Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/simd/common/t2bv_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 07:39:15 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 -n 25 -name t2bv_25 -include t2b.h -sign 1 */ | |
29 | |
30 /* | |
31 * This function contains 248 FP additions, 241 FP multiplications, | |
32 * (or, 67 additions, 60 multiplications, 181 fused multiply/add), | |
33 * 208 stack variables, 67 constants, and 50 memory accesses | |
34 */ | |
35 #include "t2b.h" | |
36 | |
37 static void t2bv_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(KP921078979, +0.921078979742360627699756128143719920817673854); | |
84 DVK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
85 DVK(KP982009705, +0.982009705009746369461829878184175962711969869); | |
86 DVK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
87 DVK(KP494780565, +0.494780565770515410344588413655324772219443730); | |
88 DVK(KP447533225, +0.447533225982656890041886979663652563063114397); | |
89 DVK(KP269969613, +0.269969613759572083574752974412347470060951301); | |
90 DVK(KP244189809, +0.244189809627953270309879511234821255780225091); | |
91 DVK(KP667278218, +0.667278218140296670899089292254759909713898805); | |
92 DVK(KP603558818, +0.603558818296015001454675132653458027918768137); | |
93 DVK(KP522847744, +0.522847744331509716623755382187077770911012542); | |
94 DVK(KP578046249, +0.578046249379945007321754579646815604023525655); | |
95 DVK(KP987388751, +0.987388751065621252324603216482382109400433949); | |
96 DVK(KP893101515, +0.893101515366181661711202267938416198338079437); | |
97 DVK(KP120146378, +0.120146378570687701782758537356596213647956445); | |
98 DVK(KP132830569, +0.132830569247582714407653942074819768844536507); | |
99 DVK(KP869845200, +0.869845200362138853122720822420327157933056305); | |
100 DVK(KP786782374, +0.786782374965295178365099601674911834788448471); | |
101 DVK(KP066152395, +0.066152395967733048213034281011006031460903353); | |
102 DVK(KP059835404, +0.059835404262124915169548397419498386427871950); | |
103 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
104 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
105 DVK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
106 { | |
107 INT m; | |
108 R *x; | |
109 x = ii; | |
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 T25, T1B, T2y, T1K, T2s, T23, T1S, T26, T20, T1X; | |
112 { | |
113 V T1O, T2X, Te, T3L, Td, T3Q, T3j, T3b, T2R, T2M, T2f, T27, T1y, T1H, T3M; | |
114 V TW, TR, TK, T2B, T3n, T3e, T2U, T2F, T2i, T2a, Tz, T1C, T3N, TQ, T11; | |
115 V T1b, T1c, T16; | |
116 { | |
117 V T1, T1g, T1i, T1p, T1k, T1m, Tb, T1N, T6, T1M; | |
118 { | |
119 V T7, T9, T2, T4, T1f, T1h, T1o; | |
120 T1 = LD(&(x[0]), ms, &(x[0])); | |
121 T7 = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
122 T9 = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
123 T2 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
124 T4 = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
125 T1f = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
126 T1h = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
127 T1o = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
128 { | |
129 V T8, Ta, T3, T5, T1j; | |
130 T1j = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
131 T8 = BYTW(&(W[TWVL * 18]), T7); | |
132 Ta = BYTW(&(W[TWVL * 28]), T9); | |
133 T3 = BYTW(&(W[TWVL * 8]), T2); | |
134 T5 = BYTW(&(W[TWVL * 38]), T4); | |
135 T1g = BYTW(&(W[TWVL * 4]), T1f); | |
136 T1i = BYTW(&(W[TWVL * 14]), T1h); | |
137 T1p = BYTW(&(W[TWVL * 34]), T1o); | |
138 T1k = BYTW(&(W[TWVL * 44]), T1j); | |
139 T1m = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
140 Tb = VADD(T8, Ta); | |
141 T1N = VSUB(T8, Ta); | |
142 T6 = VADD(T3, T5); | |
143 T1M = VSUB(T3, T5); | |
144 } | |
145 } | |
146 { | |
147 V T1v, T1l, Th, Tj, T1w, T1q, Tq, Tk, Tn, Tg; | |
148 Tg = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
149 { | |
150 V Tc, Ti, T1n, Tp; | |
151 Ti = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
152 T1v = VSUB(T1i, T1k); | |
153 T1l = VADD(T1i, T1k); | |
154 T1n = BYTW(&(W[TWVL * 24]), T1m); | |
155 Tp = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
156 T1O = VFMA(LDK(KP618033988), T1N, T1M); | |
157 T2X = VFNMS(LDK(KP618033988), T1M, T1N); | |
158 Te = VSUB(T6, Tb); | |
159 Tc = VADD(T6, Tb); | |
160 Th = BYTW(&(W[0]), Tg); | |
161 Tj = BYTW(&(W[TWVL * 10]), Ti); | |
162 T1w = VSUB(T1n, T1p); | |
163 T1q = VADD(T1n, T1p); | |
164 Tq = BYTW(&(W[TWVL * 30]), Tp); | |
165 Tk = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
166 T3L = VADD(T1, Tc); | |
167 Td = VFNMS(LDK(KP250000000), Tc, T1); | |
168 Tn = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
169 } | |
170 { | |
171 V T1x, T2K, TM, TB, Tw, Tm, Tx, Tr, TI, T2L, T1u, TD, TF, TL; | |
172 TL = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
173 { | |
174 V T1t, Tl, To, TH, T1s, T1r, TA, TC; | |
175 TA = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
176 T1r = VADD(T1l, T1q); | |
177 T1t = VSUB(T1q, T1l); | |
178 T1x = VFMA(LDK(KP618033988), T1w, T1v); | |
179 T2K = VFNMS(LDK(KP618033988), T1v, T1w); | |
180 Tl = BYTW(&(W[TWVL * 40]), Tk); | |
181 To = BYTW(&(W[TWVL * 20]), Tn); | |
182 TM = BYTW(&(W[TWVL * 6]), TL); | |
183 TB = BYTW(&(W[TWVL * 46]), TA); | |
184 TH = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
185 T1s = VFNMS(LDK(KP250000000), T1r, T1g); | |
186 T3Q = VADD(T1g, T1r); | |
187 TC = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
188 Tw = VSUB(Tj, Tl); | |
189 Tm = VADD(Tj, Tl); | |
190 Tx = VSUB(Tq, To); | |
191 Tr = VADD(To, Tq); | |
192 TI = BYTW(&(W[TWVL * 26]), TH); | |
193 T2L = VFMA(LDK(KP559016994), T1t, T1s); | |
194 T1u = VFNMS(LDK(KP559016994), T1t, T1s); | |
195 TD = BYTW(&(W[TWVL * 16]), TC); | |
196 TF = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
197 } | |
198 { | |
199 V Tu, Ty, T2E, TE, TN, TG, Tt, TV, Ts; | |
200 TV = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
201 Ts = VADD(Tm, Tr); | |
202 Tu = VSUB(Tm, Tr); | |
203 Ty = VFNMS(LDK(KP618033988), Tx, Tw); | |
204 T2E = VFMA(LDK(KP618033988), Tw, Tx); | |
205 T3j = VFNMS(LDK(KP059835404), T2K, T2L); | |
206 T3b = VFMA(LDK(KP066152395), T2L, T2K); | |
207 T2R = VFNMS(LDK(KP786782374), T2K, T2L); | |
208 T2M = VFMA(LDK(KP869845200), T2L, T2K); | |
209 T2f = VFMA(LDK(KP132830569), T1u, T1x); | |
210 T27 = VFNMS(LDK(KP120146378), T1x, T1u); | |
211 T1y = VFNMS(LDK(KP893101515), T1x, T1u); | |
212 T1H = VFMA(LDK(KP987388751), T1u, T1x); | |
213 TE = VSUB(TB, TD); | |
214 TN = VADD(TD, TB); | |
215 TG = BYTW(&(W[TWVL * 36]), TF); | |
216 Tt = VFNMS(LDK(KP250000000), Ts, Th); | |
217 T3M = VADD(Th, Ts); | |
218 TW = BYTW(&(W[TWVL * 2]), TV); | |
219 { | |
220 V TJ, TO, Tv, T2D, TY, T15, T10, T13, TP; | |
221 { | |
222 V TX, T14, TZ, T12; | |
223 TX = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
224 T14 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
225 TZ = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
226 T12 = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
227 TJ = VSUB(TG, TI); | |
228 TO = VADD(TI, TG); | |
229 Tv = VFMA(LDK(KP559016994), Tu, Tt); | |
230 T2D = VFNMS(LDK(KP559016994), Tu, Tt); | |
231 TY = BYTW(&(W[TWVL * 12]), TX); | |
232 T15 = BYTW(&(W[TWVL * 32]), T14); | |
233 T10 = BYTW(&(W[TWVL * 42]), TZ); | |
234 T13 = BYTW(&(W[TWVL * 22]), T12); | |
235 } | |
236 TP = VADD(TN, TO); | |
237 TR = VSUB(TN, TO); | |
238 TK = VFMA(LDK(KP618033988), TJ, TE); | |
239 T2B = VFNMS(LDK(KP618033988), TE, TJ); | |
240 T3n = VFMA(LDK(KP578046249), T2D, T2E); | |
241 T3e = VFNMS(LDK(KP522847744), T2E, T2D); | |
242 T2U = VFNMS(LDK(KP987388751), T2D, T2E); | |
243 T2F = VFMA(LDK(KP893101515), T2E, T2D); | |
244 T2i = VFNMS(LDK(KP603558818), Ty, Tv); | |
245 T2a = VFMA(LDK(KP667278218), Tv, Ty); | |
246 Tz = VFNMS(LDK(KP244189809), Ty, Tv); | |
247 T1C = VFMA(LDK(KP269969613), Tv, Ty); | |
248 T3N = VADD(TM, TP); | |
249 TQ = VFMS(LDK(KP250000000), TP, TM); | |
250 T11 = VADD(TY, T10); | |
251 T1b = VSUB(TY, T10); | |
252 T1c = VSUB(T15, T13); | |
253 T16 = VADD(T13, T15); | |
254 } | |
255 } | |
256 } | |
257 } | |
258 } | |
259 { | |
260 V T2z, Tf, T3W, T3O, T1d, T2H, T3m, T2j, T2b, TT, T1D, T2G, T35, T2V, T2Z; | |
261 V T3A, T3g, T2I, T1a, T3R, T3X; | |
262 T2z = VFNMS(LDK(KP559016994), Te, Td); | |
263 Tf = VFMA(LDK(KP559016994), Te, Td); | |
264 { | |
265 V TS, T2A, T17, T19; | |
266 TS = VFNMS(LDK(KP559016994), TR, TQ); | |
267 T2A = VFMA(LDK(KP559016994), TR, TQ); | |
268 T3W = VSUB(T3M, T3N); | |
269 T3O = VADD(T3M, T3N); | |
270 T1d = VFNMS(LDK(KP618033988), T1c, T1b); | |
271 T2H = VFMA(LDK(KP618033988), T1b, T1c); | |
272 T17 = VADD(T11, T16); | |
273 T19 = VSUB(T16, T11); | |
274 { | |
275 V T3f, T2T, T2C, T18, T3P; | |
276 T3m = VFMA(LDK(KP447533225), T2B, T2A); | |
277 T3f = VFNMS(LDK(KP494780565), T2A, T2B); | |
278 T2T = VFNMS(LDK(KP132830569), T2A, T2B); | |
279 T2C = VFMA(LDK(KP120146378), T2B, T2A); | |
280 T2j = VFNMS(LDK(KP786782374), TK, TS); | |
281 T2b = VFMA(LDK(KP869845200), TS, TK); | |
282 TT = VFNMS(LDK(KP667278218), TS, TK); | |
283 T1D = VFMA(LDK(KP603558818), TK, TS); | |
284 T18 = VFNMS(LDK(KP250000000), T17, TW); | |
285 T3P = VADD(TW, T17); | |
286 T2G = VFMA(LDK(KP734762448), T2F, T2C); | |
287 T35 = VFNMS(LDK(KP734762448), T2F, T2C); | |
288 T2V = VFNMS(LDK(KP734762448), T2U, T2T); | |
289 T2Z = VFMA(LDK(KP734762448), T2U, T2T); | |
290 T3A = VFMA(LDK(KP982009705), T3f, T3e); | |
291 T3g = VFNMS(LDK(KP982009705), T3f, T3e); | |
292 T2I = VFMA(LDK(KP559016994), T19, T18); | |
293 T1a = VFNMS(LDK(KP559016994), T19, T18); | |
294 T3R = VADD(T3P, T3Q); | |
295 T3X = VSUB(T3P, T3Q); | |
296 } | |
297 } | |
298 { | |
299 V T2n, T2t, T1V, T22, T2l, T2d, T1Q, T1I, T2w, T1A, T1F, T2q; | |
300 { | |
301 V T2k, T1G, T28, T2g, T3K, T3E, T3a, T34, T3x, T3H, T2c, TU, T1T, T1U, T1z; | |
302 V T3o, T3t; | |
303 T2n = VFNMS(LDK(KP912575812), T2j, T2i); | |
304 T2k = VFMA(LDK(KP912575812), T2j, T2i); | |
305 T3o = VFNMS(LDK(KP921078979), T3n, T3m); | |
306 T3t = VFMA(LDK(KP921078979), T3n, T3m); | |
307 { | |
308 V T3c, T2Q, T2J, T3k, T1e; | |
309 T3c = VFNMS(LDK(KP667278218), T2I, T2H); | |
310 T2Q = VFNMS(LDK(KP059835404), T2H, T2I); | |
311 T2J = VFMA(LDK(KP066152395), T2I, T2H); | |
312 T3k = VFMA(LDK(KP603558818), T2H, T2I); | |
313 T1G = VFMA(LDK(KP578046249), T1a, T1d); | |
314 T1e = VFNMS(LDK(KP522847744), T1d, T1a); | |
315 T28 = VFNMS(LDK(KP494780565), T1a, T1d); | |
316 T2g = VFMA(LDK(KP447533225), T1d, T1a); | |
317 { | |
318 V T3U, T3S, T40, T3Y; | |
319 T3U = VSUB(T3O, T3R); | |
320 T3S = VADD(T3O, T3R); | |
321 T40 = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T3W, T3X)); | |
322 T3Y = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T3X, T3W)); | |
323 { | |
324 V T3s, T3l, T2N, T36; | |
325 T3s = VFNMS(LDK(KP845997307), T3k, T3j); | |
326 T3l = VFMA(LDK(KP845997307), T3k, T3j); | |
327 T2N = VFNMS(LDK(KP772036680), T2M, T2J); | |
328 T36 = VFMA(LDK(KP772036680), T2M, T2J); | |
329 { | |
330 V T30, T2S, T3d, T3z, T3T; | |
331 T30 = VFNMS(LDK(KP772036680), T2R, T2Q); | |
332 T2S = VFMA(LDK(KP772036680), T2R, T2Q); | |
333 T3d = VFNMS(LDK(KP845997307), T3c, T3b); | |
334 T3z = VFMA(LDK(KP845997307), T3c, T3b); | |
335 ST(&(x[0]), VADD(T3S, T3L), ms, &(x[0])); | |
336 T3T = VFNMS(LDK(KP250000000), T3S, T3L); | |
337 { | |
338 V T3C, T3p, T2O, T37; | |
339 T3C = VFMA(LDK(KP906616052), T3o, T3l); | |
340 T3p = VFNMS(LDK(KP906616052), T3o, T3l); | |
341 T2O = VFMA(LDK(KP956723877), T2N, T2G); | |
342 T37 = VFMA(LDK(KP522616830), T2V, T36); | |
343 { | |
344 V T31, T2W, T3u, T3h; | |
345 T31 = VFNMS(LDK(KP522616830), T2G, T30); | |
346 T2W = VFMA(LDK(KP945422727), T2V, T2S); | |
347 T3u = VFNMS(LDK(KP923225144), T3g, T3d); | |
348 T3h = VFMA(LDK(KP923225144), T3g, T3d); | |
349 { | |
350 V T3I, T3B, T3V, T3Z; | |
351 T3I = VFNMS(LDK(KP669429328), T3z, T3A); | |
352 T3B = VFMA(LDK(KP570584518), T3A, T3z); | |
353 T3V = VFMA(LDK(KP559016994), T3U, T3T); | |
354 T3Z = VFNMS(LDK(KP559016994), T3U, T3T); | |
355 { | |
356 V T3y, T3q, T2P, T38; | |
357 T3y = VFMA(LDK(KP262346850), T3p, T2X); | |
358 T3q = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T2X, T3p)); | |
359 T2P = VFMA(LDK(KP992114701), T2O, T2z); | |
360 T38 = VFNMS(LDK(KP690983005), T37, T2S); | |
361 { | |
362 V T32, T2Y, T3v, T3F; | |
363 T32 = VFMA(LDK(KP763932022), T31, T2N); | |
364 T2Y = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T2X, T2W)); | |
365 T3v = VFNMS(LDK(KP997675361), T3u, T3t); | |
366 T3F = VFNMS(LDK(KP904508497), T3u, T3s); | |
367 { | |
368 V T3i, T3r, T3J, T3D; | |
369 T3i = VFMA(LDK(KP949179823), T3h, T2z); | |
370 T3r = VFNMS(LDK(KP237294955), T3h, T2z); | |
371 T3J = VFNMS(LDK(KP669429328), T3C, T3I); | |
372 T3D = VFMA(LDK(KP618033988), T3C, T3B); | |
373 ST(&(x[WS(rs, 20)]), VFNMSI(T3Y, T3V), ms, &(x[0])); | |
374 ST(&(x[WS(rs, 5)]), VFMAI(T3Y, T3V), ms, &(x[WS(rs, 1)])); | |
375 ST(&(x[WS(rs, 15)]), VFMAI(T40, T3Z), ms, &(x[WS(rs, 1)])); | |
376 ST(&(x[WS(rs, 10)]), VFNMSI(T40, T3Z), ms, &(x[0])); | |
377 { | |
378 V T39, T33, T3w, T3G; | |
379 T39 = VFMA(LDK(KP855719849), T38, T35); | |
380 T33 = VFNMS(LDK(KP855719849), T32, T2Z); | |
381 ST(&(x[WS(rs, 3)]), VFMAI(T2Y, T2P), ms, &(x[WS(rs, 1)])); | |
382 ST(&(x[WS(rs, 22)]), VFNMSI(T2Y, T2P), ms, &(x[0])); | |
383 T3w = VFMA(LDK(KP560319534), T3v, T3s); | |
384 T3G = VFNMS(LDK(KP681693190), T3F, T3t); | |
385 ST(&(x[WS(rs, 2)]), VFMAI(T3q, T3i), ms, &(x[0])); | |
386 ST(&(x[WS(rs, 23)]), VFNMSI(T3q, T3i), ms, &(x[WS(rs, 1)])); | |
387 T3K = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T3J, T3y)); | |
388 T3E = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T3D, T3y)); | |
389 T3a = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T39, T2X)); | |
390 T34 = VFMA(LDK(KP897376177), T33, T2z); | |
391 T3x = VFNMS(LDK(KP949179823), T3w, T3r); | |
392 T3H = VFNMS(LDK(KP860541664), T3G, T3r); | |
393 T2t = VFNMS(LDK(KP912575812), T2b, T2a); | |
394 T2c = VFMA(LDK(KP912575812), T2b, T2a); | |
395 TU = VFMA(LDK(KP829049696), TT, Tz); | |
396 T1T = VFNMS(LDK(KP829049696), TT, Tz); | |
397 T1U = VFNMS(LDK(KP831864738), T1y, T1e); | |
398 T1z = VFMA(LDK(KP831864738), T1y, T1e); | |
399 } | |
400 } | |
401 } | |
402 } | |
403 } | |
404 } | |
405 } | |
406 } | |
407 } | |
408 } | |
409 } | |
410 { | |
411 V T2o, T2h, T29, T2u, T2v, T2p; | |
412 T2o = VFNMS(LDK(KP958953096), T2g, T2f); | |
413 T2h = VFMA(LDK(KP958953096), T2g, T2f); | |
414 ST(&(x[WS(rs, 17)]), VFNMSI(T3a, T34), ms, &(x[WS(rs, 1)])); | |
415 ST(&(x[WS(rs, 8)]), VFMAI(T3a, T34), ms, &(x[0])); | |
416 ST(&(x[WS(rs, 13)]), VFMAI(T3E, T3x), ms, &(x[WS(rs, 1)])); | |
417 ST(&(x[WS(rs, 12)]), VFNMSI(T3E, T3x), ms, &(x[0])); | |
418 ST(&(x[WS(rs, 7)]), VFNMSI(T3K, T3H), ms, &(x[WS(rs, 1)])); | |
419 ST(&(x[WS(rs, 18)]), VFMAI(T3K, T3H), ms, &(x[0])); | |
420 T1V = VFMA(LDK(KP559154169), T1U, T1T); | |
421 T22 = VFNMS(LDK(KP683113946), T1T, T1U); | |
422 T29 = VFNMS(LDK(KP867381224), T28, T27); | |
423 T2u = VFMA(LDK(KP867381224), T28, T27); | |
424 T2l = VFMA(LDK(KP894834959), T2k, T2h); | |
425 T2v = VFMA(LDK(KP447417479), T2k, T2u); | |
426 T2d = VFNMS(LDK(KP809385824), T2c, T29); | |
427 T2p = VFMA(LDK(KP447417479), T2c, T2o); | |
428 T1Q = VFMA(LDK(KP831864738), T1H, T1G); | |
429 T1I = VFNMS(LDK(KP831864738), T1H, T1G); | |
430 T2w = VFNMS(LDK(KP763932022), T2v, T2h); | |
431 T1A = VFMA(LDK(KP904730450), T1z, TU); | |
432 T1F = VFNMS(LDK(KP904730450), T1z, TU); | |
433 T2q = VFMA(LDK(KP690983005), T2p, T29); | |
434 } | |
435 } | |
436 { | |
437 V T2e, T1E, T1P, T2m; | |
438 T2e = VFNMS(LDK(KP992114701), T2d, Tf); | |
439 T1E = VFMA(LDK(KP916574801), T1D, T1C); | |
440 T1P = VFNMS(LDK(KP916574801), T1D, T1C); | |
441 T2m = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2l, T1O)); | |
442 { | |
443 V T1J, T2r, T1R, T1W, T1Z, T2x; | |
444 T2x = VFNMS(LDK(KP999544308), T2w, T2t); | |
445 T1J = VFNMS(LDK(KP904730450), T1I, T1F); | |
446 T25 = VFMA(LDK(KP968583161), T1A, Tf); | |
447 T1B = VFNMS(LDK(KP242145790), T1A, Tf); | |
448 T2r = VFNMS(LDK(KP999544308), T2q, T2n); | |
449 T1R = VFMA(LDK(KP904730450), T1Q, T1P); | |
450 T1W = VFNMS(LDK(KP904730450), T1Q, T1P); | |
451 T1Z = VADD(T1E, T1F); | |
452 ST(&(x[WS(rs, 21)]), VFMAI(T2m, T2e), ms, &(x[WS(rs, 1)])); | |
453 ST(&(x[WS(rs, 4)]), VFNMSI(T2m, T2e), ms, &(x[0])); | |
454 T2y = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T2x, T1O)); | |
455 T1K = VFNMS(LDK(KP618033988), T1J, T1E); | |
456 T2s = VFNMS(LDK(KP803003575), T2r, Tf); | |
457 T23 = VFMA(LDK(KP617882369), T1W, T22); | |
458 T1S = VFNMS(LDK(KP242145790), T1R, T1O); | |
459 T26 = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1R, T1O)); | |
460 T20 = VFNMS(LDK(KP683113946), T1Z, T1I); | |
461 T1X = VFMA(LDK(KP559016994), T1W, T1V); | |
462 } | |
463 } | |
464 } | |
465 } | |
466 } | |
467 { | |
468 V T1L, T24, T21, T1Y; | |
469 T1L = VFNMS(LDK(KP876091699), T1K, T1B); | |
470 ST(&(x[WS(rs, 16)]), VFMAI(T2y, T2s), ms, &(x[0])); | |
471 ST(&(x[WS(rs, 9)]), VFNMSI(T2y, T2s), ms, &(x[WS(rs, 1)])); | |
472 T24 = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T23, T1S)); | |
473 ST(&(x[WS(rs, 24)]), VFNMSI(T26, T25), ms, &(x[0])); | |
474 ST(&(x[WS(rs, 1)]), VFMAI(T26, T25), ms, &(x[WS(rs, 1)])); | |
475 T21 = VFMA(LDK(KP792626838), T20, T1B); | |
476 T1Y = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1X, T1S)); | |
477 ST(&(x[WS(rs, 11)]), VFMAI(T24, T21), ms, &(x[WS(rs, 1)])); | |
478 ST(&(x[WS(rs, 14)]), VFNMSI(T24, T21), ms, &(x[0])); | |
479 ST(&(x[WS(rs, 19)]), VFNMSI(T1Y, T1L), ms, &(x[WS(rs, 1)])); | |
480 ST(&(x[WS(rs, 6)]), VFMAI(T1Y, T1L), ms, &(x[0])); | |
481 } | |
482 } | |
483 } | |
484 VLEAVE(); | |
485 } | |
486 | |
487 static const tw_instr twinstr[] = { | |
488 VTW(0, 1), | |
489 VTW(0, 2), | |
490 VTW(0, 3), | |
491 VTW(0, 4), | |
492 VTW(0, 5), | |
493 VTW(0, 6), | |
494 VTW(0, 7), | |
495 VTW(0, 8), | |
496 VTW(0, 9), | |
497 VTW(0, 10), | |
498 VTW(0, 11), | |
499 VTW(0, 12), | |
500 VTW(0, 13), | |
501 VTW(0, 14), | |
502 VTW(0, 15), | |
503 VTW(0, 16), | |
504 VTW(0, 17), | |
505 VTW(0, 18), | |
506 VTW(0, 19), | |
507 VTW(0, 20), | |
508 VTW(0, 21), | |
509 VTW(0, 22), | |
510 VTW(0, 23), | |
511 VTW(0, 24), | |
512 {TW_NEXT, VL, 0} | |
513 }; | |
514 | |
515 static const ct_desc desc = { 25, XSIMD_STRING("t2bv_25"), twinstr, &GENUS, {67, 60, 181, 0}, 0, 0, 0 }; | |
516 | |
517 void XSIMD(codelet_t2bv_25) (planner *p) { | |
518 X(kdft_dit_register) (p, t2bv_25, &desc); | |
519 } | |
520 #else /* HAVE_FMA */ | |
521 | |
522 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 25 -name t2bv_25 -include t2b.h -sign 1 */ | |
523 | |
524 /* | |
525 * This function contains 248 FP additions, 188 FP multiplications, | |
526 * (or, 171 additions, 111 multiplications, 77 fused multiply/add), | |
527 * 100 stack variables, 40 constants, and 50 memory accesses | |
528 */ | |
529 #include "t2b.h" | |
530 | |
531 static void t2bv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
532 { | |
533 DVK(KP497379774, +0.497379774329709576484567492012895936835134813); | |
534 DVK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
535 DVK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
536 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
537 DVK(KP809016994, +0.809016994374947424102293417182819058860154590); | |
538 DVK(KP309016994, +0.309016994374947424102293417182819058860154590); | |
539 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483); | |
540 DVK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
541 DVK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
542 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
543 DVK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
544 DVK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
545 DVK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
546 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691); | |
547 DVK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
548 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
549 DVK(KP851558583, +0.851558583130145297725004891488503407959946084); | |
550 DVK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
551 DVK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
552 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418); | |
553 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328); | |
554 DVK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
555 DVK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
556 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344); | |
557 DVK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
558 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
559 DVK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
560 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596); | |
561 DVK(KP125581039, +0.125581039058626752152356449131262266244969664); | |
562 DVK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
563 DVK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
564 DVK(KP250666467, +0.250666467128608490746237519633017587885836494); | |
565 DVK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
566 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465); | |
567 DVK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
568 DVK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
569 DVK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
570 DVK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
571 DVK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
572 DVK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
573 { | |
574 INT m; | |
575 R *x; | |
576 x = ii; | |
577 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)) { | |
578 V T1A, T1z, T1R, T1S, T1B, T1C, T1Q, T2L, T1l, T2v, T1i, T3e, T2u, Tb, T2i; | |
579 V Tj, T3b, T2h, Tv, T2k, TD, T3a, T2l, T11, T2s, TY, T3d, T2r; | |
580 { | |
581 V T1v, T1x, T1y, T1q, T1s, T1t, T1P; | |
582 T1A = LD(&(x[0]), ms, &(x[0])); | |
583 { | |
584 V T1u, T1w, T1p, T1r; | |
585 T1u = LD(&(x[WS(rs, 10)]), ms, &(x[0])); | |
586 T1v = BYTW(&(W[TWVL * 18]), T1u); | |
587 T1w = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)])); | |
588 T1x = BYTW(&(W[TWVL * 28]), T1w); | |
589 T1y = VADD(T1v, T1x); | |
590 T1p = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)])); | |
591 T1q = BYTW(&(W[TWVL * 8]), T1p); | |
592 T1r = LD(&(x[WS(rs, 20)]), ms, &(x[0])); | |
593 T1s = BYTW(&(W[TWVL * 38]), T1r); | |
594 T1t = VADD(T1q, T1s); | |
595 } | |
596 T1z = VMUL(LDK(KP559016994), VSUB(T1t, T1y)); | |
597 T1R = VSUB(T1v, T1x); | |
598 T1S = VMUL(LDK(KP587785252), T1R); | |
599 T1B = VADD(T1t, T1y); | |
600 T1C = VFNMS(LDK(KP250000000), T1B, T1A); | |
601 T1P = VSUB(T1q, T1s); | |
602 T1Q = VMUL(LDK(KP951056516), T1P); | |
603 T2L = VMUL(LDK(KP587785252), T1P); | |
604 } | |
605 { | |
606 V T1f, T19, T1b, T1c, T14, T16, T17, T1e; | |
607 T1e = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)])); | |
608 T1f = BYTW(&(W[TWVL * 4]), T1e); | |
609 { | |
610 V T18, T1a, T13, T15; | |
611 T18 = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)])); | |
612 T19 = BYTW(&(W[TWVL * 24]), T18); | |
613 T1a = LD(&(x[WS(rs, 18)]), ms, &(x[0])); | |
614 T1b = BYTW(&(W[TWVL * 34]), T1a); | |
615 T1c = VADD(T19, T1b); | |
616 T13 = LD(&(x[WS(rs, 8)]), ms, &(x[0])); | |
617 T14 = BYTW(&(W[TWVL * 14]), T13); | |
618 T15 = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)])); | |
619 T16 = BYTW(&(W[TWVL * 44]), T15); | |
620 T17 = VADD(T14, T16); | |
621 } | |
622 { | |
623 V T1j, T1k, T1d, T1g, T1h; | |
624 T1j = VSUB(T14, T16); | |
625 T1k = VSUB(T19, T1b); | |
626 T1l = VFMA(LDK(KP475528258), T1j, VMUL(LDK(KP293892626), T1k)); | |
627 T2v = VFNMS(LDK(KP475528258), T1k, VMUL(LDK(KP293892626), T1j)); | |
628 T1d = VMUL(LDK(KP559016994), VSUB(T17, T1c)); | |
629 T1g = VADD(T17, T1c); | |
630 T1h = VFNMS(LDK(KP250000000), T1g, T1f); | |
631 T1i = VADD(T1d, T1h); | |
632 T3e = VADD(T1f, T1g); | |
633 T2u = VSUB(T1h, T1d); | |
634 } | |
635 } | |
636 { | |
637 V Tg, T7, T9, Td, T2, T4, Tc, Tf; | |
638 Tf = LD(&(x[WS(rs, 4)]), ms, &(x[0])); | |
639 Tg = BYTW(&(W[TWVL * 6]), Tf); | |
640 { | |
641 V T6, T8, T1, T3; | |
642 T6 = LD(&(x[WS(rs, 14)]), ms, &(x[0])); | |
643 T7 = BYTW(&(W[TWVL * 26]), T6); | |
644 T8 = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)])); | |
645 T9 = BYTW(&(W[TWVL * 36]), T8); | |
646 Td = VADD(T7, T9); | |
647 T1 = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)])); | |
648 T2 = BYTW(&(W[TWVL * 16]), T1); | |
649 T3 = LD(&(x[WS(rs, 24)]), ms, &(x[0])); | |
650 T4 = BYTW(&(W[TWVL * 46]), T3); | |
651 Tc = VADD(T2, T4); | |
652 } | |
653 { | |
654 V T5, Ta, Te, Th, Ti; | |
655 T5 = VSUB(T2, T4); | |
656 Ta = VSUB(T7, T9); | |
657 Tb = VFMA(LDK(KP475528258), T5, VMUL(LDK(KP293892626), Ta)); | |
658 T2i = VFNMS(LDK(KP475528258), Ta, VMUL(LDK(KP293892626), T5)); | |
659 Te = VMUL(LDK(KP559016994), VSUB(Tc, Td)); | |
660 Th = VADD(Tc, Td); | |
661 Ti = VFNMS(LDK(KP250000000), Th, Tg); | |
662 Tj = VADD(Te, Ti); | |
663 T3b = VADD(Tg, Th); | |
664 T2h = VSUB(Ti, Te); | |
665 } | |
666 } | |
667 { | |
668 V TA, Tr, Tt, Tx, Tm, To, Tw, Tz; | |
669 Tz = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)])); | |
670 TA = BYTW(&(W[0]), Tz); | |
671 { | |
672 V Tq, Ts, Tl, Tn; | |
673 Tq = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)])); | |
674 Tr = BYTW(&(W[TWVL * 20]), Tq); | |
675 Ts = LD(&(x[WS(rs, 16)]), ms, &(x[0])); | |
676 Tt = BYTW(&(W[TWVL * 30]), Ts); | |
677 Tx = VADD(Tr, Tt); | |
678 Tl = LD(&(x[WS(rs, 6)]), ms, &(x[0])); | |
679 Tm = BYTW(&(W[TWVL * 10]), Tl); | |
680 Tn = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)])); | |
681 To = BYTW(&(W[TWVL * 40]), Tn); | |
682 Tw = VADD(Tm, To); | |
683 } | |
684 { | |
685 V Tp, Tu, Ty, TB, TC; | |
686 Tp = VSUB(Tm, To); | |
687 Tu = VSUB(Tr, Tt); | |
688 Tv = VFMA(LDK(KP475528258), Tp, VMUL(LDK(KP293892626), Tu)); | |
689 T2k = VFNMS(LDK(KP475528258), Tu, VMUL(LDK(KP293892626), Tp)); | |
690 Ty = VMUL(LDK(KP559016994), VSUB(Tw, Tx)); | |
691 TB = VADD(Tw, Tx); | |
692 TC = VFNMS(LDK(KP250000000), TB, TA); | |
693 TD = VADD(Ty, TC); | |
694 T3a = VADD(TA, TB); | |
695 T2l = VSUB(TC, Ty); | |
696 } | |
697 } | |
698 { | |
699 V TV, TP, TR, TS, TK, TM, TN, TU; | |
700 TU = LD(&(x[WS(rs, 2)]), ms, &(x[0])); | |
701 TV = BYTW(&(W[TWVL * 2]), TU); | |
702 { | |
703 V TO, TQ, TJ, TL; | |
704 TO = LD(&(x[WS(rs, 12)]), ms, &(x[0])); | |
705 TP = BYTW(&(W[TWVL * 22]), TO); | |
706 TQ = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)])); | |
707 TR = BYTW(&(W[TWVL * 32]), TQ); | |
708 TS = VADD(TP, TR); | |
709 TJ = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)])); | |
710 TK = BYTW(&(W[TWVL * 12]), TJ); | |
711 TL = LD(&(x[WS(rs, 22)]), ms, &(x[0])); | |
712 TM = BYTW(&(W[TWVL * 42]), TL); | |
713 TN = VADD(TK, TM); | |
714 } | |
715 { | |
716 V TZ, T10, TT, TW, TX; | |
717 TZ = VSUB(TK, TM); | |
718 T10 = VSUB(TP, TR); | |
719 T11 = VFMA(LDK(KP475528258), TZ, VMUL(LDK(KP293892626), T10)); | |
720 T2s = VFNMS(LDK(KP475528258), T10, VMUL(LDK(KP293892626), TZ)); | |
721 TT = VMUL(LDK(KP559016994), VSUB(TN, TS)); | |
722 TW = VADD(TN, TS); | |
723 TX = VFNMS(LDK(KP250000000), TW, TV); | |
724 TY = VADD(TT, TX); | |
725 T3d = VADD(TV, TW); | |
726 T2r = VSUB(TX, TT); | |
727 } | |
728 } | |
729 { | |
730 V T3g, T3o, T3k, T3l, T3j, T3m, T3p, T3n; | |
731 { | |
732 V T3c, T3f, T3h, T3i; | |
733 T3c = VSUB(T3a, T3b); | |
734 T3f = VSUB(T3d, T3e); | |
735 T3g = VBYI(VFMA(LDK(KP951056516), T3c, VMUL(LDK(KP587785252), T3f))); | |
736 T3o = VBYI(VFNMS(LDK(KP951056516), T3f, VMUL(LDK(KP587785252), T3c))); | |
737 T3k = VADD(T1A, T1B); | |
738 T3h = VADD(T3a, T3b); | |
739 T3i = VADD(T3d, T3e); | |
740 T3l = VADD(T3h, T3i); | |
741 T3j = VMUL(LDK(KP559016994), VSUB(T3h, T3i)); | |
742 T3m = VFNMS(LDK(KP250000000), T3l, T3k); | |
743 } | |
744 ST(&(x[0]), VADD(T3k, T3l), ms, &(x[0])); | |
745 T3p = VSUB(T3m, T3j); | |
746 ST(&(x[WS(rs, 10)]), VADD(T3o, T3p), ms, &(x[0])); | |
747 ST(&(x[WS(rs, 15)]), VSUB(T3p, T3o), ms, &(x[WS(rs, 1)])); | |
748 T3n = VADD(T3j, T3m); | |
749 ST(&(x[WS(rs, 5)]), VADD(T3g, T3n), ms, &(x[WS(rs, 1)])); | |
750 ST(&(x[WS(rs, 20)]), VSUB(T3n, T3g), ms, &(x[0])); | |
751 } | |
752 { | |
753 V T2z, T2M, T2U, T2V, T2W, T34, T35, T36, T2X, T2Y, T2Z, T31, T32, T33, T2n; | |
754 V T2N, T2E, T2K, T2y, T2H, T2A, T2G, T38, T39; | |
755 T2z = VSUB(T1C, T1z); | |
756 T2M = VFNMS(LDK(KP951056516), T1R, T2L); | |
757 T2U = VFMA(LDK(KP1_369094211), T2k, VMUL(LDK(KP728968627), T2l)); | |
758 T2V = VFNMS(LDK(KP992114701), T2h, VMUL(LDK(KP250666467), T2i)); | |
759 T2W = VADD(T2U, T2V); | |
760 T34 = VFNMS(LDK(KP125581039), T2s, VMUL(LDK(KP998026728), T2r)); | |
761 T35 = VFMA(LDK(KP1_274847979), T2v, VMUL(LDK(KP770513242), T2u)); | |
762 T36 = VADD(T34, T35); | |
763 T2X = VFMA(LDK(KP1_996053456), T2s, VMUL(LDK(KP062790519), T2r)); | |
764 T2Y = VFNMS(LDK(KP637423989), T2u, VMUL(LDK(KP1_541026485), T2v)); | |
765 T2Z = VADD(T2X, T2Y); | |
766 T31 = VFNMS(LDK(KP1_457937254), T2k, VMUL(LDK(KP684547105), T2l)); | |
767 T32 = VFMA(LDK(KP1_984229402), T2i, VMUL(LDK(KP125333233), T2h)); | |
768 T33 = VADD(T31, T32); | |
769 { | |
770 V T2j, T2m, T2I, T2C, T2D, T2J; | |
771 T2j = VFNMS(LDK(KP851558583), T2i, VMUL(LDK(KP904827052), T2h)); | |
772 T2m = VFMA(LDK(KP1_752613360), T2k, VMUL(LDK(KP481753674), T2l)); | |
773 T2I = VADD(T2m, T2j); | |
774 T2C = VFMA(LDK(KP1_071653589), T2s, VMUL(LDK(KP844327925), T2r)); | |
775 T2D = VFMA(LDK(KP125581039), T2v, VMUL(LDK(KP998026728), T2u)); | |
776 T2J = VADD(T2C, T2D); | |
777 T2n = VSUB(T2j, T2m); | |
778 T2N = VADD(T2I, T2J); | |
779 T2E = VSUB(T2C, T2D); | |
780 T2K = VMUL(LDK(KP559016994), VSUB(T2I, T2J)); | |
781 } | |
782 { | |
783 V T2o, T2p, T2q, T2t, T2w, T2x; | |
784 T2o = VFNMS(LDK(KP963507348), T2k, VMUL(LDK(KP876306680), T2l)); | |
785 T2p = VFMA(LDK(KP1_809654104), T2i, VMUL(LDK(KP425779291), T2h)); | |
786 T2q = VSUB(T2o, T2p); | |
787 T2t = VFNMS(LDK(KP1_688655851), T2s, VMUL(LDK(KP535826794), T2r)); | |
788 T2w = VFNMS(LDK(KP1_996053456), T2v, VMUL(LDK(KP062790519), T2u)); | |
789 T2x = VADD(T2t, T2w); | |
790 T2y = VMUL(LDK(KP559016994), VSUB(T2q, T2x)); | |
791 T2H = VSUB(T2t, T2w); | |
792 T2A = VADD(T2q, T2x); | |
793 T2G = VADD(T2o, T2p); | |
794 } | |
795 { | |
796 V T2S, T2T, T30, T37; | |
797 T2S = VADD(T2z, T2A); | |
798 T2T = VBYI(VADD(T2M, T2N)); | |
799 ST(&(x[WS(rs, 23)]), VSUB(T2S, T2T), ms, &(x[WS(rs, 1)])); | |
800 ST(&(x[WS(rs, 2)]), VADD(T2S, T2T), ms, &(x[0])); | |
801 T30 = VADD(T2z, VADD(T2W, T2Z)); | |
802 T37 = VBYI(VSUB(VADD(T33, T36), T2M)); | |
803 ST(&(x[WS(rs, 22)]), VSUB(T30, T37), ms, &(x[0])); | |
804 ST(&(x[WS(rs, 3)]), VADD(T30, T37), ms, &(x[WS(rs, 1)])); | |
805 } | |
806 T38 = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T2U, T2V), VFMA(LDK(KP309016994), T33, VFNMS(LDK(KP809016994), T36, VMUL(LDK(KP587785252), VSUB(T2X, T2Y))))), T2M)); | |
807 T39 = VFMA(LDK(KP309016994), T2W, VFMA(LDK(KP951056516), VSUB(T32, T31), VFMA(LDK(KP587785252), VSUB(T35, T34), VFNMS(LDK(KP809016994), T2Z, T2z)))); | |
808 ST(&(x[WS(rs, 8)]), VADD(T38, T39), ms, &(x[0])); | |
809 ST(&(x[WS(rs, 17)]), VSUB(T39, T38), ms, &(x[WS(rs, 1)])); | |
810 { | |
811 V T2F, T2Q, T2P, T2R, T2B, T2O; | |
812 T2B = VFNMS(LDK(KP250000000), T2A, T2z); | |
813 T2F = VFMA(LDK(KP951056516), T2n, VADD(T2y, VFNMS(LDK(KP587785252), T2E, T2B))); | |
814 T2Q = VFMA(LDK(KP587785252), T2n, VFMA(LDK(KP951056516), T2E, VSUB(T2B, T2y))); | |
815 T2O = VFNMS(LDK(KP250000000), T2N, T2M); | |
816 T2P = VBYI(VADD(VFMA(LDK(KP951056516), T2G, VMUL(LDK(KP587785252), T2H)), VADD(T2K, T2O))); | |
817 T2R = VBYI(VADD(VFNMS(LDK(KP951056516), T2H, VMUL(LDK(KP587785252), T2G)), VSUB(T2O, T2K))); | |
818 ST(&(x[WS(rs, 18)]), VSUB(T2F, T2P), ms, &(x[0])); | |
819 ST(&(x[WS(rs, 12)]), VADD(T2Q, T2R), ms, &(x[0])); | |
820 ST(&(x[WS(rs, 7)]), VADD(T2F, T2P), ms, &(x[WS(rs, 1)])); | |
821 ST(&(x[WS(rs, 13)]), VSUB(T2Q, T2R), ms, &(x[WS(rs, 1)])); | |
822 } | |
823 } | |
824 { | |
825 V T1D, T1T, T21, T22, T23, T2b, T2c, T2d, T24, T25, T26, T28, T29, T2a, TF; | |
826 V T1U, T1I, T1O, T1o, T1L, T1E, T1K, T2f, T2g; | |
827 T1D = VADD(T1z, T1C); | |
828 T1T = VADD(T1Q, T1S); | |
829 T21 = VFMA(LDK(KP1_688655851), Tv, VMUL(LDK(KP535826794), TD)); | |
830 T22 = VFMA(LDK(KP1_541026485), Tb, VMUL(LDK(KP637423989), Tj)); | |
831 T23 = VSUB(T21, T22); | |
832 T2b = VFMA(LDK(KP851558583), T11, VMUL(LDK(KP904827052), TY)); | |
833 T2c = VFMA(LDK(KP1_984229402), T1l, VMUL(LDK(KP125333233), T1i)); | |
834 T2d = VADD(T2b, T2c); | |
835 T24 = VFNMS(LDK(KP425779291), TY, VMUL(LDK(KP1_809654104), T11)); | |
836 T25 = VFNMS(LDK(KP992114701), T1i, VMUL(LDK(KP250666467), T1l)); | |
837 T26 = VADD(T24, T25); | |
838 T28 = VFNMS(LDK(KP1_071653589), Tv, VMUL(LDK(KP844327925), TD)); | |
839 T29 = VFNMS(LDK(KP770513242), Tj, VMUL(LDK(KP1_274847979), Tb)); | |
840 T2a = VADD(T28, T29); | |
841 { | |
842 V Tk, TE, T1M, T1G, T1H, T1N; | |
843 Tk = VFMA(LDK(KP1_071653589), Tb, VMUL(LDK(KP844327925), Tj)); | |
844 TE = VFMA(LDK(KP1_937166322), Tv, VMUL(LDK(KP248689887), TD)); | |
845 T1M = VADD(TE, Tk); | |
846 T1G = VFMA(LDK(KP1_752613360), T11, VMUL(LDK(KP481753674), TY)); | |
847 T1H = VFMA(LDK(KP1_457937254), T1l, VMUL(LDK(KP684547105), T1i)); | |
848 T1N = VADD(T1G, T1H); | |
849 TF = VSUB(Tk, TE); | |
850 T1U = VADD(T1M, T1N); | |
851 T1I = VSUB(T1G, T1H); | |
852 T1O = VMUL(LDK(KP559016994), VSUB(T1M, T1N)); | |
853 } | |
854 { | |
855 V TG, TH, TI, T12, T1m, T1n; | |
856 TG = VFNMS(LDK(KP497379774), Tv, VMUL(LDK(KP968583161), TD)); | |
857 TH = VFNMS(LDK(KP1_688655851), Tb, VMUL(LDK(KP535826794), Tj)); | |
858 TI = VADD(TG, TH); | |
859 T12 = VFNMS(LDK(KP963507348), T11, VMUL(LDK(KP876306680), TY)); | |
860 T1m = VFNMS(LDK(KP1_369094211), T1l, VMUL(LDK(KP728968627), T1i)); | |
861 T1n = VADD(T12, T1m); | |
862 T1o = VMUL(LDK(KP559016994), VSUB(TI, T1n)); | |
863 T1L = VSUB(T12, T1m); | |
864 T1E = VADD(TI, T1n); | |
865 T1K = VSUB(TG, TH); | |
866 } | |
867 { | |
868 V T1Z, T20, T27, T2e; | |
869 T1Z = VADD(T1D, T1E); | |
870 T20 = VBYI(VADD(T1T, T1U)); | |
871 ST(&(x[WS(rs, 24)]), VSUB(T1Z, T20), ms, &(x[0])); | |
872 ST(&(x[WS(rs, 1)]), VADD(T1Z, T20), ms, &(x[WS(rs, 1)])); | |
873 T27 = VADD(T1D, VADD(T23, T26)); | |
874 T2e = VBYI(VSUB(VADD(T2a, T2d), T1T)); | |
875 ST(&(x[WS(rs, 21)]), VSUB(T27, T2e), ms, &(x[WS(rs, 1)])); | |
876 ST(&(x[WS(rs, 4)]), VADD(T27, T2e), ms, &(x[0])); | |
877 } | |
878 T2f = VBYI(VSUB(VFMA(LDK(KP309016994), T2a, VFMA(LDK(KP951056516), VADD(T21, T22), VFNMS(LDK(KP809016994), T2d, VMUL(LDK(KP587785252), VSUB(T24, T25))))), T1T)); | |
879 T2g = VFMA(LDK(KP951056516), VSUB(T29, T28), VFMA(LDK(KP309016994), T23, VFMA(LDK(KP587785252), VSUB(T2c, T2b), VFNMS(LDK(KP809016994), T26, T1D)))); | |
880 ST(&(x[WS(rs, 9)]), VADD(T2f, T2g), ms, &(x[WS(rs, 1)])); | |
881 ST(&(x[WS(rs, 16)]), VSUB(T2g, T2f), ms, &(x[0])); | |
882 { | |
883 V T1J, T1X, T1W, T1Y, T1F, T1V; | |
884 T1F = VFNMS(LDK(KP250000000), T1E, T1D); | |
885 T1J = VFMA(LDK(KP951056516), TF, VADD(T1o, VFNMS(LDK(KP587785252), T1I, T1F))); | |
886 T1X = VFMA(LDK(KP587785252), TF, VFMA(LDK(KP951056516), T1I, VSUB(T1F, T1o))); | |
887 T1V = VFNMS(LDK(KP250000000), T1U, T1T); | |
888 T1W = VBYI(VADD(VFMA(LDK(KP951056516), T1K, VMUL(LDK(KP587785252), T1L)), VADD(T1O, T1V))); | |
889 T1Y = VBYI(VADD(VFNMS(LDK(KP951056516), T1L, VMUL(LDK(KP587785252), T1K)), VSUB(T1V, T1O))); | |
890 ST(&(x[WS(rs, 19)]), VSUB(T1J, T1W), ms, &(x[WS(rs, 1)])); | |
891 ST(&(x[WS(rs, 11)]), VADD(T1X, T1Y), ms, &(x[WS(rs, 1)])); | |
892 ST(&(x[WS(rs, 6)]), VADD(T1J, T1W), ms, &(x[0])); | |
893 ST(&(x[WS(rs, 14)]), VSUB(T1X, T1Y), ms, &(x[0])); | |
894 } | |
895 } | |
896 } | |
897 } | |
898 VLEAVE(); | |
899 } | |
900 | |
901 static const tw_instr twinstr[] = { | |
902 VTW(0, 1), | |
903 VTW(0, 2), | |
904 VTW(0, 3), | |
905 VTW(0, 4), | |
906 VTW(0, 5), | |
907 VTW(0, 6), | |
908 VTW(0, 7), | |
909 VTW(0, 8), | |
910 VTW(0, 9), | |
911 VTW(0, 10), | |
912 VTW(0, 11), | |
913 VTW(0, 12), | |
914 VTW(0, 13), | |
915 VTW(0, 14), | |
916 VTW(0, 15), | |
917 VTW(0, 16), | |
918 VTW(0, 17), | |
919 VTW(0, 18), | |
920 VTW(0, 19), | |
921 VTW(0, 20), | |
922 VTW(0, 21), | |
923 VTW(0, 22), | |
924 VTW(0, 23), | |
925 VTW(0, 24), | |
926 {TW_NEXT, VL, 0} | |
927 }; | |
928 | |
929 static const ct_desc desc = { 25, XSIMD_STRING("t2bv_25"), twinstr, &GENUS, {171, 111, 77, 0}, 0, 0, 0 }; | |
930 | |
931 void XSIMD(codelet_t2bv_25) (planner *p) { | |
932 X(kdft_dit_register) (p, t2bv_25, &desc); | |
933 } | |
934 #endif /* HAVE_FMA */ |