Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/dft/simd/common/n1bv_25.c @ 167:bd3cc4d1df30
Add FFTW 3.3.8 source, and a Linux build
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Tue, 19 Nov 2019 14:52:55 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
166:cbd6d7e562c7 | 167:bd3cc4d1df30 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-14 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-14 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 Thu May 24 08:05:03 EDT 2018 */ | |
23 | |
24 #include "dft/codelet-dft.h" | |
25 | |
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA) | |
27 | |
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 25 -name n1bv_25 -include dft/simd/n1b.h */ | |
29 | |
30 /* | |
31 * This function contains 224 FP additions, 193 FP multiplications, | |
32 * (or, 43 additions, 12 multiplications, 181 fused multiply/add), | |
33 * 140 stack variables, 67 constants, and 50 memory accesses | |
34 */ | |
35 #include "dft/simd/n1b.h" | |
36 | |
37 static void n1bv_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
38 { | |
39 DVK(KP803003575, +0.803003575438660414833440593570376004635464850); | |
40 DVK(KP999544308, +0.999544308746292983948881682379742149196758193); | |
41 DVK(KP792626838, +0.792626838241819413632131824093538848057784557); | |
42 DVK(KP876091699, +0.876091699473550838204498029706869638173524346); | |
43 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
44 DVK(KP617882369, +0.617882369114440893914546919006756321695042882); | |
45 DVK(KP242145790, +0.242145790282157779872542093866183953459003101); | |
46 DVK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
47 DVK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
48 DVK(KP829049696, +0.829049696159252993975487806364305442437946767); | |
49 DVK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
50 DVK(KP831864738, +0.831864738706457140726048799369896829771167132); | |
51 DVK(KP916574801, +0.916574801383451584742370439148878693530976769); | |
52 DVK(KP894834959, +0.894834959464455102997960030820114611498661386); | |
53 DVK(KP809385824, +0.809385824416008241660603814668679683846476688); | |
54 DVK(KP447417479, +0.447417479732227551498980015410057305749330693); | |
55 DVK(KP958953096, +0.958953096729998668045963838399037225970891871); | |
56 DVK(KP867381224, +0.867381224396525206773171885031575671309956167); | |
57 DVK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
58 DVK(KP897376177, +0.897376177523557693138608077137219684419427330); | |
59 DVK(KP855719849, +0.855719849902058969314654733608091555096772472); | |
60 DVK(KP860541664, +0.860541664367944677098261680920518816412804187); | |
61 DVK(KP681693190, +0.681693190061530575150324149145440022633095390); | |
62 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
63 DVK(KP560319534, +0.560319534973832390111614715371676131169633784); | |
64 DVK(KP949179823, +0.949179823508441261575555465843363271711583843); | |
65 DVK(KP952936919, +0.952936919628306576880750665357914584765951388); | |
66 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
67 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
68 DVK(KP997675361, +0.997675361079556513670859573984492383596555031); | |
69 DVK(KP904508497, +0.904508497187473712051146708591409529430077295); | |
70 DVK(KP237294955, +0.237294955877110315393888866460840817927895961); | |
71 DVK(KP923225144, +0.923225144846402650453449441572664695995209956); | |
72 DVK(KP262346850, +0.262346850930607871785420028382979691334784273); | |
73 DVK(KP669429328, +0.669429328479476605641803240971985825917022098); | |
74 DVK(KP570584518, +0.570584518783621657366766175430996792655723863); | |
75 DVK(KP982009705, +0.982009705009746369461829878184175962711969869); | |
76 DVK(KP906616052, +0.906616052148196230441134447086066874408359177); | |
77 DVK(KP921078979, +0.921078979742360627699756128143719920817673854); | |
78 DVK(KP845997307, +0.845997307939530944175097360758058292389769300); | |
79 DVK(KP690983005, +0.690983005625052575897706582817180941139845410); | |
80 DVK(KP945422727, +0.945422727388575946270360266328811958657216298); | |
81 DVK(KP763932022, +0.763932022500210303590826331268723764559381640); | |
82 DVK(KP956723877, +0.956723877038460305821989399535483155872969262); | |
83 DVK(KP522616830, +0.522616830205754336872861364785224694908468440); | |
84 DVK(KP772036680, +0.772036680810363904029489473607579825330539880); | |
85 DVK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
86 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
87 DVK(KP269969613, +0.269969613759572083574752974412347470060951301); | |
88 DVK(KP244189809, +0.244189809627953270309879511234821255780225091); | |
89 DVK(KP447533225, +0.447533225982656890041886979663652563063114397); | |
90 DVK(KP494780565, +0.494780565770515410344588413655324772219443730); | |
91 DVK(KP578046249, +0.578046249379945007321754579646815604023525655); | |
92 DVK(KP603558818, +0.603558818296015001454675132653458027918768137); | |
93 DVK(KP667278218, +0.667278218140296670899089292254759909713898805); | |
94 DVK(KP522847744, +0.522847744331509716623755382187077770911012542); | |
95 DVK(KP132830569, +0.132830569247582714407653942074819768844536507); | |
96 DVK(KP786782374, +0.786782374965295178365099601674911834788448471); | |
97 DVK(KP869845200, +0.869845200362138853122720822420327157933056305); | |
98 DVK(KP120146378, +0.120146378570687701782758537356596213647956445); | |
99 DVK(KP987388751, +0.987388751065621252324603216482382109400433949); | |
100 DVK(KP059835404, +0.059835404262124915169548397419498386427871950); | |
101 DVK(KP066152395, +0.066152395967733048213034281011006031460903353); | |
102 DVK(KP893101515, +0.893101515366181661711202267938416198338079437); | |
103 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
104 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
105 DVK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
106 { | |
107 INT i; | |
108 const R *xi; | |
109 R *xo; | |
110 xi = ii; | |
111 xo = io; | |
112 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(50, is), MAKE_VOLATILE_STRIDE(50, os)) { | |
113 V Ta, T2z, T1q, T9, T3n, T3r, T3s, T3t, T1a, T2N, T2V, T1j, T1J, T2o, T2t; | |
114 V T1R, TV, T2O, T2W, T1i, T1K, T2l, T2s, T1S, T3o, T3p, T3q, TF, T2R, T2Y; | |
115 V T1f, T1N, T2e, T2v, T1V, Tq, T2Q, T2Z, T1e, T1M, T2h, T2w, T1U; | |
116 { | |
117 V T1, T7, T1p, T4, T1o, T8; | |
118 T1 = LD(&(xi[0]), ivs, &(xi[0])); | |
119 { | |
120 V T5, T6, T2, T3; | |
121 T5 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); | |
122 T6 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); | |
123 T7 = VADD(T5, T6); | |
124 T1p = VSUB(T5, T6); | |
125 T2 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); | |
126 T3 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); | |
127 T4 = VADD(T2, T3); | |
128 T1o = VSUB(T2, T3); | |
129 } | |
130 Ta = VSUB(T4, T7); | |
131 T2z = VFNMS(LDK(KP618033988), T1o, T1p); | |
132 T1q = VFMA(LDK(KP618033988), T1p, T1o); | |
133 T8 = VADD(T4, T7); | |
134 T9 = VFNMS(LDK(KP250000000), T8, T1); | |
135 T3n = VADD(T1, T8); | |
136 } | |
137 { | |
138 V TH, TW, TO, TT, TQ, TS, T13, T18, T15, T17; | |
139 TH = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); | |
140 TW = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); | |
141 { | |
142 V TI, TJ, TK, TL, TM, TN; | |
143 TI = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); | |
144 TJ = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); | |
145 TK = VADD(TI, TJ); | |
146 TL = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); | |
147 TM = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); | |
148 TN = VADD(TL, TM); | |
149 TO = VADD(TK, TN); | |
150 TT = VSUB(TM, TL); | |
151 TQ = VSUB(TN, TK); | |
152 TS = VSUB(TI, TJ); | |
153 } | |
154 { | |
155 V TX, TY, TZ, T10, T11, T12; | |
156 TX = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); | |
157 TY = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); | |
158 TZ = VADD(TX, TY); | |
159 T10 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); | |
160 T11 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); | |
161 T12 = VADD(T10, T11); | |
162 T13 = VADD(TZ, T12); | |
163 T18 = VSUB(T10, T11); | |
164 T15 = VSUB(T12, TZ); | |
165 T17 = VSUB(TX, TY); | |
166 } | |
167 T3r = VADD(TH, TO); | |
168 T3s = VADD(TW, T13); | |
169 T3t = VADD(T3r, T3s); | |
170 { | |
171 V T19, T2m, T16, T2n, T14; | |
172 T19 = VFMA(LDK(KP618033988), T18, T17); | |
173 T2m = VFNMS(LDK(KP618033988), T17, T18); | |
174 T14 = VFNMS(LDK(KP250000000), T13, TW); | |
175 T16 = VFNMS(LDK(KP559016994), T15, T14); | |
176 T2n = VFMA(LDK(KP559016994), T15, T14); | |
177 T1a = VFNMS(LDK(KP893101515), T19, T16); | |
178 T2N = VFMA(LDK(KP066152395), T2n, T2m); | |
179 T2V = VFNMS(LDK(KP059835404), T2m, T2n); | |
180 T1j = VFMA(LDK(KP987388751), T16, T19); | |
181 T1J = VFNMS(LDK(KP120146378), T19, T16); | |
182 T2o = VFMA(LDK(KP869845200), T2n, T2m); | |
183 T2t = VFNMS(LDK(KP786782374), T2m, T2n); | |
184 T1R = VFMA(LDK(KP132830569), T16, T19); | |
185 } | |
186 { | |
187 V TU, T2j, TR, T2k, TP; | |
188 TU = VFNMS(LDK(KP618033988), TT, TS); | |
189 T2j = VFMA(LDK(KP618033988), TS, TT); | |
190 TP = VFNMS(LDK(KP250000000), TO, TH); | |
191 TR = VFNMS(LDK(KP559016994), TQ, TP); | |
192 T2k = VFMA(LDK(KP559016994), TQ, TP); | |
193 TV = VFNMS(LDK(KP522847744), TU, TR); | |
194 T2O = VFNMS(LDK(KP667278218), T2k, T2j); | |
195 T2W = VFMA(LDK(KP603558818), T2j, T2k); | |
196 T1i = VFMA(LDK(KP578046249), TR, TU); | |
197 T1K = VFNMS(LDK(KP494780565), TR, TU); | |
198 T2l = VFMA(LDK(KP066152395), T2k, T2j); | |
199 T2s = VFNMS(LDK(KP059835404), T2j, T2k); | |
200 T1S = VFMA(LDK(KP447533225), TU, TR); | |
201 } | |
202 } | |
203 { | |
204 V Tc, Ty, Tj, To, Tl, Tn, Tt, TD, Tw, TB; | |
205 Tc = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); | |
206 Ty = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); | |
207 { | |
208 V Td, Te, Tf, Tg, Th, Ti; | |
209 Td = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); | |
210 Te = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); | |
211 Tf = VADD(Td, Te); | |
212 Tg = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); | |
213 Th = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); | |
214 Ti = VADD(Tg, Th); | |
215 Tj = VADD(Tf, Ti); | |
216 To = VSUB(Th, Tg); | |
217 Tl = VSUB(Tf, Ti); | |
218 Tn = VSUB(Td, Te); | |
219 } | |
220 { | |
221 V Tr, Ts, Tz, Tu, Tv, TA; | |
222 Tr = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); | |
223 Ts = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); | |
224 Tz = VADD(Ts, Tr); | |
225 Tu = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); | |
226 Tv = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); | |
227 TA = VADD(Tv, Tu); | |
228 Tt = VSUB(Tr, Ts); | |
229 TD = VSUB(Tz, TA); | |
230 Tw = VSUB(Tu, Tv); | |
231 TB = VADD(Tz, TA); | |
232 } | |
233 T3o = VADD(Tc, Tj); | |
234 T3p = VADD(Ty, TB); | |
235 T3q = VADD(T3o, T3p); | |
236 { | |
237 V Tx, T2d, TE, T2c, TC; | |
238 Tx = VFMA(LDK(KP618033988), Tw, Tt); | |
239 T2d = VFNMS(LDK(KP618033988), Tt, Tw); | |
240 TC = VFMS(LDK(KP250000000), TB, Ty); | |
241 TE = VFNMS(LDK(KP559016994), TD, TC); | |
242 T2c = VFMA(LDK(KP559016994), TD, TC); | |
243 TF = VFNMS(LDK(KP667278218), TE, Tx); | |
244 T2R = VFNMS(LDK(KP494780565), T2c, T2d); | |
245 T2Y = VFMA(LDK(KP447533225), T2d, T2c); | |
246 T1f = VFMA(LDK(KP603558818), Tx, TE); | |
247 T1N = VFMA(LDK(KP869845200), TE, Tx); | |
248 T2e = VFMA(LDK(KP120146378), T2d, T2c); | |
249 T2v = VFNMS(LDK(KP132830569), T2c, T2d); | |
250 T1V = VFNMS(LDK(KP786782374), Tx, TE); | |
251 } | |
252 { | |
253 V Tp, T2g, Tm, T2f, Tk; | |
254 Tp = VFNMS(LDK(KP618033988), To, Tn); | |
255 T2g = VFMA(LDK(KP618033988), Tn, To); | |
256 Tk = VFNMS(LDK(KP250000000), Tj, Tc); | |
257 Tm = VFMA(LDK(KP559016994), Tl, Tk); | |
258 T2f = VFNMS(LDK(KP559016994), Tl, Tk); | |
259 Tq = VFNMS(LDK(KP244189809), Tp, Tm); | |
260 T2Q = VFNMS(LDK(KP522847744), T2g, T2f); | |
261 T2Z = VFMA(LDK(KP578046249), T2f, T2g); | |
262 T1e = VFMA(LDK(KP269969613), Tm, Tp); | |
263 T1M = VFMA(LDK(KP667278218), Tm, Tp); | |
264 T2h = VFMA(LDK(KP893101515), T2g, T2f); | |
265 T2w = VFNMS(LDK(KP987388751), T2f, T2g); | |
266 T1U = VFNMS(LDK(KP603558818), Tp, Tm); | |
267 } | |
268 } | |
269 { | |
270 V T3w, T3u, T3v, T3A, T3C, T3y, T3z, T3B, T3x; | |
271 T3w = VSUB(T3q, T3t); | |
272 T3u = VADD(T3q, T3t); | |
273 T3v = VFNMS(LDK(KP250000000), T3u, T3n); | |
274 T3y = VSUB(T3o, T3p); | |
275 T3z = VSUB(T3r, T3s); | |
276 T3A = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T3z, T3y)); | |
277 T3C = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T3y, T3z)); | |
278 ST(&(xo[0]), VADD(T3u, T3n), ovs, &(xo[0])); | |
279 T3B = VFNMS(LDK(KP559016994), T3w, T3v); | |
280 ST(&(xo[WS(os, 10)]), VFNMSI(T3C, T3B), ovs, &(xo[0])); | |
281 ST(&(xo[WS(os, 15)]), VFMAI(T3C, T3B), ovs, &(xo[WS(os, 1)])); | |
282 T3x = VFMA(LDK(KP559016994), T3w, T3v); | |
283 ST(&(xo[WS(os, 5)]), VFMAI(T3A, T3x), ovs, &(xo[WS(os, 1)])); | |
284 ST(&(xo[WS(os, 20)]), VFNMSI(T3A, T3x), ovs, &(xo[0])); | |
285 } | |
286 { | |
287 V T2B, T2H, T2q, T2E, T2y, T2K, T31, T3a, T3l, T3f, T2b, T35, T34, T2T, T33; | |
288 V T3h, T37; | |
289 T2B = VFMA(LDK(KP734762448), T2w, T2v); | |
290 T2H = VFNMS(LDK(KP734762448), T2h, T2e); | |
291 { | |
292 V T2p, T2i, T2D, T2C; | |
293 T2p = VFNMS(LDK(KP772036680), T2o, T2l); | |
294 T2i = VFMA(LDK(KP734762448), T2h, T2e); | |
295 T2C = VFNMS(LDK(KP772036680), T2t, T2s); | |
296 T2D = VFNMS(LDK(KP522616830), T2i, T2C); | |
297 T2q = VFMA(LDK(KP956723877), T2p, T2i); | |
298 T2E = VFMA(LDK(KP763932022), T2D, T2p); | |
299 } | |
300 { | |
301 V T2u, T2x, T2J, T2I; | |
302 T2u = VFMA(LDK(KP772036680), T2t, T2s); | |
303 T2x = VFNMS(LDK(KP734762448), T2w, T2v); | |
304 T2I = VFMA(LDK(KP772036680), T2o, T2l); | |
305 T2J = VFMA(LDK(KP522616830), T2x, T2I); | |
306 T2y = VFMA(LDK(KP945422727), T2x, T2u); | |
307 T2K = VFNMS(LDK(KP690983005), T2J, T2u); | |
308 } | |
309 { | |
310 V T3e, T3d, T3k, T36, T2P, T2S; | |
311 { | |
312 V T2X, T30, T3b, T3c; | |
313 T2X = VFMA(LDK(KP845997307), T2W, T2V); | |
314 T30 = VFNMS(LDK(KP921078979), T2Z, T2Y); | |
315 T31 = VFNMS(LDK(KP906616052), T30, T2X); | |
316 T3e = VFMA(LDK(KP906616052), T30, T2X); | |
317 T3b = VFMA(LDK(KP845997307), T2O, T2N); | |
318 T3c = VFMA(LDK(KP982009705), T2R, T2Q); | |
319 T3d = VFMA(LDK(KP570584518), T3c, T3b); | |
320 T3k = VFNMS(LDK(KP669429328), T3b, T3c); | |
321 } | |
322 T3a = VFMA(LDK(KP262346850), T31, T2z); | |
323 T3l = VFNMS(LDK(KP669429328), T3e, T3k); | |
324 T3f = VFMA(LDK(KP618033988), T3e, T3d); | |
325 T2b = VFNMS(LDK(KP559016994), Ta, T9); | |
326 T35 = VFMA(LDK(KP921078979), T2Z, T2Y); | |
327 T34 = VFNMS(LDK(KP845997307), T2W, T2V); | |
328 T2P = VFNMS(LDK(KP845997307), T2O, T2N); | |
329 T2S = VFNMS(LDK(KP982009705), T2R, T2Q); | |
330 T2T = VFMA(LDK(KP923225144), T2S, T2P); | |
331 T36 = VFNMS(LDK(KP923225144), T2S, T2P); | |
332 T33 = VFNMS(LDK(KP237294955), T2T, T2b); | |
333 T3h = VFNMS(LDK(KP904508497), T36, T34); | |
334 T37 = VFNMS(LDK(KP997675361), T36, T35); | |
335 } | |
336 { | |
337 V T2r, T2A, T2U, T32; | |
338 T2r = VFMA(LDK(KP992114701), T2q, T2b); | |
339 T2A = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T2z, T2y)); | |
340 ST(&(xo[WS(os, 22)]), VFNMSI(T2A, T2r), ovs, &(xo[0])); | |
341 ST(&(xo[WS(os, 3)]), VFMAI(T2A, T2r), ovs, &(xo[WS(os, 1)])); | |
342 T2U = VFMA(LDK(KP949179823), T2T, T2b); | |
343 T32 = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T2z, T31)); | |
344 ST(&(xo[WS(os, 23)]), VFNMSI(T32, T2U), ovs, &(xo[WS(os, 1)])); | |
345 ST(&(xo[WS(os, 2)]), VFMAI(T32, T2U), ovs, &(xo[0])); | |
346 } | |
347 { | |
348 V T3g, T39, T38, T3m, T3j, T3i; | |
349 T3g = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T3f, T3a)); | |
350 T38 = VFMA(LDK(KP560319534), T37, T34); | |
351 T39 = VFNMS(LDK(KP949179823), T38, T33); | |
352 ST(&(xo[WS(os, 12)]), VFNMSI(T3g, T39), ovs, &(xo[0])); | |
353 ST(&(xo[WS(os, 13)]), VFMAI(T3g, T39), ovs, &(xo[WS(os, 1)])); | |
354 T3m = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T3l, T3a)); | |
355 T3i = VFNMS(LDK(KP681693190), T3h, T35); | |
356 T3j = VFNMS(LDK(KP860541664), T3i, T33); | |
357 ST(&(xo[WS(os, 7)]), VFNMSI(T3m, T3j), ovs, &(xo[WS(os, 1)])); | |
358 ST(&(xo[WS(os, 18)]), VFMAI(T3m, T3j), ovs, &(xo[0])); | |
359 { | |
360 V T2G, T2M, T2F, T2L; | |
361 T2F = VFNMS(LDK(KP855719849), T2E, T2B); | |
362 T2G = VFMA(LDK(KP897376177), T2F, T2b); | |
363 T2L = VFMA(LDK(KP855719849), T2K, T2H); | |
364 T2M = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2L, T2z)); | |
365 ST(&(xo[WS(os, 8)]), VFMAI(T2M, T2G), ovs, &(xo[0])); | |
366 ST(&(xo[WS(os, 17)]), VFNMSI(T2M, T2G), ovs, &(xo[WS(os, 1)])); | |
367 } | |
368 } | |
369 } | |
370 { | |
371 V T1Z, T25, T1P, T22, T1X, T28, T1t, T1u, T1F, T1z, Tb, T1k, T1g, T1c, T1d; | |
372 V T1B, T1l; | |
373 T1Z = VFNMS(LDK(KP912575812), T1V, T1U); | |
374 T25 = VFNMS(LDK(KP912575812), T1N, T1M); | |
375 { | |
376 V T1L, T1O, T21, T20; | |
377 T1L = VFNMS(LDK(KP867381224), T1K, T1J); | |
378 T1O = VFMA(LDK(KP912575812), T1N, T1M); | |
379 T20 = VFNMS(LDK(KP958953096), T1S, T1R); | |
380 T21 = VFMA(LDK(KP447417479), T1O, T20); | |
381 T1P = VFNMS(LDK(KP809385824), T1O, T1L); | |
382 T22 = VFMA(LDK(KP690983005), T21, T1L); | |
383 } | |
384 { | |
385 V T1T, T1W, T27, T26; | |
386 T1T = VFMA(LDK(KP958953096), T1S, T1R); | |
387 T1W = VFMA(LDK(KP912575812), T1V, T1U); | |
388 T26 = VFMA(LDK(KP867381224), T1K, T1J); | |
389 T27 = VFMA(LDK(KP447417479), T1W, T26); | |
390 T1X = VFMA(LDK(KP894834959), T1W, T1T); | |
391 T28 = VFNMS(LDK(KP763932022), T27, T1T); | |
392 } | |
393 { | |
394 V T1y, T1x, T1E, T1h, TG, T1b; | |
395 { | |
396 V T1r, T1s, T1v, T1w; | |
397 T1r = VFNMS(LDK(KP916574801), T1f, T1e); | |
398 T1s = VFMA(LDK(KP831864738), T1j, T1i); | |
399 T1t = VFMA(LDK(KP904730450), T1s, T1r); | |
400 T1y = VFNMS(LDK(KP904730450), T1s, T1r); | |
401 T1v = VFNMS(LDK(KP829049696), TF, Tq); | |
402 T1w = VFNMS(LDK(KP831864738), T1a, TV); | |
403 T1x = VFMA(LDK(KP559154169), T1w, T1v); | |
404 T1E = VFNMS(LDK(KP683113946), T1v, T1w); | |
405 } | |
406 T1u = VFNMS(LDK(KP242145790), T1t, T1q); | |
407 T1F = VFMA(LDK(KP617882369), T1y, T1E); | |
408 T1z = VFMA(LDK(KP559016994), T1y, T1x); | |
409 Tb = VFMA(LDK(KP559016994), Ta, T9); | |
410 T1k = VFNMS(LDK(KP831864738), T1j, T1i); | |
411 T1g = VFMA(LDK(KP916574801), T1f, T1e); | |
412 TG = VFMA(LDK(KP829049696), TF, Tq); | |
413 T1b = VFMA(LDK(KP831864738), T1a, TV); | |
414 T1c = VFMA(LDK(KP904730450), T1b, TG); | |
415 T1h = VFNMS(LDK(KP904730450), T1b, TG); | |
416 T1d = VFNMS(LDK(KP242145790), T1c, Tb); | |
417 T1B = VADD(T1g, T1h); | |
418 T1l = VFNMS(LDK(KP904730450), T1k, T1h); | |
419 } | |
420 { | |
421 V T1H, T1I, T1Q, T1Y; | |
422 T1H = VFMA(LDK(KP968583161), T1c, Tb); | |
423 T1I = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1t, T1q)); | |
424 ST(&(xo[WS(os, 1)]), VFMAI(T1I, T1H), ovs, &(xo[WS(os, 1)])); | |
425 ST(&(xo[WS(os, 24)]), VFNMSI(T1I, T1H), ovs, &(xo[0])); | |
426 T1Q = VFNMS(LDK(KP992114701), T1P, Tb); | |
427 T1Y = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T1X, T1q)); | |
428 ST(&(xo[WS(os, 4)]), VFNMSI(T1Y, T1Q), ovs, &(xo[0])); | |
429 ST(&(xo[WS(os, 21)]), VFMAI(T1Y, T1Q), ovs, &(xo[WS(os, 1)])); | |
430 } | |
431 { | |
432 V T1A, T1n, T1m, T1G, T1D, T1C; | |
433 T1A = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1z, T1u)); | |
434 T1m = VFNMS(LDK(KP618033988), T1l, T1g); | |
435 T1n = VFNMS(LDK(KP876091699), T1m, T1d); | |
436 ST(&(xo[WS(os, 6)]), VFMAI(T1A, T1n), ovs, &(xo[0])); | |
437 ST(&(xo[WS(os, 19)]), VFNMSI(T1A, T1n), ovs, &(xo[WS(os, 1)])); | |
438 T1G = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T1F, T1u)); | |
439 T1C = VFNMS(LDK(KP683113946), T1B, T1k); | |
440 T1D = VFMA(LDK(KP792626838), T1C, T1d); | |
441 ST(&(xo[WS(os, 11)]), VFMAI(T1G, T1D), ovs, &(xo[WS(os, 1)])); | |
442 ST(&(xo[WS(os, 14)]), VFNMSI(T1G, T1D), ovs, &(xo[0])); | |
443 { | |
444 V T24, T2a, T23, T29; | |
445 T23 = VFNMS(LDK(KP999544308), T22, T1Z); | |
446 T24 = VFNMS(LDK(KP803003575), T23, Tb); | |
447 T29 = VFNMS(LDK(KP999544308), T28, T25); | |
448 T2a = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T29, T1q)); | |
449 ST(&(xo[WS(os, 9)]), VFNMSI(T2a, T24), ovs, &(xo[WS(os, 1)])); | |
450 ST(&(xo[WS(os, 16)]), VFMAI(T2a, T24), ovs, &(xo[0])); | |
451 } | |
452 } | |
453 } | |
454 } | |
455 } | |
456 VLEAVE(); | |
457 } | |
458 | |
459 static const kdft_desc desc = { 25, XSIMD_STRING("n1bv_25"), {43, 12, 181, 0}, &GENUS, 0, 0, 0, 0 }; | |
460 | |
461 void XSIMD(codelet_n1bv_25) (planner *p) { | |
462 X(kdft_register) (p, n1bv_25, &desc); | |
463 } | |
464 | |
465 #else | |
466 | |
467 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 25 -name n1bv_25 -include dft/simd/n1b.h */ | |
468 | |
469 /* | |
470 * This function contains 224 FP additions, 140 FP multiplications, | |
471 * (or, 147 additions, 63 multiplications, 77 fused multiply/add), | |
472 * 115 stack variables, 40 constants, and 50 memory accesses | |
473 */ | |
474 #include "dft/simd/n1b.h" | |
475 | |
476 static void n1bv_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
477 { | |
478 DVK(KP809016994, +0.809016994374947424102293417182819058860154590); | |
479 DVK(KP309016994, +0.309016994374947424102293417182819058860154590); | |
480 DVK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
481 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344); | |
482 DVK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
483 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418); | |
484 DVK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
485 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596); | |
486 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
487 DVK(KP250666467, +0.250666467128608490746237519633017587885836494); | |
488 DVK(KP851558583, +0.851558583130145297725004891488503407959946084); | |
489 DVK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
490 DVK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
491 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
492 DVK(KP497379774, +0.497379774329709576484567492012895936835134813); | |
493 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
494 DVK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
495 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
496 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483); | |
497 DVK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
498 DVK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
499 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
500 DVK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
501 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691); | |
502 DVK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
503 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
504 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
505 DVK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
506 DVK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
507 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328); | |
508 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
509 DVK(KP125581039, +0.125581039058626752152356449131262266244969664); | |
510 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465); | |
511 DVK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
512 DVK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
513 DVK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
514 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
515 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
516 DVK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
517 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
518 { | |
519 INT i; | |
520 const R *xi; | |
521 R *xo; | |
522 xi = ii; | |
523 xo = io; | |
524 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(50, is), MAKE_VOLATILE_STRIDE(50, os)) { | |
525 V T1b, T2o, T1v, T1e, T2W, T2P, T2Q, T2U, T11, T27, TY, T26, T12, T2f, T1j; | |
526 V T28, TM, T24, TJ, T23, TN, T2e, T1i, T25, T2M, T2N, T2T, Tm, T1W, Tt; | |
527 V T1X, Tu, T20, Tw, T1Y, T7, T1U, Te, T1T, Tf, T21, Tx, T1V; | |
528 { | |
529 V T1c, T1a, T1t, T17, T1r; | |
530 T1c = LD(&(xi[0]), ivs, &(xi[0])); | |
531 { | |
532 V T18, T19, T15, T16; | |
533 T18 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); | |
534 T19 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); | |
535 T1a = VADD(T18, T19); | |
536 T1t = VSUB(T18, T19); | |
537 T15 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); | |
538 T16 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); | |
539 T17 = VADD(T15, T16); | |
540 T1r = VSUB(T15, T16); | |
541 } | |
542 { | |
543 V T2n, T1s, T1u, T1d; | |
544 T1b = VMUL(LDK(KP559016994), VSUB(T17, T1a)); | |
545 T2n = VMUL(LDK(KP587785252), T1r); | |
546 T2o = VFNMS(LDK(KP951056516), T1t, T2n); | |
547 T1s = VMUL(LDK(KP951056516), T1r); | |
548 T1u = VMUL(LDK(KP587785252), T1t); | |
549 T1v = VADD(T1s, T1u); | |
550 T1d = VADD(T17, T1a); | |
551 T1e = VFNMS(LDK(KP250000000), T1d, T1c); | |
552 T2W = VADD(T1c, T1d); | |
553 } | |
554 } | |
555 { | |
556 V TG, TV, TF, TL, TH, TK, TU, T10, TW, TZ, TX, TI; | |
557 TG = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); | |
558 TV = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); | |
559 { | |
560 V Tz, TA, TB, TC, TD, TE; | |
561 Tz = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); | |
562 TA = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); | |
563 TB = VADD(Tz, TA); | |
564 TC = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); | |
565 TD = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); | |
566 TE = VADD(TC, TD); | |
567 TF = VMUL(LDK(KP559016994), VSUB(TB, TE)); | |
568 TL = VSUB(TC, TD); | |
569 TH = VADD(TB, TE); | |
570 TK = VSUB(Tz, TA); | |
571 } | |
572 { | |
573 V TO, TP, TQ, TR, TS, TT; | |
574 TO = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); | |
575 TP = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); | |
576 TQ = VADD(TO, TP); | |
577 TR = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); | |
578 TS = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); | |
579 TT = VADD(TR, TS); | |
580 TU = VMUL(LDK(KP559016994), VSUB(TQ, TT)); | |
581 T10 = VSUB(TR, TS); | |
582 TW = VADD(TQ, TT); | |
583 TZ = VSUB(TO, TP); | |
584 } | |
585 T2P = VADD(TG, TH); | |
586 T2Q = VADD(TV, TW); | |
587 T2U = VADD(T2P, T2Q); | |
588 T11 = VFMA(LDK(KP475528258), TZ, VMUL(LDK(KP293892626), T10)); | |
589 T27 = VFNMS(LDK(KP475528258), T10, VMUL(LDK(KP293892626), TZ)); | |
590 TX = VFNMS(LDK(KP250000000), TW, TV); | |
591 TY = VADD(TU, TX); | |
592 T26 = VSUB(TX, TU); | |
593 T12 = VFNMS(LDK(KP1_369094211), T11, VMUL(LDK(KP728968627), TY)); | |
594 T2f = VFMA(LDK(KP125581039), T27, VMUL(LDK(KP998026728), T26)); | |
595 T1j = VFMA(LDK(KP1_457937254), T11, VMUL(LDK(KP684547105), TY)); | |
596 T28 = VFNMS(LDK(KP1_996053456), T27, VMUL(LDK(KP062790519), T26)); | |
597 TM = VFMA(LDK(KP475528258), TK, VMUL(LDK(KP293892626), TL)); | |
598 T24 = VFNMS(LDK(KP475528258), TL, VMUL(LDK(KP293892626), TK)); | |
599 TI = VFNMS(LDK(KP250000000), TH, TG); | |
600 TJ = VADD(TF, TI); | |
601 T23 = VSUB(TI, TF); | |
602 TN = VFNMS(LDK(KP963507348), TM, VMUL(LDK(KP876306680), TJ)); | |
603 T2e = VFMA(LDK(KP1_071653589), T24, VMUL(LDK(KP844327925), T23)); | |
604 T1i = VFMA(LDK(KP1_752613360), TM, VMUL(LDK(KP481753674), TJ)); | |
605 T25 = VFNMS(LDK(KP1_688655851), T24, VMUL(LDK(KP535826794), T23)); | |
606 } | |
607 { | |
608 V Tb, Tq, T3, Tc, T6, Ta, Ti, Tr, Tl, Tp, Ts, Td; | |
609 Tb = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); | |
610 Tq = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); | |
611 { | |
612 V T1, T2, T8, T4, T5, T9; | |
613 T1 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); | |
614 T2 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); | |
615 T8 = VADD(T1, T2); | |
616 T4 = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); | |
617 T5 = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); | |
618 T9 = VADD(T4, T5); | |
619 T3 = VSUB(T1, T2); | |
620 Tc = VADD(T8, T9); | |
621 T6 = VSUB(T4, T5); | |
622 Ta = VMUL(LDK(KP559016994), VSUB(T8, T9)); | |
623 } | |
624 { | |
625 V Tg, Th, Tn, Tj, Tk, To; | |
626 Tg = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); | |
627 Th = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); | |
628 Tn = VADD(Tg, Th); | |
629 Tj = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); | |
630 Tk = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); | |
631 To = VADD(Tj, Tk); | |
632 Ti = VSUB(Tg, Th); | |
633 Tr = VADD(Tn, To); | |
634 Tl = VSUB(Tj, Tk); | |
635 Tp = VMUL(LDK(KP559016994), VSUB(Tn, To)); | |
636 } | |
637 T2M = VADD(Tq, Tr); | |
638 T2N = VADD(Tb, Tc); | |
639 T2T = VADD(T2M, T2N); | |
640 Tm = VFMA(LDK(KP475528258), Ti, VMUL(LDK(KP293892626), Tl)); | |
641 T1W = VFNMS(LDK(KP475528258), Tl, VMUL(LDK(KP293892626), Ti)); | |
642 Ts = VFNMS(LDK(KP250000000), Tr, Tq); | |
643 Tt = VADD(Tp, Ts); | |
644 T1X = VSUB(Ts, Tp); | |
645 Tu = VFMA(LDK(KP1_937166322), Tm, VMUL(LDK(KP248689887), Tt)); | |
646 T20 = VFNMS(LDK(KP963507348), T1W, VMUL(LDK(KP876306680), T1X)); | |
647 Tw = VFNMS(LDK(KP497379774), Tm, VMUL(LDK(KP968583161), Tt)); | |
648 T1Y = VFMA(LDK(KP1_752613360), T1W, VMUL(LDK(KP481753674), T1X)); | |
649 T7 = VFMA(LDK(KP475528258), T3, VMUL(LDK(KP293892626), T6)); | |
650 T1U = VFNMS(LDK(KP475528258), T6, VMUL(LDK(KP293892626), T3)); | |
651 Td = VFNMS(LDK(KP250000000), Tc, Tb); | |
652 Te = VADD(Ta, Td); | |
653 T1T = VSUB(Td, Ta); | |
654 Tf = VFMA(LDK(KP1_071653589), T7, VMUL(LDK(KP844327925), Te)); | |
655 T21 = VFMA(LDK(KP1_809654104), T1U, VMUL(LDK(KP425779291), T1T)); | |
656 Tx = VFNMS(LDK(KP1_688655851), T7, VMUL(LDK(KP535826794), Te)); | |
657 T1V = VFNMS(LDK(KP851558583), T1U, VMUL(LDK(KP904827052), T1T)); | |
658 } | |
659 { | |
660 V T2V, T2X, T2Y, T2S, T30, T2O, T2R, T31, T2Z; | |
661 T2V = VMUL(LDK(KP559016994), VSUB(T2T, T2U)); | |
662 T2X = VADD(T2T, T2U); | |
663 T2Y = VFNMS(LDK(KP250000000), T2X, T2W); | |
664 T2O = VSUB(T2M, T2N); | |
665 T2R = VSUB(T2P, T2Q); | |
666 T2S = VBYI(VFMA(LDK(KP951056516), T2O, VMUL(LDK(KP587785252), T2R))); | |
667 T30 = VBYI(VFNMS(LDK(KP951056516), T2R, VMUL(LDK(KP587785252), T2O))); | |
668 ST(&(xo[0]), VADD(T2W, T2X), ovs, &(xo[0])); | |
669 T31 = VSUB(T2Y, T2V); | |
670 ST(&(xo[WS(os, 10)]), VADD(T30, T31), ovs, &(xo[0])); | |
671 ST(&(xo[WS(os, 15)]), VSUB(T31, T30), ovs, &(xo[WS(os, 1)])); | |
672 T2Z = VADD(T2V, T2Y); | |
673 ST(&(xo[WS(os, 5)]), VADD(T2S, T2Z), ovs, &(xo[WS(os, 1)])); | |
674 ST(&(xo[WS(os, 20)]), VSUB(T2Z, T2S), ovs, &(xo[0])); | |
675 } | |
676 { | |
677 V T1Z, T2i, T2j, T2g, T2w, T2x, T2y, T2G, T2H, T2I, T2D, T2E, T2F, T2z, T2A; | |
678 V T2B, T2p, T2m, T2q, T2b, T2c, T2a, T2d, T2h, T2r; | |
679 T1Z = VSUB(T1V, T1Y); | |
680 T2i = VADD(T20, T21); | |
681 T2j = VSUB(T25, T28); | |
682 T2g = VSUB(T2e, T2f); | |
683 T2w = VFMA(LDK(KP1_369094211), T1W, VMUL(LDK(KP728968627), T1X)); | |
684 T2x = VFNMS(LDK(KP992114701), T1T, VMUL(LDK(KP250666467), T1U)); | |
685 T2y = VADD(T2w, T2x); | |
686 T2G = VFNMS(LDK(KP125581039), T24, VMUL(LDK(KP998026728), T23)); | |
687 T2H = VFMA(LDK(KP1_274847979), T27, VMUL(LDK(KP770513242), T26)); | |
688 T2I = VADD(T2G, T2H); | |
689 T2D = VFNMS(LDK(KP1_457937254), T1W, VMUL(LDK(KP684547105), T1X)); | |
690 T2E = VFMA(LDK(KP1_984229402), T1U, VMUL(LDK(KP125333233), T1T)); | |
691 T2F = VADD(T2D, T2E); | |
692 T2z = VFMA(LDK(KP1_996053456), T24, VMUL(LDK(KP062790519), T23)); | |
693 T2A = VFNMS(LDK(KP637423989), T26, VMUL(LDK(KP1_541026485), T27)); | |
694 T2B = VADD(T2z, T2A); | |
695 { | |
696 V T2k, T2l, T22, T29; | |
697 T2k = VADD(T1Y, T1V); | |
698 T2l = VADD(T2e, T2f); | |
699 T2p = VADD(T2k, T2l); | |
700 T2m = VMUL(LDK(KP559016994), VSUB(T2k, T2l)); | |
701 T2q = VFNMS(LDK(KP250000000), T2p, T2o); | |
702 T2b = VSUB(T1e, T1b); | |
703 T22 = VSUB(T20, T21); | |
704 T29 = VADD(T25, T28); | |
705 T2c = VADD(T22, T29); | |
706 T2a = VMUL(LDK(KP559016994), VSUB(T22, T29)); | |
707 T2d = VFNMS(LDK(KP250000000), T2c, T2b); | |
708 } | |
709 { | |
710 V T2u, T2v, T2C, T2J; | |
711 T2u = VADD(T2b, T2c); | |
712 T2v = VBYI(VADD(T2o, T2p)); | |
713 ST(&(xo[WS(os, 23)]), VSUB(T2u, T2v), ovs, &(xo[WS(os, 1)])); | |
714 ST(&(xo[WS(os, 2)]), VADD(T2u, T2v), ovs, &(xo[0])); | |
715 T2C = VADD(T2b, VADD(T2y, T2B)); | |
716 T2J = VBYI(VSUB(VADD(T2F, T2I), T2o)); | |
717 ST(&(xo[WS(os, 22)]), VSUB(T2C, T2J), ovs, &(xo[0])); | |
718 ST(&(xo[WS(os, 3)]), VADD(T2C, T2J), ovs, &(xo[WS(os, 1)])); | |
719 } | |
720 T2h = VFMA(LDK(KP951056516), T1Z, VADD(T2a, VFNMS(LDK(KP587785252), T2g, T2d))); | |
721 T2r = VBYI(VADD(VFMA(LDK(KP951056516), T2i, VMUL(LDK(KP587785252), T2j)), VADD(T2m, T2q))); | |
722 ST(&(xo[WS(os, 18)]), VSUB(T2h, T2r), ovs, &(xo[0])); | |
723 ST(&(xo[WS(os, 7)]), VADD(T2h, T2r), ovs, &(xo[WS(os, 1)])); | |
724 { | |
725 V T2s, T2t, T2K, T2L; | |
726 T2s = VFMA(LDK(KP587785252), T1Z, VFMA(LDK(KP951056516), T2g, VSUB(T2d, T2a))); | |
727 T2t = VBYI(VADD(VFNMS(LDK(KP951056516), T2j, VMUL(LDK(KP587785252), T2i)), VSUB(T2q, T2m))); | |
728 ST(&(xo[WS(os, 13)]), VSUB(T2s, T2t), ovs, &(xo[WS(os, 1)])); | |
729 ST(&(xo[WS(os, 12)]), VADD(T2s, T2t), ovs, &(xo[0])); | |
730 T2K = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T2w, T2x), VFMA(LDK(KP309016994), T2F, VFNMS(LDK(KP809016994), T2I, VMUL(LDK(KP587785252), VSUB(T2z, T2A))))), T2o)); | |
731 T2L = VFMA(LDK(KP309016994), T2y, VFMA(LDK(KP951056516), VSUB(T2E, T2D), VFMA(LDK(KP587785252), VSUB(T2H, T2G), VFNMS(LDK(KP809016994), T2B, T2b)))); | |
732 ST(&(xo[WS(os, 8)]), VADD(T2K, T2L), ovs, &(xo[0])); | |
733 ST(&(xo[WS(os, 17)]), VSUB(T2L, T2K), ovs, &(xo[WS(os, 1)])); | |
734 } | |
735 } | |
736 { | |
737 V Tv, T1m, T1n, T1k, T1D, T1E, T1F, T1N, T1O, T1P, T1K, T1L, T1M, T1G, T1H; | |
738 V T1I, T1w, T1q, T1x, T1f, T1g, T14, T1h, T1l, T1y; | |
739 Tv = VSUB(Tf, Tu); | |
740 T1m = VSUB(Tw, Tx); | |
741 T1n = VSUB(TN, T12); | |
742 T1k = VSUB(T1i, T1j); | |
743 T1D = VFMA(LDK(KP1_688655851), Tm, VMUL(LDK(KP535826794), Tt)); | |
744 T1E = VFMA(LDK(KP1_541026485), T7, VMUL(LDK(KP637423989), Te)); | |
745 T1F = VSUB(T1D, T1E); | |
746 T1N = VFMA(LDK(KP851558583), TM, VMUL(LDK(KP904827052), TJ)); | |
747 T1O = VFMA(LDK(KP1_984229402), T11, VMUL(LDK(KP125333233), TY)); | |
748 T1P = VADD(T1N, T1O); | |
749 T1K = VFNMS(LDK(KP1_071653589), Tm, VMUL(LDK(KP844327925), Tt)); | |
750 T1L = VFNMS(LDK(KP770513242), Te, VMUL(LDK(KP1_274847979), T7)); | |
751 T1M = VADD(T1K, T1L); | |
752 T1G = VFNMS(LDK(KP425779291), TJ, VMUL(LDK(KP1_809654104), TM)); | |
753 T1H = VFNMS(LDK(KP992114701), TY, VMUL(LDK(KP250666467), T11)); | |
754 T1I = VADD(T1G, T1H); | |
755 { | |
756 V T1o, T1p, Ty, T13; | |
757 T1o = VADD(Tu, Tf); | |
758 T1p = VADD(T1i, T1j); | |
759 T1w = VADD(T1o, T1p); | |
760 T1q = VMUL(LDK(KP559016994), VSUB(T1o, T1p)); | |
761 T1x = VFNMS(LDK(KP250000000), T1w, T1v); | |
762 T1f = VADD(T1b, T1e); | |
763 Ty = VADD(Tw, Tx); | |
764 T13 = VADD(TN, T12); | |
765 T1g = VADD(Ty, T13); | |
766 T14 = VMUL(LDK(KP559016994), VSUB(Ty, T13)); | |
767 T1h = VFNMS(LDK(KP250000000), T1g, T1f); | |
768 } | |
769 { | |
770 V T1B, T1C, T1J, T1Q; | |
771 T1B = VADD(T1f, T1g); | |
772 T1C = VBYI(VADD(T1v, T1w)); | |
773 ST(&(xo[WS(os, 24)]), VSUB(T1B, T1C), ovs, &(xo[0])); | |
774 ST(&(xo[WS(os, 1)]), VADD(T1B, T1C), ovs, &(xo[WS(os, 1)])); | |
775 T1J = VADD(T1f, VADD(T1F, T1I)); | |
776 T1Q = VBYI(VSUB(VADD(T1M, T1P), T1v)); | |
777 ST(&(xo[WS(os, 21)]), VSUB(T1J, T1Q), ovs, &(xo[WS(os, 1)])); | |
778 ST(&(xo[WS(os, 4)]), VADD(T1J, T1Q), ovs, &(xo[0])); | |
779 } | |
780 T1l = VFMA(LDK(KP951056516), Tv, VADD(T14, VFNMS(LDK(KP587785252), T1k, T1h))); | |
781 T1y = VBYI(VADD(VFMA(LDK(KP951056516), T1m, VMUL(LDK(KP587785252), T1n)), VADD(T1q, T1x))); | |
782 ST(&(xo[WS(os, 19)]), VSUB(T1l, T1y), ovs, &(xo[WS(os, 1)])); | |
783 ST(&(xo[WS(os, 6)]), VADD(T1l, T1y), ovs, &(xo[0])); | |
784 { | |
785 V T1z, T1A, T1R, T1S; | |
786 T1z = VFMA(LDK(KP587785252), Tv, VFMA(LDK(KP951056516), T1k, VSUB(T1h, T14))); | |
787 T1A = VBYI(VADD(VFNMS(LDK(KP951056516), T1n, VMUL(LDK(KP587785252), T1m)), VSUB(T1x, T1q))); | |
788 ST(&(xo[WS(os, 14)]), VSUB(T1z, T1A), ovs, &(xo[0])); | |
789 ST(&(xo[WS(os, 11)]), VADD(T1z, T1A), ovs, &(xo[WS(os, 1)])); | |
790 T1R = VBYI(VSUB(VFMA(LDK(KP309016994), T1M, VFMA(LDK(KP951056516), VADD(T1D, T1E), VFNMS(LDK(KP809016994), T1P, VMUL(LDK(KP587785252), VSUB(T1G, T1H))))), T1v)); | |
791 T1S = VFMA(LDK(KP951056516), VSUB(T1L, T1K), VFMA(LDK(KP309016994), T1F, VFMA(LDK(KP587785252), VSUB(T1O, T1N), VFNMS(LDK(KP809016994), T1I, T1f)))); | |
792 ST(&(xo[WS(os, 9)]), VADD(T1R, T1S), ovs, &(xo[WS(os, 1)])); | |
793 ST(&(xo[WS(os, 16)]), VSUB(T1S, T1R), ovs, &(xo[0])); | |
794 } | |
795 } | |
796 } | |
797 } | |
798 VLEAVE(); | |
799 } | |
800 | |
801 static const kdft_desc desc = { 25, XSIMD_STRING("n1bv_25"), {147, 63, 77, 0}, &GENUS, 0, 0, 0, 0 }; | |
802 | |
803 void XSIMD(codelet_n1bv_25) (planner *p) { | |
804 X(kdft_register) (p, n1bv_25, &desc); | |
805 } | |
806 | |
807 #endif |