Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/dft/simd/common/t3bv_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 08:06:09 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 t3bv_25 -include dft/simd/t3b.h -sign 1 */ | |
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/t3b.h" | |
36 | |
37 static void t3bv_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 = ii; | |
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 = VZMUL(Te, Tf); | |
149 Ti = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
150 Tj = VZMUL(Th, Ti); | |
151 Tk = VADD(Tg, Tj); | |
152 T7 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
153 T8 = VZMUL(T6, T7); | |
154 Tb = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
155 Tc = VZMUL(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 = VZMUL(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 = VZMUL(TC, T1C); | |
176 T1M = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
177 T1N = VZMUL(T1L, T1M); | |
178 T1F = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
179 T1G = VZMUL(T1E, T1F); | |
180 T1J = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
181 T1K = VZMUL(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 = VZMUL(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 = VZMUL(Tr, Ts); | |
213 TE = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
214 TF = VZMUL(TD, TE); | |
215 Tv = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
216 Tw = VZMUL(Tu, Tv); | |
217 TA = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
218 TB = VZMUL(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 = VZMUL(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 = VZMUL(T9, TP); | |
250 TY = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
251 TZ = VZMUL(TX, TY); | |
252 TR = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
253 TS = VZMUL(T5, TR); | |
254 TV = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
255 TW = VZMUL(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 = VZMUL(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 = VZMUL(T1e, T1f); | |
287 T1p = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
288 T1q = VZMUL(T1o, T1p); | |
289 T1i = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
290 T1j = VZMUL(T1h, T1i); | |
291 T1m = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
292 T1n = VZMUL(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)]), VFNMSI(T4o, T4n), ms, &(x[0])); | |
332 ST(&(x[WS(rs, 15)]), VFMAI(T4o, T4n), ms, &(x[WS(rs, 1)])); | |
333 T4j = VFMA(LDK(KP559016994), T4i, T4h); | |
334 ST(&(x[WS(rs, 5)]), VFMAI(T4m, T4j), ms, &(x[WS(rs, 1)])); | |
335 ST(&(x[WS(rs, 20)]), VFNMSI(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, 22)]), VFNMSI(T3m, T3d), ms, &(x[0])); | |
393 ST(&(x[WS(rs, 3)]), VFMAI(T3m, T3d), ms, &(x[WS(rs, 1)])); | |
394 T3G = VFMA(LDK(KP949179823), T3F, T2X); | |
395 T3O = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T3l, T3N)); | |
396 ST(&(x[WS(rs, 23)]), VFNMSI(T3O, T3G), ms, &(x[WS(rs, 1)])); | |
397 ST(&(x[WS(rs, 2)]), VFMAI(T3O, T3G), ms, &(x[0])); | |
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)]), VFMAI(T3y, T3s), ms, &(x[0])); | |
406 ST(&(x[WS(rs, 17)]), VFNMSI(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, 12)]), VFNMSI(T42, T3V), ms, &(x[0])); | |
423 ST(&(x[WS(rs, 18)]), VFMAI(T48, T45), ms, &(x[0])); | |
424 ST(&(x[WS(rs, 13)]), VFMAI(T42, T3V), ms, &(x[WS(rs, 1)])); | |
425 ST(&(x[WS(rs, 7)]), VFNMSI(T48, T45), ms, &(x[WS(rs, 1)])); | |
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)]), VFMAI(T2u, T2t), ms, &(x[WS(rs, 1)])); | |
484 ST(&(x[WS(rs, 24)]), VFNMSI(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)]), VFNMSI(T2K, T2C), ms, &(x[0])); | |
488 ST(&(x[WS(rs, 21)]), VFMAI(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, 9)]), VFNMSI(T2W, T2Q), ms, &(x[WS(rs, 1)])); | |
497 ST(&(x[WS(rs, 16)]), VFMAI(T2W, T2Q), ms, &(x[0])); | |
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)]), VFMAI(T2m, T29), ms, &(x[0])); | |
514 ST(&(x[WS(rs, 14)]), VFNMSI(T2s, T2p), ms, &(x[0])); | |
515 ST(&(x[WS(rs, 19)]), VFNMSI(T2m, T29), ms, &(x[WS(rs, 1)])); | |
516 ST(&(x[WS(rs, 11)]), VFMAI(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("t3bv_25"), twinstr, &GENUS, {87, 100, 181, 0}, 0, 0, 0 }; | |
534 | |
535 void XSIMD(codelet_t3bv_25) (planner *p) { | |
536 X(kdft_dit_register) (p, t3bv_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 t3bv_25 -include dft/simd/t3b.h -sign 1 */ | |
541 | |
542 /* | |
543 * This function contains 268 FP additions, 228 FP multiplications, | |
544 * (or, 191 additions, 151 multiplications, 77 fused multiply/add), | |
545 * 124 stack variables, 40 constants, and 50 memory accesses | |
546 */ | |
547 #include "dft/simd/t3b.h" | |
548 | |
549 static void t3bv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
550 { | |
551 DVK(KP497379774, +0.497379774329709576484567492012895936835134813); | |
552 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
553 DVK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
554 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
555 DVK(KP809016994, +0.809016994374947424102293417182819058860154590); | |
556 DVK(KP309016994, +0.309016994374947424102293417182819058860154590); | |
557 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483); | |
558 DVK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
559 DVK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
560 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
561 DVK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
562 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
563 DVK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
564 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691); | |
565 DVK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
566 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
567 DVK(KP851558583, +0.851558583130145297725004891488503407959946084); | |
568 DVK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
569 DVK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
570 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418); | |
571 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328); | |
572 DVK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
573 DVK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
574 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344); | |
575 DVK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
576 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
577 DVK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
578 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596); | |
579 DVK(KP125581039, +0.125581039058626752152356449131262266244969664); | |
580 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
581 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
582 DVK(KP250666467, +0.250666467128608490746237519633017587885836494); | |
583 DVK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
584 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465); | |
585 DVK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
586 DVK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
587 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
588 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
589 DVK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
590 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
591 { | |
592 INT m; | |
593 R *x; | |
594 x = ii; | |
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, Td, T8, T9, TF, Te, Tu, TB, TC, T1s, T15, Tf, TY, T4, Ta; | |
597 V Tx, T1T, Tg, T1N, T1v, T18, TG, T1o, T11; | |
598 T1 = LDW(&(W[TWVL * 4])); | |
599 Td = LDW(&(W[TWVL * 2])); | |
600 T8 = LDW(&(W[0])); | |
601 T9 = VZMUL(T8, T1); | |
602 TF = VZMULJ(T8, T1); | |
603 Te = VZMUL(T8, Td); | |
604 Tu = VZMULJ(Td, T1); | |
605 TB = VZMULJ(T8, Td); | |
606 TC = VZMUL(TB, T1); | |
607 T1s = VZMUL(Te, T1); | |
608 T15 = VZMUL(Td, T1); | |
609 Tf = VZMULJ(Te, T1); | |
610 TY = VZMULJ(TB, T1); | |
611 T4 = LDW(&(W[TWVL * 6])); | |
612 Ta = VZMULJ(T9, T4); | |
613 Tx = VZMULJ(Td, T4); | |
614 T1T = VZMULJ(T1, T4); | |
615 Tg = VZMULJ(Tf, T4); | |
616 T1N = VZMULJ(Te, T4); | |
617 T1v = VZMULJ(Tu, T4); | |
618 T18 = VZMULJ(TY, T4); | |
619 TG = VZMULJ(TF, T4); | |
620 T1o = VZMULJ(T8, T4); | |
621 T11 = VZMULJ(TB, T4); | |
622 { | |
623 V T1Y, T1X, T2f, T2g, T1Z, T20, T2e, T39, T1H, T2T, T1E, T3C, T2S, Tk, T2G; | |
624 V Ts, T3z, T2F, TK, T2I, TS, T3y, T2J, T1k, T2Q, T1h, T3B, T2P; | |
625 { | |
626 V T1S, T1V, T1W, T1M, T1P, T1Q, T2d; | |
627 T1Y = LD(&(x[0]), ms, &(x[0])); | |
628 { | |
629 V T1R, T1U, T1L, T1O; | |
630 T1R = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
631 T1S = VZMUL(T9, T1R); | |
632 T1U = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
633 T1V = VZMUL(T1T, T1U); | |
634 T1W = VADD(T1S, T1V); | |
635 T1L = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
636 T1M = VZMUL(Tf, T1L); | |
637 T1O = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
638 T1P = VZMUL(T1N, T1O); | |
639 T1Q = VADD(T1M, T1P); | |
640 } | |
641 T1X = VMUL(LDK(KP559016994), VSUB(T1Q, T1W)); | |
642 T2f = VSUB(T1S, T1V); | |
643 T2g = VMUL(LDK(KP587785252), T2f); | |
644 T1Z = VADD(T1Q, T1W); | |
645 T20 = VFNMS(LDK(KP250000000), T1Z, T1Y); | |
646 T2d = VSUB(T1M, T1P); | |
647 T2e = VMUL(LDK(KP951056516), T2d); | |
648 T39 = VMUL(LDK(KP587785252), T2d); | |
649 } | |
650 { | |
651 V T1B, T1u, T1x, T1y, T1n, T1q, T1r, T1A; | |
652 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
653 T1B = VZMUL(Td, T1A); | |
654 { | |
655 V T1t, T1w, T1m, T1p; | |
656 T1t = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
657 T1u = VZMUL(T1s, T1t); | |
658 T1w = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
659 T1x = VZMUL(T1v, T1w); | |
660 T1y = VADD(T1u, T1x); | |
661 T1m = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
662 T1n = VZMUL(TF, T1m); | |
663 T1p = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
664 T1q = VZMUL(T1o, T1p); | |
665 T1r = VADD(T1n, T1q); | |
666 } | |
667 { | |
668 V T1F, T1G, T1z, T1C, T1D; | |
669 T1F = VSUB(T1n, T1q); | |
670 T1G = VSUB(T1u, T1x); | |
671 T1H = VFMA(LDK(KP475528258), T1F, VMUL(LDK(KP293892626), T1G)); | |
672 T2T = VFNMS(LDK(KP475528258), T1G, VMUL(LDK(KP293892626), T1F)); | |
673 T1z = VMUL(LDK(KP559016994), VSUB(T1r, T1y)); | |
674 T1C = VADD(T1r, T1y); | |
675 T1D = VFNMS(LDK(KP250000000), T1C, T1B); | |
676 T1E = VADD(T1z, T1D); | |
677 T3C = VADD(T1B, T1C); | |
678 T2S = VSUB(T1D, T1z); | |
679 } | |
680 } | |
681 { | |
682 V Tp, Tc, Ti, Tm, T3, T6, Tl, To; | |
683 To = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
684 Tp = VZMUL(Te, To); | |
685 { | |
686 V Tb, Th, T2, T5; | |
687 Tb = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
688 Tc = VZMUL(Ta, Tb); | |
689 Th = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
690 Ti = VZMUL(Tg, Th); | |
691 Tm = VADD(Tc, Ti); | |
692 T2 = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
693 T3 = VZMUL(T1, T2); | |
694 T5 = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
695 T6 = VZMUL(T4, T5); | |
696 Tl = VADD(T3, T6); | |
697 } | |
698 { | |
699 V T7, Tj, Tn, Tq, Tr; | |
700 T7 = VSUB(T3, T6); | |
701 Tj = VSUB(Tc, Ti); | |
702 Tk = VFMA(LDK(KP475528258), T7, VMUL(LDK(KP293892626), Tj)); | |
703 T2G = VFNMS(LDK(KP475528258), Tj, VMUL(LDK(KP293892626), T7)); | |
704 Tn = VMUL(LDK(KP559016994), VSUB(Tl, Tm)); | |
705 Tq = VADD(Tl, Tm); | |
706 Tr = VFNMS(LDK(KP250000000), Tq, Tp); | |
707 Ts = VADD(Tn, Tr); | |
708 T3z = VADD(Tp, Tq); | |
709 T2F = VSUB(Tr, Tn); | |
710 } | |
711 } | |
712 { | |
713 V TP, TE, TI, TM, Tw, Tz, TL, TO; | |
714 TO = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
715 TP = VZMUL(T8, TO); | |
716 { | |
717 V TD, TH, Tv, Ty; | |
718 TD = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
719 TE = VZMUL(TC, TD); | |
720 TH = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
721 TI = VZMUL(TG, TH); | |
722 TM = VADD(TE, TI); | |
723 Tv = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
724 Tw = VZMUL(Tu, Tv); | |
725 Ty = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
726 Tz = VZMUL(Tx, Ty); | |
727 TL = VADD(Tw, Tz); | |
728 } | |
729 { | |
730 V TA, TJ, TN, TQ, TR; | |
731 TA = VSUB(Tw, Tz); | |
732 TJ = VSUB(TE, TI); | |
733 TK = VFMA(LDK(KP475528258), TA, VMUL(LDK(KP293892626), TJ)); | |
734 T2I = VFNMS(LDK(KP475528258), TJ, VMUL(LDK(KP293892626), TA)); | |
735 TN = VMUL(LDK(KP559016994), VSUB(TL, TM)); | |
736 TQ = VADD(TL, TM); | |
737 TR = VFNMS(LDK(KP250000000), TQ, TP); | |
738 TS = VADD(TN, TR); | |
739 T3y = VADD(TP, TQ); | |
740 T2J = VSUB(TR, TN); | |
741 } | |
742 } | |
743 { | |
744 V T1e, T17, T1a, T1b, T10, T13, T14, T1d; | |
745 T1d = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
746 T1e = VZMUL(TB, T1d); | |
747 { | |
748 V T16, T19, TZ, T12; | |
749 T16 = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
750 T17 = VZMUL(T15, T16); | |
751 T19 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
752 T1a = VZMUL(T18, T19); | |
753 T1b = VADD(T17, T1a); | |
754 TZ = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
755 T10 = VZMUL(TY, TZ); | |
756 T12 = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
757 T13 = VZMUL(T11, T12); | |
758 T14 = VADD(T10, T13); | |
759 } | |
760 { | |
761 V T1i, T1j, T1c, T1f, T1g; | |
762 T1i = VSUB(T10, T13); | |
763 T1j = VSUB(T17, T1a); | |
764 T1k = VFMA(LDK(KP475528258), T1i, VMUL(LDK(KP293892626), T1j)); | |
765 T2Q = VFNMS(LDK(KP475528258), T1j, VMUL(LDK(KP293892626), T1i)); | |
766 T1c = VMUL(LDK(KP559016994), VSUB(T14, T1b)); | |
767 T1f = VADD(T14, T1b); | |
768 T1g = VFNMS(LDK(KP250000000), T1f, T1e); | |
769 T1h = VADD(T1c, T1g); | |
770 T3B = VADD(T1e, T1f); | |
771 T2P = VSUB(T1g, T1c); | |
772 } | |
773 } | |
774 { | |
775 V T3E, T3M, T3I, T3J, T3H, T3K, T3N, T3L; | |
776 { | |
777 V T3A, T3D, T3F, T3G; | |
778 T3A = VSUB(T3y, T3z); | |
779 T3D = VSUB(T3B, T3C); | |
780 T3E = VBYI(VFMA(LDK(KP951056516), T3A, VMUL(LDK(KP587785252), T3D))); | |
781 T3M = VBYI(VFNMS(LDK(KP951056516), T3D, VMUL(LDK(KP587785252), T3A))); | |
782 T3I = VADD(T1Y, T1Z); | |
783 T3F = VADD(T3y, T3z); | |
784 T3G = VADD(T3B, T3C); | |
785 T3J = VADD(T3F, T3G); | |
786 T3H = VMUL(LDK(KP559016994), VSUB(T3F, T3G)); | |
787 T3K = VFNMS(LDK(KP250000000), T3J, T3I); | |
788 } | |
789 ST(&(x[0]), VADD(T3I, T3J), ms, &(x[0])); | |
790 T3N = VSUB(T3K, T3H); | |
791 ST(&(x[WS(rs, 10)]), VADD(T3M, T3N), ms, &(x[0])); | |
792 ST(&(x[WS(rs, 15)]), VSUB(T3N, T3M), ms, &(x[WS(rs, 1)])); | |
793 T3L = VADD(T3H, T3K); | |
794 ST(&(x[WS(rs, 5)]), VADD(T3E, T3L), ms, &(x[WS(rs, 1)])); | |
795 ST(&(x[WS(rs, 20)]), VSUB(T3L, T3E), ms, &(x[0])); | |
796 } | |
797 { | |
798 V T2X, T3a, T3i, T3j, T3k, T3s, T3t, T3u, T3l, T3m, T3n, T3p, T3q, T3r, T2L; | |
799 V T3b, T32, T38, T2W, T35, T2Y, T34, T3w, T3x; | |
800 T2X = VSUB(T20, T1X); | |
801 T3a = VFNMS(LDK(KP951056516), T2f, T39); | |
802 T3i = VFMA(LDK(KP1_369094211), T2I, VMUL(LDK(KP728968627), T2J)); | |
803 T3j = VFNMS(LDK(KP992114701), T2F, VMUL(LDK(KP250666467), T2G)); | |
804 T3k = VADD(T3i, T3j); | |
805 T3s = VFNMS(LDK(KP125581039), T2Q, VMUL(LDK(KP998026728), T2P)); | |
806 T3t = VFMA(LDK(KP1_274847979), T2T, VMUL(LDK(KP770513242), T2S)); | |
807 T3u = VADD(T3s, T3t); | |
808 T3l = VFMA(LDK(KP1_996053456), T2Q, VMUL(LDK(KP062790519), T2P)); | |
809 T3m = VFNMS(LDK(KP637423989), T2S, VMUL(LDK(KP1_541026485), T2T)); | |
810 T3n = VADD(T3l, T3m); | |
811 T3p = VFNMS(LDK(KP1_457937254), T2I, VMUL(LDK(KP684547105), T2J)); | |
812 T3q = VFMA(LDK(KP1_984229402), T2G, VMUL(LDK(KP125333233), T2F)); | |
813 T3r = VADD(T3p, T3q); | |
814 { | |
815 V T2H, T2K, T36, T30, T31, T37; | |
816 T2H = VFNMS(LDK(KP851558583), T2G, VMUL(LDK(KP904827052), T2F)); | |
817 T2K = VFMA(LDK(KP1_752613360), T2I, VMUL(LDK(KP481753674), T2J)); | |
818 T36 = VADD(T2K, T2H); | |
819 T30 = VFMA(LDK(KP1_071653589), T2Q, VMUL(LDK(KP844327925), T2P)); | |
820 T31 = VFMA(LDK(KP125581039), T2T, VMUL(LDK(KP998026728), T2S)); | |
821 T37 = VADD(T30, T31); | |
822 T2L = VSUB(T2H, T2K); | |
823 T3b = VADD(T36, T37); | |
824 T32 = VSUB(T30, T31); | |
825 T38 = VMUL(LDK(KP559016994), VSUB(T36, T37)); | |
826 } | |
827 { | |
828 V T2M, T2N, T2O, T2R, T2U, T2V; | |
829 T2M = VFNMS(LDK(KP963507348), T2I, VMUL(LDK(KP876306680), T2J)); | |
830 T2N = VFMA(LDK(KP1_809654104), T2G, VMUL(LDK(KP425779291), T2F)); | |
831 T2O = VSUB(T2M, T2N); | |
832 T2R = VFNMS(LDK(KP1_688655851), T2Q, VMUL(LDK(KP535826794), T2P)); | |
833 T2U = VFNMS(LDK(KP1_996053456), T2T, VMUL(LDK(KP062790519), T2S)); | |
834 T2V = VADD(T2R, T2U); | |
835 T2W = VMUL(LDK(KP559016994), VSUB(T2O, T2V)); | |
836 T35 = VSUB(T2R, T2U); | |
837 T2Y = VADD(T2O, T2V); | |
838 T34 = VADD(T2M, T2N); | |
839 } | |
840 { | |
841 V T3g, T3h, T3o, T3v; | |
842 T3g = VADD(T2X, T2Y); | |
843 T3h = VBYI(VADD(T3a, T3b)); | |
844 ST(&(x[WS(rs, 23)]), VSUB(T3g, T3h), ms, &(x[WS(rs, 1)])); | |
845 ST(&(x[WS(rs, 2)]), VADD(T3g, T3h), ms, &(x[0])); | |
846 T3o = VADD(T2X, VADD(T3k, T3n)); | |
847 T3v = VBYI(VSUB(VADD(T3r, T3u), T3a)); | |
848 ST(&(x[WS(rs, 22)]), VSUB(T3o, T3v), ms, &(x[0])); | |
849 ST(&(x[WS(rs, 3)]), VADD(T3o, T3v), ms, &(x[WS(rs, 1)])); | |
850 } | |
851 T3w = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T3i, T3j), VFMA(LDK(KP309016994), T3r, VFNMS(LDK(KP809016994), T3u, VMUL(LDK(KP587785252), VSUB(T3l, T3m))))), T3a)); | |
852 T3x = VFMA(LDK(KP309016994), T3k, VFMA(LDK(KP951056516), VSUB(T3q, T3p), VFMA(LDK(KP587785252), VSUB(T3t, T3s), VFNMS(LDK(KP809016994), T3n, T2X)))); | |
853 ST(&(x[WS(rs, 8)]), VADD(T3w, T3x), ms, &(x[0])); | |
854 ST(&(x[WS(rs, 17)]), VSUB(T3x, T3w), ms, &(x[WS(rs, 1)])); | |
855 { | |
856 V T33, T3e, T3d, T3f, T2Z, T3c; | |
857 T2Z = VFNMS(LDK(KP250000000), T2Y, T2X); | |
858 T33 = VFMA(LDK(KP951056516), T2L, VADD(T2W, VFNMS(LDK(KP587785252), T32, T2Z))); | |
859 T3e = VFMA(LDK(KP587785252), T2L, VFMA(LDK(KP951056516), T32, VSUB(T2Z, T2W))); | |
860 T3c = VFNMS(LDK(KP250000000), T3b, T3a); | |
861 T3d = VBYI(VADD(VFMA(LDK(KP951056516), T34, VMUL(LDK(KP587785252), T35)), VADD(T38, T3c))); | |
862 T3f = VBYI(VADD(VFNMS(LDK(KP951056516), T35, VMUL(LDK(KP587785252), T34)), VSUB(T3c, T38))); | |
863 ST(&(x[WS(rs, 18)]), VSUB(T33, T3d), ms, &(x[0])); | |
864 ST(&(x[WS(rs, 12)]), VADD(T3e, T3f), ms, &(x[0])); | |
865 ST(&(x[WS(rs, 7)]), VADD(T33, T3d), ms, &(x[WS(rs, 1)])); | |
866 ST(&(x[WS(rs, 13)]), VSUB(T3e, T3f), ms, &(x[WS(rs, 1)])); | |
867 } | |
868 } | |
869 { | |
870 V T21, T2h, T2p, T2q, T2r, T2z, T2A, T2B, T2s, T2t, T2u, T2w, T2x, T2y, TU; | |
871 V T2i, T26, T2c, T1K, T29, T22, T28, T2D, T2E; | |
872 T21 = VADD(T1X, T20); | |
873 T2h = VADD(T2e, T2g); | |
874 T2p = VFMA(LDK(KP1_688655851), TK, VMUL(LDK(KP535826794), TS)); | |
875 T2q = VFMA(LDK(KP1_541026485), Tk, VMUL(LDK(KP637423989), Ts)); | |
876 T2r = VSUB(T2p, T2q); | |
877 T2z = VFMA(LDK(KP851558583), T1k, VMUL(LDK(KP904827052), T1h)); | |
878 T2A = VFMA(LDK(KP1_984229402), T1H, VMUL(LDK(KP125333233), T1E)); | |
879 T2B = VADD(T2z, T2A); | |
880 T2s = VFNMS(LDK(KP425779291), T1h, VMUL(LDK(KP1_809654104), T1k)); | |
881 T2t = VFNMS(LDK(KP992114701), T1E, VMUL(LDK(KP250666467), T1H)); | |
882 T2u = VADD(T2s, T2t); | |
883 T2w = VFNMS(LDK(KP1_071653589), TK, VMUL(LDK(KP844327925), TS)); | |
884 T2x = VFNMS(LDK(KP770513242), Ts, VMUL(LDK(KP1_274847979), Tk)); | |
885 T2y = VADD(T2w, T2x); | |
886 { | |
887 V Tt, TT, T2a, T24, T25, T2b; | |
888 Tt = VFMA(LDK(KP1_071653589), Tk, VMUL(LDK(KP844327925), Ts)); | |
889 TT = VFMA(LDK(KP1_937166322), TK, VMUL(LDK(KP248689887), TS)); | |
890 T2a = VADD(TT, Tt); | |
891 T24 = VFMA(LDK(KP1_752613360), T1k, VMUL(LDK(KP481753674), T1h)); | |
892 T25 = VFMA(LDK(KP1_457937254), T1H, VMUL(LDK(KP684547105), T1E)); | |
893 T2b = VADD(T24, T25); | |
894 TU = VSUB(Tt, TT); | |
895 T2i = VADD(T2a, T2b); | |
896 T26 = VSUB(T24, T25); | |
897 T2c = VMUL(LDK(KP559016994), VSUB(T2a, T2b)); | |
898 } | |
899 { | |
900 V TV, TW, TX, T1l, T1I, T1J; | |
901 TV = VFNMS(LDK(KP497379774), TK, VMUL(LDK(KP968583161), TS)); | |
902 TW = VFNMS(LDK(KP1_688655851), Tk, VMUL(LDK(KP535826794), Ts)); | |
903 TX = VADD(TV, TW); | |
904 T1l = VFNMS(LDK(KP963507348), T1k, VMUL(LDK(KP876306680), T1h)); | |
905 T1I = VFNMS(LDK(KP1_369094211), T1H, VMUL(LDK(KP728968627), T1E)); | |
906 T1J = VADD(T1l, T1I); | |
907 T1K = VMUL(LDK(KP559016994), VSUB(TX, T1J)); | |
908 T29 = VSUB(T1l, T1I); | |
909 T22 = VADD(TX, T1J); | |
910 T28 = VSUB(TV, TW); | |
911 } | |
912 { | |
913 V T2n, T2o, T2v, T2C; | |
914 T2n = VADD(T21, T22); | |
915 T2o = VBYI(VADD(T2h, T2i)); | |
916 ST(&(x[WS(rs, 24)]), VSUB(T2n, T2o), ms, &(x[0])); | |
917 ST(&(x[WS(rs, 1)]), VADD(T2n, T2o), ms, &(x[WS(rs, 1)])); | |
918 T2v = VADD(T21, VADD(T2r, T2u)); | |
919 T2C = VBYI(VSUB(VADD(T2y, T2B), T2h)); | |
920 ST(&(x[WS(rs, 21)]), VSUB(T2v, T2C), ms, &(x[WS(rs, 1)])); | |
921 ST(&(x[WS(rs, 4)]), VADD(T2v, T2C), ms, &(x[0])); | |
922 } | |
923 T2D = VBYI(VSUB(VFMA(LDK(KP309016994), T2y, VFMA(LDK(KP951056516), VADD(T2p, T2q), VFNMS(LDK(KP809016994), T2B, VMUL(LDK(KP587785252), VSUB(T2s, T2t))))), T2h)); | |
924 T2E = VFMA(LDK(KP951056516), VSUB(T2x, T2w), VFMA(LDK(KP309016994), T2r, VFMA(LDK(KP587785252), VSUB(T2A, T2z), VFNMS(LDK(KP809016994), T2u, T21)))); | |
925 ST(&(x[WS(rs, 9)]), VADD(T2D, T2E), ms, &(x[WS(rs, 1)])); | |
926 ST(&(x[WS(rs, 16)]), VSUB(T2E, T2D), ms, &(x[0])); | |
927 { | |
928 V T27, T2l, T2k, T2m, T23, T2j; | |
929 T23 = VFNMS(LDK(KP250000000), T22, T21); | |
930 T27 = VFMA(LDK(KP951056516), TU, VADD(T1K, VFNMS(LDK(KP587785252), T26, T23))); | |
931 T2l = VFMA(LDK(KP587785252), TU, VFMA(LDK(KP951056516), T26, VSUB(T23, T1K))); | |
932 T2j = VFNMS(LDK(KP250000000), T2i, T2h); | |
933 T2k = VBYI(VADD(VFMA(LDK(KP951056516), T28, VMUL(LDK(KP587785252), T29)), VADD(T2c, T2j))); | |
934 T2m = VBYI(VADD(VFNMS(LDK(KP951056516), T29, VMUL(LDK(KP587785252), T28)), VSUB(T2j, T2c))); | |
935 ST(&(x[WS(rs, 19)]), VSUB(T27, T2k), ms, &(x[WS(rs, 1)])); | |
936 ST(&(x[WS(rs, 11)]), VADD(T2l, T2m), ms, &(x[WS(rs, 1)])); | |
937 ST(&(x[WS(rs, 6)]), VADD(T27, T2k), ms, &(x[0])); | |
938 ST(&(x[WS(rs, 14)]), VSUB(T2l, T2m), ms, &(x[0])); | |
939 } | |
940 } | |
941 } | |
942 } | |
943 } | |
944 VLEAVE(); | |
945 } | |
946 | |
947 static const tw_instr twinstr[] = { | |
948 VTW(0, 1), | |
949 VTW(0, 3), | |
950 VTW(0, 9), | |
951 VTW(0, 24), | |
952 {TW_NEXT, VL, 0} | |
953 }; | |
954 | |
955 static const ct_desc desc = { 25, XSIMD_STRING("t3bv_25"), twinstr, &GENUS, {191, 151, 77, 0}, 0, 0, 0 }; | |
956 | |
957 void XSIMD(codelet_t3bv_25) (planner *p) { | |
958 X(kdft_dit_register) (p, t3bv_25, &desc); | |
959 } | |
960 #endif |