Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/simd/common/t3fv_25.c @ 10:37bf6b4a2645
Add FFTW3
author | Chris Cannam |
---|---|
date | Wed, 20 Mar 2013 15:35:50 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
9:c0fb53affa76 | 10:37bf6b4a2645 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-11 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Sun Nov 25 07:38:56 EST 2012 */ | |
23 | |
24 #include "codelet-dft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_twiddle_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 25 -name t3fv_25 -include 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 * 223 stack variables, 67 constants, and 50 memory accesses | |
34 */ | |
35 #include "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(KP792626838, +0.792626838241819413632131824093538848057784557); | |
40 DVK(KP876091699, +0.876091699473550838204498029706869638173524346); | |
41 DVK(KP617882369, +0.617882369114440893914546919006756321695042882); | |
42 DVK(KP803003575, +0.803003575438660414833440593570376004635464850); | |
43 DVK(KP242145790, +0.242145790282157779872542093866183953459003101); | |
44 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
45 DVK(KP999544308, +0.999544308746292983948881682379742149196758193); | |
46 DVK(KP916574801, +0.916574801383451584742370439148878693530976769); | |
47 DVK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
48 DVK(KP809385824, +0.809385824416008241660603814668679683846476688); | |
49 DVK(KP447417479, +0.447417479732227551498980015410057305749330693); | |
50 DVK(KP894834959, +0.894834959464455102997960030820114611498661386); | |
51 DVK(KP867381224, +0.867381224396525206773171885031575671309956167); | |
52 DVK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
53 DVK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
54 DVK(KP958953096, +0.958953096729998668045963838399037225970891871); | |
55 DVK(KP831864738, +0.831864738706457140726048799369896829771167132); | |
56 DVK(KP829049696, +0.829049696159252993975487806364305442437946767); | |
57 DVK(KP860541664, +0.860541664367944677098261680920518816412804187); | |
58 DVK(KP897376177, +0.897376177523557693138608077137219684419427330); | |
59 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
60 DVK(KP681693190, +0.681693190061530575150324149145440022633095390); | |
61 DVK(KP560319534, +0.560319534973832390111614715371676131169633784); | |
62 DVK(KP855719849, +0.855719849902058969314654733608091555096772472); | |
63 DVK(KP237294955, +0.237294955877110315393888866460840817927895961); | |
64 DVK(KP949179823, +0.949179823508441261575555465843363271711583843); | |
65 DVK(KP904508497, +0.904508497187473712051146708591409529430077295); | |
66 DVK(KP997675361, +0.997675361079556513670859573984492383596555031); | |
67 DVK(KP763932022, +0.763932022500210303590826331268723764559381640); | |
68 DVK(KP690983005, +0.690983005625052575897706582817180941139845410); | |
69 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
70 DVK(KP952936919, +0.952936919628306576880750665357914584765951388); | |
71 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
72 DVK(KP262346850, +0.262346850930607871785420028382979691334784273); | |
73 DVK(KP570584518, +0.570584518783621657366766175430996792655723863); | |
74 DVK(KP669429328, +0.669429328479476605641803240971985825917022098); | |
75 DVK(KP923225144, +0.923225144846402650453449441572664695995209956); | |
76 DVK(KP945422727, +0.945422727388575946270360266328811958657216298); | |
77 DVK(KP522616830, +0.522616830205754336872861364785224694908468440); | |
78 DVK(KP956723877, +0.956723877038460305821989399535483155872969262); | |
79 DVK(KP906616052, +0.906616052148196230441134447086066874408359177); | |
80 DVK(KP772036680, +0.772036680810363904029489473607579825330539880); | |
81 DVK(KP845997307, +0.845997307939530944175097360758058292389769300); | |
82 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
83 DVK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
84 DVK(KP921078979, +0.921078979742360627699756128143719920817673854); | |
85 DVK(KP982009705, +0.982009705009746369461829878184175962711969869); | |
86 DVK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
87 DVK(KP494780565, +0.494780565770515410344588413655324772219443730); | |
88 DVK(KP447533225, +0.447533225982656890041886979663652563063114397); | |
89 DVK(KP603558818, +0.603558818296015001454675132653458027918768137); | |
90 DVK(KP667278218, +0.667278218140296670899089292254759909713898805); | |
91 DVK(KP244189809, +0.244189809627953270309879511234821255780225091); | |
92 DVK(KP269969613, +0.269969613759572083574752974412347470060951301); | |
93 DVK(KP578046249, +0.578046249379945007321754579646815604023525655); | |
94 DVK(KP522847744, +0.522847744331509716623755382187077770911012542); | |
95 DVK(KP132830569, +0.132830569247582714407653942074819768844536507); | |
96 DVK(KP120146378, +0.120146378570687701782758537356596213647956445); | |
97 DVK(KP893101515, +0.893101515366181661711202267938416198338079437); | |
98 DVK(KP987388751, +0.987388751065621252324603216482382109400433949); | |
99 DVK(KP059835404, +0.059835404262124915169548397419498386427871950); | |
100 DVK(KP066152395, +0.066152395967733048213034281011006031460903353); | |
101 DVK(KP786782374, +0.786782374965295178365099601674911834788448471); | |
102 DVK(KP869845200, +0.869845200362138853122720822420327157933056305); | |
103 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
104 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
105 DVK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
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 T2t, T1Z, T2W, T28, T2Q, T2r, T2g, T2u, T2o, T2l; | |
112 { | |
113 V T2, T5, T3, T9; | |
114 T2 = LDW(&(W[0])); | |
115 T5 = LDW(&(W[TWVL * 4])); | |
116 T3 = LDW(&(W[TWVL * 2])); | |
117 T9 = LDW(&(W[TWVL * 6])); | |
118 { | |
119 V T2c, T3l, Tn, T49, Tm, T4e, TN, T32, T1d, T3a, T3f, T3z, T3H, T25, T1W; | |
120 V T2v, T2D, T4a, T1g, T18, T2Z, T11, T31, TK, T1q, T1j, T1n, T4b, T17; | |
121 { | |
122 V T1, T1l, Tr, T4, Ty, T1E, Tu, TX, TD, T1h, Tz, T1e, T1I, T1o, TU; | |
123 V Tk, T2b, T1B, T1D, T1N, T1F, Td, T2a, T1J; | |
124 { | |
125 V T7, Tb, TC, Tg, T1L, Ta, T6, Tj, T1A; | |
126 T1 = LD(&(x[0]), ms, &(x[0])); | |
127 { | |
128 V Tf, Ti, Te, Th; | |
129 Tf = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
130 Ti = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
131 T7 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
132 Tb = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
133 Te = VZMUL(T2, T5); | |
134 TC = VZMULJ(T2, T5); | |
135 T1l = VZMUL(T3, T5); | |
136 Tr = VZMULJ(T3, T5); | |
137 T4 = VZMUL(T2, T3); | |
138 Ty = VZMULJ(T2, T3); | |
139 T1E = VZMULJ(T2, T9); | |
140 Th = VZMULJ(T5, T9); | |
141 Tu = VZMULJ(T3, T9); | |
142 Tg = VZMULJ(Te, Tf); | |
143 TX = VZMULJ(Te, T9); | |
144 TD = VZMULJ(TC, T9); | |
145 T1h = VZMULJ(Ty, T9); | |
146 Tz = VZMUL(Ty, T5); | |
147 T1e = VZMULJ(Ty, T5); | |
148 T1L = VZMULJ(Tr, T9); | |
149 Ta = VZMULJ(T4, T9); | |
150 T1I = VZMUL(T4, T5); | |
151 T6 = VZMULJ(T4, T5); | |
152 Tj = VZMULJ(Th, Ti); | |
153 } | |
154 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
155 T1o = VZMULJ(T1e, T9); | |
156 { | |
157 V Tc, T8, T1C, T1M; | |
158 T1C = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
159 T1M = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
160 Tc = VZMULJ(Ta, Tb); | |
161 T8 = VZMULJ(T6, T7); | |
162 TU = VZMULJ(T6, T9); | |
163 Tk = VADD(Tg, Tj); | |
164 T2b = VSUB(Tg, Tj); | |
165 T1B = VZMULJ(T3, T1A); | |
166 T1D = VZMULJ(TC, T1C); | |
167 T1N = VZMULJ(T1L, T1M); | |
168 T1F = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
169 Td = VADD(T8, Tc); | |
170 T2a = VSUB(T8, Tc); | |
171 T1J = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
172 } | |
173 } | |
174 { | |
175 V Tq, Tt, TF, T1T, T1H, Tw, T1U, T1O, TA, Tp, Ts, TE; | |
176 Tp = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
177 Ts = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
178 TE = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
179 { | |
180 V T1K, Tv, T1G, Tl; | |
181 Tv = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
182 T1G = VZMULJ(T1E, T1F); | |
183 T2c = VFMA(LDK(KP618033988), T2b, T2a); | |
184 T3l = VFNMS(LDK(KP618033988), T2a, T2b); | |
185 Tn = VSUB(Td, Tk); | |
186 Tl = VADD(Td, Tk); | |
187 T1K = VZMULJ(T1I, T1J); | |
188 Tq = VZMULJ(T2, Tp); | |
189 Tt = VZMULJ(Tr, Ts); | |
190 TF = VZMULJ(TD, TE); | |
191 T1T = VSUB(T1D, T1G); | |
192 T1H = VADD(T1D, T1G); | |
193 T49 = VADD(T1, Tl); | |
194 Tm = VFNMS(LDK(KP250000000), Tl, T1); | |
195 Tw = VZMULJ(Tu, Tv); | |
196 T1U = VSUB(T1K, T1N); | |
197 T1O = VADD(T1K, T1N); | |
198 TA = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
199 } | |
200 { | |
201 V Tx, TL, T1R, T38, T1V, T13, TQ, TZ, TS, T1Q, TV, TG, TM, T12, T1c; | |
202 V T16; | |
203 T12 = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
204 { | |
205 V TP, TY, T1P, TB, TR; | |
206 TP = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
207 TY = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
208 TR = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
209 Tx = VADD(Tt, Tw); | |
210 TL = VSUB(Tt, Tw); | |
211 T1R = VSUB(T1O, T1H); | |
212 T1P = VADD(T1H, T1O); | |
213 T38 = VFNMS(LDK(KP618033988), T1T, T1U); | |
214 T1V = VFMA(LDK(KP618033988), T1U, T1T); | |
215 TB = VZMULJ(Tz, TA); | |
216 T13 = VZMULJ(T4, T12); | |
217 TQ = VZMULJ(T9, TP); | |
218 TZ = VZMULJ(TX, TY); | |
219 TS = VZMULJ(T5, TR); | |
220 T4e = VADD(T1B, T1P); | |
221 T1Q = VFNMS(LDK(KP250000000), T1P, T1B); | |
222 TV = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
223 TG = VADD(TB, TF); | |
224 TM = VSUB(TF, TB); | |
225 } | |
226 T1c = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
227 { | |
228 V T14, TT, TJ, T15, T10, TI, T1p, T1f, T1i, T1m; | |
229 T1f = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
230 T14 = VADD(TS, TQ); | |
231 TT = VSUB(TQ, TS); | |
232 { | |
233 V T39, T1S, TW, TH; | |
234 T39 = VFMA(LDK(KP559016994), T1R, T1Q); | |
235 T1S = VFNMS(LDK(KP559016994), T1R, T1Q); | |
236 TW = VZMULJ(TU, TV); | |
237 TH = VADD(Tx, TG); | |
238 TJ = VSUB(Tx, TG); | |
239 TN = VFNMS(LDK(KP618033988), TM, TL); | |
240 T32 = VFMA(LDK(KP618033988), TL, TM); | |
241 T1d = VZMULJ(Ty, T1c); | |
242 T3a = VFMA(LDK(KP869845200), T39, T38); | |
243 T3f = VFNMS(LDK(KP786782374), T38, T39); | |
244 T3z = VFMA(LDK(KP066152395), T39, T38); | |
245 T3H = VFNMS(LDK(KP059835404), T38, T39); | |
246 T25 = VFMA(LDK(KP987388751), T1S, T1V); | |
247 T1W = VFNMS(LDK(KP893101515), T1V, T1S); | |
248 T2v = VFNMS(LDK(KP120146378), T1V, T1S); | |
249 T2D = VFMA(LDK(KP132830569), T1S, T1V); | |
250 T15 = VADD(TZ, TW); | |
251 T10 = VSUB(TW, TZ); | |
252 TI = VFNMS(LDK(KP250000000), TH, Tq); | |
253 T4a = VADD(Tq, TH); | |
254 T1g = VZMULJ(T1e, T1f); | |
255 } | |
256 T1p = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
257 T1i = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
258 T1m = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
259 T18 = VSUB(T14, T15); | |
260 T16 = VADD(T14, T15); | |
261 T2Z = VFNMS(LDK(KP618033988), TT, T10); | |
262 T11 = VFMA(LDK(KP618033988), T10, TT); | |
263 T31 = VFNMS(LDK(KP559016994), TJ, TI); | |
264 TK = VFMA(LDK(KP559016994), TJ, TI); | |
265 T1q = VZMULJ(T1o, T1p); | |
266 T1j = VZMULJ(T1h, T1i); | |
267 T1n = VZMULJ(T1l, T1m); | |
268 } | |
269 T4b = VADD(T13, T16); | |
270 T17 = VFMS(LDK(KP250000000), T16, T13); | |
271 } | |
272 } | |
273 } | |
274 { | |
275 V T33, T3i, T3C, T3L, T20, TO, T2y, T2G, T1k, T1w, T1r, T1x, T2Y, T19, T4k; | |
276 V T4c; | |
277 T33 = VFMA(LDK(KP893101515), T32, T31); | |
278 T3i = VFNMS(LDK(KP987388751), T31, T32); | |
279 T3C = VFNMS(LDK(KP522847744), T32, T31); | |
280 T3L = VFMA(LDK(KP578046249), T31, T32); | |
281 T20 = VFMA(LDK(KP269969613), TK, TN); | |
282 TO = VFNMS(LDK(KP244189809), TN, TK); | |
283 T2y = VFMA(LDK(KP667278218), TK, TN); | |
284 T2G = VFNMS(LDK(KP603558818), TN, TK); | |
285 T1k = VADD(T1g, T1j); | |
286 T1w = VSUB(T1g, T1j); | |
287 T1r = VADD(T1n, T1q); | |
288 T1x = VSUB(T1q, T1n); | |
289 T2Y = VFMA(LDK(KP559016994), T18, T17); | |
290 T19 = VFNMS(LDK(KP559016994), T18, T17); | |
291 T4k = VSUB(T4a, T4b); | |
292 T4c = VADD(T4a, T4b); | |
293 { | |
294 V T2X, To, T35, T1y, T2H, T2z, T1a, T21, T3t, T34, T3n, T3j, T3E, T3Y, T3M; | |
295 V T3R, T1v, T36, T4l, T4f, T1u, T1s; | |
296 T2X = VFNMS(LDK(KP559016994), Tn, Tm); | |
297 To = VFMA(LDK(KP559016994), Tn, Tm); | |
298 T1u = VSUB(T1r, T1k); | |
299 T1s = VADD(T1k, T1r); | |
300 T35 = VFMA(LDK(KP618033988), T1w, T1x); | |
301 T1y = VFNMS(LDK(KP618033988), T1x, T1w); | |
302 { | |
303 V T3K, T30, T3h, T3D, T4d, T1t; | |
304 T3K = VFMA(LDK(KP447533225), T2Z, T2Y); | |
305 T30 = VFMA(LDK(KP120146378), T2Z, T2Y); | |
306 T3h = VFNMS(LDK(KP132830569), T2Y, T2Z); | |
307 T3D = VFNMS(LDK(KP494780565), T2Y, T2Z); | |
308 T2H = VFNMS(LDK(KP786782374), T11, T19); | |
309 T2z = VFMA(LDK(KP869845200), T19, T11); | |
310 T1a = VFNMS(LDK(KP667278218), T19, T11); | |
311 T21 = VFMA(LDK(KP603558818), T11, T19); | |
312 T4d = VADD(T1d, T1s); | |
313 T1t = VFNMS(LDK(KP250000000), T1s, T1d); | |
314 T3t = VFNMS(LDK(KP734762448), T33, T30); | |
315 T34 = VFMA(LDK(KP734762448), T33, T30); | |
316 T3n = VFMA(LDK(KP734762448), T3i, T3h); | |
317 T3j = VFNMS(LDK(KP734762448), T3i, T3h); | |
318 T3E = VFNMS(LDK(KP982009705), T3D, T3C); | |
319 T3Y = VFMA(LDK(KP982009705), T3D, T3C); | |
320 T3M = VFNMS(LDK(KP921078979), T3L, T3K); | |
321 T3R = VFMA(LDK(KP921078979), T3L, T3K); | |
322 T1v = VFNMS(LDK(KP559016994), T1u, T1t); | |
323 T36 = VFMA(LDK(KP559016994), T1u, T1t); | |
324 T4l = VSUB(T4d, T4e); | |
325 T4f = VADD(T4d, T4e); | |
326 } | |
327 { | |
328 V T2L, T2R, T2j, T2q, T2J, T2B, T2e, T26, T2U, T1Y, T23, T2O; | |
329 { | |
330 V T2I, T24, T2w, T2E, T48, T42, T3y, T3s, T3V, T45, T2A, T1b, T2h, T2i, T1X; | |
331 T2L = VFNMS(LDK(KP912575812), T2H, T2G); | |
332 T2I = VFMA(LDK(KP912575812), T2H, T2G); | |
333 { | |
334 V T3A, T3e, T37, T3I, T1z; | |
335 T3A = VFNMS(LDK(KP667278218), T36, T35); | |
336 T3e = VFNMS(LDK(KP059835404), T35, T36); | |
337 T37 = VFMA(LDK(KP066152395), T36, T35); | |
338 T3I = VFMA(LDK(KP603558818), T35, T36); | |
339 T24 = VFMA(LDK(KP578046249), T1v, T1y); | |
340 T1z = VFNMS(LDK(KP522847744), T1y, T1v); | |
341 T2w = VFNMS(LDK(KP494780565), T1v, T1y); | |
342 T2E = VFMA(LDK(KP447533225), T1y, T1v); | |
343 { | |
344 V T4i, T4g, T4o, T4m; | |
345 T4i = VSUB(T4c, T4f); | |
346 T4g = VADD(T4c, T4f); | |
347 T4o = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T4k, T4l)); | |
348 T4m = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T4l, T4k)); | |
349 { | |
350 V T3Q, T3J, T3b, T3u; | |
351 T3Q = VFNMS(LDK(KP845997307), T3I, T3H); | |
352 T3J = VFMA(LDK(KP845997307), T3I, T3H); | |
353 T3b = VFNMS(LDK(KP772036680), T3a, T37); | |
354 T3u = VFMA(LDK(KP772036680), T3a, T37); | |
355 { | |
356 V T3o, T3g, T3B, T3X, T4h; | |
357 T3o = VFNMS(LDK(KP772036680), T3f, T3e); | |
358 T3g = VFMA(LDK(KP772036680), T3f, T3e); | |
359 T3B = VFNMS(LDK(KP845997307), T3A, T3z); | |
360 T3X = VFMA(LDK(KP845997307), T3A, T3z); | |
361 ST(&(x[0]), VADD(T4g, T49), ms, &(x[0])); | |
362 T4h = VFNMS(LDK(KP250000000), T4g, T49); | |
363 { | |
364 V T40, T3N, T3c, T3v; | |
365 T40 = VFMA(LDK(KP906616052), T3M, T3J); | |
366 T3N = VFNMS(LDK(KP906616052), T3M, T3J); | |
367 T3c = VFMA(LDK(KP956723877), T3b, T34); | |
368 T3v = VFMA(LDK(KP522616830), T3j, T3u); | |
369 { | |
370 V T3p, T3k, T3S, T3F; | |
371 T3p = VFNMS(LDK(KP522616830), T34, T3o); | |
372 T3k = VFMA(LDK(KP945422727), T3j, T3g); | |
373 T3S = VFNMS(LDK(KP923225144), T3E, T3B); | |
374 T3F = VFMA(LDK(KP923225144), T3E, T3B); | |
375 { | |
376 V T46, T3Z, T4j, T4n; | |
377 T46 = VFNMS(LDK(KP669429328), T3X, T3Y); | |
378 T3Z = VFMA(LDK(KP570584518), T3Y, T3X); | |
379 T4j = VFMA(LDK(KP559016994), T4i, T4h); | |
380 T4n = VFNMS(LDK(KP559016994), T4i, T4h); | |
381 { | |
382 V T3W, T3O, T3d, T3w; | |
383 T3W = VFMA(LDK(KP262346850), T3N, T3l); | |
384 T3O = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T3l, T3N)); | |
385 T3d = VFMA(LDK(KP992114701), T3c, T2X); | |
386 T3w = VFNMS(LDK(KP690983005), T3v, T3g); | |
387 { | |
388 V T3q, T3m, T3T, T43; | |
389 T3q = VFMA(LDK(KP763932022), T3p, T3b); | |
390 T3m = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T3l, T3k)); | |
391 T3T = VFNMS(LDK(KP997675361), T3S, T3R); | |
392 T43 = VFNMS(LDK(KP904508497), T3S, T3Q); | |
393 { | |
394 V T3G, T3P, T47, T41; | |
395 T3G = VFMA(LDK(KP949179823), T3F, T2X); | |
396 T3P = VFNMS(LDK(KP237294955), T3F, T2X); | |
397 T47 = VFNMS(LDK(KP669429328), T40, T46); | |
398 T41 = VFMA(LDK(KP618033988), T40, T3Z); | |
399 ST(&(x[WS(rs, 20)]), VFMAI(T4m, T4j), ms, &(x[0])); | |
400 ST(&(x[WS(rs, 5)]), VFNMSI(T4m, T4j), ms, &(x[WS(rs, 1)])); | |
401 ST(&(x[WS(rs, 15)]), VFNMSI(T4o, T4n), ms, &(x[WS(rs, 1)])); | |
402 ST(&(x[WS(rs, 10)]), VFMAI(T4o, T4n), ms, &(x[0])); | |
403 { | |
404 V T3x, T3r, T3U, T44; | |
405 T3x = VFMA(LDK(KP855719849), T3w, T3t); | |
406 T3r = VFNMS(LDK(KP855719849), T3q, T3n); | |
407 ST(&(x[WS(rs, 22)]), VFMAI(T3m, T3d), ms, &(x[0])); | |
408 ST(&(x[WS(rs, 3)]), VFNMSI(T3m, T3d), ms, &(x[WS(rs, 1)])); | |
409 T3U = VFMA(LDK(KP560319534), T3T, T3Q); | |
410 T44 = VFNMS(LDK(KP681693190), T43, T3R); | |
411 ST(&(x[WS(rs, 23)]), VFMAI(T3O, T3G), ms, &(x[WS(rs, 1)])); | |
412 ST(&(x[WS(rs, 2)]), VFNMSI(T3O, T3G), ms, &(x[0])); | |
413 T48 = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T47, T3W)); | |
414 T42 = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T41, T3W)); | |
415 T3y = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T3x, T3l)); | |
416 T3s = VFMA(LDK(KP897376177), T3r, T2X); | |
417 T3V = VFNMS(LDK(KP949179823), T3U, T3P); | |
418 T45 = VFNMS(LDK(KP860541664), T44, T3P); | |
419 T2R = VFNMS(LDK(KP912575812), T2z, T2y); | |
420 T2A = VFMA(LDK(KP912575812), T2z, T2y); | |
421 T1b = VFMA(LDK(KP829049696), T1a, TO); | |
422 T2h = VFNMS(LDK(KP829049696), T1a, TO); | |
423 T2i = VFNMS(LDK(KP831864738), T1W, T1z); | |
424 T1X = VFMA(LDK(KP831864738), T1W, T1z); | |
425 } | |
426 } | |
427 } | |
428 } | |
429 } | |
430 } | |
431 } | |
432 } | |
433 } | |
434 } | |
435 } | |
436 { | |
437 V T2M, T2F, T2x, T2S, T2T, T2N; | |
438 T2M = VFNMS(LDK(KP958953096), T2E, T2D); | |
439 T2F = VFMA(LDK(KP958953096), T2E, T2D); | |
440 ST(&(x[WS(rs, 17)]), VFMAI(T3y, T3s), ms, &(x[WS(rs, 1)])); | |
441 ST(&(x[WS(rs, 8)]), VFNMSI(T3y, T3s), ms, &(x[0])); | |
442 ST(&(x[WS(rs, 12)]), VFMAI(T42, T3V), ms, &(x[0])); | |
443 ST(&(x[WS(rs, 13)]), VFNMSI(T42, T3V), ms, &(x[WS(rs, 1)])); | |
444 ST(&(x[WS(rs, 18)]), VFNMSI(T48, T45), ms, &(x[0])); | |
445 ST(&(x[WS(rs, 7)]), VFMAI(T48, T45), ms, &(x[WS(rs, 1)])); | |
446 T2j = VFMA(LDK(KP559154169), T2i, T2h); | |
447 T2q = VFNMS(LDK(KP683113946), T2h, T2i); | |
448 T2x = VFNMS(LDK(KP867381224), T2w, T2v); | |
449 T2S = VFMA(LDK(KP867381224), T2w, T2v); | |
450 T2J = VFMA(LDK(KP894834959), T2I, T2F); | |
451 T2T = VFMA(LDK(KP447417479), T2I, T2S); | |
452 T2B = VFNMS(LDK(KP809385824), T2A, T2x); | |
453 T2N = VFMA(LDK(KP447417479), T2A, T2M); | |
454 T2e = VFMA(LDK(KP831864738), T25, T24); | |
455 T26 = VFNMS(LDK(KP831864738), T25, T24); | |
456 T2U = VFNMS(LDK(KP763932022), T2T, T2F); | |
457 T1Y = VFMA(LDK(KP904730450), T1X, T1b); | |
458 T23 = VFNMS(LDK(KP904730450), T1X, T1b); | |
459 T2O = VFMA(LDK(KP690983005), T2N, T2x); | |
460 } | |
461 } | |
462 { | |
463 V T2C, T22, T2d, T2K; | |
464 T2C = VFNMS(LDK(KP992114701), T2B, To); | |
465 T22 = VFMA(LDK(KP916574801), T21, T20); | |
466 T2d = VFNMS(LDK(KP916574801), T21, T20); | |
467 T2K = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2J, T2c)); | |
468 { | |
469 V T27, T2P, T2f, T2k, T2n, T2V; | |
470 T2V = VFNMS(LDK(KP999544308), T2U, T2R); | |
471 T27 = VFNMS(LDK(KP904730450), T26, T23); | |
472 T2t = VFMA(LDK(KP968583161), T1Y, To); | |
473 T1Z = VFNMS(LDK(KP242145790), T1Y, To); | |
474 T2P = VFNMS(LDK(KP999544308), T2O, T2L); | |
475 T2f = VFMA(LDK(KP904730450), T2e, T2d); | |
476 T2k = VFNMS(LDK(KP904730450), T2e, T2d); | |
477 T2n = VADD(T22, T23); | |
478 ST(&(x[WS(rs, 21)]), VFNMSI(T2K, T2C), ms, &(x[WS(rs, 1)])); | |
479 ST(&(x[WS(rs, 4)]), VFMAI(T2K, T2C), ms, &(x[0])); | |
480 T2W = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T2V, T2c)); | |
481 T28 = VFNMS(LDK(KP618033988), T27, T22); | |
482 T2Q = VFNMS(LDK(KP803003575), T2P, To); | |
483 T2r = VFMA(LDK(KP617882369), T2k, T2q); | |
484 T2g = VFNMS(LDK(KP242145790), T2f, T2c); | |
485 T2u = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T2f, T2c)); | |
486 T2o = VFNMS(LDK(KP683113946), T2n, T26); | |
487 T2l = VFMA(LDK(KP559016994), T2k, T2j); | |
488 } | |
489 } | |
490 } | |
491 } | |
492 } | |
493 } | |
494 } | |
495 { | |
496 V T29, T2s, T2p, T2m; | |
497 T29 = VFNMS(LDK(KP876091699), T28, T1Z); | |
498 ST(&(x[WS(rs, 9)]), VFMAI(T2W, T2Q), ms, &(x[WS(rs, 1)])); | |
499 ST(&(x[WS(rs, 16)]), VFNMSI(T2W, T2Q), ms, &(x[0])); | |
500 T2s = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T2r, T2g)); | |
501 ST(&(x[WS(rs, 24)]), VFMAI(T2u, T2t), ms, &(x[0])); | |
502 ST(&(x[WS(rs, 1)]), VFNMSI(T2u, T2t), ms, &(x[WS(rs, 1)])); | |
503 T2p = VFMA(LDK(KP792626838), T2o, T1Z); | |
504 T2m = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T2l, T2g)); | |
505 ST(&(x[WS(rs, 11)]), VFNMSI(T2s, T2p), ms, &(x[WS(rs, 1)])); | |
506 ST(&(x[WS(rs, 14)]), VFMAI(T2s, T2p), ms, &(x[0])); | |
507 ST(&(x[WS(rs, 19)]), VFMAI(T2m, T29), ms, &(x[WS(rs, 1)])); | |
508 ST(&(x[WS(rs, 6)]), VFNMSI(T2m, T29), ms, &(x[0])); | |
509 } | |
510 } | |
511 } | |
512 VLEAVE(); | |
513 } | |
514 | |
515 static const tw_instr twinstr[] = { | |
516 VTW(0, 1), | |
517 VTW(0, 3), | |
518 VTW(0, 9), | |
519 VTW(0, 24), | |
520 {TW_NEXT, VL, 0} | |
521 }; | |
522 | |
523 static const ct_desc desc = { 25, XSIMD_STRING("t3fv_25"), twinstr, &GENUS, {87, 100, 181, 0}, 0, 0, 0 }; | |
524 | |
525 void XSIMD(codelet_t3fv_25) (planner *p) { | |
526 X(kdft_dit_register) (p, t3fv_25, &desc); | |
527 } | |
528 #else /* HAVE_FMA */ | |
529 | |
530 /* 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 t3f.h */ | |
531 | |
532 /* | |
533 * This function contains 268 FP additions, 228 FP multiplications, | |
534 * (or, 190 additions, 150 multiplications, 78 fused multiply/add), | |
535 * 123 stack variables, 40 constants, and 50 memory accesses | |
536 */ | |
537 #include "t3f.h" | |
538 | |
539 static void t3fv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
540 { | |
541 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
542 DVK(KP125581039, +0.125581039058626752152356449131262266244969664); | |
543 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
544 DVK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
545 DVK(KP809016994, +0.809016994374947424102293417182819058860154590); | |
546 DVK(KP309016994, +0.309016994374947424102293417182819058860154590); | |
547 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465); | |
548 DVK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
549 DVK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
550 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
551 DVK(KP497379774, +0.497379774329709576484567492012895936835134813); | |
552 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
553 DVK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
554 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328); | |
555 DVK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
556 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
557 DVK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
558 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
559 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
560 DVK(KP250666467, +0.250666467128608490746237519633017587885836494); | |
561 DVK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
562 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
563 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596); | |
564 DVK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
565 DVK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
566 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691); | |
567 DVK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
568 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418); | |
569 DVK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
570 DVK(KP851558583, +0.851558583130145297725004891488503407959946084); | |
571 DVK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
572 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344); | |
573 DVK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
574 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483); | |
575 DVK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
576 DVK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
577 DVK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
578 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
579 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
580 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
581 { | |
582 INT m; | |
583 R *x; | |
584 x = ri; | |
585 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)) { | |
586 V T1, T4, T2, T3, TA, Td, Tp, Tw, Tx, T1G, T1j, T5, T1c, T8, T9; | |
587 V Ts, T1J, Tg, T1C, T1m, TX, TB, T1f, TU; | |
588 T1 = LDW(&(W[0])); | |
589 T4 = LDW(&(W[TWVL * 4])); | |
590 T2 = LDW(&(W[TWVL * 2])); | |
591 T3 = VZMUL(T1, T2); | |
592 TA = VZMULJ(T1, T4); | |
593 Td = VZMUL(T1, T4); | |
594 Tp = VZMULJ(T2, T4); | |
595 Tw = VZMULJ(T1, T2); | |
596 Tx = VZMUL(Tw, T4); | |
597 T1G = VZMUL(T3, T4); | |
598 T1j = VZMUL(T2, T4); | |
599 T5 = VZMULJ(T3, T4); | |
600 T1c = VZMULJ(Tw, T4); | |
601 T8 = LDW(&(W[TWVL * 6])); | |
602 T9 = VZMULJ(T3, T8); | |
603 Ts = VZMULJ(T2, T8); | |
604 T1J = VZMULJ(Tp, T8); | |
605 Tg = VZMULJ(T4, T8); | |
606 T1C = VZMULJ(T1, T8); | |
607 T1m = VZMULJ(T1c, T8); | |
608 TX = VZMULJ(T5, T8); | |
609 TB = VZMULJ(TA, T8); | |
610 T1f = VZMULJ(Tw, T8); | |
611 TU = VZMULJ(Td, T8); | |
612 { | |
613 V Tl, Tk, Tm, Tn, T20, T2R, T22, T1V, T2K, T1S, T3A, T2L, TN, T2G, TK; | |
614 V T3w, T2H, T19, T2D, T16, T3x, T2E, T1y, T2N, T1v, T3z, T2O; | |
615 { | |
616 V Tf, Ti, Tj, T7, Tb, Tc, T21; | |
617 Tl = LD(&(x[0]), ms, &(x[0])); | |
618 { | |
619 V Te, Th, T6, Ta; | |
620 Te = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
621 Tf = VZMULJ(Td, Te); | |
622 Th = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
623 Ti = VZMULJ(Tg, Th); | |
624 Tj = VADD(Tf, Ti); | |
625 T6 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
626 T7 = VZMULJ(T5, T6); | |
627 Ta = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
628 Tb = VZMULJ(T9, Ta); | |
629 Tc = VADD(T7, Tb); | |
630 } | |
631 Tk = VMUL(LDK(KP559016994), VSUB(Tc, Tj)); | |
632 Tm = VADD(Tc, Tj); | |
633 Tn = VFNMS(LDK(KP250000000), Tm, Tl); | |
634 T20 = VSUB(T7, Tb); | |
635 T21 = VSUB(Tf, Ti); | |
636 T2R = VMUL(LDK(KP951056516), T21); | |
637 T22 = VFMA(LDK(KP951056516), T20, VMUL(LDK(KP587785252), T21)); | |
638 } | |
639 { | |
640 V T1P, T1I, T1L, T1M, T1B, T1E, T1F, T1O; | |
641 T1O = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
642 T1P = VZMULJ(T2, T1O); | |
643 { | |
644 V T1H, T1K, T1A, T1D; | |
645 T1H = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
646 T1I = VZMULJ(T1G, T1H); | |
647 T1K = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
648 T1L = VZMULJ(T1J, T1K); | |
649 T1M = VADD(T1I, T1L); | |
650 T1A = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
651 T1B = VZMULJ(TA, T1A); | |
652 T1D = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
653 T1E = VZMULJ(T1C, T1D); | |
654 T1F = VADD(T1B, T1E); | |
655 } | |
656 { | |
657 V T1T, T1U, T1N, T1Q, T1R; | |
658 T1T = VSUB(T1B, T1E); | |
659 T1U = VSUB(T1I, T1L); | |
660 T1V = VFMA(LDK(KP475528258), T1T, VMUL(LDK(KP293892626), T1U)); | |
661 T2K = VFNMS(LDK(KP293892626), T1T, VMUL(LDK(KP475528258), T1U)); | |
662 T1N = VMUL(LDK(KP559016994), VSUB(T1F, T1M)); | |
663 T1Q = VADD(T1F, T1M); | |
664 T1R = VFNMS(LDK(KP250000000), T1Q, T1P); | |
665 T1S = VADD(T1N, T1R); | |
666 T3A = VADD(T1P, T1Q); | |
667 T2L = VSUB(T1R, T1N); | |
668 } | |
669 } | |
670 { | |
671 V TH, Tz, TD, TE, Tr, Tu, Tv, TG; | |
672 TG = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
673 TH = VZMULJ(T1, TG); | |
674 { | |
675 V Ty, TC, Tq, Tt; | |
676 Ty = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
677 Tz = VZMULJ(Tx, Ty); | |
678 TC = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
679 TD = VZMULJ(TB, TC); | |
680 TE = VADD(Tz, TD); | |
681 Tq = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
682 Tr = VZMULJ(Tp, Tq); | |
683 Tt = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
684 Tu = VZMULJ(Ts, Tt); | |
685 Tv = VADD(Tr, Tu); | |
686 } | |
687 { | |
688 V TL, TM, TF, TI, TJ; | |
689 TL = VSUB(Tr, Tu); | |
690 TM = VSUB(Tz, TD); | |
691 TN = VFMA(LDK(KP475528258), TL, VMUL(LDK(KP293892626), TM)); | |
692 T2G = VFNMS(LDK(KP293892626), TL, VMUL(LDK(KP475528258), TM)); | |
693 TF = VMUL(LDK(KP559016994), VSUB(Tv, TE)); | |
694 TI = VADD(Tv, TE); | |
695 TJ = VFNMS(LDK(KP250000000), TI, TH); | |
696 TK = VADD(TF, TJ); | |
697 T3w = VADD(TH, TI); | |
698 T2H = VSUB(TJ, TF); | |
699 } | |
700 } | |
701 { | |
702 V T13, TW, TZ, T10, TQ, TS, TT, T12; | |
703 T12 = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
704 T13 = VZMULJ(T3, T12); | |
705 { | |
706 V TV, TY, TP, TR; | |
707 TV = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
708 TW = VZMULJ(TU, TV); | |
709 TY = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
710 TZ = VZMULJ(TX, TY); | |
711 T10 = VADD(TW, TZ); | |
712 TP = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
713 TQ = VZMULJ(T4, TP); | |
714 TR = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
715 TS = VZMULJ(T8, TR); | |
716 TT = VADD(TQ, TS); | |
717 } | |
718 { | |
719 V T17, T18, T11, T14, T15; | |
720 T17 = VSUB(TQ, TS); | |
721 T18 = VSUB(TW, TZ); | |
722 T19 = VFMA(LDK(KP475528258), T17, VMUL(LDK(KP293892626), T18)); | |
723 T2D = VFNMS(LDK(KP293892626), T17, VMUL(LDK(KP475528258), T18)); | |
724 T11 = VMUL(LDK(KP559016994), VSUB(TT, T10)); | |
725 T14 = VADD(TT, T10); | |
726 T15 = VFNMS(LDK(KP250000000), T14, T13); | |
727 T16 = VADD(T11, T15); | |
728 T3x = VADD(T13, T14); | |
729 T2E = VSUB(T15, T11); | |
730 } | |
731 } | |
732 { | |
733 V T1s, T1l, T1o, T1p, T1e, T1h, T1i, T1r; | |
734 T1r = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
735 T1s = VZMULJ(Tw, T1r); | |
736 { | |
737 V T1k, T1n, T1d, T1g; | |
738 T1k = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
739 T1l = VZMULJ(T1j, T1k); | |
740 T1n = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
741 T1o = VZMULJ(T1m, T1n); | |
742 T1p = VADD(T1l, T1o); | |
743 T1d = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
744 T1e = VZMULJ(T1c, T1d); | |
745 T1g = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
746 T1h = VZMULJ(T1f, T1g); | |
747 T1i = VADD(T1e, T1h); | |
748 } | |
749 { | |
750 V T1w, T1x, T1q, T1t, T1u; | |
751 T1w = VSUB(T1e, T1h); | |
752 T1x = VSUB(T1l, T1o); | |
753 T1y = VFMA(LDK(KP475528258), T1w, VMUL(LDK(KP293892626), T1x)); | |
754 T2N = VFNMS(LDK(KP293892626), T1w, VMUL(LDK(KP475528258), T1x)); | |
755 T1q = VMUL(LDK(KP559016994), VSUB(T1i, T1p)); | |
756 T1t = VADD(T1i, T1p); | |
757 T1u = VFNMS(LDK(KP250000000), T1t, T1s); | |
758 T1v = VADD(T1q, T1u); | |
759 T3z = VADD(T1s, T1t); | |
760 T2O = VSUB(T1u, T1q); | |
761 } | |
762 } | |
763 { | |
764 V T3J, T3K, T3D, T3E, T3C, T3F, T3L, T3G; | |
765 { | |
766 V T3H, T3I, T3y, T3B; | |
767 T3H = VSUB(T3w, T3x); | |
768 T3I = VSUB(T3z, T3A); | |
769 T3J = VBYI(VFMA(LDK(KP951056516), T3H, VMUL(LDK(KP587785252), T3I))); | |
770 T3K = VBYI(VFNMS(LDK(KP587785252), T3H, VMUL(LDK(KP951056516), T3I))); | |
771 T3D = VADD(Tl, Tm); | |
772 T3y = VADD(T3w, T3x); | |
773 T3B = VADD(T3z, T3A); | |
774 T3E = VADD(T3y, T3B); | |
775 T3C = VMUL(LDK(KP559016994), VSUB(T3y, T3B)); | |
776 T3F = VFNMS(LDK(KP250000000), T3E, T3D); | |
777 } | |
778 ST(&(x[0]), VADD(T3D, T3E), ms, &(x[0])); | |
779 T3L = VSUB(T3F, T3C); | |
780 ST(&(x[WS(rs, 10)]), VADD(T3K, T3L), ms, &(x[0])); | |
781 ST(&(x[WS(rs, 15)]), VSUB(T3L, T3K), ms, &(x[WS(rs, 1)])); | |
782 T3G = VADD(T3C, T3F); | |
783 ST(&(x[WS(rs, 5)]), VSUB(T3G, T3J), ms, &(x[WS(rs, 1)])); | |
784 ST(&(x[WS(rs, 20)]), VADD(T3J, T3G), ms, &(x[0])); | |
785 } | |
786 { | |
787 V To, T2n, T2o, T2p, T2x, T2y, T2z, T2u, T2v, T2w, T2q, T2r, T2s, T29, T2i; | |
788 V T2e, T2g, T1Y, T2j, T2b, T2c, T2B, T2C; | |
789 To = VADD(Tk, Tn); | |
790 T2n = VFMA(LDK(KP1_688655851), TN, VMUL(LDK(KP535826794), TK)); | |
791 T2o = VFMA(LDK(KP1_541026485), T19, VMUL(LDK(KP637423989), T16)); | |
792 T2p = VSUB(T2n, T2o); | |
793 T2x = VFMA(LDK(KP851558583), T1y, VMUL(LDK(KP904827052), T1v)); | |
794 T2y = VFMA(LDK(KP1_984229402), T1V, VMUL(LDK(KP125333233), T1S)); | |
795 T2z = VADD(T2x, T2y); | |
796 T2u = VFNMS(LDK(KP844327925), TK, VMUL(LDK(KP1_071653589), TN)); | |
797 T2v = VFNMS(LDK(KP1_274847979), T19, VMUL(LDK(KP770513242), T16)); | |
798 T2w = VADD(T2u, T2v); | |
799 T2q = VFNMS(LDK(KP425779291), T1v, VMUL(LDK(KP1_809654104), T1y)); | |
800 T2r = VFNMS(LDK(KP992114701), T1S, VMUL(LDK(KP250666467), T1V)); | |
801 T2s = VADD(T2q, T2r); | |
802 { | |
803 V T23, T24, T25, T26, T27, T28; | |
804 T23 = VFMA(LDK(KP1_937166322), TN, VMUL(LDK(KP248689887), TK)); | |
805 T24 = VFMA(LDK(KP1_071653589), T19, VMUL(LDK(KP844327925), T16)); | |
806 T25 = VADD(T23, T24); | |
807 T26 = VFMA(LDK(KP1_752613360), T1y, VMUL(LDK(KP481753674), T1v)); | |
808 T27 = VFMA(LDK(KP1_457937254), T1V, VMUL(LDK(KP684547105), T1S)); | |
809 T28 = VADD(T26, T27); | |
810 T29 = VADD(T25, T28); | |
811 T2i = VSUB(T27, T26); | |
812 T2e = VMUL(LDK(KP559016994), VSUB(T28, T25)); | |
813 T2g = VSUB(T24, T23); | |
814 } | |
815 { | |
816 V TO, T1a, T1b, T1z, T1W, T1X; | |
817 TO = VFNMS(LDK(KP497379774), TN, VMUL(LDK(KP968583161), TK)); | |
818 T1a = VFNMS(LDK(KP1_688655851), T19, VMUL(LDK(KP535826794), T16)); | |
819 T1b = VADD(TO, T1a); | |
820 T1z = VFNMS(LDK(KP963507348), T1y, VMUL(LDK(KP876306680), T1v)); | |
821 T1W = VFNMS(LDK(KP1_369094211), T1V, VMUL(LDK(KP728968627), T1S)); | |
822 T1X = VADD(T1z, T1W); | |
823 T1Y = VADD(T1b, T1X); | |
824 T2j = VMUL(LDK(KP559016994), VSUB(T1b, T1X)); | |
825 T2b = VSUB(T1a, TO); | |
826 T2c = VSUB(T1z, T1W); | |
827 } | |
828 { | |
829 V T1Z, T2a, T2t, T2A; | |
830 T1Z = VADD(To, T1Y); | |
831 T2a = VBYI(VADD(T22, T29)); | |
832 ST(&(x[WS(rs, 1)]), VSUB(T1Z, T2a), ms, &(x[WS(rs, 1)])); | |
833 ST(&(x[WS(rs, 24)]), VADD(T1Z, T2a), ms, &(x[0])); | |
834 T2t = VADD(To, VADD(T2p, T2s)); | |
835 T2A = VBYI(VADD(T22, VSUB(T2w, T2z))); | |
836 ST(&(x[WS(rs, 21)]), VSUB(T2t, T2A), ms, &(x[WS(rs, 1)])); | |
837 ST(&(x[WS(rs, 4)]), VADD(T2t, T2A), ms, &(x[0])); | |
838 } | |
839 T2B = VBYI(VADD(T22, VFMA(LDK(KP309016994), T2w, VFMA(LDK(KP587785252), VSUB(T2r, T2q), VFNMS(LDK(KP951056516), VADD(T2n, T2o), VMUL(LDK(KP809016994), T2z)))))); | |
840 T2C = VFMA(LDK(KP309016994), T2p, VFMA(LDK(KP951056516), VSUB(T2u, T2v), VFMA(LDK(KP587785252), VSUB(T2y, T2x), VFNMS(LDK(KP809016994), T2s, To)))); | |
841 ST(&(x[WS(rs, 9)]), VADD(T2B, T2C), ms, &(x[WS(rs, 1)])); | |
842 ST(&(x[WS(rs, 16)]), VSUB(T2C, T2B), ms, &(x[0])); | |
843 { | |
844 V T2f, T2l, T2k, T2m, T2d, T2h; | |
845 T2d = VFMS(LDK(KP250000000), T29, T22); | |
846 T2f = VBYI(VADD(VFMA(LDK(KP587785252), T2b, VMUL(LDK(KP951056516), T2c)), VSUB(T2d, T2e))); | |
847 T2l = VBYI(VADD(VFNMS(LDK(KP587785252), T2c, VMUL(LDK(KP951056516), T2b)), VADD(T2d, T2e))); | |
848 T2h = VFNMS(LDK(KP250000000), T1Y, To); | |
849 T2k = VFMA(LDK(KP587785252), T2g, VFNMS(LDK(KP951056516), T2i, VSUB(T2h, T2j))); | |
850 T2m = VFMA(LDK(KP951056516), T2g, VADD(T2j, VFMA(LDK(KP587785252), T2i, T2h))); | |
851 ST(&(x[WS(rs, 11)]), VADD(T2f, T2k), ms, &(x[WS(rs, 1)])); | |
852 ST(&(x[WS(rs, 19)]), VSUB(T2m, T2l), ms, &(x[WS(rs, 1)])); | |
853 ST(&(x[WS(rs, 14)]), VSUB(T2k, T2f), ms, &(x[0])); | |
854 ST(&(x[WS(rs, 6)]), VADD(T2l, T2m), ms, &(x[0])); | |
855 } | |
856 } | |
857 { | |
858 V T2S, T2U, T2F, T2I, T2J, T2Y, T2Z, T30, T2M, T2P, T2Q, T2V, T2W, T2X, T3a; | |
859 V T3l, T3b, T3k, T3f, T3p, T3i, T3o, T32, T33; | |
860 T2S = VFNMS(LDK(KP587785252), T20, T2R); | |
861 T2U = VSUB(Tn, Tk); | |
862 T2F = VFNMS(LDK(KP125333233), T2E, VMUL(LDK(KP1_984229402), T2D)); | |
863 T2I = VFMA(LDK(KP1_457937254), T2G, VMUL(LDK(KP684547105), T2H)); | |
864 T2J = VSUB(T2F, T2I); | |
865 T2Y = VFNMS(LDK(KP1_996053456), T2N, VMUL(LDK(KP062790519), T2O)); | |
866 T2Z = VFMA(LDK(KP1_541026485), T2K, VMUL(LDK(KP637423989), T2L)); | |
867 T30 = VSUB(T2Y, T2Z); | |
868 T2M = VFNMS(LDK(KP770513242), T2L, VMUL(LDK(KP1_274847979), T2K)); | |
869 T2P = VFMA(LDK(KP125581039), T2N, VMUL(LDK(KP998026728), T2O)); | |
870 T2Q = VSUB(T2M, T2P); | |
871 T2V = VFNMS(LDK(KP1_369094211), T2G, VMUL(LDK(KP728968627), T2H)); | |
872 T2W = VFMA(LDK(KP250666467), T2D, VMUL(LDK(KP992114701), T2E)); | |
873 T2X = VSUB(T2V, T2W); | |
874 { | |
875 V T34, T35, T36, T37, T38, T39; | |
876 T34 = VFNMS(LDK(KP481753674), T2H, VMUL(LDK(KP1_752613360), T2G)); | |
877 T35 = VFMA(LDK(KP851558583), T2D, VMUL(LDK(KP904827052), T2E)); | |
878 T36 = VSUB(T34, T35); | |
879 T37 = VFNMS(LDK(KP844327925), T2O, VMUL(LDK(KP1_071653589), T2N)); | |
880 T38 = VFNMS(LDK(KP998026728), T2L, VMUL(LDK(KP125581039), T2K)); | |
881 T39 = VADD(T37, T38); | |
882 T3a = VMUL(LDK(KP559016994), VSUB(T36, T39)); | |
883 T3l = VSUB(T37, T38); | |
884 T3b = VADD(T36, T39); | |
885 T3k = VADD(T34, T35); | |
886 } | |
887 { | |
888 V T3d, T3e, T3m, T3g, T3h, T3n; | |
889 T3d = VFNMS(LDK(KP425779291), T2E, VMUL(LDK(KP1_809654104), T2D)); | |
890 T3e = VFMA(LDK(KP963507348), T2G, VMUL(LDK(KP876306680), T2H)); | |
891 T3m = VADD(T3e, T3d); | |
892 T3g = VFMA(LDK(KP1_688655851), T2N, VMUL(LDK(KP535826794), T2O)); | |
893 T3h = VFMA(LDK(KP1_996053456), T2K, VMUL(LDK(KP062790519), T2L)); | |
894 T3n = VADD(T3g, T3h); | |
895 T3f = VSUB(T3d, T3e); | |
896 T3p = VADD(T3m, T3n); | |
897 T3i = VSUB(T3g, T3h); | |
898 T3o = VMUL(LDK(KP559016994), VSUB(T3m, T3n)); | |
899 } | |
900 { | |
901 V T3u, T3v, T2T, T31; | |
902 T3u = VBYI(VADD(T2S, T3b)); | |
903 T3v = VADD(T2U, T3p); | |
904 ST(&(x[WS(rs, 2)]), VADD(T3u, T3v), ms, &(x[0])); | |
905 ST(&(x[WS(rs, 23)]), VSUB(T3v, T3u), ms, &(x[WS(rs, 1)])); | |
906 T2T = VBYI(VSUB(VADD(T2J, T2Q), T2S)); | |
907 T31 = VADD(T2U, VADD(T2X, T30)); | |
908 ST(&(x[WS(rs, 3)]), VADD(T2T, T31), ms, &(x[WS(rs, 1)])); | |
909 ST(&(x[WS(rs, 22)]), VSUB(T31, T2T), ms, &(x[0])); | |
910 } | |
911 T32 = VFMA(LDK(KP309016994), T2X, VFNMS(LDK(KP809016994), T30, VFNMS(LDK(KP587785252), VADD(T2P, T2M), VFNMS(LDK(KP951056516), VADD(T2I, T2F), T2U)))); | |
912 T33 = VBYI(VSUB(VFNMS(LDK(KP587785252), VADD(T2Y, T2Z), VFNMS(LDK(KP809016994), T2Q, VFNMS(LDK(KP951056516), VADD(T2V, T2W), VMUL(LDK(KP309016994), T2J)))), T2S)); | |
913 ST(&(x[WS(rs, 17)]), VSUB(T32, T33), ms, &(x[WS(rs, 1)])); | |
914 ST(&(x[WS(rs, 8)]), VADD(T32, T33), ms, &(x[0])); | |
915 { | |
916 V T3j, T3s, T3r, T3t, T3c, T3q; | |
917 T3c = VFNMS(LDK(KP250000000), T3b, T2S); | |
918 T3j = VBYI(VADD(T3a, VADD(T3c, VFNMS(LDK(KP587785252), T3i, VMUL(LDK(KP951056516), T3f))))); | |
919 T3s = VBYI(VADD(T3c, VSUB(VFMA(LDK(KP587785252), T3f, VMUL(LDK(KP951056516), T3i)), T3a))); | |
920 T3q = VFNMS(LDK(KP250000000), T3p, T2U); | |
921 T3r = VFMA(LDK(KP951056516), T3k, VFMA(LDK(KP587785252), T3l, VADD(T3o, T3q))); | |
922 T3t = VFMA(LDK(KP587785252), T3k, VSUB(VFNMS(LDK(KP951056516), T3l, T3q), T3o)); | |
923 ST(&(x[WS(rs, 7)]), VADD(T3j, T3r), ms, &(x[WS(rs, 1)])); | |
924 ST(&(x[WS(rs, 13)]), VSUB(T3t, T3s), ms, &(x[WS(rs, 1)])); | |
925 ST(&(x[WS(rs, 18)]), VSUB(T3r, T3j), ms, &(x[0])); | |
926 ST(&(x[WS(rs, 12)]), VADD(T3s, T3t), ms, &(x[0])); | |
927 } | |
928 } | |
929 } | |
930 } | |
931 } | |
932 VLEAVE(); | |
933 } | |
934 | |
935 static const tw_instr twinstr[] = { | |
936 VTW(0, 1), | |
937 VTW(0, 3), | |
938 VTW(0, 9), | |
939 VTW(0, 24), | |
940 {TW_NEXT, VL, 0} | |
941 }; | |
942 | |
943 static const ct_desc desc = { 25, XSIMD_STRING("t3fv_25"), twinstr, &GENUS, {190, 150, 78, 0}, 0, 0, 0 }; | |
944 | |
945 void XSIMD(codelet_t3fv_25) (planner *p) { | |
946 X(kdft_dit_register) (p, t3fv_25, &desc); | |
947 } | |
948 #endif /* HAVE_FMA */ |