Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/rdft/scalar/r2cf/r2cfII_25.c @ 167:bd3cc4d1df30
Add FFTW 3.3.8 source, and a Linux build
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Tue, 19 Nov 2019 14:52:55 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
166:cbd6d7e562c7 | 167:bd3cc4d1df30 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-14 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Thu May 24 08:06:49 EDT 2018 */ | |
23 | |
24 #include "rdft/codelet-rdft.h" | |
25 | |
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA) | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cfII_25 -dft-II -include rdft/scalar/r2cfII.h */ | |
29 | |
30 /* | |
31 * This function contains 212 FP additions, 177 FP multiplications, | |
32 * (or, 47 additions, 12 multiplications, 165 fused multiply/add), | |
33 * 131 stack variables, 67 constants, and 50 memory accesses | |
34 */ | |
35 #include "rdft/scalar/r2cfII.h" | |
36 | |
37 static void r2cfII_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP876091699, +0.876091699473550838204498029706869638173524346); | |
40 DK(KP792626838, +0.792626838241819413632131824093538848057784557); | |
41 DK(KP690668130, +0.690668130712929053565177988380887884042527623); | |
42 DK(KP809385824, +0.809385824416008241660603814668679683846476688); | |
43 DK(KP860541664, +0.860541664367944677098261680920518816412804187); | |
44 DK(KP681693190, +0.681693190061530575150324149145440022633095390); | |
45 DK(KP560319534, +0.560319534973832390111614715371676131169633784); | |
46 DK(KP237294955, +0.237294955877110315393888866460840817927895961); | |
47 DK(KP897376177, +0.897376177523557693138608077137219684419427330); | |
48 DK(KP584303379, +0.584303379262766050358567120694562180043261496); | |
49 DK(KP653711795, +0.653711795629256296299985401753308353544378892); | |
50 DK(KP997675361, +0.997675361079556513670859573984492383596555031); | |
51 DK(KP645989928, +0.645989928319777763844272876603899665178054552); | |
52 DK(KP591287873, +0.591287873858343558732323717242372865934480959); | |
53 DK(KP952936919, +0.952936919628306576880750665357914584765951388); | |
54 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
55 DK(KP956723877, +0.956723877038460305821989399535483155872969262); | |
56 DK(KP945422727, +0.945422727388575946270360266328811958657216298); | |
57 DK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
58 DK(KP772036680, +0.772036680810363904029489473607579825330539880); | |
59 DK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
60 DK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
61 DK(KP242145790, +0.242145790282157779872542093866183953459003101); | |
62 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
63 DK(KP999754674, +0.999754674276473633366203429228112409535557487); | |
64 DK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
65 DK(KP916574801, +0.916574801383451584742370439148878693530976769); | |
66 DK(KP829049696, +0.829049696159252993975487806364305442437946767); | |
67 DK(KP831864738, +0.831864738706457140726048799369896829771167132); | |
68 DK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
69 DK(KP949179823, +0.949179823508441261575555465843363271711583843); | |
70 DK(KP669429328, +0.669429328479476605641803240971985825917022098); | |
71 DK(KP262346850, +0.262346850930607871785420028382979691334784273); | |
72 DK(KP923225144, +0.923225144846402650453449441572664695995209956); | |
73 DK(KP906616052, +0.906616052148196230441134447086066874408359177); | |
74 DK(KP921078979, +0.921078979742360627699756128143719920817673854); | |
75 DK(KP982009705, +0.982009705009746369461829878184175962711969869); | |
76 DK(KP845997307, +0.845997307939530944175097360758058292389769300); | |
77 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
78 DK(KP803003575, +0.803003575438660414833440593570376004635464850); | |
79 DK(KP763583905, +0.763583905359130246362948588764067237776594106); | |
80 DK(KP248028675, +0.248028675328619457762448260696444630363259177); | |
81 DK(KP904508497, +0.904508497187473712051146708591409529430077295); | |
82 DK(KP894834959, +0.894834959464455102997960030820114611498661386); | |
83 DK(KP958953096, +0.958953096729998668045963838399037225970891871); | |
84 DK(KP867381224, +0.867381224396525206773171885031575671309956167); | |
85 DK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
86 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
87 DK(KP869845200, +0.869845200362138853122720822420327157933056305); | |
88 DK(KP120146378, +0.120146378570687701782758537356596213647956445); | |
89 DK(KP132830569, +0.132830569247582714407653942074819768844536507); | |
90 DK(KP786782374, +0.786782374965295178365099601674911834788448471); | |
91 DK(KP893101515, +0.893101515366181661711202267938416198338079437); | |
92 DK(KP987388751, +0.987388751065621252324603216482382109400433949); | |
93 DK(KP244189809, +0.244189809627953270309879511234821255780225091); | |
94 DK(KP269969613, +0.269969613759572083574752974412347470060951301); | |
95 DK(KP494780565, +0.494780565770515410344588413655324772219443730); | |
96 DK(KP066152395, +0.066152395967733048213034281011006031460903353); | |
97 DK(KP059835404, +0.059835404262124915169548397419498386427871950); | |
98 DK(KP447533225, +0.447533225982656890041886979663652563063114397); | |
99 DK(KP522847744, +0.522847744331509716623755382187077770911012542); | |
100 DK(KP667278218, +0.667278218140296670899089292254759909713898805); | |
101 DK(KP603558818, +0.603558818296015001454675132653458027918768137); | |
102 DK(KP578046249, +0.578046249379945007321754579646815604023525655); | |
103 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
104 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
105 DK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
106 { | |
107 INT i; | |
108 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(100, rs), MAKE_VOLATILE_STRIDE(100, csr), MAKE_VOLATILE_STRIDE(100, csi)) { | |
109 E T2v, TJ, T2A, T1K, T2y, T2z, TB, T15, T2d, T2l, T1g, T1s, T1N, T21, T1D; | |
110 E T9, TQ, T2g, T2o, T1j, T1u, T1X, T25, T1z, Ti, TX, T2f, T2p, T1k, T1v; | |
111 E T1U, T24, T1A, Ts, T1c, T2c, T2k, T1h, T1r, T1Q, T22, T1C, Tj, TC; | |
112 { | |
113 E TI, T2x, TF, T2w; | |
114 T2v = R0[0]; | |
115 { | |
116 E TG, TH, TD, TE; | |
117 TG = R0[WS(rs, 10)]; | |
118 TH = R1[WS(rs, 2)]; | |
119 TI = TG + TH; | |
120 T2x = TG - TH; | |
121 TD = R0[WS(rs, 5)]; | |
122 TE = R1[WS(rs, 7)]; | |
123 TF = TD + TE; | |
124 T2w = TD - TE; | |
125 } | |
126 TJ = FMA(KP618033988, TI, TF); | |
127 T2A = T2w - T2x; | |
128 T1K = FNMS(KP618033988, TF, TI); | |
129 T2y = T2w + T2x; | |
130 T2z = FNMS(KP250000000, T2y, T2v); | |
131 } | |
132 { | |
133 E Tt, TA, T13, TZ, T10; | |
134 Tt = R0[WS(rs, 2)]; | |
135 { | |
136 E Tu, Tv, Tw, Tx, Ty, Tz; | |
137 Tu = R0[WS(rs, 7)]; | |
138 Tv = R1[WS(rs, 9)]; | |
139 Tw = Tu - Tv; | |
140 Tx = R0[WS(rs, 12)]; | |
141 Ty = R1[WS(rs, 4)]; | |
142 Tz = Tx - Ty; | |
143 TA = Tw + Tz; | |
144 T13 = Tz - Tw; | |
145 TZ = Tu + Tv; | |
146 T10 = Tx + Ty; | |
147 } | |
148 TB = Tt + TA; | |
149 { | |
150 E T11, T1M, T14, T1L, T12; | |
151 T11 = FMA(KP618033988, T10, TZ); | |
152 T1M = FNMS(KP618033988, TZ, T10); | |
153 T12 = FNMS(KP250000000, TA, Tt); | |
154 T14 = FNMS(KP559016994, T13, T12); | |
155 T1L = FMA(KP559016994, T13, T12); | |
156 T15 = FMA(KP578046249, T14, T11); | |
157 T2d = FNMS(KP603558818, T1M, T1L); | |
158 T2l = FMA(KP667278218, T1L, T1M); | |
159 T1g = FNMS(KP522847744, T11, T14); | |
160 T1s = FMA(KP447533225, T11, T14); | |
161 T1N = FMA(KP059835404, T1M, T1L); | |
162 T21 = FNMS(KP066152395, T1L, T1M); | |
163 T1D = FNMS(KP494780565, T14, T11); | |
164 } | |
165 } | |
166 { | |
167 E T1, T8, TO, TK, TL; | |
168 T1 = R0[WS(rs, 1)]; | |
169 { | |
170 E T2, T3, T4, T5, T6, T7; | |
171 T2 = R0[WS(rs, 6)]; | |
172 T3 = R1[WS(rs, 8)]; | |
173 T4 = T2 - T3; | |
174 T5 = R0[WS(rs, 11)]; | |
175 T6 = R1[WS(rs, 3)]; | |
176 T7 = T5 - T6; | |
177 T8 = T4 + T7; | |
178 TO = T4 - T7; | |
179 TK = T2 + T3; | |
180 TL = T5 + T6; | |
181 } | |
182 T9 = T1 + T8; | |
183 { | |
184 E TM, T1V, TP, T1W, TN; | |
185 TM = FMA(KP618033988, TL, TK); | |
186 T1V = FNMS(KP618033988, TK, TL); | |
187 TN = FNMS(KP250000000, T8, T1); | |
188 TP = FMA(KP559016994, TO, TN); | |
189 T1W = FNMS(KP559016994, TO, TN); | |
190 TQ = FMA(KP269969613, TP, TM); | |
191 T2g = FNMS(KP578046249, T1W, T1V); | |
192 T2o = FMA(KP522847744, T1V, T1W); | |
193 T1j = FNMS(KP244189809, TM, TP); | |
194 T1u = FNMS(KP603558818, TM, TP); | |
195 T1X = FMA(KP987388751, T1W, T1V); | |
196 T25 = FNMS(KP893101515, T1V, T1W); | |
197 T1z = FMA(KP667278218, TP, TM); | |
198 } | |
199 } | |
200 { | |
201 E Th, Tg, TV, TS, TU; | |
202 Th = R0[WS(rs, 4)]; | |
203 { | |
204 E Ta, Tb, Tc, Td, Te, Tf; | |
205 Ta = R0[WS(rs, 9)]; | |
206 Tb = R1[WS(rs, 11)]; | |
207 Tc = Ta - Tb; | |
208 Td = R1[WS(rs, 6)]; | |
209 Te = R1[WS(rs, 1)]; | |
210 Tf = Td + Te; | |
211 Tg = Tc - Tf; | |
212 TV = Te - Td; | |
213 TS = Tc + Tf; | |
214 TU = Ta + Tb; | |
215 } | |
216 Ti = Tg + Th; | |
217 { | |
218 E TW, T1S, TT, T1T, TR; | |
219 TW = FNMS(KP618033988, TV, TU); | |
220 T1S = FMA(KP618033988, TU, TV); | |
221 TR = FNMS(KP250000000, Tg, Th); | |
222 TT = FMA(KP559016994, TS, TR); | |
223 T1T = FNMS(KP559016994, TS, TR); | |
224 TX = FMA(KP603558818, TW, TT); | |
225 T2f = FNMS(KP447533225, T1S, T1T); | |
226 T2p = FMA(KP494780565, T1T, T1S); | |
227 T1k = FNMS(KP667278218, TT, TW); | |
228 T1v = FNMS(KP786782374, TW, TT); | |
229 T1U = FMA(KP132830569, T1T, T1S); | |
230 T24 = FNMS(KP120146378, T1S, T1T); | |
231 T1A = FMA(KP869845200, TT, TW); | |
232 } | |
233 } | |
234 { | |
235 E Tk, Tr, T1a, T16, T17; | |
236 Tk = R0[WS(rs, 3)]; | |
237 { | |
238 E Tl, Tm, Tn, To, Tp, Tq; | |
239 Tl = R0[WS(rs, 8)]; | |
240 Tm = R1[WS(rs, 10)]; | |
241 Tn = Tl - Tm; | |
242 To = R1[0]; | |
243 Tp = R1[WS(rs, 5)]; | |
244 Tq = To + Tp; | |
245 Tr = Tn - Tq; | |
246 T1a = Tn + Tq; | |
247 T16 = Tl + Tm; | |
248 T17 = Tp - To; | |
249 } | |
250 Ts = Tk + Tr; | |
251 { | |
252 E T18, T1P, T1b, T1O, T19; | |
253 T18 = FMA(KP618033988, T17, T16); | |
254 T1P = FNMS(KP618033988, T16, T17); | |
255 T19 = FNMS(KP250000000, Tr, Tk); | |
256 T1b = FMA(KP559016994, T1a, T19); | |
257 T1O = FNMS(KP559016994, T1a, T19); | |
258 T1c = FMA(KP987388751, T1b, T18); | |
259 T2c = FNMS(KP059835404, T1P, T1O); | |
260 T2k = FMA(KP066152395, T1O, T1P); | |
261 T1h = FNMS(KP893101515, T18, T1b); | |
262 T1r = FMA(KP132830569, T1b, T18); | |
263 T1Q = FNMS(KP786782374, T1P, T1O); | |
264 T22 = FMA(KP869845200, T1O, T1P); | |
265 T1C = FNMS(KP120146378, T18, T1b); | |
266 } | |
267 } | |
268 Tj = T9 - Ti; | |
269 TC = Ts - TB; | |
270 Ci[WS(csi, 2)] = -(KP951056516 * (FNMS(KP618033988, TC, Tj))); | |
271 Ci[WS(csi, 7)] = KP951056516 * (FMA(KP618033988, Tj, TC)); | |
272 { | |
273 E T3l, T3o, T3q, T3m, T3n, T3p; | |
274 T3l = T2v + T2y; | |
275 T3m = T9 + Ti; | |
276 T3n = TB + Ts; | |
277 T3o = T3m + T3n; | |
278 T3q = T3m - T3n; | |
279 Cr[WS(csr, 12)] = T3o + T3l; | |
280 T3p = FNMS(KP250000000, T3o, T3l); | |
281 Cr[WS(csr, 2)] = FMA(KP559016994, T3q, T3p); | |
282 Cr[WS(csr, 7)] = FNMS(KP559016994, T3q, T3p); | |
283 } | |
284 { | |
285 E T1B, T1E, T1x, T1I, T1G, T1t, T1w, T1F, T1y, T1J, T1H; | |
286 T1B = FMA(KP912575812, T1A, T1z); | |
287 T1E = FMA(KP867381224, T1D, T1C); | |
288 T1t = FMA(KP958953096, T1s, T1r); | |
289 T1w = FNMS(KP912575812, T1v, T1u); | |
290 T1F = FNMS(KP894834959, T1w, T1t); | |
291 T1x = FMA(KP894834959, T1w, T1t); | |
292 T1I = FNMS(KP894834959, T1B, T1F); | |
293 T1G = FNMS(KP904508497, T1F, T1E); | |
294 T1y = FMA(KP248028675, T1x, TJ); | |
295 T1J = FMA(KP559016994, T1I, T1E); | |
296 T1H = FMA(KP763583905, T1G, T1B); | |
297 Ci[WS(csi, 4)] = KP951056516 * (FNMS(KP803003575, T1H, T1y)); | |
298 Ci[WS(csi, 9)] = KP951056516 * (FNMS(KP992114701, T1J, T1y)); | |
299 } | |
300 { | |
301 E T2m, T2q, T2i, T2t, T2r, T2e, T2h, T2n, T2j, T2u, T2s; | |
302 T2m = FNMS(KP845997307, T2l, T2k); | |
303 T2q = FMA(KP982009705, T2p, T2o); | |
304 T2e = FMA(KP845997307, T2d, T2c); | |
305 T2h = FNMS(KP921078979, T2g, T2f); | |
306 T2n = FNMS(KP906616052, T2h, T2e); | |
307 T2i = FMA(KP906616052, T2h, T2e); | |
308 T2t = T2m + T2n; | |
309 T2r = FNMS(KP923225144, T2q, T2n); | |
310 T2j = FMA(KP262346850, T2i, T1K); | |
311 T2u = FNMS(KP669429328, T2t, T2q); | |
312 T2s = FNMS(KP618033988, T2r, T2m); | |
313 Ci[WS(csi, 8)] = KP951056516 * (FMA(KP949179823, T2s, T2j)); | |
314 Ci[WS(csi, 3)] = KP951056516 * (FNMS(KP876306680, T2u, T2j)); | |
315 } | |
316 { | |
317 E T1i, T1l, T1e, T1p, T1n, TY, T1d, T1m, T1f, T1q, T1o; | |
318 T1i = FNMS(KP831864738, T1h, T1g); | |
319 T1l = FMA(KP829049696, T1k, T1j); | |
320 TY = FMA(KP916574801, TX, TQ); | |
321 T1d = FMA(KP831864738, T1c, T15); | |
322 T1m = FNMS(KP904730450, T1d, TY); | |
323 T1e = FMA(KP904730450, T1d, TY); | |
324 T1p = FNMS(KP999754674, T1m, T1i); | |
325 T1n = FNMS(KP904508497, T1m, T1l); | |
326 Ci[0] = -(KP951056516 * (FMA(KP968583161, T1e, TJ))); | |
327 T1f = FNMS(KP242145790, T1e, TJ); | |
328 T1q = FMA(KP559154169, T1p, T1l); | |
329 T1o = FNMS(KP683113946, T1n, T1i); | |
330 Ci[WS(csi, 5)] = -(KP951056516 * (FNMS(KP876306680, T1o, T1f))); | |
331 Ci[WS(csi, 10)] = -(KP951056516 * (FNMS(KP968583161, T1q, T1f))); | |
332 } | |
333 { | |
334 E T23, T26, T1Z, T2a, T28, T1R, T1Y, T27, T20, T2b, T29; | |
335 T23 = FNMS(KP772036680, T22, T21); | |
336 T26 = FMA(KP734762448, T25, T24); | |
337 T1R = FMA(KP772036680, T1Q, T1N); | |
338 T1Y = FMA(KP734762448, T1X, T1U); | |
339 T27 = FNMS(KP945422727, T1Y, T1R); | |
340 T1Z = FMA(KP945422727, T1Y, T1R); | |
341 T2a = T27 - T23; | |
342 T28 = FMA(KP956723877, T27, T26); | |
343 Ci[WS(csi, 1)] = -(KP998026728 * (FMA(KP952936919, T1K, T1Z))); | |
344 T20 = FNMS(KP262346850, T1Z, T1K); | |
345 T2b = FMA(KP591287873, T2a, T26); | |
346 T29 = FMA(KP645989928, T28, T23); | |
347 Ci[WS(csi, 6)] = -(KP951056516 * (FMA(KP949179823, T29, T20))); | |
348 Ci[WS(csi, 11)] = -(KP951056516 * (FNMS(KP992114701, T2b, T20))); | |
349 } | |
350 { | |
351 E T2Y, T33, T31, T38, T36, T3e, T3f, T3c, T3j, T3h, T3a, T3b, T3g; | |
352 T2Y = FNMS(KP559016994, T2A, T2z); | |
353 T33 = FNMS(KP772036680, T1Q, T1N); | |
354 { | |
355 E T34, T2Z, T30, T35; | |
356 T34 = FNMS(KP734762448, T1X, T1U); | |
357 T2Z = FNMS(KP734762448, T25, T24); | |
358 T30 = FMA(KP772036680, T22, T21); | |
359 T35 = FNMS(KP956723877, T30, T2Z); | |
360 T31 = FMA(KP956723877, T30, T2Z); | |
361 T38 = FMA(KP618033988, T35, T34); | |
362 T36 = T34 + T35; | |
363 } | |
364 T3e = FMA(KP921078979, T2g, T2f); | |
365 T3f = FNMS(KP845997307, T2d, T2c); | |
366 T3a = FMA(KP845997307, T2l, T2k); | |
367 T3b = FNMS(KP982009705, T2p, T2o); | |
368 T3g = FNMS(KP923225144, T3b, T3a); | |
369 T3c = FMA(KP923225144, T3b, T3a); | |
370 T3j = FNMS(KP997675361, T3g, T3e); | |
371 T3h = FNMS(KP904508497, T3g, T3f); | |
372 Cr[WS(csr, 1)] = FNMS(KP992114701, T31, T2Y); | |
373 { | |
374 E T32, T39, T37, T3d, T3k, T3i; | |
375 T32 = FMA(KP248028675, T31, T2Y); | |
376 T39 = FNMS(KP653711795, T33, T38); | |
377 T37 = FMA(KP584303379, T36, T33); | |
378 Cr[WS(csr, 6)] = FMA(KP949179823, T37, T32); | |
379 Cr[WS(csr, 11)] = FNMS(KP897376177, T39, T32); | |
380 T3d = FNMS(KP237294955, T3c, T2Y); | |
381 T3k = FNMS(KP560319534, T3j, T3f); | |
382 T3i = FMA(KP681693190, T3h, T3e); | |
383 Cr[WS(csr, 3)] = FMA(KP860541664, T3i, T3d); | |
384 Cr[WS(csr, 8)] = FMA(KP949179823, T3k, T3d); | |
385 } | |
386 } | |
387 { | |
388 E T2B, T2R, T2T, T2P, T2W, T2U, T2G, T2H, T2E, T2L, T2J; | |
389 T2B = FMA(KP559016994, T2A, T2z); | |
390 { | |
391 E T2N, T2O, T2S, T2C, T2D, T2I; | |
392 T2R = FNMS(KP958953096, T1s, T1r); | |
393 T2T = FMA(KP912575812, T1v, T1u); | |
394 T2N = FNMS(KP867381224, T1D, T1C); | |
395 T2O = FNMS(KP912575812, T1A, T1z); | |
396 T2S = FMA(KP809385824, T2O, T2N); | |
397 T2P = FNMS(KP809385824, T2O, T2N); | |
398 T2W = T2R + T2S; | |
399 T2U = FNMS(KP894834959, T2T, T2S); | |
400 T2G = FNMS(KP831864738, T1c, T15); | |
401 T2H = FNMS(KP916574801, TX, TQ); | |
402 T2C = FNMS(KP829049696, T1k, T1j); | |
403 T2D = FMA(KP831864738, T1h, T1g); | |
404 T2I = FNMS(KP904730450, T2D, T2C); | |
405 T2E = FMA(KP904730450, T2D, T2C); | |
406 T2L = FMA(KP904730450, T2G, T2I); | |
407 T2J = T2H + T2I; | |
408 } | |
409 Cr[0] = FMA(KP968583161, T2E, T2B); | |
410 { | |
411 E T2Q, T2X, T2V, T2F, T2M, T2K; | |
412 T2Q = FMA(KP248028675, T2P, T2B); | |
413 T2X = FNMS(KP690668130, T2W, T2T); | |
414 T2V = FNMS(KP618033988, T2U, T2R); | |
415 Cr[WS(csr, 9)] = FMA(KP897376177, T2V, T2Q); | |
416 Cr[WS(csr, 4)] = FNMS(KP803003575, T2X, T2Q); | |
417 T2F = FNMS(KP242145790, T2E, T2B); | |
418 T2M = FMA(KP618033988, T2L, T2H); | |
419 T2K = FNMS(KP683113946, T2J, T2G); | |
420 Cr[WS(csr, 5)] = FMA(KP792626838, T2K, T2F); | |
421 Cr[WS(csr, 10)] = FMA(KP876091699, T2M, T2F); | |
422 } | |
423 } | |
424 } | |
425 } | |
426 } | |
427 | |
428 static const kr2c_desc desc = { 25, "r2cfII_25", {47, 12, 165, 0}, &GENUS }; | |
429 | |
430 void X(codelet_r2cfII_25) (planner *p) { | |
431 X(kr2c_register) (p, r2cfII_25, &desc); | |
432 } | |
433 | |
434 #else | |
435 | |
436 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cfII_25 -dft-II -include rdft/scalar/r2cfII.h */ | |
437 | |
438 /* | |
439 * This function contains 213 FP additions, 148 FP multiplications, | |
440 * (or, 126 additions, 61 multiplications, 87 fused multiply/add), | |
441 * 94 stack variables, 38 constants, and 50 memory accesses | |
442 */ | |
443 #include "rdft/scalar/r2cfII.h" | |
444 | |
445 static void r2cfII_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
446 { | |
447 DK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
448 DK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
449 DK(KP125581039, +0.125581039058626752152356449131262266244969664); | |
450 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
451 DK(KP1_369094211, +1.369094211857377347464566715242418539779038465); | |
452 DK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
453 DK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
454 DK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
455 DK(KP497379774, +0.497379774329709576484567492012895936835134813); | |
456 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
457 DK(KP1_457937254, +1.457937254842823046293460638110518222745143328); | |
458 DK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
459 DK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
460 DK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
461 DK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
462 DK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
463 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
464 DK(KP250666467, +0.250666467128608490746237519633017587885836494); | |
465 DK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
466 DK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
467 DK(KP1_541026485, +1.541026485551578461606019272792355694543335344); | |
468 DK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
469 DK(KP1_688655851, +1.688655851004030157097116127933363010763318483); | |
470 DK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
471 DK(KP851558583, +0.851558583130145297725004891488503407959946084); | |
472 DK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
473 DK(KP1_984229402, +1.984229402628955662099586085571557042906073418); | |
474 DK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
475 DK(KP1_274847979, +1.274847979497379420353425623352032390869834596); | |
476 DK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
477 DK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
478 DK(KP1_071653589, +1.071653589957993236542617535735279956127150691); | |
479 DK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
480 DK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
481 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
482 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
483 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
484 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
485 { | |
486 INT i; | |
487 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(100, rs), MAKE_VOLATILE_STRIDE(100, csr), MAKE_VOLATILE_STRIDE(100, csi)) { | |
488 E TE, TR, T2i, T1z, TL, TS, TB, T2d, T1l, T1i, T2c, T9, T23, TZ, TW; | |
489 E T22, Ti, T26, T16, T13, T25, Ts, T2a, T1e, T1b, T29, TP, TQ; | |
490 { | |
491 E TK, T1y, TH, T1x; | |
492 TE = R0[0]; | |
493 { | |
494 E TI, TJ, TF, TG; | |
495 TI = R0[WS(rs, 10)]; | |
496 TJ = R1[WS(rs, 2)]; | |
497 TK = TI - TJ; | |
498 T1y = TI + TJ; | |
499 TF = R0[WS(rs, 5)]; | |
500 TG = R1[WS(rs, 7)]; | |
501 TH = TF - TG; | |
502 T1x = TF + TG; | |
503 } | |
504 TR = KP559016994 * (TH - TK); | |
505 T2i = FNMS(KP587785252, T1x, KP951056516 * T1y); | |
506 T1z = FMA(KP951056516, T1x, KP587785252 * T1y); | |
507 TL = TH + TK; | |
508 TS = FNMS(KP250000000, TL, TE); | |
509 } | |
510 { | |
511 E Tt, Tw, Tz, TA, T1k, T1j, T1g, T1h; | |
512 Tt = R0[WS(rs, 3)]; | |
513 { | |
514 E Tu, Tv, Tx, Ty; | |
515 Tu = R0[WS(rs, 8)]; | |
516 Tv = R1[WS(rs, 10)]; | |
517 Tw = Tu - Tv; | |
518 Tx = R1[0]; | |
519 Ty = R1[WS(rs, 5)]; | |
520 Tz = Tx + Ty; | |
521 TA = Tw - Tz; | |
522 T1k = Ty - Tx; | |
523 T1j = Tu + Tv; | |
524 } | |
525 TB = Tt + TA; | |
526 T2d = FNMS(KP293892626, T1j, KP475528258 * T1k); | |
527 T1l = FMA(KP475528258, T1j, KP293892626 * T1k); | |
528 T1g = FNMS(KP250000000, TA, Tt); | |
529 T1h = KP559016994 * (Tw + Tz); | |
530 T1i = T1g + T1h; | |
531 T2c = T1g - T1h; | |
532 } | |
533 { | |
534 E T1, T4, T7, T8, TY, TX, TU, TV; | |
535 T1 = R0[WS(rs, 1)]; | |
536 { | |
537 E T2, T3, T5, T6; | |
538 T2 = R0[WS(rs, 6)]; | |
539 T3 = R1[WS(rs, 8)]; | |
540 T4 = T2 - T3; | |
541 T5 = R0[WS(rs, 11)]; | |
542 T6 = R1[WS(rs, 3)]; | |
543 T7 = T5 - T6; | |
544 T8 = T4 + T7; | |
545 TY = T5 + T6; | |
546 TX = T2 + T3; | |
547 } | |
548 T9 = T1 + T8; | |
549 T23 = FNMS(KP293892626, TX, KP475528258 * TY); | |
550 TZ = FMA(KP475528258, TX, KP293892626 * TY); | |
551 TU = KP559016994 * (T4 - T7); | |
552 TV = FNMS(KP250000000, T8, T1); | |
553 TW = TU + TV; | |
554 T22 = TV - TU; | |
555 } | |
556 { | |
557 E Ta, Td, Tg, Th, T15, T14, T11, T12; | |
558 Ta = R0[WS(rs, 4)]; | |
559 { | |
560 E Tb, Tc, Te, Tf; | |
561 Tb = R0[WS(rs, 9)]; | |
562 Tc = R1[WS(rs, 11)]; | |
563 Td = Tb - Tc; | |
564 Te = R1[WS(rs, 1)]; | |
565 Tf = R1[WS(rs, 6)]; | |
566 Tg = Te + Tf; | |
567 Th = Td - Tg; | |
568 T15 = Tf - Te; | |
569 T14 = Tb + Tc; | |
570 } | |
571 Ti = Ta + Th; | |
572 T26 = FNMS(KP293892626, T14, KP475528258 * T15); | |
573 T16 = FMA(KP475528258, T14, KP293892626 * T15); | |
574 T11 = FNMS(KP250000000, Th, Ta); | |
575 T12 = KP559016994 * (Td + Tg); | |
576 T13 = T11 + T12; | |
577 T25 = T11 - T12; | |
578 } | |
579 { | |
580 E Tk, Tn, Tq, Tr, T1d, T1c, T19, T1a; | |
581 Tk = R0[WS(rs, 2)]; | |
582 { | |
583 E Tl, Tm, To, Tp; | |
584 Tl = R0[WS(rs, 7)]; | |
585 Tm = R1[WS(rs, 9)]; | |
586 Tn = Tl - Tm; | |
587 To = R0[WS(rs, 12)]; | |
588 Tp = R1[WS(rs, 4)]; | |
589 Tq = To - Tp; | |
590 Tr = Tn + Tq; | |
591 T1d = To + Tp; | |
592 T1c = Tl + Tm; | |
593 } | |
594 Ts = Tk + Tr; | |
595 T2a = FNMS(KP293892626, T1c, KP475528258 * T1d); | |
596 T1e = FMA(KP475528258, T1c, KP293892626 * T1d); | |
597 T19 = KP559016994 * (Tn - Tq); | |
598 T1a = FNMS(KP250000000, Tr, Tk); | |
599 T1b = T19 + T1a; | |
600 T29 = T1a - T19; | |
601 } | |
602 TP = TB - Ts; | |
603 TQ = T9 - Ti; | |
604 Ci[WS(csi, 2)] = FNMS(KP951056516, TQ, KP587785252 * TP); | |
605 Ci[WS(csi, 7)] = FMA(KP587785252, TQ, KP951056516 * TP); | |
606 { | |
607 E TM, TD, TN, Tj, TC, TO; | |
608 TM = TE + TL; | |
609 Tj = T9 + Ti; | |
610 TC = Ts + TB; | |
611 TD = KP559016994 * (Tj - TC); | |
612 TN = Tj + TC; | |
613 Cr[WS(csr, 12)] = TM + TN; | |
614 TO = FNMS(KP250000000, TN, TM); | |
615 Cr[WS(csr, 2)] = TD + TO; | |
616 Cr[WS(csr, 7)] = TO - TD; | |
617 } | |
618 { | |
619 E TT, T1J, T1Y, T1U, T1X, T1P, T1V, T1M, T1W, T1A, T1B, T1r, T1C, T1v, T18; | |
620 E T1n, T1o, T1G, T1D; | |
621 TT = TR + TS; | |
622 { | |
623 E T1H, T1I, T1S, T1T; | |
624 T1H = FNMS(KP844327925, TW, KP1_071653589 * TZ); | |
625 T1I = FNMS(KP1_274847979, T16, KP770513242 * T13); | |
626 T1J = T1H - T1I; | |
627 T1Y = T1H + T1I; | |
628 T1S = FMA(KP125333233, T1i, KP1_984229402 * T1l); | |
629 T1T = FMA(KP904827052, T1b, KP851558583 * T1e); | |
630 T1U = T1S - T1T; | |
631 T1X = T1T + T1S; | |
632 } | |
633 { | |
634 E T1N, T1O, T1K, T1L; | |
635 T1N = FMA(KP535826794, TW, KP1_688655851 * TZ); | |
636 T1O = FMA(KP637423989, T13, KP1_541026485 * T16); | |
637 T1P = T1N - T1O; | |
638 T1V = T1N + T1O; | |
639 T1K = FNMS(KP1_809654104, T1e, KP425779291 * T1b); | |
640 T1L = FNMS(KP992114701, T1i, KP250666467 * T1l); | |
641 T1M = T1K - T1L; | |
642 T1W = T1K + T1L; | |
643 } | |
644 { | |
645 E T1p, T1q, T1t, T1u; | |
646 T1p = FMA(KP844327925, T13, KP1_071653589 * T16); | |
647 T1q = FMA(KP248689887, TW, KP1_937166322 * TZ); | |
648 T1A = T1q + T1p; | |
649 T1t = FMA(KP481753674, T1b, KP1_752613360 * T1e); | |
650 T1u = FMA(KP684547105, T1i, KP1_457937254 * T1l); | |
651 T1B = T1t + T1u; | |
652 T1r = T1p - T1q; | |
653 T1C = T1A + T1B; | |
654 T1v = T1t - T1u; | |
655 } | |
656 { | |
657 E T10, T17, T1f, T1m; | |
658 T10 = FNMS(KP497379774, TZ, KP968583161 * TW); | |
659 T17 = FNMS(KP1_688655851, T16, KP535826794 * T13); | |
660 T18 = T10 + T17; | |
661 T1f = FNMS(KP963507348, T1e, KP876306680 * T1b); | |
662 T1m = FNMS(KP1_369094211, T1l, KP728968627 * T1i); | |
663 T1n = T1f + T1m; | |
664 T1o = T18 + T1n; | |
665 T1G = T10 - T17; | |
666 T1D = T1f - T1m; | |
667 } | |
668 { | |
669 E T1R, T1Q, T20, T1Z; | |
670 Cr[0] = TT + T1o; | |
671 Ci[0] = -(T1z + T1C); | |
672 T1R = KP559016994 * (T1P + T1M); | |
673 T1Q = FMA(KP250000000, T1M - T1P, TT); | |
674 Cr[WS(csr, 4)] = FMA(KP951056516, T1J, T1Q) + FMA(KP587785252, T1U, T1R); | |
675 Cr[WS(csr, 9)] = FMA(KP951056516, T1U, T1Q) + FNMA(KP587785252, T1J, T1R); | |
676 T20 = KP559016994 * (T1Y + T1X); | |
677 T1Z = FMA(KP250000000, T1X - T1Y, T1z); | |
678 Ci[WS(csi, 9)] = FMA(KP587785252, T1V, KP951056516 * T1W) + T1Z - T20; | |
679 Ci[WS(csi, 4)] = FMA(KP587785252, T1W, T1Z) + FNMS(KP951056516, T1V, T20); | |
680 { | |
681 E T1E, T1F, T1s, T1w; | |
682 T1E = FMS(KP250000000, T1C, T1z); | |
683 T1F = KP559016994 * (T1B - T1A); | |
684 Ci[WS(csi, 5)] = FMA(KP951056516, T1D, T1E) + FNMA(KP587785252, T1G, T1F); | |
685 Ci[WS(csi, 10)] = FMA(KP951056516, T1G, KP587785252 * T1D) + T1E + T1F; | |
686 T1s = FNMS(KP250000000, T1o, TT); | |
687 T1w = KP559016994 * (T18 - T1n); | |
688 Cr[WS(csr, 5)] = FMA(KP587785252, T1r, T1s) + FMS(KP951056516, T1v, T1w); | |
689 Cr[WS(csr, 10)] = T1w + FMA(KP587785252, T1v, T1s) - (KP951056516 * T1r); | |
690 } | |
691 } | |
692 } | |
693 { | |
694 E T21, T2z, T2L, T2K, T2M, T2F, T2P, T2C, T2Q, T2l, T2o, T2p, T2w, T2u, T28; | |
695 E T2f, T2g, T2s, T2h; | |
696 T21 = TS - TR; | |
697 { | |
698 E T2x, T2y, T2I, T2J; | |
699 T2x = FNMS(KP844327925, T29, KP1_071653589 * T2a); | |
700 T2y = FNMS(KP125581039, T2d, KP998026728 * T2c); | |
701 T2z = T2x + T2y; | |
702 T2L = T2y - T2x; | |
703 T2I = FNMS(KP481753674, T22, KP1_752613360 * T23); | |
704 T2J = FMA(KP904827052, T25, KP851558583 * T26); | |
705 T2K = T2I + T2J; | |
706 T2M = T2I - T2J; | |
707 } | |
708 { | |
709 E T2D, T2E, T2A, T2B; | |
710 T2D = FMA(KP535826794, T29, KP1_688655851 * T2a); | |
711 T2E = FMA(KP062790519, T2c, KP1_996053456 * T2d); | |
712 T2F = T2D + T2E; | |
713 T2P = T2E - T2D; | |
714 T2A = FMA(KP876306680, T22, KP963507348 * T23); | |
715 T2B = FNMS(KP425779291, T25, KP1_809654104 * T26); | |
716 T2C = T2A + T2B; | |
717 T2Q = T2A - T2B; | |
718 } | |
719 { | |
720 E T2j, T2k, T2m, T2n; | |
721 T2j = FNMS(KP125333233, T25, KP1_984229402 * T26); | |
722 T2k = FMA(KP684547105, T22, KP1_457937254 * T23); | |
723 T2l = T2j - T2k; | |
724 T2m = FNMS(KP770513242, T2c, KP1_274847979 * T2d); | |
725 T2n = FMA(KP998026728, T29, KP125581039 * T2a); | |
726 T2o = T2m - T2n; | |
727 T2p = T2l + T2o; | |
728 T2w = T2k + T2j; | |
729 T2u = T2n + T2m; | |
730 } | |
731 { | |
732 E T24, T27, T2b, T2e; | |
733 T24 = FNMS(KP1_369094211, T23, KP728968627 * T22); | |
734 T27 = FMA(KP992114701, T25, KP250666467 * T26); | |
735 T28 = T24 - T27; | |
736 T2b = FNMS(KP1_996053456, T2a, KP062790519 * T29); | |
737 T2e = FMA(KP637423989, T2c, KP1_541026485 * T2d); | |
738 T2f = T2b - T2e; | |
739 T2g = T28 + T2f; | |
740 T2s = T24 + T27; | |
741 T2h = T2b + T2e; | |
742 } | |
743 { | |
744 E T2H, T2G, T2O, T2N; | |
745 Cr[WS(csr, 1)] = T21 + T2g; | |
746 Ci[WS(csi, 1)] = T2p - T2i; | |
747 T2H = KP559016994 * (T2C - T2F); | |
748 T2G = FNMS(KP250000000, T2C + T2F, T21); | |
749 Cr[WS(csr, 8)] = FMA(KP951056516, T2z, T2G) + FNMA(KP587785252, T2K, T2H); | |
750 Cr[WS(csr, 3)] = FMA(KP951056516, T2K, KP587785252 * T2z) + T2G + T2H; | |
751 T2O = KP559016994 * (T2M + T2L); | |
752 T2N = FMA(KP250000000, T2L - T2M, T2i); | |
753 Ci[WS(csi, 3)] = T2N + FMA(KP587785252, T2P, T2O) - (KP951056516 * T2Q); | |
754 Ci[WS(csi, 8)] = FMA(KP587785252, T2Q, T2N) + FMS(KP951056516, T2P, T2O); | |
755 { | |
756 E T2t, T2v, T2q, T2r; | |
757 T2t = FNMS(KP250000000, T2g, T21); | |
758 T2v = KP559016994 * (T28 - T2f); | |
759 Cr[WS(csr, 6)] = FMA(KP951056516, T2u, T2t) + FNMA(KP587785252, T2w, T2v); | |
760 Cr[WS(csr, 11)] = FMA(KP951056516, T2w, T2v) + FMA(KP587785252, T2u, T2t); | |
761 T2q = KP250000000 * T2p; | |
762 T2r = KP559016994 * (T2l - T2o); | |
763 Ci[WS(csi, 6)] = FMS(KP951056516, T2h, T2i + T2q) + FNMA(KP587785252, T2s, T2r); | |
764 Ci[WS(csi, 11)] = FMA(KP951056516, T2s, KP587785252 * T2h) + T2r - (T2i + T2q); | |
765 } | |
766 } | |
767 } | |
768 } | |
769 } | |
770 } | |
771 | |
772 static const kr2c_desc desc = { 25, "r2cfII_25", {126, 61, 87, 0}, &GENUS }; | |
773 | |
774 void X(codelet_r2cfII_25) (planner *p) { | |
775 X(kr2c_register) (p, r2cfII_25, &desc); | |
776 } | |
777 | |
778 #endif |