Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/dft/simd/common/t3fv_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 08:05:54 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_twiddle_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 25 -name t3fv_25 -include dft/simd/t3f.h */ | |
29 | |
30 /* | |
31 * This function contains 268 FP additions, 281 FP multiplications, | |
32 * (or, 87 additions, 100 multiplications, 181 fused multiply/add), | |
33 * 171 stack variables, 67 constants, and 50 memory accesses | |
34 */ | |
35 #include "dft/simd/t3f.h" | |
36 | |
37 static void t3fv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DVK(KP617882369, +0.617882369114440893914546919006756321695042882); | |
40 DVK(KP792626838, +0.792626838241819413632131824093538848057784557); | |
41 DVK(KP876091699, +0.876091699473550838204498029706869638173524346); | |
42 DVK(KP803003575, +0.803003575438660414833440593570376004635464850); | |
43 DVK(KP999544308, +0.999544308746292983948881682379742149196758193); | |
44 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
45 DVK(KP242145790, +0.242145790282157779872542093866183953459003101); | |
46 DVK(KP916574801, +0.916574801383451584742370439148878693530976769); | |
47 DVK(KP269969613, +0.269969613759572083574752974412347470060951301); | |
48 DVK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
49 DVK(KP809385824, +0.809385824416008241660603814668679683846476688); | |
50 DVK(KP894834959, +0.894834959464455102997960030820114611498661386); | |
51 DVK(KP447417479, +0.447417479732227551498980015410057305749330693); | |
52 DVK(KP867381224, +0.867381224396525206773171885031575671309956167); | |
53 DVK(KP958953096, +0.958953096729998668045963838399037225970891871); | |
54 DVK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
55 DVK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
56 DVK(KP831864738, +0.831864738706457140726048799369896829771167132); | |
57 DVK(KP829049696, +0.829049696159252993975487806364305442437946767); | |
58 DVK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
59 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
60 DVK(KP262346850, +0.262346850930607871785420028382979691334784273); | |
61 DVK(KP860541664, +0.860541664367944677098261680920518816412804187); | |
62 DVK(KP681693190, +0.681693190061530575150324149145440022633095390); | |
63 DVK(KP560319534, +0.560319534973832390111614715371676131169633784); | |
64 DVK(KP897376177, +0.897376177523557693138608077137219684419427330); | |
65 DVK(KP855719849, +0.855719849902058969314654733608091555096772472); | |
66 DVK(KP949179823, +0.949179823508441261575555465843363271711583843); | |
67 DVK(KP952936919, +0.952936919628306576880750665357914584765951388); | |
68 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
69 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
70 DVK(KP997675361, +0.997675361079556513670859573984492383596555031); | |
71 DVK(KP237294955, +0.237294955877110315393888866460840817927895961); | |
72 DVK(KP904508497, +0.904508497187473712051146708591409529430077295); | |
73 DVK(KP906616052, +0.906616052148196230441134447086066874408359177); | |
74 DVK(KP923225144, +0.923225144846402650453449441572664695995209956); | |
75 DVK(KP921078979, +0.921078979742360627699756128143719920817673854); | |
76 DVK(KP578046249, +0.578046249379945007321754579646815604023525655); | |
77 DVK(KP763932022, +0.763932022500210303590826331268723764559381640); | |
78 DVK(KP956723877, +0.956723877038460305821989399535483155872969262); | |
79 DVK(KP690983005, +0.690983005625052575897706582817180941139845410); | |
80 DVK(KP945422727, +0.945422727388575946270360266328811958657216298); | |
81 DVK(KP522616830, +0.522616830205754336872861364785224694908468440); | |
82 DVK(KP772036680, +0.772036680810363904029489473607579825330539880); | |
83 DVK(KP669429328, +0.669429328479476605641803240971985825917022098); | |
84 DVK(KP570584518, +0.570584518783621657366766175430996792655723863); | |
85 DVK(KP982009705, +0.982009705009746369461829878184175962711969869); | |
86 DVK(KP845997307, +0.845997307939530944175097360758058292389769300); | |
87 DVK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
88 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
89 DVK(KP447533225, +0.447533225982656890041886979663652563063114397); | |
90 DVK(KP059835404, +0.059835404262124915169548397419498386427871950); | |
91 DVK(KP494780565, +0.494780565770515410344588413655324772219443730); | |
92 DVK(KP603558818, +0.603558818296015001454675132653458027918768137); | |
93 DVK(KP987388751, +0.987388751065621252324603216482382109400433949); | |
94 DVK(KP522847744, +0.522847744331509716623755382187077770911012542); | |
95 DVK(KP667278218, +0.667278218140296670899089292254759909713898805); | |
96 DVK(KP244189809, +0.244189809627953270309879511234821255780225091); | |
97 DVK(KP132830569, +0.132830569247582714407653942074819768844536507); | |
98 DVK(KP869845200, +0.869845200362138853122720822420327157933056305); | |
99 DVK(KP786782374, +0.786782374965295178365099601674911834788448471); | |
100 DVK(KP066152395, +0.066152395967733048213034281011006031460903353); | |
101 DVK(KP120146378, +0.120146378570687701782758537356596213647956445); | |
102 DVK(KP893101515, +0.893101515366181661711202267938416198338079437); | |
103 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
104 DVK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
105 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
106 { | |
107 INT m; | |
108 R *x; | |
109 x = ri; | |
110 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(25, rs)) { | |
111 V T2, T5, T3, T4, TC, Te, Tr, Ty, Tz, T1I, T1l, T6, T1e, T9, Ta; | |
112 V Tu, T1L, Th, T1E, T1o, TX, TD, T1h, TU; | |
113 T2 = LDW(&(W[0])); | |
114 T5 = LDW(&(W[TWVL * 4])); | |
115 T3 = LDW(&(W[TWVL * 2])); | |
116 T4 = VZMUL(T2, T3); | |
117 TC = VZMULJ(T2, T5); | |
118 Te = VZMUL(T2, T5); | |
119 Tr = VZMULJ(T3, T5); | |
120 Ty = VZMULJ(T2, T3); | |
121 Tz = VZMUL(Ty, T5); | |
122 T1I = VZMUL(T4, T5); | |
123 T1l = VZMUL(T3, T5); | |
124 T6 = VZMULJ(T4, T5); | |
125 T1e = VZMULJ(Ty, T5); | |
126 T9 = LDW(&(W[TWVL * 6])); | |
127 Ta = VZMULJ(T4, T9); | |
128 Tu = VZMULJ(T3, T9); | |
129 T1L = VZMULJ(Tr, T9); | |
130 Th = VZMULJ(T5, T9); | |
131 T1E = VZMULJ(T2, T9); | |
132 T1o = VZMULJ(T1e, T9); | |
133 TX = VZMULJ(Te, T9); | |
134 TD = VZMULJ(TC, T9); | |
135 T1h = VZMULJ(Ty, T9); | |
136 TU = VZMULJ(T6, T9); | |
137 { | |
138 V T1, Tn, Tl, Tm, T2c, T3l, T4e, T1V, T38, T1S, T39, T1W, T2v, T3z, T3f; | |
139 V T3a, T2D, T4a, TN, T32, TK, T31, TO, T2y, T3C, T3i, T33, T2G, T4b, T11; | |
140 V T2Z, T19, T2Y, T1a, T2z, T3D, T3h, T30, T2H, T4d, T1y, T35, T1v, T36, T1z; | |
141 V T2w, T3A, T3e, T37, T2E; | |
142 { | |
143 V Tg, Tj, Tk, T8, Tc, Td, T2a, T2b; | |
144 T1 = LD(&(x[0]), ms, &(x[0])); | |
145 { | |
146 V Tf, Ti, T7, Tb; | |
147 Tf = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
148 Tg = VZMULJ(Te, Tf); | |
149 Ti = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
150 Tj = VZMULJ(Th, Ti); | |
151 Tk = VADD(Tg, Tj); | |
152 T7 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
153 T8 = VZMULJ(T6, T7); | |
154 Tb = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
155 Tc = VZMULJ(Ta, Tb); | |
156 Td = VADD(T8, Tc); | |
157 } | |
158 Tn = VSUB(Td, Tk); | |
159 Tl = VADD(Td, Tk); | |
160 Tm = VFNMS(LDK(KP250000000), Tl, T1); | |
161 T2a = VSUB(T8, Tc); | |
162 T2b = VSUB(Tg, Tj); | |
163 T2c = VFMA(LDK(KP618033988), T2b, T2a); | |
164 T3l = VFNMS(LDK(KP618033988), T2a, T2b); | |
165 } | |
166 { | |
167 V T1B, T1T, T1U, T1H, T1O, T1P, T1A, T1Q, T1R; | |
168 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
169 T1B = VZMULJ(T3, T1A); | |
170 { | |
171 V T1D, T1N, T1G, T1K; | |
172 { | |
173 V T1C, T1M, T1F, T1J; | |
174 T1C = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
175 T1D = VZMULJ(TC, T1C); | |
176 T1M = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
177 T1N = VZMULJ(T1L, T1M); | |
178 T1F = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
179 T1G = VZMULJ(T1E, T1F); | |
180 T1J = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
181 T1K = VZMULJ(T1I, T1J); | |
182 } | |
183 T1T = VSUB(T1D, T1G); | |
184 T1U = VSUB(T1K, T1N); | |
185 T1H = VADD(T1D, T1G); | |
186 T1O = VADD(T1K, T1N); | |
187 T1P = VADD(T1H, T1O); | |
188 } | |
189 T4e = VADD(T1B, T1P); | |
190 T1V = VFMA(LDK(KP618033988), T1U, T1T); | |
191 T38 = VFNMS(LDK(KP618033988), T1T, T1U); | |
192 T1Q = VFNMS(LDK(KP250000000), T1P, T1B); | |
193 T1R = VSUB(T1O, T1H); | |
194 T1S = VFNMS(LDK(KP559016994), T1R, T1Q); | |
195 T39 = VFMA(LDK(KP559016994), T1R, T1Q); | |
196 T1W = VFNMS(LDK(KP893101515), T1V, T1S); | |
197 T2v = VFNMS(LDK(KP120146378), T1V, T1S); | |
198 T3z = VFMA(LDK(KP066152395), T39, T38); | |
199 T3f = VFNMS(LDK(KP786782374), T38, T39); | |
200 T3a = VFMA(LDK(KP869845200), T39, T38); | |
201 T2D = VFMA(LDK(KP132830569), T1S, T1V); | |
202 } | |
203 { | |
204 V Tq, TL, TM, Tx, TG, TH, Tp, TI, TJ; | |
205 Tp = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
206 Tq = VZMULJ(T2, Tp); | |
207 { | |
208 V Tt, TF, Tw, TB; | |
209 { | |
210 V Ts, TE, Tv, TA; | |
211 Ts = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
212 Tt = VZMULJ(Tr, Ts); | |
213 TE = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
214 TF = VZMULJ(TD, TE); | |
215 Tv = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
216 Tw = VZMULJ(Tu, Tv); | |
217 TA = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
218 TB = VZMULJ(Tz, TA); | |
219 } | |
220 TL = VSUB(Tt, Tw); | |
221 TM = VSUB(TF, TB); | |
222 Tx = VADD(Tt, Tw); | |
223 TG = VADD(TB, TF); | |
224 TH = VADD(Tx, TG); | |
225 } | |
226 T4a = VADD(Tq, TH); | |
227 TN = VFNMS(LDK(KP618033988), TM, TL); | |
228 T32 = VFMA(LDK(KP618033988), TL, TM); | |
229 TI = VFNMS(LDK(KP250000000), TH, Tq); | |
230 TJ = VSUB(Tx, TG); | |
231 TK = VFMA(LDK(KP559016994), TJ, TI); | |
232 T31 = VFNMS(LDK(KP559016994), TJ, TI); | |
233 TO = VFNMS(LDK(KP244189809), TN, TK); | |
234 T2y = VFMA(LDK(KP667278218), TK, TN); | |
235 T3C = VFNMS(LDK(KP522847744), T32, T31); | |
236 T3i = VFNMS(LDK(KP987388751), T31, T32); | |
237 T33 = VFMA(LDK(KP893101515), T32, T31); | |
238 T2G = VFNMS(LDK(KP603558818), TN, TK); | |
239 } | |
240 { | |
241 V T13, TT, T10, T14, T15, T16, T12, T17, T18; | |
242 T12 = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
243 T13 = VZMULJ(T4, T12); | |
244 { | |
245 V TQ, TZ, TS, TW; | |
246 { | |
247 V TP, TY, TR, TV; | |
248 TP = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
249 TQ = VZMULJ(T9, TP); | |
250 TY = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
251 TZ = VZMULJ(TX, TY); | |
252 TR = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
253 TS = VZMULJ(T5, TR); | |
254 TV = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
255 TW = VZMULJ(TU, TV); | |
256 } | |
257 TT = VSUB(TQ, TS); | |
258 T10 = VSUB(TW, TZ); | |
259 T14 = VADD(TS, TQ); | |
260 T15 = VADD(TZ, TW); | |
261 T16 = VADD(T14, T15); | |
262 } | |
263 T4b = VADD(T13, T16); | |
264 T11 = VFMA(LDK(KP618033988), T10, TT); | |
265 T2Z = VFNMS(LDK(KP618033988), TT, T10); | |
266 T17 = VFMS(LDK(KP250000000), T16, T13); | |
267 T18 = VSUB(T14, T15); | |
268 T19 = VFNMS(LDK(KP559016994), T18, T17); | |
269 T2Y = VFMA(LDK(KP559016994), T18, T17); | |
270 T1a = VFNMS(LDK(KP667278218), T19, T11); | |
271 T2z = VFMA(LDK(KP869845200), T19, T11); | |
272 T3D = VFNMS(LDK(KP494780565), T2Y, T2Z); | |
273 T3h = VFNMS(LDK(KP132830569), T2Y, T2Z); | |
274 T30 = VFMA(LDK(KP120146378), T2Z, T2Y); | |
275 T2H = VFNMS(LDK(KP786782374), T11, T19); | |
276 } | |
277 { | |
278 V T1d, T1w, T1x, T1k, T1r, T1s, T1c, T1t, T1u; | |
279 T1c = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
280 T1d = VZMULJ(Ty, T1c); | |
281 { | |
282 V T1g, T1q, T1j, T1n; | |
283 { | |
284 V T1f, T1p, T1i, T1m; | |
285 T1f = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
286 T1g = VZMULJ(T1e, T1f); | |
287 T1p = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
288 T1q = VZMULJ(T1o, T1p); | |
289 T1i = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
290 T1j = VZMULJ(T1h, T1i); | |
291 T1m = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
292 T1n = VZMULJ(T1l, T1m); | |
293 } | |
294 T1w = VSUB(T1g, T1j); | |
295 T1x = VSUB(T1q, T1n); | |
296 T1k = VADD(T1g, T1j); | |
297 T1r = VADD(T1n, T1q); | |
298 T1s = VADD(T1k, T1r); | |
299 } | |
300 T4d = VADD(T1d, T1s); | |
301 T1y = VFNMS(LDK(KP618033988), T1x, T1w); | |
302 T35 = VFMA(LDK(KP618033988), T1w, T1x); | |
303 T1t = VFNMS(LDK(KP250000000), T1s, T1d); | |
304 T1u = VSUB(T1r, T1k); | |
305 T1v = VFNMS(LDK(KP559016994), T1u, T1t); | |
306 T36 = VFMA(LDK(KP559016994), T1u, T1t); | |
307 T1z = VFNMS(LDK(KP522847744), T1y, T1v); | |
308 T2w = VFNMS(LDK(KP494780565), T1v, T1y); | |
309 T3A = VFNMS(LDK(KP667278218), T36, T35); | |
310 T3e = VFNMS(LDK(KP059835404), T35, T36); | |
311 T37 = VFMA(LDK(KP066152395), T36, T35); | |
312 T2E = VFMA(LDK(KP447533225), T1y, T1v); | |
313 } | |
314 { | |
315 V T4m, T4o, T49, T4g, T4h, T4i, T4n, T4j; | |
316 { | |
317 V T4k, T4l, T4c, T4f; | |
318 T4k = VSUB(T4a, T4b); | |
319 T4l = VSUB(T4d, T4e); | |
320 T4m = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T4l, T4k)); | |
321 T4o = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T4k, T4l)); | |
322 T49 = VADD(T1, Tl); | |
323 T4c = VADD(T4a, T4b); | |
324 T4f = VADD(T4d, T4e); | |
325 T4g = VADD(T4c, T4f); | |
326 T4h = VFNMS(LDK(KP250000000), T4g, T49); | |
327 T4i = VSUB(T4c, T4f); | |
328 } | |
329 ST(&(x[0]), VADD(T4g, T49), ms, &(x[0])); | |
330 T4n = VFNMS(LDK(KP559016994), T4i, T4h); | |
331 ST(&(x[WS(rs, 10)]), VFMAI(T4o, T4n), ms, &(x[0])); | |
332 ST(&(x[WS(rs, 15)]), VFNMSI(T4o, T4n), ms, &(x[WS(rs, 1)])); | |
333 T4j = VFMA(LDK(KP559016994), T4i, T4h); | |
334 ST(&(x[WS(rs, 5)]), VFNMSI(T4m, T4j), ms, &(x[WS(rs, 1)])); | |
335 ST(&(x[WS(rs, 20)]), VFMAI(T4m, T4j), ms, &(x[0])); | |
336 } | |
337 { | |
338 V T3n, T3t, T3Z, T46, T3k, T3w, T3c, T3q, T2X, T3R, T3F, T3Q, T3N, T43, T3P; | |
339 V T3T, T40, T3X, T3Y; | |
340 T3n = VFMA(LDK(KP734762448), T3i, T3h); | |
341 T3t = VFNMS(LDK(KP734762448), T33, T30); | |
342 T3X = VFMA(LDK(KP845997307), T3A, T3z); | |
343 T3Y = VFMA(LDK(KP982009705), T3D, T3C); | |
344 T3Z = VFMA(LDK(KP570584518), T3Y, T3X); | |
345 T46 = VFNMS(LDK(KP669429328), T3X, T3Y); | |
346 { | |
347 V T3g, T3j, T3v, T3u; | |
348 T3g = VFMA(LDK(KP772036680), T3f, T3e); | |
349 T3j = VFNMS(LDK(KP734762448), T3i, T3h); | |
350 T3u = VFMA(LDK(KP772036680), T3a, T37); | |
351 T3v = VFMA(LDK(KP522616830), T3j, T3u); | |
352 T3k = VFMA(LDK(KP945422727), T3j, T3g); | |
353 T3w = VFNMS(LDK(KP690983005), T3v, T3g); | |
354 } | |
355 { | |
356 V T3b, T34, T3p, T3o; | |
357 T3b = VFNMS(LDK(KP772036680), T3a, T37); | |
358 T34 = VFMA(LDK(KP734762448), T33, T30); | |
359 T3o = VFNMS(LDK(KP772036680), T3f, T3e); | |
360 T3p = VFNMS(LDK(KP522616830), T34, T3o); | |
361 T3c = VFMA(LDK(KP956723877), T3b, T34); | |
362 T3q = VFMA(LDK(KP763932022), T3p, T3b); | |
363 } | |
364 { | |
365 V T3M, T3S, T3J, T3K, T3L; | |
366 T2X = VFNMS(LDK(KP559016994), Tn, Tm); | |
367 T3K = VFMA(LDK(KP447533225), T2Z, T2Y); | |
368 T3L = VFMA(LDK(KP578046249), T31, T32); | |
369 T3M = VFNMS(LDK(KP921078979), T3L, T3K); | |
370 T3R = VFMA(LDK(KP921078979), T3L, T3K); | |
371 { | |
372 V T3B, T3E, T3H, T3I; | |
373 T3B = VFNMS(LDK(KP845997307), T3A, T3z); | |
374 T3E = VFNMS(LDK(KP982009705), T3D, T3C); | |
375 T3F = VFMA(LDK(KP923225144), T3E, T3B); | |
376 T3S = VFNMS(LDK(KP923225144), T3E, T3B); | |
377 T3H = VFNMS(LDK(KP059835404), T38, T39); | |
378 T3I = VFMA(LDK(KP603558818), T35, T36); | |
379 T3J = VFMA(LDK(KP845997307), T3I, T3H); | |
380 T3Q = VFNMS(LDK(KP845997307), T3I, T3H); | |
381 } | |
382 T3N = VFNMS(LDK(KP906616052), T3M, T3J); | |
383 T43 = VFNMS(LDK(KP904508497), T3S, T3Q); | |
384 T3P = VFNMS(LDK(KP237294955), T3F, T2X); | |
385 T3T = VFNMS(LDK(KP997675361), T3S, T3R); | |
386 T40 = VFMA(LDK(KP906616052), T3M, T3J); | |
387 } | |
388 { | |
389 V T3d, T3m, T3G, T3O; | |
390 T3d = VFMA(LDK(KP992114701), T3c, T2X); | |
391 T3m = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T3l, T3k)); | |
392 ST(&(x[WS(rs, 3)]), VFNMSI(T3m, T3d), ms, &(x[WS(rs, 1)])); | |
393 ST(&(x[WS(rs, 22)]), VFMAI(T3m, T3d), ms, &(x[0])); | |
394 T3G = VFMA(LDK(KP949179823), T3F, T2X); | |
395 T3O = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T3l, T3N)); | |
396 ST(&(x[WS(rs, 2)]), VFNMSI(T3O, T3G), ms, &(x[0])); | |
397 ST(&(x[WS(rs, 23)]), VFMAI(T3O, T3G), ms, &(x[WS(rs, 1)])); | |
398 } | |
399 { | |
400 V T3s, T3y, T3r, T3x; | |
401 T3r = VFNMS(LDK(KP855719849), T3q, T3n); | |
402 T3s = VFMA(LDK(KP897376177), T3r, T2X); | |
403 T3x = VFMA(LDK(KP855719849), T3w, T3t); | |
404 T3y = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T3x, T3l)); | |
405 ST(&(x[WS(rs, 8)]), VFNMSI(T3y, T3s), ms, &(x[0])); | |
406 ST(&(x[WS(rs, 17)]), VFMAI(T3y, T3s), ms, &(x[WS(rs, 1)])); | |
407 } | |
408 { | |
409 V T3V, T45, T42, T48, T3U; | |
410 T3U = VFMA(LDK(KP560319534), T3T, T3Q); | |
411 T3V = VFNMS(LDK(KP949179823), T3U, T3P); | |
412 { | |
413 V T44, T3W, T47, T41; | |
414 T44 = VFNMS(LDK(KP681693190), T43, T3R); | |
415 T45 = VFNMS(LDK(KP860541664), T44, T3P); | |
416 T3W = VFMA(LDK(KP262346850), T3N, T3l); | |
417 T47 = VFNMS(LDK(KP669429328), T40, T46); | |
418 T41 = VFMA(LDK(KP618033988), T40, T3Z); | |
419 T42 = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T41, T3W)); | |
420 T48 = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T47, T3W)); | |
421 } | |
422 ST(&(x[WS(rs, 13)]), VFNMSI(T42, T3V), ms, &(x[WS(rs, 1)])); | |
423 ST(&(x[WS(rs, 7)]), VFMAI(T48, T45), ms, &(x[WS(rs, 1)])); | |
424 ST(&(x[WS(rs, 12)]), VFMAI(T42, T3V), ms, &(x[0])); | |
425 ST(&(x[WS(rs, 18)]), VFNMSI(T48, T45), ms, &(x[0])); | |
426 } | |
427 } | |
428 { | |
429 V T2L, T2R, T2j, T2q, T2J, T2U, T2B, T2O, To, T26, T1Y, T22, T1Z, T2n, T27; | |
430 V T2f, T2k, T2h, T2i; | |
431 T2L = VFNMS(LDK(KP912575812), T2H, T2G); | |
432 T2R = VFNMS(LDK(KP912575812), T2z, T2y); | |
433 T2h = VFNMS(LDK(KP829049696), T1a, TO); | |
434 T2i = VFNMS(LDK(KP831864738), T1W, T1z); | |
435 T2j = VFMA(LDK(KP559154169), T2i, T2h); | |
436 T2q = VFNMS(LDK(KP683113946), T2h, T2i); | |
437 { | |
438 V T2F, T2I, T2T, T2S; | |
439 T2F = VFMA(LDK(KP958953096), T2E, T2D); | |
440 T2I = VFMA(LDK(KP912575812), T2H, T2G); | |
441 T2S = VFMA(LDK(KP867381224), T2w, T2v); | |
442 T2T = VFMA(LDK(KP447417479), T2I, T2S); | |
443 T2J = VFMA(LDK(KP894834959), T2I, T2F); | |
444 T2U = VFNMS(LDK(KP763932022), T2T, T2F); | |
445 } | |
446 { | |
447 V T2x, T2A, T2N, T2M; | |
448 T2x = VFNMS(LDK(KP867381224), T2w, T2v); | |
449 T2A = VFMA(LDK(KP912575812), T2z, T2y); | |
450 T2M = VFNMS(LDK(KP958953096), T2E, T2D); | |
451 T2N = VFMA(LDK(KP447417479), T2A, T2M); | |
452 T2B = VFNMS(LDK(KP809385824), T2A, T2x); | |
453 T2O = VFMA(LDK(KP690983005), T2N, T2x); | |
454 } | |
455 { | |
456 V T2e, T23, T2d, T24, T25; | |
457 To = VFMA(LDK(KP559016994), Tn, Tm); | |
458 T24 = VFMA(LDK(KP578046249), T1v, T1y); | |
459 T25 = VFMA(LDK(KP987388751), T1S, T1V); | |
460 T26 = VFNMS(LDK(KP831864738), T25, T24); | |
461 T2e = VFMA(LDK(KP831864738), T25, T24); | |
462 { | |
463 V T1b, T1X, T20, T21; | |
464 T1b = VFMA(LDK(KP829049696), T1a, TO); | |
465 T1X = VFMA(LDK(KP831864738), T1W, T1z); | |
466 T1Y = VFMA(LDK(KP904730450), T1X, T1b); | |
467 T23 = VFNMS(LDK(KP904730450), T1X, T1b); | |
468 T20 = VFMA(LDK(KP269969613), TK, TN); | |
469 T21 = VFMA(LDK(KP603558818), T11, T19); | |
470 T22 = VFMA(LDK(KP916574801), T21, T20); | |
471 T2d = VFNMS(LDK(KP916574801), T21, T20); | |
472 } | |
473 T1Z = VFNMS(LDK(KP242145790), T1Y, To); | |
474 T2n = VADD(T22, T23); | |
475 T27 = VFNMS(LDK(KP904730450), T26, T23); | |
476 T2f = VFMA(LDK(KP904730450), T2e, T2d); | |
477 T2k = VFNMS(LDK(KP904730450), T2e, T2d); | |
478 } | |
479 { | |
480 V T2t, T2u, T2C, T2K; | |
481 T2t = VFMA(LDK(KP968583161), T1Y, To); | |
482 T2u = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T2f, T2c)); | |
483 ST(&(x[WS(rs, 1)]), VFNMSI(T2u, T2t), ms, &(x[WS(rs, 1)])); | |
484 ST(&(x[WS(rs, 24)]), VFMAI(T2u, T2t), ms, &(x[0])); | |
485 T2C = VFNMS(LDK(KP992114701), T2B, To); | |
486 T2K = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2J, T2c)); | |
487 ST(&(x[WS(rs, 4)]), VFMAI(T2K, T2C), ms, &(x[0])); | |
488 ST(&(x[WS(rs, 21)]), VFNMSI(T2K, T2C), ms, &(x[WS(rs, 1)])); | |
489 } | |
490 { | |
491 V T2Q, T2W, T2P, T2V; | |
492 T2P = VFNMS(LDK(KP999544308), T2O, T2L); | |
493 T2Q = VFNMS(LDK(KP803003575), T2P, To); | |
494 T2V = VFNMS(LDK(KP999544308), T2U, T2R); | |
495 T2W = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T2V, T2c)); | |
496 ST(&(x[WS(rs, 16)]), VFNMSI(T2W, T2Q), ms, &(x[0])); | |
497 ST(&(x[WS(rs, 9)]), VFMAI(T2W, T2Q), ms, &(x[WS(rs, 1)])); | |
498 } | |
499 { | |
500 V T29, T2p, T2m, T2s, T28; | |
501 T28 = VFNMS(LDK(KP618033988), T27, T22); | |
502 T29 = VFNMS(LDK(KP876091699), T28, T1Z); | |
503 { | |
504 V T2o, T2g, T2r, T2l; | |
505 T2o = VFNMS(LDK(KP683113946), T2n, T26); | |
506 T2p = VFMA(LDK(KP792626838), T2o, T1Z); | |
507 T2g = VFNMS(LDK(KP242145790), T2f, T2c); | |
508 T2r = VFMA(LDK(KP617882369), T2k, T2q); | |
509 T2l = VFMA(LDK(KP559016994), T2k, T2j); | |
510 T2m = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T2l, T2g)); | |
511 T2s = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T2r, T2g)); | |
512 } | |
513 ST(&(x[WS(rs, 6)]), VFNMSI(T2m, T29), ms, &(x[0])); | |
514 ST(&(x[WS(rs, 14)]), VFMAI(T2s, T2p), ms, &(x[0])); | |
515 ST(&(x[WS(rs, 19)]), VFMAI(T2m, T29), ms, &(x[WS(rs, 1)])); | |
516 ST(&(x[WS(rs, 11)]), VFNMSI(T2s, T2p), ms, &(x[WS(rs, 1)])); | |
517 } | |
518 } | |
519 } | |
520 } | |
521 } | |
522 VLEAVE(); | |
523 } | |
524 | |
525 static const tw_instr twinstr[] = { | |
526 VTW(0, 1), | |
527 VTW(0, 3), | |
528 VTW(0, 9), | |
529 VTW(0, 24), | |
530 {TW_NEXT, VL, 0} | |
531 }; | |
532 | |
533 static const ct_desc desc = { 25, XSIMD_STRING("t3fv_25"), twinstr, &GENUS, {87, 100, 181, 0}, 0, 0, 0 }; | |
534 | |
535 void XSIMD(codelet_t3fv_25) (planner *p) { | |
536 X(kdft_dit_register) (p, t3fv_25, &desc); | |
537 } | |
538 #else | |
539 | |
540 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 25 -name t3fv_25 -include dft/simd/t3f.h */ | |
541 | |
542 /* | |
543 * This function contains 268 FP additions, 228 FP multiplications, | |
544 * (or, 190 additions, 150 multiplications, 78 fused multiply/add), | |
545 * 123 stack variables, 40 constants, and 50 memory accesses | |
546 */ | |
547 #include "dft/simd/t3f.h" | |
548 | |
549 static void t3fv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
550 { | |
551 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
552 DVK(KP125581039, +0.125581039058626752152356449131262266244969664); | |
553 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
554 DVK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
555 DVK(KP809016994, +0.809016994374947424102293417182819058860154590); | |
556 DVK(KP309016994, +0.309016994374947424102293417182819058860154590); | |
557 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465); | |
558 DVK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
559 DVK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
560 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
561 DVK(KP497379774, +0.497379774329709576484567492012895936835134813); | |
562 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
563 DVK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
564 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328); | |
565 DVK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
566 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
567 DVK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
568 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
569 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
570 DVK(KP250666467, +0.250666467128608490746237519633017587885836494); | |
571 DVK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
572 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
573 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596); | |
574 DVK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
575 DVK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
576 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691); | |
577 DVK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
578 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418); | |
579 DVK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
580 DVK(KP851558583, +0.851558583130145297725004891488503407959946084); | |
581 DVK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
582 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344); | |
583 DVK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
584 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483); | |
585 DVK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
586 DVK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
587 DVK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
588 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
589 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
590 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
591 { | |
592 INT m; | |
593 R *x; | |
594 x = ri; | |
595 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(25, rs)) { | |
596 V T1, T4, T2, T3, TA, Td, Tp, Tw, Tx, T1G, T1j, T5, T1c, T8, T9; | |
597 V Ts, T1J, Tg, T1C, T1m, TX, TB, T1f, TU; | |
598 T1 = LDW(&(W[0])); | |
599 T4 = LDW(&(W[TWVL * 4])); | |
600 T2 = LDW(&(W[TWVL * 2])); | |
601 T3 = VZMUL(T1, T2); | |
602 TA = VZMULJ(T1, T4); | |
603 Td = VZMUL(T1, T4); | |
604 Tp = VZMULJ(T2, T4); | |
605 Tw = VZMULJ(T1, T2); | |
606 Tx = VZMUL(Tw, T4); | |
607 T1G = VZMUL(T3, T4); | |
608 T1j = VZMUL(T2, T4); | |
609 T5 = VZMULJ(T3, T4); | |
610 T1c = VZMULJ(Tw, T4); | |
611 T8 = LDW(&(W[TWVL * 6])); | |
612 T9 = VZMULJ(T3, T8); | |
613 Ts = VZMULJ(T2, T8); | |
614 T1J = VZMULJ(Tp, T8); | |
615 Tg = VZMULJ(T4, T8); | |
616 T1C = VZMULJ(T1, T8); | |
617 T1m = VZMULJ(T1c, T8); | |
618 TX = VZMULJ(T5, T8); | |
619 TB = VZMULJ(TA, T8); | |
620 T1f = VZMULJ(Tw, T8); | |
621 TU = VZMULJ(Td, T8); | |
622 { | |
623 V Tl, Tk, Tm, Tn, T20, T2R, T22, T1V, T2K, T1S, T3A, T2L, TN, T2G, TK; | |
624 V T3w, T2H, T19, T2D, T16, T3x, T2E, T1y, T2N, T1v, T3z, T2O; | |
625 { | |
626 V Tf, Ti, Tj, T7, Tb, Tc, T21; | |
627 Tl = LD(&(x[0]), ms, &(x[0])); | |
628 { | |
629 V Te, Th, T6, Ta; | |
630 Te = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
631 Tf = VZMULJ(Td, Te); | |
632 Th = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
633 Ti = VZMULJ(Tg, Th); | |
634 Tj = VADD(Tf, Ti); | |
635 T6 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
636 T7 = VZMULJ(T5, T6); | |
637 Ta = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
638 Tb = VZMULJ(T9, Ta); | |
639 Tc = VADD(T7, Tb); | |
640 } | |
641 Tk = VMUL(LDK(KP559016994), VSUB(Tc, Tj)); | |
642 Tm = VADD(Tc, Tj); | |
643 Tn = VFNMS(LDK(KP250000000), Tm, Tl); | |
644 T20 = VSUB(T7, Tb); | |
645 T21 = VSUB(Tf, Ti); | |
646 T2R = VMUL(LDK(KP951056516), T21); | |
647 T22 = VFMA(LDK(KP951056516), T20, VMUL(LDK(KP587785252), T21)); | |
648 } | |
649 { | |
650 V T1P, T1I, T1L, T1M, T1B, T1E, T1F, T1O; | |
651 T1O = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
652 T1P = VZMULJ(T2, T1O); | |
653 { | |
654 V T1H, T1K, T1A, T1D; | |
655 T1H = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
656 T1I = VZMULJ(T1G, T1H); | |
657 T1K = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
658 T1L = VZMULJ(T1J, T1K); | |
659 T1M = VADD(T1I, T1L); | |
660 T1A = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
661 T1B = VZMULJ(TA, T1A); | |
662 T1D = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
663 T1E = VZMULJ(T1C, T1D); | |
664 T1F = VADD(T1B, T1E); | |
665 } | |
666 { | |
667 V T1T, T1U, T1N, T1Q, T1R; | |
668 T1T = VSUB(T1B, T1E); | |
669 T1U = VSUB(T1I, T1L); | |
670 T1V = VFMA(LDK(KP475528258), T1T, VMUL(LDK(KP293892626), T1U)); | |
671 T2K = VFNMS(LDK(KP293892626), T1T, VMUL(LDK(KP475528258), T1U)); | |
672 T1N = VMUL(LDK(KP559016994), VSUB(T1F, T1M)); | |
673 T1Q = VADD(T1F, T1M); | |
674 T1R = VFNMS(LDK(KP250000000), T1Q, T1P); | |
675 T1S = VADD(T1N, T1R); | |
676 T3A = VADD(T1P, T1Q); | |
677 T2L = VSUB(T1R, T1N); | |
678 } | |
679 } | |
680 { | |
681 V TH, Tz, TD, TE, Tr, Tu, Tv, TG; | |
682 TG = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
683 TH = VZMULJ(T1, TG); | |
684 { | |
685 V Ty, TC, Tq, Tt; | |
686 Ty = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
687 Tz = VZMULJ(Tx, Ty); | |
688 TC = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
689 TD = VZMULJ(TB, TC); | |
690 TE = VADD(Tz, TD); | |
691 Tq = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
692 Tr = VZMULJ(Tp, Tq); | |
693 Tt = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
694 Tu = VZMULJ(Ts, Tt); | |
695 Tv = VADD(Tr, Tu); | |
696 } | |
697 { | |
698 V TL, TM, TF, TI, TJ; | |
699 TL = VSUB(Tr, Tu); | |
700 TM = VSUB(Tz, TD); | |
701 TN = VFMA(LDK(KP475528258), TL, VMUL(LDK(KP293892626), TM)); | |
702 T2G = VFNMS(LDK(KP293892626), TL, VMUL(LDK(KP475528258), TM)); | |
703 TF = VMUL(LDK(KP559016994), VSUB(Tv, TE)); | |
704 TI = VADD(Tv, TE); | |
705 TJ = VFNMS(LDK(KP250000000), TI, TH); | |
706 TK = VADD(TF, TJ); | |
707 T3w = VADD(TH, TI); | |
708 T2H = VSUB(TJ, TF); | |
709 } | |
710 } | |
711 { | |
712 V T13, TW, TZ, T10, TQ, TS, TT, T12; | |
713 T12 = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
714 T13 = VZMULJ(T3, T12); | |
715 { | |
716 V TV, TY, TP, TR; | |
717 TV = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
718 TW = VZMULJ(TU, TV); | |
719 TY = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
720 TZ = VZMULJ(TX, TY); | |
721 T10 = VADD(TW, TZ); | |
722 TP = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
723 TQ = VZMULJ(T4, TP); | |
724 TR = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
725 TS = VZMULJ(T8, TR); | |
726 TT = VADD(TQ, TS); | |
727 } | |
728 { | |
729 V T17, T18, T11, T14, T15; | |
730 T17 = VSUB(TQ, TS); | |
731 T18 = VSUB(TW, TZ); | |
732 T19 = VFMA(LDK(KP475528258), T17, VMUL(LDK(KP293892626), T18)); | |
733 T2D = VFNMS(LDK(KP293892626), T17, VMUL(LDK(KP475528258), T18)); | |
734 T11 = VMUL(LDK(KP559016994), VSUB(TT, T10)); | |
735 T14 = VADD(TT, T10); | |
736 T15 = VFNMS(LDK(KP250000000), T14, T13); | |
737 T16 = VADD(T11, T15); | |
738 T3x = VADD(T13, T14); | |
739 T2E = VSUB(T15, T11); | |
740 } | |
741 } | |
742 { | |
743 V T1s, T1l, T1o, T1p, T1e, T1h, T1i, T1r; | |
744 T1r = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
745 T1s = VZMULJ(Tw, T1r); | |
746 { | |
747 V T1k, T1n, T1d, T1g; | |
748 T1k = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
749 T1l = VZMULJ(T1j, T1k); | |
750 T1n = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
751 T1o = VZMULJ(T1m, T1n); | |
752 T1p = VADD(T1l, T1o); | |
753 T1d = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
754 T1e = VZMULJ(T1c, T1d); | |
755 T1g = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
756 T1h = VZMULJ(T1f, T1g); | |
757 T1i = VADD(T1e, T1h); | |
758 } | |
759 { | |
760 V T1w, T1x, T1q, T1t, T1u; | |
761 T1w = VSUB(T1e, T1h); | |
762 T1x = VSUB(T1l, T1o); | |
763 T1y = VFMA(LDK(KP475528258), T1w, VMUL(LDK(KP293892626), T1x)); | |
764 T2N = VFNMS(LDK(KP293892626), T1w, VMUL(LDK(KP475528258), T1x)); | |
765 T1q = VMUL(LDK(KP559016994), VSUB(T1i, T1p)); | |
766 T1t = VADD(T1i, T1p); | |
767 T1u = VFNMS(LDK(KP250000000), T1t, T1s); | |
768 T1v = VADD(T1q, T1u); | |
769 T3z = VADD(T1s, T1t); | |
770 T2O = VSUB(T1u, T1q); | |
771 } | |
772 } | |
773 { | |
774 V T3J, T3K, T3D, T3E, T3C, T3F, T3L, T3G; | |
775 { | |
776 V T3H, T3I, T3y, T3B; | |
777 T3H = VSUB(T3w, T3x); | |
778 T3I = VSUB(T3z, T3A); | |
779 T3J = VBYI(VFMA(LDK(KP951056516), T3H, VMUL(LDK(KP587785252), T3I))); | |
780 T3K = VBYI(VFNMS(LDK(KP587785252), T3H, VMUL(LDK(KP951056516), T3I))); | |
781 T3D = VADD(Tl, Tm); | |
782 T3y = VADD(T3w, T3x); | |
783 T3B = VADD(T3z, T3A); | |
784 T3E = VADD(T3y, T3B); | |
785 T3C = VMUL(LDK(KP559016994), VSUB(T3y, T3B)); | |
786 T3F = VFNMS(LDK(KP250000000), T3E, T3D); | |
787 } | |
788 ST(&(x[0]), VADD(T3D, T3E), ms, &(x[0])); | |
789 T3L = VSUB(T3F, T3C); | |
790 ST(&(x[WS(rs, 10)]), VADD(T3K, T3L), ms, &(x[0])); | |
791 ST(&(x[WS(rs, 15)]), VSUB(T3L, T3K), ms, &(x[WS(rs, 1)])); | |
792 T3G = VADD(T3C, T3F); | |
793 ST(&(x[WS(rs, 5)]), VSUB(T3G, T3J), ms, &(x[WS(rs, 1)])); | |
794 ST(&(x[WS(rs, 20)]), VADD(T3J, T3G), ms, &(x[0])); | |
795 } | |
796 { | |
797 V To, T2n, T2o, T2p, T2x, T2y, T2z, T2u, T2v, T2w, T2q, T2r, T2s, T29, T2i; | |
798 V T2e, T2g, T1Y, T2j, T2b, T2c, T2B, T2C; | |
799 To = VADD(Tk, Tn); | |
800 T2n = VFMA(LDK(KP1_688655851), TN, VMUL(LDK(KP535826794), TK)); | |
801 T2o = VFMA(LDK(KP1_541026485), T19, VMUL(LDK(KP637423989), T16)); | |
802 T2p = VSUB(T2n, T2o); | |
803 T2x = VFMA(LDK(KP851558583), T1y, VMUL(LDK(KP904827052), T1v)); | |
804 T2y = VFMA(LDK(KP1_984229402), T1V, VMUL(LDK(KP125333233), T1S)); | |
805 T2z = VADD(T2x, T2y); | |
806 T2u = VFNMS(LDK(KP844327925), TK, VMUL(LDK(KP1_071653589), TN)); | |
807 T2v = VFNMS(LDK(KP1_274847979), T19, VMUL(LDK(KP770513242), T16)); | |
808 T2w = VADD(T2u, T2v); | |
809 T2q = VFNMS(LDK(KP425779291), T1v, VMUL(LDK(KP1_809654104), T1y)); | |
810 T2r = VFNMS(LDK(KP992114701), T1S, VMUL(LDK(KP250666467), T1V)); | |
811 T2s = VADD(T2q, T2r); | |
812 { | |
813 V T23, T24, T25, T26, T27, T28; | |
814 T23 = VFMA(LDK(KP1_937166322), TN, VMUL(LDK(KP248689887), TK)); | |
815 T24 = VFMA(LDK(KP1_071653589), T19, VMUL(LDK(KP844327925), T16)); | |
816 T25 = VADD(T23, T24); | |
817 T26 = VFMA(LDK(KP1_752613360), T1y, VMUL(LDK(KP481753674), T1v)); | |
818 T27 = VFMA(LDK(KP1_457937254), T1V, VMUL(LDK(KP684547105), T1S)); | |
819 T28 = VADD(T26, T27); | |
820 T29 = VADD(T25, T28); | |
821 T2i = VSUB(T27, T26); | |
822 T2e = VMUL(LDK(KP559016994), VSUB(T28, T25)); | |
823 T2g = VSUB(T24, T23); | |
824 } | |
825 { | |
826 V TO, T1a, T1b, T1z, T1W, T1X; | |
827 TO = VFNMS(LDK(KP497379774), TN, VMUL(LDK(KP968583161), TK)); | |
828 T1a = VFNMS(LDK(KP1_688655851), T19, VMUL(LDK(KP535826794), T16)); | |
829 T1b = VADD(TO, T1a); | |
830 T1z = VFNMS(LDK(KP963507348), T1y, VMUL(LDK(KP876306680), T1v)); | |
831 T1W = VFNMS(LDK(KP1_369094211), T1V, VMUL(LDK(KP728968627), T1S)); | |
832 T1X = VADD(T1z, T1W); | |
833 T1Y = VADD(T1b, T1X); | |
834 T2j = VMUL(LDK(KP559016994), VSUB(T1b, T1X)); | |
835 T2b = VSUB(T1a, TO); | |
836 T2c = VSUB(T1z, T1W); | |
837 } | |
838 { | |
839 V T1Z, T2a, T2t, T2A; | |
840 T1Z = VADD(To, T1Y); | |
841 T2a = VBYI(VADD(T22, T29)); | |
842 ST(&(x[WS(rs, 1)]), VSUB(T1Z, T2a), ms, &(x[WS(rs, 1)])); | |
843 ST(&(x[WS(rs, 24)]), VADD(T1Z, T2a), ms, &(x[0])); | |
844 T2t = VADD(To, VADD(T2p, T2s)); | |
845 T2A = VBYI(VADD(T22, VSUB(T2w, T2z))); | |
846 ST(&(x[WS(rs, 21)]), VSUB(T2t, T2A), ms, &(x[WS(rs, 1)])); | |
847 ST(&(x[WS(rs, 4)]), VADD(T2t, T2A), ms, &(x[0])); | |
848 } | |
849 T2B = VBYI(VADD(T22, VFMA(LDK(KP309016994), T2w, VFMA(LDK(KP587785252), VSUB(T2r, T2q), VFNMS(LDK(KP951056516), VADD(T2n, T2o), VMUL(LDK(KP809016994), T2z)))))); | |
850 T2C = VFMA(LDK(KP309016994), T2p, VFMA(LDK(KP951056516), VSUB(T2u, T2v), VFMA(LDK(KP587785252), VSUB(T2y, T2x), VFNMS(LDK(KP809016994), T2s, To)))); | |
851 ST(&(x[WS(rs, 9)]), VADD(T2B, T2C), ms, &(x[WS(rs, 1)])); | |
852 ST(&(x[WS(rs, 16)]), VSUB(T2C, T2B), ms, &(x[0])); | |
853 { | |
854 V T2f, T2l, T2k, T2m, T2d, T2h; | |
855 T2d = VFMS(LDK(KP250000000), T29, T22); | |
856 T2f = VBYI(VADD(VFMA(LDK(KP587785252), T2b, VMUL(LDK(KP951056516), T2c)), VSUB(T2d, T2e))); | |
857 T2l = VBYI(VADD(VFNMS(LDK(KP587785252), T2c, VMUL(LDK(KP951056516), T2b)), VADD(T2d, T2e))); | |
858 T2h = VFNMS(LDK(KP250000000), T1Y, To); | |
859 T2k = VFMA(LDK(KP587785252), T2g, VFNMS(LDK(KP951056516), T2i, VSUB(T2h, T2j))); | |
860 T2m = VFMA(LDK(KP951056516), T2g, VADD(T2j, VFMA(LDK(KP587785252), T2i, T2h))); | |
861 ST(&(x[WS(rs, 11)]), VADD(T2f, T2k), ms, &(x[WS(rs, 1)])); | |
862 ST(&(x[WS(rs, 19)]), VSUB(T2m, T2l), ms, &(x[WS(rs, 1)])); | |
863 ST(&(x[WS(rs, 14)]), VSUB(T2k, T2f), ms, &(x[0])); | |
864 ST(&(x[WS(rs, 6)]), VADD(T2l, T2m), ms, &(x[0])); | |
865 } | |
866 } | |
867 { | |
868 V T2S, T2U, T2F, T2I, T2J, T2Y, T2Z, T30, T2M, T2P, T2Q, T2V, T2W, T2X, T3a; | |
869 V T3l, T3b, T3k, T3f, T3p, T3i, T3o, T32, T33; | |
870 T2S = VFNMS(LDK(KP587785252), T20, T2R); | |
871 T2U = VSUB(Tn, Tk); | |
872 T2F = VFNMS(LDK(KP125333233), T2E, VMUL(LDK(KP1_984229402), T2D)); | |
873 T2I = VFMA(LDK(KP1_457937254), T2G, VMUL(LDK(KP684547105), T2H)); | |
874 T2J = VSUB(T2F, T2I); | |
875 T2Y = VFNMS(LDK(KP1_996053456), T2N, VMUL(LDK(KP062790519), T2O)); | |
876 T2Z = VFMA(LDK(KP1_541026485), T2K, VMUL(LDK(KP637423989), T2L)); | |
877 T30 = VSUB(T2Y, T2Z); | |
878 T2M = VFNMS(LDK(KP770513242), T2L, VMUL(LDK(KP1_274847979), T2K)); | |
879 T2P = VFMA(LDK(KP125581039), T2N, VMUL(LDK(KP998026728), T2O)); | |
880 T2Q = VSUB(T2M, T2P); | |
881 T2V = VFNMS(LDK(KP1_369094211), T2G, VMUL(LDK(KP728968627), T2H)); | |
882 T2W = VFMA(LDK(KP250666467), T2D, VMUL(LDK(KP992114701), T2E)); | |
883 T2X = VSUB(T2V, T2W); | |
884 { | |
885 V T34, T35, T36, T37, T38, T39; | |
886 T34 = VFNMS(LDK(KP481753674), T2H, VMUL(LDK(KP1_752613360), T2G)); | |
887 T35 = VFMA(LDK(KP851558583), T2D, VMUL(LDK(KP904827052), T2E)); | |
888 T36 = VSUB(T34, T35); | |
889 T37 = VFNMS(LDK(KP844327925), T2O, VMUL(LDK(KP1_071653589), T2N)); | |
890 T38 = VFNMS(LDK(KP998026728), T2L, VMUL(LDK(KP125581039), T2K)); | |
891 T39 = VADD(T37, T38); | |
892 T3a = VMUL(LDK(KP559016994), VSUB(T36, T39)); | |
893 T3l = VSUB(T37, T38); | |
894 T3b = VADD(T36, T39); | |
895 T3k = VADD(T34, T35); | |
896 } | |
897 { | |
898 V T3d, T3e, T3m, T3g, T3h, T3n; | |
899 T3d = VFNMS(LDK(KP425779291), T2E, VMUL(LDK(KP1_809654104), T2D)); | |
900 T3e = VFMA(LDK(KP963507348), T2G, VMUL(LDK(KP876306680), T2H)); | |
901 T3m = VADD(T3e, T3d); | |
902 T3g = VFMA(LDK(KP1_688655851), T2N, VMUL(LDK(KP535826794), T2O)); | |
903 T3h = VFMA(LDK(KP1_996053456), T2K, VMUL(LDK(KP062790519), T2L)); | |
904 T3n = VADD(T3g, T3h); | |
905 T3f = VSUB(T3d, T3e); | |
906 T3p = VADD(T3m, T3n); | |
907 T3i = VSUB(T3g, T3h); | |
908 T3o = VMUL(LDK(KP559016994), VSUB(T3m, T3n)); | |
909 } | |
910 { | |
911 V T3u, T3v, T2T, T31; | |
912 T3u = VBYI(VADD(T2S, T3b)); | |
913 T3v = VADD(T2U, T3p); | |
914 ST(&(x[WS(rs, 2)]), VADD(T3u, T3v), ms, &(x[0])); | |
915 ST(&(x[WS(rs, 23)]), VSUB(T3v, T3u), ms, &(x[WS(rs, 1)])); | |
916 T2T = VBYI(VSUB(VADD(T2J, T2Q), T2S)); | |
917 T31 = VADD(T2U, VADD(T2X, T30)); | |
918 ST(&(x[WS(rs, 3)]), VADD(T2T, T31), ms, &(x[WS(rs, 1)])); | |
919 ST(&(x[WS(rs, 22)]), VSUB(T31, T2T), ms, &(x[0])); | |
920 } | |
921 T32 = VFMA(LDK(KP309016994), T2X, VFNMS(LDK(KP809016994), T30, VFNMS(LDK(KP587785252), VADD(T2P, T2M), VFNMS(LDK(KP951056516), VADD(T2I, T2F), T2U)))); | |
922 T33 = VBYI(VSUB(VFNMS(LDK(KP587785252), VADD(T2Y, T2Z), VFNMS(LDK(KP809016994), T2Q, VFNMS(LDK(KP951056516), VADD(T2V, T2W), VMUL(LDK(KP309016994), T2J)))), T2S)); | |
923 ST(&(x[WS(rs, 17)]), VSUB(T32, T33), ms, &(x[WS(rs, 1)])); | |
924 ST(&(x[WS(rs, 8)]), VADD(T32, T33), ms, &(x[0])); | |
925 { | |
926 V T3j, T3s, T3r, T3t, T3c, T3q; | |
927 T3c = VFNMS(LDK(KP250000000), T3b, T2S); | |
928 T3j = VBYI(VADD(T3a, VADD(T3c, VFNMS(LDK(KP587785252), T3i, VMUL(LDK(KP951056516), T3f))))); | |
929 T3s = VBYI(VADD(T3c, VSUB(VFMA(LDK(KP587785252), T3f, VMUL(LDK(KP951056516), T3i)), T3a))); | |
930 T3q = VFNMS(LDK(KP250000000), T3p, T2U); | |
931 T3r = VFMA(LDK(KP951056516), T3k, VFMA(LDK(KP587785252), T3l, VADD(T3o, T3q))); | |
932 T3t = VFMA(LDK(KP587785252), T3k, VSUB(VFNMS(LDK(KP951056516), T3l, T3q), T3o)); | |
933 ST(&(x[WS(rs, 7)]), VADD(T3j, T3r), ms, &(x[WS(rs, 1)])); | |
934 ST(&(x[WS(rs, 13)]), VSUB(T3t, T3s), ms, &(x[WS(rs, 1)])); | |
935 ST(&(x[WS(rs, 18)]), VSUB(T3r, T3j), ms, &(x[0])); | |
936 ST(&(x[WS(rs, 12)]), VADD(T3s, T3t), ms, &(x[0])); | |
937 } | |
938 } | |
939 } | |
940 } | |
941 } | |
942 VLEAVE(); | |
943 } | |
944 | |
945 static const tw_instr twinstr[] = { | |
946 VTW(0, 1), | |
947 VTW(0, 3), | |
948 VTW(0, 9), | |
949 VTW(0, 24), | |
950 {TW_NEXT, VL, 0} | |
951 }; | |
952 | |
953 static const ct_desc desc = { 25, XSIMD_STRING("t3fv_25"), twinstr, &GENUS, {190, 150, 78, 0}, 0, 0, 0 }; | |
954 | |
955 void XSIMD(codelet_t3fv_25) (planner *p) { | |
956 X(kdft_dit_register) (p, t3fv_25, &desc); | |
957 } | |
958 #endif |