); | |
77 T1R = VADD(T1, T2); | |
78 T3 = VSUB(T1, T2); | |
79 T1S = VADD(T12, T13); | |
80 T14 = VSUB(T12, T13); | |
81 T1U = VADD(T4, T5); | |
82 T6 = VSUB(T4, T5); | |
83 T1V = VADD(T7, T8); | |
84 T9 = VSUB(T7, T8); | |
85 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); | |
86 TU = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); | |
87 T2x = VSUB(T1R, T1S); | |
88 T1T = VADD(T1R, T1S); | |
89 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); | |
90 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); | |
91 T2K = VSUB(T1V, T1U); | |
92 T1W = VADD(T1U, T1V); | |
93 Ta = VADD(T6, T9); | |
94 T15 = VSUB(T9, T6); | |
95 T25 = VADD(TJ, TK); | |
96 TL = VSUB(TJ, TK); | |
97 T26 = VADD(TV, TU); | |
98 TW = VSUB(TU, TV); | |
99 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); | |
100 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); | |
101 T1p = VFNMS(LDK(KP707106781), Ta, T3); | |
102 Tb = VFMA(LDK(KP707106781), Ta, T3); | |
103 T1A = VFMA(LDK(KP707106781), T15, T14); | |
104 T16 = VFNMS(LDK(KP707106781), T15, T14); | |
105 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); | |
106 } | |
107 T2F = VSUB(T25, T26); | |
108 T27 = VADD(T25, T26); | |
109 T28 = VADD(TM, TN); | |
110 TO = VSUB(TM, TN); | |
111 } | |
112 } | |
113 { | |
114 V Ty, T21, Tx, Tz, T1Y, T1Z; | |
115 { | |
116 V Ts, Tt, TD, T29, TR, TE, Tv, Tw; | |
117 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); | |
118 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); | |
119 TD = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); | |
120 T29 = VADD(TP, TQ); | |
121 TR = VSUB(TP, TQ); | |
122 TE = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); | |
123 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); | |
124 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); | |
125 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); | |
126 T1Y = VADD(Ts, Tt); | |
127 Tu = VSUB(Ts, Tt); | |
128 { | |
129 V T2G, T2a, TX, TS; | |
130 T2G = VSUB(T29, T28); | |
131 T2a = VADD(T28, T29); | |
132 TX = VSUB(TR, TO); | |
133 TS = VADD(TO, TR); | |
134 T1Z = VADD(TD, TE); | |
135 TF = VSUB(TD, TE); | |
136 T21 = VADD(Tv, Tw); | |
137 Tx = VSUB(Tv, Tw); | |
138 T2N = VFMA(LDK(KP414213562), T2F, T2G); | |
139 T2H = VFNMS(LDK(KP414213562), T2G, T2F); | |
140 T2b = VSUB(T27, T2a); | |
141 T2t = VADD(T27, T2a); | |
142 TY = VFMA(LDK(KP707106781), TX, TW); | |
143 T1w = VFNMS(LDK(KP707106781), TX, TW); | |
144 TT = VFMA(LDK(KP707106781), TS, TL); | |
145 T1v = VFNMS(LDK(KP707106781), TS, TL); | |
146 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); | |
147 } | |
148 } | |
149 T20 = VADD(T1Y, T1Z); | |
150 T2C = VSUB(T1Y, T1Z); | |
151 { | |
152 V Tc, Td, Tm, Tn; | |
153 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); | |
154 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); | |
155 Tm = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); | |
156 Tn = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); | |
157 { | |
158 V Tf, TA, T22, Tg; | |
159 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); | |
160 TA = VSUB(Ty, Tz); | |
161 T22 = VADD(Ty, Tz); | |
162 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); | |
163 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); | |
164 Te = VSUB(Tc, Td); | |
165 T2h = VADD(Tc, Td); | |
166 To = VSUB(Tm, Tn); | |
167 T2f = VADD(Tn, Tm); | |
168 T23 = VADD(T21, T22); | |
169 T2D = VSUB(T21, T22); | |
170 TB = VADD(Tx, TA); | |
171 TG = VSUB(Tx, TA); | |
172 Th = VSUB(Tf, Tg); | |
173 T2i = VADD(Tf, Tg); | |
174 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); | |
175 } | |
176 } | |
177 } | |
178 } | |
179 { | |
180 V T1t, TH, T1s, TC, T2P, T2U, T2n, T2d, T2w, T2u, T1q, T19, T1B, Tq, T2W; | |
181 V T2M, T2B, T2T, T2v, T2r, T2o, T2m, T2X, T2I; | |
182 { | |
183 V T1X, T2p, T2E, T2O, T2s, T2y, T2j, T17, Ti, T2e, Tl, T2c, T2l, T24; | |
184 T1X = VSUB(T1T, T1W); | |
185 T2p = VADD(T1T, T1W); | |
186 T2E = VFNMS(LDK(KP414213562), T2D, T2C); | |
187 T2O = VFMA(LDK(KP414213562), T2C, T2D); | |
188 T2s = VADD(T20, T23); | |
189 T24 = VSUB(T20, T23); | |
190 T1t = VFNMS(LDK(KP707106781), TG, TF); | |
191 TH = VFMA(LDK(KP707106781), TG, TF); | |
192 T1s = VFNMS(LDK(KP707106781), TB, Tu); | |
193 TC = VFMA(LDK(KP707106781), TB, Tu); | |
194 T2y = VSUB(T2h, T2i); | |
195 T2j = VADD(T2h, T2i); | |
196 T17 = VFMA(LDK(KP414213562), Te, Th); | |
197 Ti = VFNMS(LDK(KP414213562), Th, Te); | |
198 T2e = VADD(Tj, Tk); | |
199 Tl = VSUB(Tj, Tk); | |
200 T2c = VADD(T24, T2b); | |
201 T2l = VSUB(T2b, T24); | |
202 { | |
203 V T2L, T2A, T2q, T2k; | |
204 T2P = VSUB(T2N, T2O); | |
205 T2U = VADD(T2O, T2N); | |
206 { | |
207 V T2z, T2g, T18, Tp; | |
208 T2z = VSUB(T2e, T2f); | |
209 T2g = VADD(T2e, T2f); | |
210 T18 = VFMA(LDK(KP414213562), Tl, To); | |
211 Tp = VFNMS(LDK(KP414213562), To, Tl); | |
212 T2n = VFMA(LDK(KP707106781), T2c, T1X); | |
213 T2d = VFNMS(LDK(KP707106781), T2c, T1X); | |
214 T2w = VSUB(T2t, T2s); | |
215 T2u = VADD(T2s, T2t); | |
216 T2L = VSUB(T2z, T2y); | |
217 T2A = VADD(T2y, T2z); | |
218 T2q = VADD(T2j, T2g); | |
219 T2k = VSUB(T2g, T2j); | |
220 T1q = VADD(T17, T18); | |
221 T19 = VSUB(T17, T18); | |
222 T1B = VSUB(Tp, Ti); | |
223 Tq = VADD(Ti, Tp); | |
224 } | |
225 T2W = VFNMS(LDK(KP707106781), T2L, T2K); | |
226 T2M = VFMA(LDK(KP707106781), T2L, T2K); | |
227 T2B = VFMA(LDK(KP707106781), T2A, T2x); | |
228 T2T = VFNMS(LDK(KP707106781), T2A, T2x); | |
229 T2v = VSUB(T2p, T2q); | |
230 T2r = VADD(T2p, T2q); | |
231 T2o = VFMA(LDK(KP707106781), T2l, T2k); | |
232 T2m = VFNMS(LDK(KP707106781), T2l, T2k); | |
233 T2X = VSUB(T2H, T2E); | |
234 T2I = VADD(T2E, T2H); | |
235 } | |
236 } | |
237 { | |
238 V T2V, T2Z, T2Y, T30, T2R, T2J; | |
239 T2V = VFNMS(LDK(KP923879532), T2U, T2T); | |
240 T2Z = VFMA(LDK(KP923879532), T2U, T2T); | |
241 T31 = VFNMSI(T2w, T2v); | |
242 STM2(&(xo[48]), T31, ovs, &(xo[0])); | |
243 T32 = VFMAI(T2w, T2v); | |
244 STM2(&(xo[16]), T32, ovs, &(xo[0])); | |
245 T33 = VADD(T2r, T2u); | |
246 STM2(&(xo[0]), T33, ovs, &(xo[0])); | |
247 T34 = VSUB(T2r, T2u); | |
248 STM2(&(xo[32]), T34, ovs, &(xo[0])); | |
249 T35 = VFNMSI(T2o, T2n); | |
250 STM2(&(xo[56]), T35, ovs, &(xo[0])); | |
251 T36 = VFMAI(T2o, T2n); | |
252 STM2(&(xo[8]), T36, ovs, &(xo[0])); | |
253 T37 = VFMAI(T2m, T2d); | |
254 STM2(&(xo[40]), T37, ovs, &(xo[0])); | |
255 T38 = VFNMSI(T2m, T2d); | |
256 STM2(&(xo[24]), T38, ovs, &(xo[0])); | |
257 T2Y = VFMA(LDK(KP923879532), T2X, T2W); | |
258 T30 = VFNMS(LDK(KP923879532), T2X, T2W); | |
259 T2R = VFMA(LDK(KP923879532), T2I, T2B); | |
260 T2J = VFNMS(LDK(KP923879532), T2I, T2B); | |
261 { | |
262 V T1J, T1r, T1C, T1M, T2S, T2Q, T1u, T1D, T1E, T1x; | |
263 T1J = VFNMS(LDK(KP923879532), T1q, T1p); | |
264 T1r = VFMA(LDK(KP923879532), T1q, T1p); | |
265 T1C = VFMA(LDK(KP923879532), T1B, T1A); | |
266 T1M = VFNMS(LDK(KP923879532), T1B, T1A); | |
267 T39 = VFNMSI(T30, T2Z); | |
268 STM2(&(xo[12]), T39, ovs, &(xo[0])); | |
269 T3a = VFMAI(T30, T2Z); | |
270 STM2(&(xo[52]), T3a, ovs, &(xo[0])); | |
271 T3b = VFNMSI(T2Y, T2V); | |
272 STM2(&(xo[44]), T3b, ovs, &(xo[0])); | |
273 T3c = VFMAI(T2Y, T2V); | |
274 STM2(&(xo[20]), T3c, ovs, &(xo[0])); | |
275 T2S = VFMA(LDK(KP923879532), T2P, T2M); | |
276 T2Q = VFNMS(LDK(KP923879532), T2P, T2M); | |
277 T1u = VFMA(LDK(KP668178637), T1t, T1s); | |
278 T1D = VFNMS(LDK(KP668178637), T1s, T1t); | |
279 T1E = VFNMS(LDK(KP668178637), T1v, T1w); | |
280 T1x = VFMA(LDK(KP668178637), T1w, T1v); | |
281 { | |
282 V T1K, T1F, T1N, T1y; | |
283 T1h = VFNMS(LDK(KP923879532), Tq, Tb); | |
284 Tr = VFMA(LDK(KP923879532), Tq, Tb); | |
285 T3d = VFNMSI(T2S, T2R); | |
286 STM2(&(xo[60]), T3d, ovs, &(xo[0])); | |
287 T3e = VFMAI(T2S, T2R); | |
288 STM2(&(xo[4]), T3e, ovs, &(xo[0])); | |
289 T3f = VFMAI(T2Q, T2J); | |
290 STM2(&(xo[36]), T3f, ovs, &(xo[0])); | |
291 T3g = VFNMSI(T2Q, T2J); | |
292 STM2(&(xo[28]), T3g, ovs, &(xo[0])); | |
293 T1K = VADD(T1D, T1E); | |
294 T1F = VSUB(T1D, T1E); | |
295 T1N = VSUB(T1x, T1u); | |
296 T1y = VADD(T1u, T1x); | |
297 T1a = VFMA(LDK(KP923879532), T19, T16); | |
298 T1k = VFNMS(LDK(KP923879532), T19, T16); | |
299 TI = VFNMS(LDK(KP198912367), TH, TC); | |
300 T1b = VFMA(LDK(KP198912367), TC, TH); | |
301 T1L = VFMA(LDK(KP831469612), T1K, T1J); | |
302 T1P = VFNMS(LDK(KP831469612), T1K, T1J); | |
303 T1I = VFMA(LDK(KP831469612), T1F, T1C); | |
304 T1G = VFNMS(LDK(KP831469612), T1F, T1C); | |
305 T1O = VFMA(LDK(KP831469612), T1N, T1M); | |
306 T1Q = VFNMS(LDK(KP831469612), T1N, T1M); | |
307 T1H = VFMA(LDK(KP831469612), T1y, T1r); | |
308 T1z = VFNMS(LDK(KP831469612), T1y, T1r); | |
309 T1c = VFMA(LDK(KP198912367), TT, TY); | |
310 TZ = VFNMS(LDK(KP198912367), TY, TT); | |
311 } | |
312 } | |
313 } | |
314 } | |
315 } | |
316 { | |
317 V T1d, T1i, T10, T1l; | |
318 { | |
319 V T3h, T3i, T3j, T3k; | |
320 T3h = VFNMSI(T1O, T1L); | |
321 STM2(&(xo[42]), T3h, ovs, &(xo[2])); | |
322 STN2(&(xo[40]), T37, T3h, ovs); | |
323 T3i = VFMAI(T1O, T1L); | |
324 STM2(&(xo[22]), T3i, ovs, &(xo[2])); | |
325 STN2(&(xo[20]), T3c, T3i, ovs); | |
326 T3j = VFMAI(T1Q, T1P); | |
327 STM2(&(xo[54]), T3j, ovs, &(xo[2])); | |
328 STN2(&(xo[52]), T3a, T3j, ovs); | |
329 T3k = VFNMSI(T1Q, T1P); | |
330 STM2(&(xo[10]), T3k, ovs, &(xo[2])); | |
331 STN2(&(xo[8]), T36, T3k, ovs); | |
332 { | |
333 V T3l, T3m, T3n, T3o; | |
334 T3l = VFMAI(T1I, T1H); | |
335 STM2(&(xo[6]), T3l, ovs, &(xo[2])); | |
336 STN2(&(xo[4]), T3e, T3l, ovs); | |
337 T3m = VFNMSI(T1I, T1H); | |
338 STM2(&(xo[58]), T3m, ovs, &(xo[2])); | |
339 STN2(&(xo[56]), T35, T3m, ovs); | |
340 T3n = VFMAI(T1G, T1z); | |
341 STM2(&(xo[38]), T3n, ovs, &(xo[2])); | |
342 STN2(&(xo[36]), T3f, T3n, ovs); | |
343 T3o = VFNMSI(T1G, T1z); | |
344 STM2(&(xo[26]), T3o, ovs, &(xo[2])); | |
345 STN2(&(xo[24]), T38, T3o, ovs); | |
346 T1d = VSUB(T1b, T1c); | |
347 T1i = VADD(T1b, T1c); | |
348 T10 = VADD(TI, TZ); | |
349 T1l = VSUB(TZ, TI); | |
350 } | |
351 } | |
352 { | |
353 V T1n, T1j, T1e, T1g, T1o, T1m, T11, T1f; | |
354 T1n = VFMA(LDK(KP980785280), T1i, T1h); | |
355 T1j = VFNMS(LDK(KP980785280), T1i, T1h); | |
356 T1e = VFNMS(LDK(KP980785280), T1d, T1a); | |
357 T1g = VFMA(LDK(KP980785280), T1d, T1a); | |
358 T1o = VFMA(LDK(KP980785280), T1l, T1k); | |
359 T1m = VFNMS(LDK(KP980785280), T1l, T1k); | |
360 T11 = VFNMS(LDK(KP980785280), T10, Tr); | |
361 T1f = VFMA(LDK(KP980785280), T10, Tr); | |
362 { | |
363 V T3p, T3q, T3r, T3s; | |
364 T3p = VFMAI(T1m, T1j); | |
365 STM2(&(xo[46]), T3p, ovs, &(xo[2])); | |
366 STN2(&(xo[44]), T3b, T3p, ovs); | |
367 T3q = VFNMSI(T1m, T1j); | |
368 STM2(&(xo[18]), T3q, ovs, &(xo[2])); | |
369 STN2(&(xo[16]), T32, T3q, ovs); | |
370 T3r = VFNMSI(T1o, T1n); | |
371 STM2(&(xo[50]), T3r, ovs, &(xo[2])); | |
372 STN2(&(xo[48]), T31, T3r, ovs); | |
373 T3s = VFMAI(T1o, T1n); | |
374 STM2(&(xo[14]), T3s, ovs, &(xo[2])); | |
375 STN2(&(xo[12]), T39, T3s, ovs); | |
376 { | |
377 V T3t, T3u, T3v, T3w; | |
378 T3t = VFMAI(T1g, T1f); | |
379 STM2(&(xo[62]), T3t, ovs, &(xo[2])); | |
380 STN2(&(xo[60]), T3d, T3t, ovs); | |
381 T3u = VFNMSI(T1g, T1f); | |
382 STM2(&(xo[2]), T3u, ovs, &(xo[2])); | |
383 STN2(&(xo[0]), T33, T3u, ovs); | |
384 T3v = VFMAI(T1e, T11); | |
385 STM2(&(xo[30]), T3v, ovs, &(xo[2])); | |
386 STN2(&(xo[28]), T3g, T3v, ovs); | |
387 T3w = VFNMSI(T1e, T11); | |
388 STM2(&(xo[34]), T3w, ovs, &(xo[2])); | |
389 STN2(&(xo[32]), T34, T3w, ovs); | |
390 } | |
391 } | |
392 } | |
393 } | |
394 } | |
395 } | |
396 VLEAVE(); | |
397 } | |
398 | |
399 static const kdft_desc desc = { 32, XSIMD_STRING("n2fv_32"), {88, 0, 98, 0}, &GENUS, 0, 2, 0, 0 }; | |
400 | |
401 void XSIMD(codelet_n2fv_32) (planner *p) { | |
402 X(kdft_register) (p, n2fv_32, &desc); | |
403 } | |
404 | |
405 #else /* HAVE_FMA */ | |
406 | |
407 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2fv_32 -with-ostride 2 -include n2f.h -store-multiple 2 */ | |
408 | |
409 /* | |
410 * This function contains 186 FP additions, 42 FP multiplications, | |
411 * (or, 170 additions, 26 multiplications, 16 fused multiply/add), | |
412 * 72 stack variables, 7 constants, and 80 memory accesses | |
413 */ | |
414 #include "n2f.h" | |
415 | |
416 static void n2fv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
417 { | |
418 DVK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
419 DVK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
420 DVK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
421 DVK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
422 DVK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
423 DVK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
424 DVK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
425 { | |
426 INT i; | |
427 const R *xi; | |
428 R *xo; | |
429 xi = ri; | |
430 xo = ro; | |
431 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(64, is), MAKE_VOLATILE_STRIDE(64, os)) { | |
432 V T1T, T1W, T2K, T2x, T16, T1A, Tb, T1p, TT, T1v, TY, T1w, T27, T2a, T2b; | |
433 V T2H, T2O, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2N, T2g, T2j, Tq, T1B; | |
434 V T19, T1q, T2A, T2L; | |
435 { | |
436 V T3, T1R, T15, T1S, T6, T1U, T9, T1V, T12, Ta; | |
437 { | |
438 V T1, T2, T13, T14; | |
439 T1 = LD(&(xi[0]), ivs, &(xi[0])); | |
440 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); | |
441 T3 = VSUB(T1, T2); | |
442 T1R = VADD(T1, T2); | |
443 T13 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); | |
444 T14 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); | |
445 T15 = VSUB(T13, T14); | |
446 T1S = VADD(T13, T14); | |
447 } | |
448 { | |
449 V T4, T5, T7, T8; | |
450 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); | |
451 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); | |
452 T6 = VSUB(T4, T5); | |
453 T1U = VADD(T4, T5); | |
454 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); | |
455 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); | |
456 T9 = VSUB(T7, T8); | |
457 T1V = VADD(T7, T8); | |
458 } | |
459 T1T = VADD(T1R, T1S); | |
460 T1W = VADD(T1U, T1V); | |
461 T2K = VSUB(T1V, T1U); | |
462 T2x = VSUB(T1R, T1S); | |
463 T12 = VMUL(LDK(KP707106781), VSUB(T9, T6)); | |
464 T16 = VSUB(T12, T15); | |
465 T1A = VADD(T15, T12); | |
466 Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); | |
467 Tb = VADD(T3, Ta); | |
468 T1p = VSUB(T3, Ta); | |
469 } | |
470 { | |
471 V TL, T25, TX, T26, TO, T28, TR, T29; | |
472 { | |
473 V TJ, TK, TV, TW; | |
474 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); | |
475 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); | |
476 TL = VSUB(TJ, TK); | |
477 T25 = VADD(TJ, TK); | |
478 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); | |
479 TW = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); | |
480 TX = VSUB(TV, TW); | |
481 T26 = VADD(TV, TW); | |
482 } | |
483 { | |
484 V TM, TN, TP, TQ; | |
485 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); | |
486 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); | |
487 TO = VSUB(TM, TN); | |
488 T28 = VADD(TM, TN); | |
489 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); | |
490 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); | |
491 TR = VSUB(TP, TQ); | |
492 T29 = VADD(TP, TQ); | |
493 } | |
494 { | |
495 V TS, TU, T2F, T2G; | |
496 TS = VMUL(LDK(KP707106781), VADD(TO, TR)); | |
497 TT = VADD(TL, TS); | |
498 T1v = VSUB(TL, TS); | |
499 TU = VMUL(LDK(KP707106781), VSUB(TR, TO)); | |
500 TY = VSUB(TU, TX); | |
501 T1w = VADD(TX, TU); | |
502 T27 = VADD(T25, T26); | |
503 T2a = VADD(T28, T29); | |
504 T2b = VSUB(T27, T2a); | |
505 T2F = VSUB(T25, T26); | |
506 T2G = VSUB(T29, T28); | |
507 T2H = VFNMS(LDK(KP382683432), T2G, VMUL(LDK(KP923879532), T2F)); | |
508 T2O = VFMA(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2G)); | |
509 } | |
510 } | |
511 { | |
512 V Tu, T1Y, TG, T1Z, Tx, T21, TA, T22; | |
513 { | |
514 V Ts, Tt, TE, TF; | |
515 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); | |
516 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); | |
517 Tu = VSUB(Ts, Tt); | |
518 T1Y = VADD(Ts, Tt); | |
519 TE = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); | |
520 TF = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); | |
521 TG = VSUB(TE, TF); | |
522 T1Z = VADD(TE, TF); | |
523 } | |
524 { | |
525 V Tv, Tw, Ty, Tz; | |
526 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); | |
527 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); | |
528 Tx = VSUB(Tv, Tw); | |
529 T21 = VADD(Tv, Tw); | |
530 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); | |
531 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); | |
532 TA = VSUB(Ty, Tz); | |
533 T22 = VADD(Ty, Tz); | |
534 } | |
535 { | |
536 V TB, TD, T2C, T2D; | |
537 TB = VMUL(LDK(KP707106781), VADD(Tx, TA)); | |
538 TC = VADD(Tu, TB); | |
539 T1s = VSUB(Tu, TB); | |
540 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx)); | |
541 TH = VSUB(TD, TG); | |
542 T1t = VADD(TG, TD); | |
543 T20 = VADD(T1Y, T1Z); | |
544 T23 = VADD(T21, T22); | |
545 T24 = VSUB(T20, T23); | |
546 T2C = VSUB(T1Y, T1Z); | |
547 T2D = VSUB(T22, T21); | |
548 T2E = VFMA(LDK(KP923879532), T2C, VMUL(LDK(KP382683432), T2D)); | |
549 T2N = VFNMS(LDK(KP382683432), T2C, VMUL(LDK(KP923879532), T2D)); | |
550 } | |
551 } | |
552 { | |
553 V Te, T2h, To, T2f, Th, T2i, Tl, T2e, Ti, Tp; | |
554 { | |
555 V Tc, Td, Tm, Tn; | |
556 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); | |
557 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); | |
558 Te = VSUB(Tc, Td); | |
559 T2h = VADD(Tc, Td); | |
560 Tm = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); | |
561 Tn = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); | |
562 To = VSUB(Tm, Tn); | |
563 T2f = VADD(Tm, Tn); | |
564 } | |
565 { | |
566 V Tf, Tg, Tj, Tk; | |
567 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); | |
568 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); | |
569 Th = VSUB(Tf, Tg); | |
570 T2i = VADD(Tf, Tg); | |
571 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); | |
572 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); | |
573 Tl = VSUB(Tj, Tk); | |
574 T2e = VADD(Tj, Tk); | |
575 } | |
576 T2g = VADD(T2e, T2f); | |
577 T2j = VADD(T2h, T2i); | |
578 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); | |
579 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To)); | |
580 Tq = VADD(Ti, Tp); | |
581 T1B = VSUB(Tp, Ti); | |
582 { | |
583 V T17, T18, T2y, T2z; | |
584 T17 = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); | |
585 T18 = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); | |
586 T19 = VSUB(T17, T18); | |
587 T1q = VADD(T18, T17); | |
588 T2y = VSUB(T2h, T2i); | |
589 T2z = VSUB(T2e, T2f); | |
590 T2A = VMUL(LDK(KP707106781), VADD(T2y, T2z)); | |
591 T2L = VMUL(LDK(KP707106781), VSUB(T2z, T2y)); | |
592 } | |
593 } | |
594 { | |
595 V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c; | |
596 { | |
597 V T2d, T2n, T2m, T2o; | |
598 { | |
599 V T1X, T2c, T2k, T2l; | |
600 T1X = VSUB(T1T, T1W); | |
601 T2c = VMUL(LDK(KP707106781), VADD(T24, T2b)); | |
602 T2d = VADD(T1X, T2c); | |
603 T2n = VSUB(T1X, T2c); | |
604 T2k = VSUB(T2g, T2j); | |
605 T2l = VMUL(LDK(KP707106781), VSUB(T2b, T24)); | |
606 T2m = VBYI(VADD(T2k, T2l)); | |
607 T2o = VBYI(VSUB(T2l, T2k)); | |
608 } | |
609 T31 = VSUB(T2d, T2m); | |
610 STM2(&(xo[56]), T31, ovs, &(xo[0])); | |
611 T32 = VADD(T2n, T2o); | |
612 STM2(&(xo[24]), T32, ovs, &(xo[0])); | |
613 T33 = VADD(T2d, T2m); | |
614 STM2(&(xo[8]), T33, ovs, &(xo[0])); | |
615 T34 = VSUB(T2n, T2o); | |
616 STM2(&(xo[40]), T34, ovs, &(xo[0])); | |
617 } | |
618 { | |
619 V T2r, T2v, T2u, T2w; | |
620 { | |
621 V T2p, T2q, T2s, T2t; | |
622 T2p = VADD(T1T, T1W); | |
623 T2q = VADD(T2j, T2g); | |
624 T2r = VADD(T2p, T2q); | |
625 T2v = VSUB(T2p, T2q); | |
626 T2s = VADD(T20, T23); | |
627 T2t = VADD(T27, T2a); | |
628 T2u = VADD(T2s, T2t); | |
629 T2w = VBYI(VSUB(T2t, T2s)); | |
630 } | |
631 T35 = VSUB(T2r, T2u); | |
632 STM2(&(xo[32]), T35, ovs, &(xo[0])); | |
633 T36 = VADD(T2v, T2w); | |
634 STM2(&(xo[16]), T36, ovs, &(xo[0])); | |
635 T37 = VADD(T2r, T2u); | |
636 STM2(&(xo[0]), T37, ovs, &(xo[0])); | |
637 T38 = VSUB(T2v, T2w); | |
638 STM2(&(xo[48]), T38, ovs, &(xo[0])); | |
639 } | |
640 { | |
641 V T2V, T2Z, T2Y, T30; | |
642 { | |
643 V T2T, T2U, T2W, T2X; | |
644 T2T = VSUB(T2H, T2E); | |
645 T2U = VSUB(T2L, T2K); | |
646 T2V = VBYI(VSUB(T2T, T2U)); | |
647 T2Z = VBYI(VADD(T2U, T2T)); | |
648 T2W = VSUB(T2x, T2A); | |
649 T2X = VSUB(T2O, T2N); | |
650 T2Y = VSUB(T2W, T2X); | |
651 T30 = VADD(T2W, T2X); | |
652 } | |
653 T39 = VADD(T2V, T2Y); | |
654 STM2(&(xo[20]), T39, ovs, &(xo[0])); | |
655 T3a = VSUB(T30, T2Z); | |
656 STM2(&(xo[52]), T3a, ovs, &(xo[0])); | |
657 T3b = VSUB(T2Y, T2V); | |
658 STM2(&(xo[44]), T3b, ovs, &(xo[0])); | |
659 T3c = VADD(T2Z, T30); | |
660 STM2(&(xo[12]), T3c, ovs, &(xo[0])); | |
661 } | |
662 { | |
663 V T3d, T3e, T3f, T3g; | |
664 { | |
665 V T2J, T2R, T2Q, T2S; | |
666 { | |
667 V T2B, T2I, T2M, T2P; | |
668 T2B = VADD(T2x, T2A); | |
669 T2I = VADD(T2E, T2H); | |
670 T2J = VADD(T2B, T2I); | |
671 T2R = VSUB(T2B, T2I); | |
672 T2M = VADD(T2K, T2L); | |
673 T2P = VADD(T2N, T2O); | |
674 T2Q = VBYI(VADD(T2M, T2P)); | |
675 T2S = VBYI(VSUB(T2P, T2M)); | |
676 } | |
677 T3d = VSUB(T2J, T2Q); | |
678 STM2(&(xo[60]), T3d, ovs, &(xo[0])); | |
679 T3e = VADD(T2R, T2S); | |
680 STM2(&(xo[28]), T3e, ovs, &(xo[0])); | |
681 T3f = VADD(T2J, T2Q); | |
682 STM2(&(xo[4]), T3f, ovs, &(xo[0])); | |
683 T3g = VSUB(T2R, T2S); | |
684 STM2(&(xo[36]), T3g, ovs, &(xo[0])); | |
685 } | |
686 { | |
687 V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J; | |
688 T1r = VADD(T1p, T1q); | |
689 T1C = VADD(T1A, T1B); | |
690 T1M = VSUB(T1p, T1q); | |
691 T1K = VSUB(T1B, T1A); | |
692 { | |
693 V T1D, T1E, T1u, T1x; | |
694 T1D = VFNMS(LDK(KP555570233), T1s, VMUL(LDK(KP831469612), T1t)); | |
695 T1E = VFMA(LDK(KP555570233), T1v, VMUL(LDK(KP831469612), T1w)); | |
696 T1F = VADD(T1D, T1E); | |
697 T1N = VSUB(T1E, T1D); | |
698 T1u = VFMA(LDK(KP831469612), T1s, VMUL(LDK(KP555570233), T1t)); | |
699 T1x = VFNMS(LDK(KP555570233), T1w, VMUL(LDK(KP831469612), T1v)); | |
700 T1y = VADD(T1u, T1x); | |
701 T1J = VSUB(T1x, T1u); | |
702 } | |
703 { | |
704 V T1z, T1G, T3h, T3i; | |
705 T1z = VADD(T1r, T1y); | |
706 T1G = VBYI(VADD(T1C, T1F)); | |
707 T3h = VSUB(T1z, T1G); | |
708 STM2(&(xo[58]), T3h, ovs, &(xo[2])); | |
709 STN2(&(xo[56]), T31, T3h, ovs); | |
710 T3i = VADD(T1z, T1G); | |
711 STM2(&(xo[6]), T3i, ovs, &(xo[2])); | |
712 STN2(&(xo[4]), T3f, T3i, ovs); | |
713 } | |
714 { | |
715 V T1P, T1Q, T3j, T3k; | |
716 T1P = VBYI(VADD(T1K, T1J)); | |
717 T1Q = VADD(T1M, T1N); | |
718 T3j = VADD(T1P, T1Q); | |
719 STM2(&(xo[10]), T3j, ovs, &(xo[2])); | |
720 STN2(&(xo[8]), T33, T3j, ovs); | |
721 T3k = VSUB(T1Q, T1P); | |
722 STM2(&(xo[54]), T3k, ovs, &(xo[2])); | |
723 STN2(&(xo[52]), T3a, T3k, ovs); | |
724 } | |
725 { | |
726 V T1H, T1I, T3l, T3m; | |
727 T1H = VSUB(T1r, T1y); | |
728 T1I = VBYI(VSUB(T1F, T1C)); | |
729 T3l = VSUB(T1H, T1I); | |
730 STM2(&(xo[38]), T3l, ovs, &(xo[2])); | |
731 STN2(&(xo[36]), T3g, T3l, ovs); | |
732 T3m = VADD(T1H, T1I); | |
733 STM2(&(xo[26]), T3m, ovs, &(xo[2])); | |
734 STN2(&(xo[24]), T32, T3m, ovs); | |
735 } | |
736 { | |
737 V T1L, T1O, T3n, T3o; | |
738 T1L = VBYI(VSUB(T1J, T1K)); | |
739 T1O = VSUB(T1M, T1N); | |
740 T3n = VADD(T1L, T1O); | |
741 STM2(&(xo[22]), T3n, ovs, &(xo[2])); | |
742 STN2(&(xo[20]), T39, T3n, ovs); | |
743 T3o = VSUB(T1O, T1L); | |
744 STM2(&(xo[42]), T3o, ovs, &(xo[2])); | |
745 STN2(&(xo[40]), T34, T3o, ovs); | |
746 } | |
747 } | |
748 { | |
749 V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h; | |
750 Tr = VADD(Tb, Tq); | |
751 T1a = VADD(T16, T19); | |
752 T1k = VSUB(Tb, Tq); | |
753 T1i = VSUB(T19, T16); | |
754 { | |
755 V T1b, T1c, TI, TZ; | |
756 T1b = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH)); | |
757 T1c = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY)); | |
758 T1d = VADD(T1b, T1c); | |
759 T1l = VSUB(T1c, T1b); | |
760 TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH)); | |
761 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT)); | |
762 T10 = VADD(TI, TZ); | |
763 T1h = VSUB(TZ, TI); | |
764 } | |
765 { | |
766 V T11, T1e, T3p, T3q; | |
767 T11 = VADD(Tr, T10); | |
768 T1e = VBYI(VADD(T1a, T1d)); | |
769 T3p = VSUB(T11, T1e); | |
770 STM2(&(xo[62]), T3p, ovs, &(xo[2])); | |
771 STN2(&(xo[60]), T3d, T3p, ovs); | |
772 T3q = VADD(T11, T1e); | |
773 STM2(&(xo[2]), T3q, ovs, &(xo[2])); | |
774 STN2(&(xo[0]), T37, T3q, ovs); | |
775 } | |
776 { | |
777 V T1n, T1o, T3r, T3s; | |
778 T1n = VBYI(VADD(T1i, T1h)); | |
779 T1o = VADD(T1k, T1l); | |
780 T3r = VADD(T1n, T1o); | |
781 STM2(&(xo[14]), T3r, ovs, &(xo[2])); | |
782 STN2(&(xo[12]), T3c, T3r, ovs); | |
783 T3s = VSUB(T1o, T1n); | |
784 STM2(&(xo[50]), T3s, ovs, &(xo[2])); | |
785 STN2(&(xo[48]), T38, T3s, ovs); | |
786 } | |
787 { | |
788 V T1f, T1g, T3t, T3u; | |
789 T1f = VSUB(Tr, T10); | |
790 T1g = VBYI(VSUB(T1d, T1a)); | |
791 T3t = VSUB(T1f, T1g); | |
792 STM2(&(xo[34]), T3t, ovs, &(xo[2])); | |
793 STN2(&(xo[32]), T35, T3t, ovs); | |
794 T3u = VADD(T1f, T1g); | |
795 STM2(&(xo[30]), T3u, ovs, &(xo[2])); | |
796 STN2(&(xo[28]), T3e, T3u, ovs); | |
797 } | |
798 { | |
799 V T1j, T1m, T3v, T3w; | |
800 T1j = VBYI(VSUB(T1h, T1i)); | |
801 T1m = VSUB(T1k, T1l); | |
802 T3v = VADD(T1j, T1m); | |
803 STM2(&(xo[18]), T3v, ovs, &(xo[2])); | |
804 STN2(&(xo[16]), T36, T3v, ovs); | |
805 T3w = VSUB(T1m, T1j); | |
806 STM2(&(xo[46]), T3w, ovs, &(xo[2])); | |
807 STN2(&(xo[44]), T3b, T3w, ovs); | |
808 } | |
809 } | |
810 } | |
811 } | |
812 } | |
813 } | |
814 VLEAVE(); | |
815 } | |
816 | |
817 static const kdft_desc desc = { 32, XSIMD_STRING("n2fv_32"), {170, 26, 16, 0}, &GENUS, 0, 2, 0, 0 }; | |
818 | |
819 void XSIMD(codelet_n2fv_32) (planner *p) { | |
820 X(kdft_register) (p, n2fv_32, &desc); | |
821 } | |
822 | |
823 #endif /* HAVE_FMA */ |