Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/dft/simd/common/t1fv_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 08:05:42 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 -n 25 -name t1fv_25 -include dft/simd/t1f.h */ | |
29 | |
30 /* | |
31 * This function contains 248 FP additions, 241 FP multiplications, | |
32 * (or, 67 additions, 60 multiplications, 181 fused multiply/add), | |
33 * 147 stack variables, 67 constants, and 50 memory accesses | |
34 */ | |
35 #include "dft/simd/t1f.h" | |
36 | |
37 static void t1fv_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) * 48)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 48), MAKE_VOLATILE_STRIDE(25, rs)) { | |
111 V T1, Te, Tc, Td, T1O, T2X, T3Q, T1x, T2K, T1u, T2L, T1y, T27, T3b, T2R; | |
112 V T2M, T2f, T3M, Ty, T2E, Tv, T2D, Tz, T2a, T3e, T2U, T2F, T2i, T3N, TK; | |
113 V T2B, TS, T2A, TT, T2b, T3f, T2T, T2C, T2j, T3P, T1d, T2H, T1a, T2I, T1e; | |
114 V T28, T3c, T2Q, T2J, T2g; | |
115 { | |
116 V T8, Ta, Tb, T3, T5, T6, T1M, T1N; | |
117 T1 = LD(&(x[0]), ms, &(x[0])); | |
118 { | |
119 V T7, T9, T2, T4; | |
120 T7 = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
121 T8 = BYTWJ(&(W[TWVL * 18]), T7); | |
122 T9 = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
123 Ta = BYTWJ(&(W[TWVL * 28]), T9); | |
124 Tb = VADD(T8, Ta); | |
125 T2 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
126 T3 = BYTWJ(&(W[TWVL * 8]), T2); | |
127 T4 = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
128 T5 = BYTWJ(&(W[TWVL * 38]), T4); | |
129 T6 = VADD(T3, T5); | |
130 } | |
131 Te = VSUB(T6, Tb); | |
132 Tc = VADD(T6, Tb); | |
133 Td = VFNMS(LDK(KP250000000), Tc, T1); | |
134 T1M = VSUB(T3, T5); | |
135 T1N = VSUB(T8, Ta); | |
136 T1O = VFMA(LDK(KP618033988), T1N, T1M); | |
137 T2X = VFNMS(LDK(KP618033988), T1M, T1N); | |
138 } | |
139 { | |
140 V T1g, T1v, T1w, T1l, T1q, T1r, T1f, T1s, T1t; | |
141 T1f = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
142 T1g = BYTWJ(&(W[TWVL * 4]), T1f); | |
143 { | |
144 V T1i, T1p, T1k, T1n; | |
145 { | |
146 V T1h, T1o, T1j, T1m; | |
147 T1h = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
148 T1i = BYTWJ(&(W[TWVL * 14]), T1h); | |
149 T1o = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
150 T1p = BYTWJ(&(W[TWVL * 34]), T1o); | |
151 T1j = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
152 T1k = BYTWJ(&(W[TWVL * 44]), T1j); | |
153 T1m = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
154 T1n = BYTWJ(&(W[TWVL * 24]), T1m); | |
155 } | |
156 T1v = VSUB(T1i, T1k); | |
157 T1w = VSUB(T1n, T1p); | |
158 T1l = VADD(T1i, T1k); | |
159 T1q = VADD(T1n, T1p); | |
160 T1r = VADD(T1l, T1q); | |
161 } | |
162 T3Q = VADD(T1g, T1r); | |
163 T1x = VFMA(LDK(KP618033988), T1w, T1v); | |
164 T2K = VFNMS(LDK(KP618033988), T1v, T1w); | |
165 T1s = VFNMS(LDK(KP250000000), T1r, T1g); | |
166 T1t = VSUB(T1q, T1l); | |
167 T1u = VFNMS(LDK(KP559016994), T1t, T1s); | |
168 T2L = VFMA(LDK(KP559016994), T1t, T1s); | |
169 T1y = VFNMS(LDK(KP893101515), T1x, T1u); | |
170 T27 = VFNMS(LDK(KP120146378), T1x, T1u); | |
171 T3b = VFMA(LDK(KP066152395), T2L, T2K); | |
172 T2R = VFNMS(LDK(KP786782374), T2K, T2L); | |
173 T2M = VFMA(LDK(KP869845200), T2L, T2K); | |
174 T2f = VFMA(LDK(KP132830569), T1u, T1x); | |
175 } | |
176 { | |
177 V Th, Tw, Tx, Tm, Tr, Ts, Tg, Tt, Tu; | |
178 Tg = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
179 Th = BYTWJ(&(W[0]), Tg); | |
180 { | |
181 V Tj, Tq, Tl, To; | |
182 { | |
183 V Ti, Tp, Tk, Tn; | |
184 Ti = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
185 Tj = BYTWJ(&(W[TWVL * 10]), Ti); | |
186 Tp = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
187 Tq = BYTWJ(&(W[TWVL * 30]), Tp); | |
188 Tk = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
189 Tl = BYTWJ(&(W[TWVL * 40]), Tk); | |
190 Tn = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
191 To = BYTWJ(&(W[TWVL * 20]), Tn); | |
192 } | |
193 Tw = VSUB(Tj, Tl); | |
194 Tx = VSUB(Tq, To); | |
195 Tm = VADD(Tj, Tl); | |
196 Tr = VADD(To, Tq); | |
197 Ts = VADD(Tm, Tr); | |
198 } | |
199 T3M = VADD(Th, Ts); | |
200 Ty = VFNMS(LDK(KP618033988), Tx, Tw); | |
201 T2E = VFMA(LDK(KP618033988), Tw, Tx); | |
202 Tt = VFNMS(LDK(KP250000000), Ts, Th); | |
203 Tu = VSUB(Tm, Tr); | |
204 Tv = VFMA(LDK(KP559016994), Tu, Tt); | |
205 T2D = VFNMS(LDK(KP559016994), Tu, Tt); | |
206 Tz = VFNMS(LDK(KP244189809), Ty, Tv); | |
207 T2a = VFMA(LDK(KP667278218), Tv, Ty); | |
208 T3e = VFNMS(LDK(KP522847744), T2E, T2D); | |
209 T2U = VFNMS(LDK(KP987388751), T2D, T2E); | |
210 T2F = VFMA(LDK(KP893101515), T2E, T2D); | |
211 T2i = VFNMS(LDK(KP603558818), Ty, Tv); | |
212 } | |
213 { | |
214 V TM, TE, TJ, TN, TO, TP, TL, TQ, TR; | |
215 TL = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
216 TM = BYTWJ(&(W[TWVL * 6]), TL); | |
217 { | |
218 V TB, TI, TD, TG; | |
219 { | |
220 V TA, TH, TC, TF; | |
221 TA = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
222 TB = BYTWJ(&(W[TWVL * 46]), TA); | |
223 TH = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
224 TI = BYTWJ(&(W[TWVL * 26]), TH); | |
225 TC = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
226 TD = BYTWJ(&(W[TWVL * 16]), TC); | |
227 TF = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
228 TG = BYTWJ(&(W[TWVL * 36]), TF); | |
229 } | |
230 TE = VSUB(TB, TD); | |
231 TJ = VSUB(TG, TI); | |
232 TN = VADD(TD, TB); | |
233 TO = VADD(TI, TG); | |
234 TP = VADD(TN, TO); | |
235 } | |
236 T3N = VADD(TM, TP); | |
237 TK = VFMA(LDK(KP618033988), TJ, TE); | |
238 T2B = VFNMS(LDK(KP618033988), TE, TJ); | |
239 TQ = VFMS(LDK(KP250000000), TP, TM); | |
240 TR = VSUB(TN, TO); | |
241 TS = VFNMS(LDK(KP559016994), TR, TQ); | |
242 T2A = VFMA(LDK(KP559016994), TR, TQ); | |
243 TT = VFNMS(LDK(KP667278218), TS, TK); | |
244 T2b = VFMA(LDK(KP869845200), TS, TK); | |
245 T3f = VFNMS(LDK(KP494780565), T2A, T2B); | |
246 T2T = VFNMS(LDK(KP132830569), T2A, T2B); | |
247 T2C = VFMA(LDK(KP120146378), T2B, T2A); | |
248 T2j = VFNMS(LDK(KP786782374), TK, TS); | |
249 } | |
250 { | |
251 V TW, T1b, T1c, T11, T16, T17, TV, T18, T19; | |
252 TV = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
253 TW = BYTWJ(&(W[TWVL * 2]), TV); | |
254 { | |
255 V TY, T15, T10, T13; | |
256 { | |
257 V TX, T14, TZ, T12; | |
258 TX = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
259 TY = BYTWJ(&(W[TWVL * 12]), TX); | |
260 T14 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
261 T15 = BYTWJ(&(W[TWVL * 32]), T14); | |
262 TZ = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
263 T10 = BYTWJ(&(W[TWVL * 42]), TZ); | |
264 T12 = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
265 T13 = BYTWJ(&(W[TWVL * 22]), T12); | |
266 } | |
267 T1b = VSUB(TY, T10); | |
268 T1c = VSUB(T15, T13); | |
269 T11 = VADD(TY, T10); | |
270 T16 = VADD(T13, T15); | |
271 T17 = VADD(T11, T16); | |
272 } | |
273 T3P = VADD(TW, T17); | |
274 T1d = VFNMS(LDK(KP618033988), T1c, T1b); | |
275 T2H = VFMA(LDK(KP618033988), T1b, T1c); | |
276 T18 = VFNMS(LDK(KP250000000), T17, TW); | |
277 T19 = VSUB(T16, T11); | |
278 T1a = VFNMS(LDK(KP559016994), T19, T18); | |
279 T2I = VFMA(LDK(KP559016994), T19, T18); | |
280 T1e = VFNMS(LDK(KP522847744), T1d, T1a); | |
281 T28 = VFNMS(LDK(KP494780565), T1a, T1d); | |
282 T3c = VFNMS(LDK(KP667278218), T2I, T2H); | |
283 T2Q = VFNMS(LDK(KP059835404), T2H, T2I); | |
284 T2J = VFMA(LDK(KP066152395), T2I, T2H); | |
285 T2g = VFMA(LDK(KP447533225), T1d, T1a); | |
286 } | |
287 { | |
288 V T3Y, T40, T3L, T3S, T3T, T3U, T3Z, T3V; | |
289 { | |
290 V T3W, T3X, T3O, T3R; | |
291 T3W = VSUB(T3M, T3N); | |
292 T3X = VSUB(T3P, T3Q); | |
293 T3Y = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T3X, T3W)); | |
294 T40 = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T3W, T3X)); | |
295 T3L = VADD(T1, Tc); | |
296 T3O = VADD(T3M, T3N); | |
297 T3R = VADD(T3P, T3Q); | |
298 T3S = VADD(T3O, T3R); | |
299 T3T = VFNMS(LDK(KP250000000), T3S, T3L); | |
300 T3U = VSUB(T3O, T3R); | |
301 } | |
302 ST(&(x[0]), VADD(T3S, T3L), ms, &(x[0])); | |
303 T3Z = VFNMS(LDK(KP559016994), T3U, T3T); | |
304 ST(&(x[WS(rs, 10)]), VFMAI(T40, T3Z), ms, &(x[0])); | |
305 ST(&(x[WS(rs, 15)]), VFNMSI(T40, T3Z), ms, &(x[WS(rs, 1)])); | |
306 T3V = VFMA(LDK(KP559016994), T3U, T3T); | |
307 ST(&(x[WS(rs, 5)]), VFNMSI(T3Y, T3V), ms, &(x[WS(rs, 1)])); | |
308 ST(&(x[WS(rs, 20)]), VFMAI(T3Y, T3V), ms, &(x[0])); | |
309 } | |
310 { | |
311 V T2Z, T35, T3B, T3I, T2W, T38, T2O, T32, T2z, T3t, T3h, T3s, T3p, T3F, T3r; | |
312 V T3v, T3C, T3z, T3A; | |
313 T2Z = VFMA(LDK(KP734762448), T2U, T2T); | |
314 T35 = VFNMS(LDK(KP734762448), T2F, T2C); | |
315 T3z = VFMA(LDK(KP845997307), T3c, T3b); | |
316 T3A = VFMA(LDK(KP982009705), T3f, T3e); | |
317 T3B = VFMA(LDK(KP570584518), T3A, T3z); | |
318 T3I = VFNMS(LDK(KP669429328), T3z, T3A); | |
319 { | |
320 V T2S, T2V, T37, T36; | |
321 T2S = VFMA(LDK(KP772036680), T2R, T2Q); | |
322 T2V = VFNMS(LDK(KP734762448), T2U, T2T); | |
323 T36 = VFMA(LDK(KP772036680), T2M, T2J); | |
324 T37 = VFMA(LDK(KP522616830), T2V, T36); | |
325 T2W = VFMA(LDK(KP945422727), T2V, T2S); | |
326 T38 = VFNMS(LDK(KP690983005), T37, T2S); | |
327 } | |
328 { | |
329 V T2N, T2G, T31, T30; | |
330 T2N = VFNMS(LDK(KP772036680), T2M, T2J); | |
331 T2G = VFMA(LDK(KP734762448), T2F, T2C); | |
332 T30 = VFNMS(LDK(KP772036680), T2R, T2Q); | |
333 T31 = VFNMS(LDK(KP522616830), T2G, T30); | |
334 T2O = VFMA(LDK(KP956723877), T2N, T2G); | |
335 T32 = VFMA(LDK(KP763932022), T31, T2N); | |
336 } | |
337 { | |
338 V T3o, T3u, T3l, T3m, T3n; | |
339 T2z = VFNMS(LDK(KP559016994), Te, Td); | |
340 T3m = VFMA(LDK(KP447533225), T2B, T2A); | |
341 T3n = VFMA(LDK(KP578046249), T2D, T2E); | |
342 T3o = VFNMS(LDK(KP921078979), T3n, T3m); | |
343 T3t = VFMA(LDK(KP921078979), T3n, T3m); | |
344 { | |
345 V T3d, T3g, T3j, T3k; | |
346 T3d = VFNMS(LDK(KP845997307), T3c, T3b); | |
347 T3g = VFNMS(LDK(KP982009705), T3f, T3e); | |
348 T3h = VFMA(LDK(KP923225144), T3g, T3d); | |
349 T3u = VFNMS(LDK(KP923225144), T3g, T3d); | |
350 T3j = VFNMS(LDK(KP059835404), T2K, T2L); | |
351 T3k = VFMA(LDK(KP603558818), T2H, T2I); | |
352 T3l = VFMA(LDK(KP845997307), T3k, T3j); | |
353 T3s = VFNMS(LDK(KP845997307), T3k, T3j); | |
354 } | |
355 T3p = VFNMS(LDK(KP906616052), T3o, T3l); | |
356 T3F = VFNMS(LDK(KP904508497), T3u, T3s); | |
357 T3r = VFNMS(LDK(KP237294955), T3h, T2z); | |
358 T3v = VFNMS(LDK(KP997675361), T3u, T3t); | |
359 T3C = VFMA(LDK(KP906616052), T3o, T3l); | |
360 } | |
361 { | |
362 V T2P, T2Y, T3i, T3q; | |
363 T2P = VFMA(LDK(KP992114701), T2O, T2z); | |
364 T2Y = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T2X, T2W)); | |
365 ST(&(x[WS(rs, 3)]), VFNMSI(T2Y, T2P), ms, &(x[WS(rs, 1)])); | |
366 ST(&(x[WS(rs, 22)]), VFMAI(T2Y, T2P), ms, &(x[0])); | |
367 T3i = VFMA(LDK(KP949179823), T3h, T2z); | |
368 T3q = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T2X, T3p)); | |
369 ST(&(x[WS(rs, 2)]), VFNMSI(T3q, T3i), ms, &(x[0])); | |
370 ST(&(x[WS(rs, 23)]), VFMAI(T3q, T3i), ms, &(x[WS(rs, 1)])); | |
371 } | |
372 { | |
373 V T34, T3a, T33, T39; | |
374 T33 = VFNMS(LDK(KP855719849), T32, T2Z); | |
375 T34 = VFMA(LDK(KP897376177), T33, T2z); | |
376 T39 = VFMA(LDK(KP855719849), T38, T35); | |
377 T3a = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T39, T2X)); | |
378 ST(&(x[WS(rs, 8)]), VFNMSI(T3a, T34), ms, &(x[0])); | |
379 ST(&(x[WS(rs, 17)]), VFMAI(T3a, T34), ms, &(x[WS(rs, 1)])); | |
380 } | |
381 { | |
382 V T3x, T3H, T3E, T3K, T3w; | |
383 T3w = VFMA(LDK(KP560319534), T3v, T3s); | |
384 T3x = VFNMS(LDK(KP949179823), T3w, T3r); | |
385 { | |
386 V T3G, T3y, T3J, T3D; | |
387 T3G = VFNMS(LDK(KP681693190), T3F, T3t); | |
388 T3H = VFNMS(LDK(KP860541664), T3G, T3r); | |
389 T3y = VFMA(LDK(KP262346850), T3p, T2X); | |
390 T3J = VFNMS(LDK(KP669429328), T3C, T3I); | |
391 T3D = VFMA(LDK(KP618033988), T3C, T3B); | |
392 T3E = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T3D, T3y)); | |
393 T3K = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T3J, T3y)); | |
394 } | |
395 ST(&(x[WS(rs, 13)]), VFNMSI(T3E, T3x), ms, &(x[WS(rs, 1)])); | |
396 ST(&(x[WS(rs, 7)]), VFMAI(T3K, T3H), ms, &(x[WS(rs, 1)])); | |
397 ST(&(x[WS(rs, 12)]), VFMAI(T3E, T3x), ms, &(x[0])); | |
398 ST(&(x[WS(rs, 18)]), VFNMSI(T3K, T3H), ms, &(x[0])); | |
399 } | |
400 } | |
401 { | |
402 V T2n, T2t, T1V, T22, T2l, T2w, T2d, T2q, Tf, T1I, T1A, T1E, T1B, T1Z, T1J; | |
403 V T1R, T1W, T1T, T1U; | |
404 T2n = VFNMS(LDK(KP912575812), T2j, T2i); | |
405 T2t = VFNMS(LDK(KP912575812), T2b, T2a); | |
406 T1T = VFNMS(LDK(KP829049696), TT, Tz); | |
407 T1U = VFNMS(LDK(KP831864738), T1y, T1e); | |
408 T1V = VFMA(LDK(KP559154169), T1U, T1T); | |
409 T22 = VFNMS(LDK(KP683113946), T1T, T1U); | |
410 { | |
411 V T2h, T2k, T2v, T2u; | |
412 T2h = VFMA(LDK(KP958953096), T2g, T2f); | |
413 T2k = VFMA(LDK(KP912575812), T2j, T2i); | |
414 T2u = VFMA(LDK(KP867381224), T28, T27); | |
415 T2v = VFMA(LDK(KP447417479), T2k, T2u); | |
416 T2l = VFMA(LDK(KP894834959), T2k, T2h); | |
417 T2w = VFNMS(LDK(KP763932022), T2v, T2h); | |
418 } | |
419 { | |
420 V T29, T2c, T2p, T2o; | |
421 T29 = VFNMS(LDK(KP867381224), T28, T27); | |
422 T2c = VFMA(LDK(KP912575812), T2b, T2a); | |
423 T2o = VFNMS(LDK(KP958953096), T2g, T2f); | |
424 T2p = VFMA(LDK(KP447417479), T2c, T2o); | |
425 T2d = VFNMS(LDK(KP809385824), T2c, T29); | |
426 T2q = VFMA(LDK(KP690983005), T2p, T29); | |
427 } | |
428 { | |
429 V T1Q, T1F, T1P, T1G, T1H; | |
430 Tf = VFMA(LDK(KP559016994), Te, Td); | |
431 T1G = VFMA(LDK(KP578046249), T1a, T1d); | |
432 T1H = VFMA(LDK(KP987388751), T1u, T1x); | |
433 T1I = VFNMS(LDK(KP831864738), T1H, T1G); | |
434 T1Q = VFMA(LDK(KP831864738), T1H, T1G); | |
435 { | |
436 V TU, T1z, T1C, T1D; | |
437 TU = VFMA(LDK(KP829049696), TT, Tz); | |
438 T1z = VFMA(LDK(KP831864738), T1y, T1e); | |
439 T1A = VFMA(LDK(KP904730450), T1z, TU); | |
440 T1F = VFNMS(LDK(KP904730450), T1z, TU); | |
441 T1C = VFMA(LDK(KP269969613), Tv, Ty); | |
442 T1D = VFMA(LDK(KP603558818), TK, TS); | |
443 T1E = VFMA(LDK(KP916574801), T1D, T1C); | |
444 T1P = VFNMS(LDK(KP916574801), T1D, T1C); | |
445 } | |
446 T1B = VFNMS(LDK(KP242145790), T1A, Tf); | |
447 T1Z = VADD(T1E, T1F); | |
448 T1J = VFNMS(LDK(KP904730450), T1I, T1F); | |
449 T1R = VFMA(LDK(KP904730450), T1Q, T1P); | |
450 T1W = VFNMS(LDK(KP904730450), T1Q, T1P); | |
451 } | |
452 { | |
453 V T25, T26, T2e, T2m; | |
454 T25 = VFMA(LDK(KP968583161), T1A, Tf); | |
455 T26 = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1R, T1O)); | |
456 ST(&(x[WS(rs, 1)]), VFNMSI(T26, T25), ms, &(x[WS(rs, 1)])); | |
457 ST(&(x[WS(rs, 24)]), VFMAI(T26, T25), ms, &(x[0])); | |
458 T2e = VFNMS(LDK(KP992114701), T2d, Tf); | |
459 T2m = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2l, T1O)); | |
460 ST(&(x[WS(rs, 4)]), VFMAI(T2m, T2e), ms, &(x[0])); | |
461 ST(&(x[WS(rs, 21)]), VFNMSI(T2m, T2e), ms, &(x[WS(rs, 1)])); | |
462 } | |
463 { | |
464 V T2s, T2y, T2r, T2x; | |
465 T2r = VFNMS(LDK(KP999544308), T2q, T2n); | |
466 T2s = VFNMS(LDK(KP803003575), T2r, Tf); | |
467 T2x = VFNMS(LDK(KP999544308), T2w, T2t); | |
468 T2y = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T2x, T1O)); | |
469 ST(&(x[WS(rs, 16)]), VFNMSI(T2y, T2s), ms, &(x[0])); | |
470 ST(&(x[WS(rs, 9)]), VFMAI(T2y, T2s), ms, &(x[WS(rs, 1)])); | |
471 } | |
472 { | |
473 V T1L, T21, T1Y, T24, T1K; | |
474 T1K = VFNMS(LDK(KP618033988), T1J, T1E); | |
475 T1L = VFNMS(LDK(KP876091699), T1K, T1B); | |
476 { | |
477 V T20, T1S, T23, T1X; | |
478 T20 = VFNMS(LDK(KP683113946), T1Z, T1I); | |
479 T21 = VFMA(LDK(KP792626838), T20, T1B); | |
480 T1S = VFNMS(LDK(KP242145790), T1R, T1O); | |
481 T23 = VFMA(LDK(KP617882369), T1W, T22); | |
482 T1X = VFMA(LDK(KP559016994), T1W, T1V); | |
483 T1Y = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1X, T1S)); | |
484 T24 = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T23, T1S)); | |
485 } | |
486 ST(&(x[WS(rs, 6)]), VFNMSI(T1Y, T1L), ms, &(x[0])); | |
487 ST(&(x[WS(rs, 14)]), VFMAI(T24, T21), ms, &(x[0])); | |
488 ST(&(x[WS(rs, 19)]), VFMAI(T1Y, T1L), ms, &(x[WS(rs, 1)])); | |
489 ST(&(x[WS(rs, 11)]), VFNMSI(T24, T21), ms, &(x[WS(rs, 1)])); | |
490 } | |
491 } | |
492 } | |
493 } | |
494 VLEAVE(); | |
495 } | |
496 | |
497 static const tw_instr twinstr[] = { | |
498 VTW(0, 1), | |
499 VTW(0, 2), | |
500 VTW(0, 3), | |
501 VTW(0, 4), | |
502 VTW(0, 5), | |
503 VTW(0, 6), | |
504 VTW(0, 7), | |
505 VTW(0, 8), | |
506 VTW(0, 9), | |
507 VTW(0, 10), | |
508 VTW(0, 11), | |
509 VTW(0, 12), | |
510 VTW(0, 13), | |
511 VTW(0, 14), | |
512 VTW(0, 15), | |
513 VTW(0, 16), | |
514 VTW(0, 17), | |
515 VTW(0, 18), | |
516 VTW(0, 19), | |
517 VTW(0, 20), | |
518 VTW(0, 21), | |
519 VTW(0, 22), | |
520 VTW(0, 23), | |
521 VTW(0, 24), | |
522 {TW_NEXT, VL, 0} | |
523 }; | |
524 | |
525 static const ct_desc desc = { 25, XSIMD_STRING("t1fv_25"), twinstr, &GENUS, {67, 60, 181, 0}, 0, 0, 0 }; | |
526 | |
527 void XSIMD(codelet_t1fv_25) (planner *p) { | |
528 X(kdft_dit_register) (p, t1fv_25, &desc); | |
529 } | |
530 #else | |
531 | |
532 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 25 -name t1fv_25 -include dft/simd/t1f.h */ | |
533 | |
534 /* | |
535 * This function contains 248 FP additions, 188 FP multiplications, | |
536 * (or, 170 additions, 110 multiplications, 78 fused multiply/add), | |
537 * 99 stack variables, 40 constants, and 50 memory accesses | |
538 */ | |
539 #include "dft/simd/t1f.h" | |
540 | |
541 static void t1fv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
542 { | |
543 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
544 DVK(KP125581039, +0.125581039058626752152356449131262266244969664); | |
545 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
546 DVK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
547 DVK(KP809016994, +0.809016994374947424102293417182819058860154590); | |
548 DVK(KP309016994, +0.309016994374947424102293417182819058860154590); | |
549 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465); | |
550 DVK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
551 DVK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
552 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
553 DVK(KP497379774, +0.497379774329709576484567492012895936835134813); | |
554 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
555 DVK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
556 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328); | |
557 DVK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
558 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
559 DVK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
560 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
561 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
562 DVK(KP250666467, +0.250666467128608490746237519633017587885836494); | |
563 DVK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
564 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
565 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596); | |
566 DVK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
567 DVK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
568 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691); | |
569 DVK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
570 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418); | |
571 DVK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
572 DVK(KP851558583, +0.851558583130145297725004891488503407959946084); | |
573 DVK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
574 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344); | |
575 DVK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
576 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483); | |
577 DVK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
578 DVK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
579 DVK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
580 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
581 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
582 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
583 { | |
584 INT m; | |
585 R *x; | |
586 x = ri; | |
587 for (m = mb, W = W + (mb * ((TWVL / VL) * 48)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 48), MAKE_VOLATILE_STRIDE(25, rs)) { | |
588 V Tc, Tb, Td, Te, T1C, T2t, T1E, T1x, T2m, T1u, T3c, T2n, Ty, T2i, Tv; | |
589 V T38, T2j, TS, T2f, TP, T39, T2g, T1d, T2p, T1a, T3b, T2q; | |
590 { | |
591 V T7, T9, Ta, T2, T4, T5, T1D; | |
592 Tc = LD(&(x[0]), ms, &(x[0])); | |
593 { | |
594 V T6, T8, T1, T3; | |
595 T6 = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
596 T7 = BYTWJ(&(W[TWVL * 18]), T6); | |
597 T8 = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
598 T9 = BYTWJ(&(W[TWVL * 28]), T8); | |
599 Ta = VADD(T7, T9); | |
600 T1 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
601 T2 = BYTWJ(&(W[TWVL * 8]), T1); | |
602 T3 = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
603 T4 = BYTWJ(&(W[TWVL * 38]), T3); | |
604 T5 = VADD(T2, T4); | |
605 } | |
606 Tb = VMUL(LDK(KP559016994), VSUB(T5, Ta)); | |
607 Td = VADD(T5, Ta); | |
608 Te = VFNMS(LDK(KP250000000), Td, Tc); | |
609 T1C = VSUB(T2, T4); | |
610 T1D = VSUB(T7, T9); | |
611 T2t = VMUL(LDK(KP951056516), T1D); | |
612 T1E = VFMA(LDK(KP951056516), T1C, VMUL(LDK(KP587785252), T1D)); | |
613 } | |
614 { | |
615 V T1r, T1l, T1n, T1o, T1g, T1i, T1j, T1q; | |
616 T1q = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
617 T1r = BYTWJ(&(W[TWVL * 4]), T1q); | |
618 { | |
619 V T1k, T1m, T1f, T1h; | |
620 T1k = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
621 T1l = BYTWJ(&(W[TWVL * 24]), T1k); | |
622 T1m = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
623 T1n = BYTWJ(&(W[TWVL * 34]), T1m); | |
624 T1o = VADD(T1l, T1n); | |
625 T1f = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
626 T1g = BYTWJ(&(W[TWVL * 14]), T1f); | |
627 T1h = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
628 T1i = BYTWJ(&(W[TWVL * 44]), T1h); | |
629 T1j = VADD(T1g, T1i); | |
630 } | |
631 { | |
632 V T1v, T1w, T1p, T1s, T1t; | |
633 T1v = VSUB(T1g, T1i); | |
634 T1w = VSUB(T1l, T1n); | |
635 T1x = VFMA(LDK(KP475528258), T1v, VMUL(LDK(KP293892626), T1w)); | |
636 T2m = VFNMS(LDK(KP293892626), T1v, VMUL(LDK(KP475528258), T1w)); | |
637 T1p = VMUL(LDK(KP559016994), VSUB(T1j, T1o)); | |
638 T1s = VADD(T1j, T1o); | |
639 T1t = VFNMS(LDK(KP250000000), T1s, T1r); | |
640 T1u = VADD(T1p, T1t); | |
641 T3c = VADD(T1r, T1s); | |
642 T2n = VSUB(T1t, T1p); | |
643 } | |
644 } | |
645 { | |
646 V Ts, Tm, To, Tp, Th, Tj, Tk, Tr; | |
647 Tr = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
648 Ts = BYTWJ(&(W[0]), Tr); | |
649 { | |
650 V Tl, Tn, Tg, Ti; | |
651 Tl = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
652 Tm = BYTWJ(&(W[TWVL * 20]), Tl); | |
653 Tn = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
654 To = BYTWJ(&(W[TWVL * 30]), Tn); | |
655 Tp = VADD(Tm, To); | |
656 Tg = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
657 Th = BYTWJ(&(W[TWVL * 10]), Tg); | |
658 Ti = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
659 Tj = BYTWJ(&(W[TWVL * 40]), Ti); | |
660 Tk = VADD(Th, Tj); | |
661 } | |
662 { | |
663 V Tw, Tx, Tq, Tt, Tu; | |
664 Tw = VSUB(Th, Tj); | |
665 Tx = VSUB(Tm, To); | |
666 Ty = VFMA(LDK(KP475528258), Tw, VMUL(LDK(KP293892626), Tx)); | |
667 T2i = VFNMS(LDK(KP293892626), Tw, VMUL(LDK(KP475528258), Tx)); | |
668 Tq = VMUL(LDK(KP559016994), VSUB(Tk, Tp)); | |
669 Tt = VADD(Tk, Tp); | |
670 Tu = VFNMS(LDK(KP250000000), Tt, Ts); | |
671 Tv = VADD(Tq, Tu); | |
672 T38 = VADD(Ts, Tt); | |
673 T2j = VSUB(Tu, Tq); | |
674 } | |
675 } | |
676 { | |
677 V TM, TG, TI, TJ, TB, TD, TE, TL; | |
678 TL = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
679 TM = BYTWJ(&(W[TWVL * 6]), TL); | |
680 { | |
681 V TF, TH, TA, TC; | |
682 TF = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
683 TG = BYTWJ(&(W[TWVL * 26]), TF); | |
684 TH = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
685 TI = BYTWJ(&(W[TWVL * 36]), TH); | |
686 TJ = VADD(TG, TI); | |
687 TA = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
688 TB = BYTWJ(&(W[TWVL * 16]), TA); | |
689 TC = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
690 TD = BYTWJ(&(W[TWVL * 46]), TC); | |
691 TE = VADD(TB, TD); | |
692 } | |
693 { | |
694 V TQ, TR, TK, TN, TO; | |
695 TQ = VSUB(TB, TD); | |
696 TR = VSUB(TG, TI); | |
697 TS = VFMA(LDK(KP475528258), TQ, VMUL(LDK(KP293892626), TR)); | |
698 T2f = VFNMS(LDK(KP293892626), TQ, VMUL(LDK(KP475528258), TR)); | |
699 TK = VMUL(LDK(KP559016994), VSUB(TE, TJ)); | |
700 TN = VADD(TE, TJ); | |
701 TO = VFNMS(LDK(KP250000000), TN, TM); | |
702 TP = VADD(TK, TO); | |
703 T39 = VADD(TM, TN); | |
704 T2g = VSUB(TO, TK); | |
705 } | |
706 } | |
707 { | |
708 V T17, T11, T13, T14, TW, TY, TZ, T16; | |
709 T16 = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
710 T17 = BYTWJ(&(W[TWVL * 2]), T16); | |
711 { | |
712 V T10, T12, TV, TX; | |
713 T10 = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
714 T11 = BYTWJ(&(W[TWVL * 22]), T10); | |
715 T12 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
716 T13 = BYTWJ(&(W[TWVL * 32]), T12); | |
717 T14 = VADD(T11, T13); | |
718 TV = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
719 TW = BYTWJ(&(W[TWVL * 12]), TV); | |
720 TX = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
721 TY = BYTWJ(&(W[TWVL * 42]), TX); | |
722 TZ = VADD(TW, TY); | |
723 } | |
724 { | |
725 V T1b, T1c, T15, T18, T19; | |
726 T1b = VSUB(TW, TY); | |
727 T1c = VSUB(T11, T13); | |
728 T1d = VFMA(LDK(KP475528258), T1b, VMUL(LDK(KP293892626), T1c)); | |
729 T2p = VFNMS(LDK(KP293892626), T1b, VMUL(LDK(KP475528258), T1c)); | |
730 T15 = VMUL(LDK(KP559016994), VSUB(TZ, T14)); | |
731 T18 = VADD(TZ, T14); | |
732 T19 = VFNMS(LDK(KP250000000), T18, T17); | |
733 T1a = VADD(T15, T19); | |
734 T3b = VADD(T17, T18); | |
735 T2q = VSUB(T19, T15); | |
736 } | |
737 } | |
738 { | |
739 V T3l, T3m, T3f, T3g, T3e, T3h, T3n, T3i; | |
740 { | |
741 V T3j, T3k, T3a, T3d; | |
742 T3j = VSUB(T38, T39); | |
743 T3k = VSUB(T3b, T3c); | |
744 T3l = VBYI(VFMA(LDK(KP951056516), T3j, VMUL(LDK(KP587785252), T3k))); | |
745 T3m = VBYI(VFNMS(LDK(KP587785252), T3j, VMUL(LDK(KP951056516), T3k))); | |
746 T3f = VADD(Tc, Td); | |
747 T3a = VADD(T38, T39); | |
748 T3d = VADD(T3b, T3c); | |
749 T3g = VADD(T3a, T3d); | |
750 T3e = VMUL(LDK(KP559016994), VSUB(T3a, T3d)); | |
751 T3h = VFNMS(LDK(KP250000000), T3g, T3f); | |
752 } | |
753 ST(&(x[0]), VADD(T3f, T3g), ms, &(x[0])); | |
754 T3n = VSUB(T3h, T3e); | |
755 ST(&(x[WS(rs, 10)]), VADD(T3m, T3n), ms, &(x[0])); | |
756 ST(&(x[WS(rs, 15)]), VSUB(T3n, T3m), ms, &(x[WS(rs, 1)])); | |
757 T3i = VADD(T3e, T3h); | |
758 ST(&(x[WS(rs, 5)]), VSUB(T3i, T3l), ms, &(x[WS(rs, 1)])); | |
759 ST(&(x[WS(rs, 20)]), VADD(T3l, T3i), ms, &(x[0])); | |
760 } | |
761 { | |
762 V Tf, T1Z, T20, T21, T29, T2a, T2b, T26, T27, T28, T22, T23, T24, T1L, T1U; | |
763 V T1Q, T1S, T1A, T1V, T1N, T1O, T2d, T2e; | |
764 Tf = VADD(Tb, Te); | |
765 T1Z = VFMA(LDK(KP1_688655851), Ty, VMUL(LDK(KP535826794), Tv)); | |
766 T20 = VFMA(LDK(KP1_541026485), TS, VMUL(LDK(KP637423989), TP)); | |
767 T21 = VSUB(T1Z, T20); | |
768 T29 = VFMA(LDK(KP851558583), T1d, VMUL(LDK(KP904827052), T1a)); | |
769 T2a = VFMA(LDK(KP1_984229402), T1x, VMUL(LDK(KP125333233), T1u)); | |
770 T2b = VADD(T29, T2a); | |
771 T26 = VFNMS(LDK(KP844327925), Tv, VMUL(LDK(KP1_071653589), Ty)); | |
772 T27 = VFNMS(LDK(KP1_274847979), TS, VMUL(LDK(KP770513242), TP)); | |
773 T28 = VADD(T26, T27); | |
774 T22 = VFNMS(LDK(KP425779291), T1a, VMUL(LDK(KP1_809654104), T1d)); | |
775 T23 = VFNMS(LDK(KP992114701), T1u, VMUL(LDK(KP250666467), T1x)); | |
776 T24 = VADD(T22, T23); | |
777 { | |
778 V T1F, T1G, T1H, T1I, T1J, T1K; | |
779 T1F = VFMA(LDK(KP1_937166322), Ty, VMUL(LDK(KP248689887), Tv)); | |
780 T1G = VFMA(LDK(KP1_071653589), TS, VMUL(LDK(KP844327925), TP)); | |
781 T1H = VADD(T1F, T1G); | |
782 T1I = VFMA(LDK(KP1_752613360), T1d, VMUL(LDK(KP481753674), T1a)); | |
783 T1J = VFMA(LDK(KP1_457937254), T1x, VMUL(LDK(KP684547105), T1u)); | |
784 T1K = VADD(T1I, T1J); | |
785 T1L = VADD(T1H, T1K); | |
786 T1U = VSUB(T1J, T1I); | |
787 T1Q = VMUL(LDK(KP559016994), VSUB(T1K, T1H)); | |
788 T1S = VSUB(T1G, T1F); | |
789 } | |
790 { | |
791 V Tz, TT, TU, T1e, T1y, T1z; | |
792 Tz = VFNMS(LDK(KP497379774), Ty, VMUL(LDK(KP968583161), Tv)); | |
793 TT = VFNMS(LDK(KP1_688655851), TS, VMUL(LDK(KP535826794), TP)); | |
794 TU = VADD(Tz, TT); | |
795 T1e = VFNMS(LDK(KP963507348), T1d, VMUL(LDK(KP876306680), T1a)); | |
796 T1y = VFNMS(LDK(KP1_369094211), T1x, VMUL(LDK(KP728968627), T1u)); | |
797 T1z = VADD(T1e, T1y); | |
798 T1A = VADD(TU, T1z); | |
799 T1V = VMUL(LDK(KP559016994), VSUB(TU, T1z)); | |
800 T1N = VSUB(TT, Tz); | |
801 T1O = VSUB(T1e, T1y); | |
802 } | |
803 { | |
804 V T1B, T1M, T25, T2c; | |
805 T1B = VADD(Tf, T1A); | |
806 T1M = VBYI(VADD(T1E, T1L)); | |
807 ST(&(x[WS(rs, 1)]), VSUB(T1B, T1M), ms, &(x[WS(rs, 1)])); | |
808 ST(&(x[WS(rs, 24)]), VADD(T1B, T1M), ms, &(x[0])); | |
809 T25 = VADD(Tf, VADD(T21, T24)); | |
810 T2c = VBYI(VADD(T1E, VSUB(T28, T2b))); | |
811 ST(&(x[WS(rs, 21)]), VSUB(T25, T2c), ms, &(x[WS(rs, 1)])); | |
812 ST(&(x[WS(rs, 4)]), VADD(T25, T2c), ms, &(x[0])); | |
813 } | |
814 T2d = VBYI(VADD(T1E, VFMA(LDK(KP309016994), T28, VFMA(LDK(KP587785252), VSUB(T23, T22), VFNMS(LDK(KP951056516), VADD(T1Z, T20), VMUL(LDK(KP809016994), T2b)))))); | |
815 T2e = VFMA(LDK(KP309016994), T21, VFMA(LDK(KP951056516), VSUB(T26, T27), VFMA(LDK(KP587785252), VSUB(T2a, T29), VFNMS(LDK(KP809016994), T24, Tf)))); | |
816 ST(&(x[WS(rs, 9)]), VADD(T2d, T2e), ms, &(x[WS(rs, 1)])); | |
817 ST(&(x[WS(rs, 16)]), VSUB(T2e, T2d), ms, &(x[0])); | |
818 { | |
819 V T1R, T1X, T1W, T1Y, T1P, T1T; | |
820 T1P = VFMS(LDK(KP250000000), T1L, T1E); | |
821 T1R = VBYI(VADD(VFMA(LDK(KP587785252), T1N, VMUL(LDK(KP951056516), T1O)), VSUB(T1P, T1Q))); | |
822 T1X = VBYI(VADD(VFNMS(LDK(KP587785252), T1O, VMUL(LDK(KP951056516), T1N)), VADD(T1P, T1Q))); | |
823 T1T = VFNMS(LDK(KP250000000), T1A, Tf); | |
824 T1W = VFMA(LDK(KP587785252), T1S, VFNMS(LDK(KP951056516), T1U, VSUB(T1T, T1V))); | |
825 T1Y = VFMA(LDK(KP951056516), T1S, VADD(T1V, VFMA(LDK(KP587785252), T1U, T1T))); | |
826 ST(&(x[WS(rs, 11)]), VADD(T1R, T1W), ms, &(x[WS(rs, 1)])); | |
827 ST(&(x[WS(rs, 19)]), VSUB(T1Y, T1X), ms, &(x[WS(rs, 1)])); | |
828 ST(&(x[WS(rs, 14)]), VSUB(T1W, T1R), ms, &(x[0])); | |
829 ST(&(x[WS(rs, 6)]), VADD(T1X, T1Y), ms, &(x[0])); | |
830 } | |
831 } | |
832 { | |
833 V T2u, T2w, T2h, T2k, T2l, T2A, T2B, T2C, T2o, T2r, T2s, T2x, T2y, T2z, T2M; | |
834 V T2X, T2N, T2W, T2R, T31, T2U, T30, T2E, T2F; | |
835 T2u = VFNMS(LDK(KP587785252), T1C, T2t); | |
836 T2w = VSUB(Te, Tb); | |
837 T2h = VFNMS(LDK(KP125333233), T2g, VMUL(LDK(KP1_984229402), T2f)); | |
838 T2k = VFMA(LDK(KP1_457937254), T2i, VMUL(LDK(KP684547105), T2j)); | |
839 T2l = VSUB(T2h, T2k); | |
840 T2A = VFNMS(LDK(KP1_996053456), T2p, VMUL(LDK(KP062790519), T2q)); | |
841 T2B = VFMA(LDK(KP1_541026485), T2m, VMUL(LDK(KP637423989), T2n)); | |
842 T2C = VSUB(T2A, T2B); | |
843 T2o = VFNMS(LDK(KP770513242), T2n, VMUL(LDK(KP1_274847979), T2m)); | |
844 T2r = VFMA(LDK(KP125581039), T2p, VMUL(LDK(KP998026728), T2q)); | |
845 T2s = VSUB(T2o, T2r); | |
846 T2x = VFNMS(LDK(KP1_369094211), T2i, VMUL(LDK(KP728968627), T2j)); | |
847 T2y = VFMA(LDK(KP250666467), T2f, VMUL(LDK(KP992114701), T2g)); | |
848 T2z = VSUB(T2x, T2y); | |
849 { | |
850 V T2G, T2H, T2I, T2J, T2K, T2L; | |
851 T2G = VFNMS(LDK(KP481753674), T2j, VMUL(LDK(KP1_752613360), T2i)); | |
852 T2H = VFMA(LDK(KP851558583), T2f, VMUL(LDK(KP904827052), T2g)); | |
853 T2I = VSUB(T2G, T2H); | |
854 T2J = VFNMS(LDK(KP844327925), T2q, VMUL(LDK(KP1_071653589), T2p)); | |
855 T2K = VFNMS(LDK(KP998026728), T2n, VMUL(LDK(KP125581039), T2m)); | |
856 T2L = VADD(T2J, T2K); | |
857 T2M = VMUL(LDK(KP559016994), VSUB(T2I, T2L)); | |
858 T2X = VSUB(T2J, T2K); | |
859 T2N = VADD(T2I, T2L); | |
860 T2W = VADD(T2G, T2H); | |
861 } | |
862 { | |
863 V T2P, T2Q, T2Y, T2S, T2T, T2Z; | |
864 T2P = VFNMS(LDK(KP425779291), T2g, VMUL(LDK(KP1_809654104), T2f)); | |
865 T2Q = VFMA(LDK(KP963507348), T2i, VMUL(LDK(KP876306680), T2j)); | |
866 T2Y = VADD(T2Q, T2P); | |
867 T2S = VFMA(LDK(KP1_688655851), T2p, VMUL(LDK(KP535826794), T2q)); | |
868 T2T = VFMA(LDK(KP1_996053456), T2m, VMUL(LDK(KP062790519), T2n)); | |
869 T2Z = VADD(T2S, T2T); | |
870 T2R = VSUB(T2P, T2Q); | |
871 T31 = VADD(T2Y, T2Z); | |
872 T2U = VSUB(T2S, T2T); | |
873 T30 = VMUL(LDK(KP559016994), VSUB(T2Y, T2Z)); | |
874 } | |
875 { | |
876 V T36, T37, T2v, T2D; | |
877 T36 = VBYI(VADD(T2u, T2N)); | |
878 T37 = VADD(T2w, T31); | |
879 ST(&(x[WS(rs, 2)]), VADD(T36, T37), ms, &(x[0])); | |
880 ST(&(x[WS(rs, 23)]), VSUB(T37, T36), ms, &(x[WS(rs, 1)])); | |
881 T2v = VBYI(VSUB(VADD(T2l, T2s), T2u)); | |
882 T2D = VADD(T2w, VADD(T2z, T2C)); | |
883 ST(&(x[WS(rs, 3)]), VADD(T2v, T2D), ms, &(x[WS(rs, 1)])); | |
884 ST(&(x[WS(rs, 22)]), VSUB(T2D, T2v), ms, &(x[0])); | |
885 } | |
886 T2E = VFMA(LDK(KP309016994), T2z, VFNMS(LDK(KP809016994), T2C, VFNMS(LDK(KP587785252), VADD(T2r, T2o), VFNMS(LDK(KP951056516), VADD(T2k, T2h), T2w)))); | |
887 T2F = VBYI(VSUB(VFNMS(LDK(KP587785252), VADD(T2A, T2B), VFNMS(LDK(KP809016994), T2s, VFNMS(LDK(KP951056516), VADD(T2x, T2y), VMUL(LDK(KP309016994), T2l)))), T2u)); | |
888 ST(&(x[WS(rs, 17)]), VSUB(T2E, T2F), ms, &(x[WS(rs, 1)])); | |
889 ST(&(x[WS(rs, 8)]), VADD(T2E, T2F), ms, &(x[0])); | |
890 { | |
891 V T2V, T34, T33, T35, T2O, T32; | |
892 T2O = VFNMS(LDK(KP250000000), T2N, T2u); | |
893 T2V = VBYI(VADD(T2M, VADD(T2O, VFNMS(LDK(KP587785252), T2U, VMUL(LDK(KP951056516), T2R))))); | |
894 T34 = VBYI(VADD(T2O, VSUB(VFMA(LDK(KP587785252), T2R, VMUL(LDK(KP951056516), T2U)), T2M))); | |
895 T32 = VFNMS(LDK(KP250000000), T31, T2w); | |
896 T33 = VFMA(LDK(KP951056516), T2W, VFMA(LDK(KP587785252), T2X, VADD(T30, T32))); | |
897 T35 = VFMA(LDK(KP587785252), T2W, VSUB(VFNMS(LDK(KP951056516), T2X, T32), T30)); | |
898 ST(&(x[WS(rs, 7)]), VADD(T2V, T33), ms, &(x[WS(rs, 1)])); | |
899 ST(&(x[WS(rs, 13)]), VSUB(T35, T34), ms, &(x[WS(rs, 1)])); | |
900 ST(&(x[WS(rs, 18)]), VSUB(T33, T2V), ms, &(x[0])); | |
901 ST(&(x[WS(rs, 12)]), VADD(T34, T35), ms, &(x[0])); | |
902 } | |
903 } | |
904 } | |
905 } | |
906 VLEAVE(); | |
907 } | |
908 | |
909 static const tw_instr twinstr[] = { | |
910 VTW(0, 1), | |
911 VTW(0, 2), | |
912 VTW(0, 3), | |
913 VTW(0, 4), | |
914 VTW(0, 5), | |
915 VTW(0, 6), | |
916 VTW(0, 7), | |
917 VTW(0, 8), | |
918 VTW(0, 9), | |
919 VTW(0, 10), | |
920 VTW(0, 11), | |
921 VTW(0, 12), | |
922 VTW(0, 13), | |
923 VTW(0, 14), | |
924 VTW(0, 15), | |
925 VTW(0, 16), | |
926 VTW(0, 17), | |
927 VTW(0, 18), | |
928 VTW(0, 19), | |
929 VTW(0, 20), | |
930 VTW(0, 21), | |
931 VTW(0, 22), | |
932 VTW(0, 23), | |
933 VTW(0, 24), | |
934 {TW_NEXT, VL, 0} | |
935 }; | |
936 | |
937 static const ct_desc desc = { 25, XSIMD_STRING("t1fv_25"), twinstr, &GENUS, {170, 110, 78, 0}, 0, 0, 0 }; | |
938 | |
939 void XSIMD(codelet_t1fv_25) (planner *p) { | |
940 X(kdft_dit_register) (p, t1fv_25, &desc); | |
941 } | |
942 #endif |