cannam@127
|
1 /*
|
cannam@127
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@127
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@127
|
4 *
|
cannam@127
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@127
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@127
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@127
|
8 * (at your option) any later version.
|
cannam@127
|
9 *
|
cannam@127
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@127
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@127
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@127
|
13 * GNU General Public License for more details.
|
cannam@127
|
14 *
|
cannam@127
|
15 * You should have received a copy of the GNU General Public License
|
cannam@127
|
16 * along with this program; if not, write to the Free Software
|
cannam@127
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@127
|
18 *
|
cannam@127
|
19 */
|
cannam@127
|
20
|
cannam@127
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@127
|
22 /* Generated on Sat Jul 30 16:46:11 EDT 2016 */
|
cannam@127
|
23
|
cannam@127
|
24 #include "codelet-rdft.h"
|
cannam@127
|
25
|
cannam@127
|
26 #ifdef HAVE_FMA
|
cannam@127
|
27
|
cannam@127
|
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cf_25 -include r2cf.h */
|
cannam@127
|
29
|
cannam@127
|
30 /*
|
cannam@127
|
31 * This function contains 200 FP additions, 168 FP multiplications,
|
cannam@127
|
32 * (or, 44 additions, 12 multiplications, 156 fused multiply/add),
|
cannam@127
|
33 * 157 stack variables, 66 constants, and 50 memory accesses
|
cannam@127
|
34 */
|
cannam@127
|
35 #include "r2cf.h"
|
cannam@127
|
36
|
cannam@127
|
37 static void r2cf_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@127
|
38 {
|
cannam@127
|
39 DK(KP792626838, +0.792626838241819413632131824093538848057784557);
|
cannam@127
|
40 DK(KP876091699, +0.876091699473550838204498029706869638173524346);
|
cannam@127
|
41 DK(KP809385824, +0.809385824416008241660603814668679683846476688);
|
cannam@127
|
42 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
cannam@127
|
43 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
cannam@127
|
44 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
cannam@127
|
45 DK(KP997675361, +0.997675361079556513670859573984492383596555031);
|
cannam@127
|
46 DK(KP237294955, +0.237294955877110315393888866460840817927895961);
|
cannam@127
|
47 DK(KP897376177, +0.897376177523557693138608077137219684419427330);
|
cannam@127
|
48 DK(KP923225144, +0.923225144846402650453449441572664695995209956);
|
cannam@127
|
49 DK(KP956723877, +0.956723877038460305821989399535483155872969262);
|
cannam@127
|
50 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
cannam@127
|
51 DK(KP669429328, +0.669429328479476605641803240971985825917022098);
|
cannam@127
|
52 DK(KP570584518, +0.570584518783621657366766175430996792655723863);
|
cannam@127
|
53 DK(KP262346850, +0.262346850930607871785420028382979691334784273);
|
cannam@127
|
54 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
cannam@127
|
55 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
cannam@127
|
56 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
cannam@127
|
57 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
cannam@127
|
58 DK(KP921078979, +0.921078979742360627699756128143719920817673854);
|
cannam@127
|
59 DK(KP904508497, +0.904508497187473712051146708591409529430077295);
|
cannam@127
|
60 DK(KP999754674, +0.999754674276473633366203429228112409535557487);
|
cannam@127
|
61 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
cannam@127
|
62 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
cannam@127
|
63 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
cannam@127
|
64 DK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
cannam@127
|
65 DK(KP855719849, +0.855719849902058969314654733608091555096772472);
|
cannam@127
|
66 DK(KP982009705, +0.982009705009746369461829878184175962711969869);
|
cannam@127
|
67 DK(KP916574801, +0.916574801383451584742370439148878693530976769);
|
cannam@127
|
68 DK(KP690983005, +0.690983005625052575897706582817180941139845410);
|
cannam@127
|
69 DK(KP952936919, +0.952936919628306576880750665357914584765951388);
|
cannam@127
|
70 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
cannam@127
|
71 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
cannam@127
|
72 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
cannam@127
|
73 DK(KP522616830, +0.522616830205754336872861364785224694908468440);
|
cannam@127
|
74 DK(KP829049696, +0.829049696159252993975487806364305442437946767);
|
cannam@127
|
75 DK(KP999544308, +0.999544308746292983948881682379742149196758193);
|
cannam@127
|
76 DK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
cannam@127
|
77 DK(KP763932022, +0.763932022500210303590826331268723764559381640);
|
cannam@127
|
78 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
cannam@127
|
79 DK(KP447417479, +0.447417479732227551498980015410057305749330693);
|
cannam@127
|
80 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
cannam@127
|
81 DK(KP894834959, +0.894834959464455102997960030820114611498661386);
|
cannam@127
|
82 DK(KP867381224, +0.867381224396525206773171885031575671309956167);
|
cannam@127
|
83 DK(KP958953096, +0.958953096729998668045963838399037225970891871);
|
cannam@127
|
84 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
cannam@127
|
85 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@127
|
86 DK(KP244189809, +0.244189809627953270309879511234821255780225091);
|
cannam@127
|
87 DK(KP269969613, +0.269969613759572083574752974412347470060951301);
|
cannam@127
|
88 DK(KP522847744, +0.522847744331509716623755382187077770911012542);
|
cannam@127
|
89 DK(KP578046249, +0.578046249379945007321754579646815604023525655);
|
cannam@127
|
90 DK(KP603558818, +0.603558818296015001454675132653458027918768137);
|
cannam@127
|
91 DK(KP667278218, +0.667278218140296670899089292254759909713898805);
|
cannam@127
|
92 DK(KP447533225, +0.447533225982656890041886979663652563063114397);
|
cannam@127
|
93 DK(KP494780565, +0.494780565770515410344588413655324772219443730);
|
cannam@127
|
94 DK(KP987388751, +0.987388751065621252324603216482382109400433949);
|
cannam@127
|
95 DK(KP893101515, +0.893101515366181661711202267938416198338079437);
|
cannam@127
|
96 DK(KP132830569, +0.132830569247582714407653942074819768844536507);
|
cannam@127
|
97 DK(KP120146378, +0.120146378570687701782758537356596213647956445);
|
cannam@127
|
98 DK(KP059835404, +0.059835404262124915169548397419498386427871950);
|
cannam@127
|
99 DK(KP066152395, +0.066152395967733048213034281011006031460903353);
|
cannam@127
|
100 DK(KP786782374, +0.786782374965295178365099601674911834788448471);
|
cannam@127
|
101 DK(KP869845200, +0.869845200362138853122720822420327157933056305);
|
cannam@127
|
102 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@127
|
103 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@127
|
104 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
cannam@127
|
105 {
|
cannam@127
|
106 INT i;
|
cannam@127
|
107 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)) {
|
cannam@127
|
108 E T2H, T2w, T2x, T2A, T2C, T2v, T2M, T2y, T2B, T2N;
|
cannam@127
|
109 {
|
cannam@127
|
110 E T2u, TJ, T1O, T39, T2t, TB, T21, T1M, T2e, T26, T1B, T1r, T1k, T1c, T9;
|
cannam@127
|
111 E T1X, T1R, T2k, T29, T1z, T1v, T1h, TX, Ti, T13, T2a, T2j, T1U, T1Y, TQ;
|
cannam@127
|
112 E T1g, T1u, T1y, T12, Ts, T11, T1I;
|
cannam@127
|
113 {
|
cannam@127
|
114 E Tt, Tw, T16, Tx, Ty;
|
cannam@127
|
115 {
|
cannam@127
|
116 E T2p, TG, TH, TD, TE, TI, T2r;
|
cannam@127
|
117 T2p = R0[0];
|
cannam@127
|
118 TG = R0[WS(rs, 5)];
|
cannam@127
|
119 TH = R1[WS(rs, 7)];
|
cannam@127
|
120 TD = R1[WS(rs, 2)];
|
cannam@127
|
121 TE = R0[WS(rs, 10)];
|
cannam@127
|
122 Tt = R1[WS(rs, 1)];
|
cannam@127
|
123 TI = TG - TH;
|
cannam@127
|
124 T2r = TG + TH;
|
cannam@127
|
125 {
|
cannam@127
|
126 E TF, T2q, Tu, Tv, T2s;
|
cannam@127
|
127 TF = TD - TE;
|
cannam@127
|
128 T2q = TD + TE;
|
cannam@127
|
129 Tu = R0[WS(rs, 4)];
|
cannam@127
|
130 Tv = R1[WS(rs, 11)];
|
cannam@127
|
131 T2u = T2q - T2r;
|
cannam@127
|
132 T2s = T2q + T2r;
|
cannam@127
|
133 TJ = FMA(KP618033988, TI, TF);
|
cannam@127
|
134 T1O = FNMS(KP618033988, TF, TI);
|
cannam@127
|
135 T39 = T2p + T2s;
|
cannam@127
|
136 T2t = FNMS(KP250000000, T2s, T2p);
|
cannam@127
|
137 Tw = Tu + Tv;
|
cannam@127
|
138 T16 = Tv - Tu;
|
cannam@127
|
139 Tx = R1[WS(rs, 6)];
|
cannam@127
|
140 Ty = R0[WS(rs, 9)];
|
cannam@127
|
141 }
|
cannam@127
|
142 }
|
cannam@127
|
143 {
|
cannam@127
|
144 E T1P, TW, TS, TR;
|
cannam@127
|
145 {
|
cannam@127
|
146 E T1, T5, T1L, T18, T1a, TA, T4, TU, T6, T19;
|
cannam@127
|
147 T1 = R0[WS(rs, 2)];
|
cannam@127
|
148 {
|
cannam@127
|
149 E T2, T17, Tz, T3;
|
cannam@127
|
150 T2 = R1[WS(rs, 4)];
|
cannam@127
|
151 T17 = Tx - Ty;
|
cannam@127
|
152 Tz = Tx + Ty;
|
cannam@127
|
153 T3 = R0[WS(rs, 12)];
|
cannam@127
|
154 T5 = R0[WS(rs, 7)];
|
cannam@127
|
155 T1L = FMA(KP618033988, T16, T17);
|
cannam@127
|
156 T18 = FNMS(KP618033988, T17, T16);
|
cannam@127
|
157 T1a = Tz - Tw;
|
cannam@127
|
158 TA = Tw + Tz;
|
cannam@127
|
159 T4 = T2 + T3;
|
cannam@127
|
160 TU = T3 - T2;
|
cannam@127
|
161 T6 = R1[WS(rs, 9)];
|
cannam@127
|
162 }
|
cannam@127
|
163 TB = Tt + TA;
|
cannam@127
|
164 T19 = FNMS(KP250000000, TA, Tt);
|
cannam@127
|
165 {
|
cannam@127
|
166 E T7, TV, T1b, T1K, T8;
|
cannam@127
|
167 T7 = T5 + T6;
|
cannam@127
|
168 TV = T5 - T6;
|
cannam@127
|
169 T1b = FNMS(KP559016994, T1a, T19);
|
cannam@127
|
170 T1K = FMA(KP559016994, T1a, T19);
|
cannam@127
|
171 T1P = FMA(KP618033988, TU, TV);
|
cannam@127
|
172 TW = FNMS(KP618033988, TV, TU);
|
cannam@127
|
173 TS = T4 - T7;
|
cannam@127
|
174 T8 = T4 + T7;
|
cannam@127
|
175 T21 = FMA(KP869845200, T1K, T1L);
|
cannam@127
|
176 T1M = FNMS(KP786782374, T1L, T1K);
|
cannam@127
|
177 T2e = FMA(KP066152395, T1K, T1L);
|
cannam@127
|
178 T26 = FNMS(KP059835404, T1L, T1K);
|
cannam@127
|
179 T1B = FMA(KP120146378, T18, T1b);
|
cannam@127
|
180 T1r = FNMS(KP132830569, T1b, T18);
|
cannam@127
|
181 T1k = FMA(KP893101515, T18, T1b);
|
cannam@127
|
182 T1c = FNMS(KP987388751, T1b, T18);
|
cannam@127
|
183 T9 = T1 + T8;
|
cannam@127
|
184 TR = FMS(KP250000000, T8, T1);
|
cannam@127
|
185 }
|
cannam@127
|
186 }
|
cannam@127
|
187 {
|
cannam@127
|
188 E Ta, Te, TK, Td, Tf;
|
cannam@127
|
189 Ta = R1[0];
|
cannam@127
|
190 {
|
cannam@127
|
191 E Tb, Tc, T1Q, TT;
|
cannam@127
|
192 Tb = R0[WS(rs, 3)];
|
cannam@127
|
193 Tc = R1[WS(rs, 10)];
|
cannam@127
|
194 T1Q = FMA(KP559016994, TS, TR);
|
cannam@127
|
195 TT = FNMS(KP559016994, TS, TR);
|
cannam@127
|
196 Te = R1[WS(rs, 5)];
|
cannam@127
|
197 TK = Tb - Tc;
|
cannam@127
|
198 Td = Tb + Tc;
|
cannam@127
|
199 T1X = FNMS(KP120146378, T1P, T1Q);
|
cannam@127
|
200 T1R = FMA(KP132830569, T1Q, T1P);
|
cannam@127
|
201 T2k = FMA(KP494780565, T1Q, T1P);
|
cannam@127
|
202 T29 = FNMS(KP447533225, T1P, T1Q);
|
cannam@127
|
203 T1z = FMA(KP869845200, TT, TW);
|
cannam@127
|
204 T1v = FNMS(KP786782374, TW, TT);
|
cannam@127
|
205 T1h = FNMS(KP667278218, TT, TW);
|
cannam@127
|
206 TX = FMA(KP603558818, TW, TT);
|
cannam@127
|
207 Tf = R0[WS(rs, 8)];
|
cannam@127
|
208 }
|
cannam@127
|
209 {
|
cannam@127
|
210 E Tk, T1S, TM, TO, Tn, TZ, TN, T10, Tq, To, Th, Tp, TP, T1T, Tr;
|
cannam@127
|
211 Tk = R0[WS(rs, 1)];
|
cannam@127
|
212 {
|
cannam@127
|
213 E Tl, TL, Tg, Tm;
|
cannam@127
|
214 Tl = R1[WS(rs, 3)];
|
cannam@127
|
215 TL = Tf - Te;
|
cannam@127
|
216 Tg = Te + Tf;
|
cannam@127
|
217 Tm = R0[WS(rs, 11)];
|
cannam@127
|
218 To = R0[WS(rs, 6)];
|
cannam@127
|
219 T1S = FMA(KP618033988, TK, TL);
|
cannam@127
|
220 TM = FNMS(KP618033988, TL, TK);
|
cannam@127
|
221 TO = Td - Tg;
|
cannam@127
|
222 Th = Td + Tg;
|
cannam@127
|
223 Tn = Tl + Tm;
|
cannam@127
|
224 TZ = Tm - Tl;
|
cannam@127
|
225 Tp = R1[WS(rs, 8)];
|
cannam@127
|
226 }
|
cannam@127
|
227 Ti = Ta + Th;
|
cannam@127
|
228 TN = FNMS(KP250000000, Th, Ta);
|
cannam@127
|
229 T10 = Tp - To;
|
cannam@127
|
230 Tq = To + Tp;
|
cannam@127
|
231 TP = FMA(KP559016994, TO, TN);
|
cannam@127
|
232 T1T = FNMS(KP559016994, TO, TN);
|
cannam@127
|
233 Tr = Tn + Tq;
|
cannam@127
|
234 T13 = Tn - Tq;
|
cannam@127
|
235 T2a = FMA(KP578046249, T1T, T1S);
|
cannam@127
|
236 T2j = FNMS(KP522847744, T1S, T1T);
|
cannam@127
|
237 T1U = FNMS(KP987388751, T1T, T1S);
|
cannam@127
|
238 T1Y = FMA(KP893101515, T1S, T1T);
|
cannam@127
|
239 TQ = FMA(KP269969613, TP, TM);
|
cannam@127
|
240 T1g = FNMS(KP244189809, TM, TP);
|
cannam@127
|
241 T1u = FNMS(KP603558818, TM, TP);
|
cannam@127
|
242 T1y = FMA(KP667278218, TP, TM);
|
cannam@127
|
243 T12 = FMS(KP250000000, Tr, Tk);
|
cannam@127
|
244 Ts = Tk + Tr;
|
cannam@127
|
245 T11 = FMA(KP618033988, T10, TZ);
|
cannam@127
|
246 T1I = FNMS(KP618033988, TZ, T10);
|
cannam@127
|
247 }
|
cannam@127
|
248 }
|
cannam@127
|
249 }
|
cannam@127
|
250 }
|
cannam@127
|
251 {
|
cannam@127
|
252 E T2f, T27, T1j, T15, T2K, T2J, T2I, T2T, T1Z, T2X, T1N, T1V, T2W, T2U, T22;
|
cannam@127
|
253 E T1G;
|
cannam@127
|
254 {
|
cannam@127
|
255 E T3a, T3b, T20, T1J, T1C, T1s;
|
cannam@127
|
256 {
|
cannam@127
|
257 E Tj, TC, T1H, T14;
|
cannam@127
|
258 T3a = T9 + Ti;
|
cannam@127
|
259 Tj = T9 - Ti;
|
cannam@127
|
260 TC = Ts - TB;
|
cannam@127
|
261 T3b = Ts + TB;
|
cannam@127
|
262 T1H = FMA(KP559016994, T13, T12);
|
cannam@127
|
263 T14 = FNMS(KP559016994, T13, T12);
|
cannam@127
|
264 Ci[WS(csi, 10)] = KP951056516 * (FMA(KP618033988, Tj, TC));
|
cannam@127
|
265 Ci[WS(csi, 5)] = KP951056516 * (FNMS(KP618033988, TC, Tj));
|
cannam@127
|
266 T20 = FNMS(KP066152395, T1H, T1I);
|
cannam@127
|
267 T1J = FMA(KP059835404, T1I, T1H);
|
cannam@127
|
268 T2f = FMA(KP667278218, T1H, T1I);
|
cannam@127
|
269 T27 = FNMS(KP603558818, T1I, T1H);
|
cannam@127
|
270 T1C = FNMS(KP494780565, T14, T11);
|
cannam@127
|
271 T1s = FMA(KP447533225, T11, T14);
|
cannam@127
|
272 T1j = FNMS(KP522847744, T11, T14);
|
cannam@127
|
273 T15 = FMA(KP578046249, T14, T11);
|
cannam@127
|
274 }
|
cannam@127
|
275 {
|
cannam@127
|
276 E T1A, T1t, T1w, T3c, T3e, T1D, T1x, T3d, T1E, T1F;
|
cannam@127
|
277 T1A = FNMS(KP912575812, T1z, T1y);
|
cannam@127
|
278 T2K = FMA(KP912575812, T1z, T1y);
|
cannam@127
|
279 T2J = FNMS(KP958953096, T1s, T1r);
|
cannam@127
|
280 T1t = FMA(KP958953096, T1s, T1r);
|
cannam@127
|
281 T1w = FMA(KP912575812, T1v, T1u);
|
cannam@127
|
282 T2H = FNMS(KP912575812, T1v, T1u);
|
cannam@127
|
283 T3c = T3a + T3b;
|
cannam@127
|
284 T3e = T3a - T3b;
|
cannam@127
|
285 T2I = FMA(KP867381224, T1C, T1B);
|
cannam@127
|
286 T1D = FNMS(KP867381224, T1C, T1B);
|
cannam@127
|
287 T1x = FNMS(KP894834959, T1w, T1t);
|
cannam@127
|
288 T2T = FMA(KP734762448, T1Y, T1X);
|
cannam@127
|
289 T1Z = FNMS(KP734762448, T1Y, T1X);
|
cannam@127
|
290 T3d = FNMS(KP250000000, T3c, T39);
|
cannam@127
|
291 Cr[0] = T3c + T39;
|
cannam@127
|
292 T1E = FMA(KP447417479, T1w, T1D);
|
cannam@127
|
293 Ci[WS(csi, 4)] = KP951056516 * (FMA(KP992114701, T1x, TJ));
|
cannam@127
|
294 Cr[WS(csr, 10)] = FNMS(KP559016994, T3e, T3d);
|
cannam@127
|
295 Cr[WS(csr, 5)] = FMA(KP559016994, T3e, T3d);
|
cannam@127
|
296 T1F = FMA(KP763932022, T1E, T1t);
|
cannam@127
|
297 T2X = FMA(KP772036680, T1M, T1J);
|
cannam@127
|
298 T1N = FNMS(KP772036680, T1M, T1J);
|
cannam@127
|
299 T1V = FMA(KP734762448, T1U, T1R);
|
cannam@127
|
300 T2W = FNMS(KP734762448, T1U, T1R);
|
cannam@127
|
301 T2U = FNMS(KP772036680, T21, T20);
|
cannam@127
|
302 T22 = FMA(KP772036680, T21, T20);
|
cannam@127
|
303 T1G = FMA(KP999544308, T1F, T1A);
|
cannam@127
|
304 }
|
cannam@127
|
305 }
|
cannam@127
|
306 {
|
cannam@127
|
307 E T1i, T1l, T2l, T2R, T2g, T2Q, T28, T32, T1f, T1n, T1p, T33, T2b;
|
cannam@127
|
308 {
|
cannam@127
|
309 E T24, TY, T1d, T1W, T23, T25, T1m, T1e;
|
cannam@127
|
310 T2w = FMA(KP829049696, T1h, T1g);
|
cannam@127
|
311 T1i = FNMS(KP829049696, T1h, T1g);
|
cannam@127
|
312 T1W = FNMS(KP992114701, T1V, T1O);
|
cannam@127
|
313 T23 = FNMS(KP522616830, T1V, T22);
|
cannam@127
|
314 Ci[WS(csi, 9)] = KP951056516 * (FNMS(KP803003575, T1G, TJ));
|
cannam@127
|
315 T2x = FNMS(KP831864738, T1k, T1j);
|
cannam@127
|
316 T1l = FMA(KP831864738, T1k, T1j);
|
cannam@127
|
317 Ci[WS(csi, 3)] = KP998026728 * (FNMS(KP952936919, T1W, T1N));
|
cannam@127
|
318 T24 = FMA(KP690983005, T23, T1N);
|
cannam@127
|
319 TY = FNMS(KP916574801, TX, TQ);
|
cannam@127
|
320 T2A = FMA(KP916574801, TX, TQ);
|
cannam@127
|
321 T2C = FNMS(KP831864738, T1c, T15);
|
cannam@127
|
322 T1d = FMA(KP831864738, T1c, T15);
|
cannam@127
|
323 T2l = FNMS(KP982009705, T2k, T2j);
|
cannam@127
|
324 T2R = FMA(KP982009705, T2k, T2j);
|
cannam@127
|
325 T25 = FNMS(KP855719849, T24, T1Z);
|
cannam@127
|
326 T2g = FMA(KP845997307, T2f, T2e);
|
cannam@127
|
327 T2Q = FNMS(KP845997307, T2f, T2e);
|
cannam@127
|
328 T1m = FMA(KP904730450, T1d, TY);
|
cannam@127
|
329 T1e = FNMS(KP904730450, T1d, TY);
|
cannam@127
|
330 Ci[WS(csi, 8)] = -(KP951056516 * (FNMS(KP992114701, T25, T1O)));
|
cannam@127
|
331 T28 = FNMS(KP845997307, T27, T26);
|
cannam@127
|
332 T32 = FMA(KP845997307, T27, T26);
|
cannam@127
|
333 T1f = FNMS(KP242145790, T1e, TJ);
|
cannam@127
|
334 Ci[WS(csi, 1)] = -(KP951056516 * (FMA(KP968583161, T1e, TJ)));
|
cannam@127
|
335 T1n = FNMS(KP999754674, T1m, T1l);
|
cannam@127
|
336 T1p = FNMS(KP904508497, T1m, T1i);
|
cannam@127
|
337 T33 = FMA(KP921078979, T2a, T29);
|
cannam@127
|
338 T2b = FNMS(KP921078979, T2a, T29);
|
cannam@127
|
339 }
|
cannam@127
|
340 {
|
cannam@127
|
341 E T2P, T2Z, T2V, T2O;
|
cannam@127
|
342 {
|
cannam@127
|
343 E T2d, T2n, T2i, T2Y, T2m, T2o;
|
cannam@127
|
344 T2P = FNMS(KP559016994, T2u, T2t);
|
cannam@127
|
345 T2v = FMA(KP559016994, T2u, T2t);
|
cannam@127
|
346 {
|
cannam@127
|
347 E T1o, T1q, T2h, T2c;
|
cannam@127
|
348 T1o = FNMS(KP559154169, T1n, T1i);
|
cannam@127
|
349 T1q = FMA(KP683113946, T1p, T1l);
|
cannam@127
|
350 T2h = FMA(KP906616052, T2b, T28);
|
cannam@127
|
351 T2c = FNMS(KP906616052, T2b, T28);
|
cannam@127
|
352 Ci[WS(csi, 6)] = -(KP951056516 * (FMA(KP968583161, T1o, T1f)));
|
cannam@127
|
353 Ci[WS(csi, 11)] = -(KP951056516 * (FMA(KP876306680, T1q, T1f)));
|
cannam@127
|
354 T2d = FMA(KP262346850, T2c, T1O);
|
cannam@127
|
355 Ci[WS(csi, 2)] = -(KP998026728 * (FNMS(KP952936919, T1O, T2c)));
|
cannam@127
|
356 T2n = T2g + T2h;
|
cannam@127
|
357 T2i = FMA(KP618033988, T2h, T2g);
|
cannam@127
|
358 }
|
cannam@127
|
359 T2m = FMA(KP570584518, T2l, T2i);
|
cannam@127
|
360 T2o = FNMS(KP669429328, T2n, T2l);
|
cannam@127
|
361 Ci[WS(csi, 12)] = KP951056516 * (FNMS(KP949179823, T2m, T2d));
|
cannam@127
|
362 Ci[WS(csi, 7)] = KP951056516 * (FNMS(KP876306680, T2o, T2d));
|
cannam@127
|
363 T2V = FMA(KP956723877, T2U, T2T);
|
cannam@127
|
364 T2Y = FMA(KP522616830, T2T, T2X);
|
cannam@127
|
365 T2Z = FNMS(KP763932022, T2Y, T2U);
|
cannam@127
|
366 }
|
cannam@127
|
367 Cr[WS(csr, 3)] = FMA(KP992114701, T2V, T2P);
|
cannam@127
|
368 {
|
cannam@127
|
369 E T30, T34, T2S, T31, T35;
|
cannam@127
|
370 T30 = FMA(KP855719849, T2Z, T2W);
|
cannam@127
|
371 T34 = FNMS(KP923225144, T2R, T2Q);
|
cannam@127
|
372 T2S = FMA(KP923225144, T2R, T2Q);
|
cannam@127
|
373 Cr[WS(csr, 8)] = FNMS(KP897376177, T30, T2P);
|
cannam@127
|
374 T31 = FNMS(KP237294955, T2S, T2P);
|
cannam@127
|
375 Cr[WS(csr, 2)] = FMA(KP949179823, T2S, T2P);
|
cannam@127
|
376 T35 = FNMS(KP997675361, T34, T33);
|
cannam@127
|
377 {
|
cannam@127
|
378 E T37, T36, T38, T2L;
|
cannam@127
|
379 T37 = FNMS(KP904508497, T34, T32);
|
cannam@127
|
380 T36 = FMA(KP560319534, T35, T32);
|
cannam@127
|
381 T38 = FNMS(KP681693190, T37, T33);
|
cannam@127
|
382 Cr[WS(csr, 12)] = FNMS(KP949179823, T36, T31);
|
cannam@127
|
383 Cr[WS(csr, 7)] = FNMS(KP860541664, T38, T31);
|
cannam@127
|
384 T2O = FNMS(KP809385824, T2K, T2I);
|
cannam@127
|
385 T2L = FNMS(KP447417479, T2K, T2J);
|
cannam@127
|
386 T2M = FNMS(KP690983005, T2L, T2I);
|
cannam@127
|
387 }
|
cannam@127
|
388 }
|
cannam@127
|
389 Cr[WS(csr, 4)] = FNMS(KP992114701, T2O, T2v);
|
cannam@127
|
390 }
|
cannam@127
|
391 }
|
cannam@127
|
392 }
|
cannam@127
|
393 }
|
cannam@127
|
394 T2y = FNMS(KP904730450, T2x, T2w);
|
cannam@127
|
395 T2B = FMA(KP904730450, T2x, T2w);
|
cannam@127
|
396 T2N = FNMS(KP999544308, T2M, T2H);
|
cannam@127
|
397 {
|
cannam@127
|
398 E T2z, T2D, T2F, T2E, T2G;
|
cannam@127
|
399 T2z = FNMS(KP242145790, T2y, T2v);
|
cannam@127
|
400 Cr[WS(csr, 1)] = FMA(KP968583161, T2y, T2v);
|
cannam@127
|
401 T2D = FMA(KP904730450, T2C, T2B);
|
cannam@127
|
402 T2F = T2A + T2B;
|
cannam@127
|
403 Cr[WS(csr, 9)] = FNMS(KP803003575, T2N, T2v);
|
cannam@127
|
404 T2E = FNMS(KP618033988, T2D, T2A);
|
cannam@127
|
405 T2G = FMA(KP683113946, T2F, T2C);
|
cannam@127
|
406 Cr[WS(csr, 6)] = FNMS(KP876091699, T2E, T2z);
|
cannam@127
|
407 Cr[WS(csr, 11)] = FNMS(KP792626838, T2G, T2z);
|
cannam@127
|
408 }
|
cannam@127
|
409 }
|
cannam@127
|
410 }
|
cannam@127
|
411 }
|
cannam@127
|
412
|
cannam@127
|
413 static const kr2c_desc desc = { 25, "r2cf_25", {44, 12, 156, 0}, &GENUS };
|
cannam@127
|
414
|
cannam@127
|
415 void X(codelet_r2cf_25) (planner *p) {
|
cannam@127
|
416 X(kr2c_register) (p, r2cf_25, &desc);
|
cannam@127
|
417 }
|
cannam@127
|
418
|
cannam@127
|
419 #else /* HAVE_FMA */
|
cannam@127
|
420
|
cannam@127
|
421 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cf_25 -include r2cf.h */
|
cannam@127
|
422
|
cannam@127
|
423 /*
|
cannam@127
|
424 * This function contains 200 FP additions, 140 FP multiplications,
|
cannam@127
|
425 * (or, 117 additions, 57 multiplications, 83 fused multiply/add),
|
cannam@127
|
426 * 101 stack variables, 40 constants, and 50 memory accesses
|
cannam@127
|
427 */
|
cannam@127
|
428 #include "r2cf.h"
|
cannam@127
|
429
|
cannam@127
|
430 static void r2cf_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@127
|
431 {
|
cannam@127
|
432 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
cannam@127
|
433 DK(KP125581039, +0.125581039058626752152356449131262266244969664);
|
cannam@127
|
434 DK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
|
cannam@127
|
435 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
cannam@127
|
436 DK(KP809016994, +0.809016994374947424102293417182819058860154590);
|
cannam@127
|
437 DK(KP309016994, +0.309016994374947424102293417182819058860154590);
|
cannam@127
|
438 DK(KP1_369094211, +1.369094211857377347464566715242418539779038465);
|
cannam@127
|
439 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
cannam@127
|
440 DK(KP963507348, +0.963507348203430549974383005744259307057084020);
|
cannam@127
|
441 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
cannam@127
|
442 DK(KP497379774, +0.497379774329709576484567492012895936835134813);
|
cannam@127
|
443 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
cannam@127
|
444 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
cannam@127
|
445 DK(KP1_457937254, +1.457937254842823046293460638110518222745143328);
|
cannam@127
|
446 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
cannam@127
|
447 DK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
|
cannam@127
|
448 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
cannam@127
|
449 DK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
|
cannam@127
|
450 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
cannam@127
|
451 DK(KP250666467, +0.250666467128608490746237519633017587885836494);
|
cannam@127
|
452 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
cannam@127
|
453 DK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
|
cannam@127
|
454 DK(KP1_274847979, +1.274847979497379420353425623352032390869834596);
|
cannam@127
|
455 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
cannam@127
|
456 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
cannam@127
|
457 DK(KP1_071653589, +1.071653589957993236542617535735279956127150691);
|
cannam@127
|
458 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
cannam@127
|
459 DK(KP1_984229402, +1.984229402628955662099586085571557042906073418);
|
cannam@127
|
460 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
cannam@127
|
461 DK(KP851558583, +0.851558583130145297725004891488503407959946084);
|
cannam@127
|
462 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
cannam@127
|
463 DK(KP1_541026485, +1.541026485551578461606019272792355694543335344);
|
cannam@127
|
464 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
cannam@127
|
465 DK(KP1_688655851, +1.688655851004030157097116127933363010763318483);
|
cannam@127
|
466 DK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
cannam@127
|
467 DK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
cannam@127
|
468 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@127
|
469 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@127
|
470 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
cannam@127
|
471 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@127
|
472 {
|
cannam@127
|
473 INT i;
|
cannam@127
|
474 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)) {
|
cannam@127
|
475 E T8, T1j, T1V, T1l, T7, T9, Ta, T12, T2u, T1O, T19, T1P, Ti, T2r, T1K;
|
cannam@127
|
476 E Tp, T1L, Tx, T2q, T1H, TE, T1I, TN, T2t, T1R, TU, T1S, T6, T1k, T3;
|
cannam@127
|
477 E T2s, T2v;
|
cannam@127
|
478 T8 = R0[0];
|
cannam@127
|
479 {
|
cannam@127
|
480 E T4, T5, T1, T2;
|
cannam@127
|
481 T4 = R0[WS(rs, 5)];
|
cannam@127
|
482 T5 = R1[WS(rs, 7)];
|
cannam@127
|
483 T6 = T4 + T5;
|
cannam@127
|
484 T1k = T4 - T5;
|
cannam@127
|
485 T1 = R1[WS(rs, 2)];
|
cannam@127
|
486 T2 = R0[WS(rs, 10)];
|
cannam@127
|
487 T3 = T1 + T2;
|
cannam@127
|
488 T1j = T1 - T2;
|
cannam@127
|
489 }
|
cannam@127
|
490 T1V = KP951056516 * T1k;
|
cannam@127
|
491 T1l = FMA(KP951056516, T1j, KP587785252 * T1k);
|
cannam@127
|
492 T7 = KP559016994 * (T3 - T6);
|
cannam@127
|
493 T9 = T3 + T6;
|
cannam@127
|
494 Ta = FNMS(KP250000000, T9, T8);
|
cannam@127
|
495 {
|
cannam@127
|
496 E T16, T13, T14, TY, T17, T11, T15, T18;
|
cannam@127
|
497 T16 = R1[WS(rs, 1)];
|
cannam@127
|
498 {
|
cannam@127
|
499 E TW, TX, TZ, T10;
|
cannam@127
|
500 TW = R0[WS(rs, 4)];
|
cannam@127
|
501 TX = R1[WS(rs, 11)];
|
cannam@127
|
502 T13 = TW + TX;
|
cannam@127
|
503 TZ = R1[WS(rs, 6)];
|
cannam@127
|
504 T10 = R0[WS(rs, 9)];
|
cannam@127
|
505 T14 = TZ + T10;
|
cannam@127
|
506 TY = TW - TX;
|
cannam@127
|
507 T17 = T13 + T14;
|
cannam@127
|
508 T11 = TZ - T10;
|
cannam@127
|
509 }
|
cannam@127
|
510 T12 = FMA(KP475528258, TY, KP293892626 * T11);
|
cannam@127
|
511 T2u = T16 + T17;
|
cannam@127
|
512 T1O = FNMS(KP293892626, TY, KP475528258 * T11);
|
cannam@127
|
513 T15 = KP559016994 * (T13 - T14);
|
cannam@127
|
514 T18 = FNMS(KP250000000, T17, T16);
|
cannam@127
|
515 T19 = T15 + T18;
|
cannam@127
|
516 T1P = T18 - T15;
|
cannam@127
|
517 }
|
cannam@127
|
518 {
|
cannam@127
|
519 E Tm, Tj, Tk, Te, Tn, Th, Tl, To;
|
cannam@127
|
520 Tm = R1[0];
|
cannam@127
|
521 {
|
cannam@127
|
522 E Tc, Td, Tf, Tg;
|
cannam@127
|
523 Tc = R0[WS(rs, 3)];
|
cannam@127
|
524 Td = R1[WS(rs, 10)];
|
cannam@127
|
525 Tj = Tc + Td;
|
cannam@127
|
526 Tf = R1[WS(rs, 5)];
|
cannam@127
|
527 Tg = R0[WS(rs, 8)];
|
cannam@127
|
528 Tk = Tf + Tg;
|
cannam@127
|
529 Te = Tc - Td;
|
cannam@127
|
530 Tn = Tj + Tk;
|
cannam@127
|
531 Th = Tf - Tg;
|
cannam@127
|
532 }
|
cannam@127
|
533 Ti = FMA(KP475528258, Te, KP293892626 * Th);
|
cannam@127
|
534 T2r = Tm + Tn;
|
cannam@127
|
535 T1K = FNMS(KP293892626, Te, KP475528258 * Th);
|
cannam@127
|
536 Tl = KP559016994 * (Tj - Tk);
|
cannam@127
|
537 To = FNMS(KP250000000, Tn, Tm);
|
cannam@127
|
538 Tp = Tl + To;
|
cannam@127
|
539 T1L = To - Tl;
|
cannam@127
|
540 }
|
cannam@127
|
541 {
|
cannam@127
|
542 E TB, Ty, Tz, Tt, TC, Tw, TA, TD;
|
cannam@127
|
543 TB = R0[WS(rs, 2)];
|
cannam@127
|
544 {
|
cannam@127
|
545 E Tr, Ts, Tu, Tv;
|
cannam@127
|
546 Tr = R1[WS(rs, 4)];
|
cannam@127
|
547 Ts = R0[WS(rs, 12)];
|
cannam@127
|
548 Ty = Tr + Ts;
|
cannam@127
|
549 Tu = R0[WS(rs, 7)];
|
cannam@127
|
550 Tv = R1[WS(rs, 9)];
|
cannam@127
|
551 Tz = Tu + Tv;
|
cannam@127
|
552 Tt = Tr - Ts;
|
cannam@127
|
553 TC = Ty + Tz;
|
cannam@127
|
554 Tw = Tu - Tv;
|
cannam@127
|
555 }
|
cannam@127
|
556 Tx = FMA(KP475528258, Tt, KP293892626 * Tw);
|
cannam@127
|
557 T2q = TB + TC;
|
cannam@127
|
558 T1H = FNMS(KP293892626, Tt, KP475528258 * Tw);
|
cannam@127
|
559 TA = KP559016994 * (Ty - Tz);
|
cannam@127
|
560 TD = FNMS(KP250000000, TC, TB);
|
cannam@127
|
561 TE = TA + TD;
|
cannam@127
|
562 T1I = TD - TA;
|
cannam@127
|
563 }
|
cannam@127
|
564 {
|
cannam@127
|
565 E TR, TO, TP, TJ, TS, TM, TQ, TT;
|
cannam@127
|
566 TR = R0[WS(rs, 1)];
|
cannam@127
|
567 {
|
cannam@127
|
568 E TH, TI, TK, TL;
|
cannam@127
|
569 TH = R1[WS(rs, 3)];
|
cannam@127
|
570 TI = R0[WS(rs, 11)];
|
cannam@127
|
571 TO = TH + TI;
|
cannam@127
|
572 TK = R0[WS(rs, 6)];
|
cannam@127
|
573 TL = R1[WS(rs, 8)];
|
cannam@127
|
574 TP = TK + TL;
|
cannam@127
|
575 TJ = TH - TI;
|
cannam@127
|
576 TS = TO + TP;
|
cannam@127
|
577 TM = TK - TL;
|
cannam@127
|
578 }
|
cannam@127
|
579 TN = FMA(KP475528258, TJ, KP293892626 * TM);
|
cannam@127
|
580 T2t = TR + TS;
|
cannam@127
|
581 T1R = FNMS(KP293892626, TJ, KP475528258 * TM);
|
cannam@127
|
582 TQ = KP559016994 * (TO - TP);
|
cannam@127
|
583 TT = FNMS(KP250000000, TS, TR);
|
cannam@127
|
584 TU = TQ + TT;
|
cannam@127
|
585 T1S = TT - TQ;
|
cannam@127
|
586 }
|
cannam@127
|
587 T2s = T2q - T2r;
|
cannam@127
|
588 T2v = T2t - T2u;
|
cannam@127
|
589 Ci[WS(csi, 5)] = FNMS(KP587785252, T2v, KP951056516 * T2s);
|
cannam@127
|
590 Ci[WS(csi, 10)] = FMA(KP587785252, T2s, KP951056516 * T2v);
|
cannam@127
|
591 {
|
cannam@127
|
592 E T2z, T2y, T2A, T2w, T2x, T2B;
|
cannam@127
|
593 T2z = T8 + T9;
|
cannam@127
|
594 T2w = T2r + T2q;
|
cannam@127
|
595 T2x = T2t + T2u;
|
cannam@127
|
596 T2y = KP559016994 * (T2w - T2x);
|
cannam@127
|
597 T2A = T2w + T2x;
|
cannam@127
|
598 Cr[0] = T2z + T2A;
|
cannam@127
|
599 T2B = FNMS(KP250000000, T2A, T2z);
|
cannam@127
|
600 Cr[WS(csr, 5)] = T2y + T2B;
|
cannam@127
|
601 Cr[WS(csr, 10)] = T2B - T2y;
|
cannam@127
|
602 }
|
cannam@127
|
603 {
|
cannam@127
|
604 E Tb, Tq, TF, TG, T1E, T1F, T1G, T1B, T1C, T1D, TV, T1a, T1b, T1o, T1r;
|
cannam@127
|
605 E T1s, T1z, T1x, T1e, T1h, T1i, T1u, T1t;
|
cannam@127
|
606 Tb = T7 + Ta;
|
cannam@127
|
607 Tq = FMA(KP1_688655851, Ti, KP535826794 * Tp);
|
cannam@127
|
608 TF = FMA(KP1_541026485, Tx, KP637423989 * TE);
|
cannam@127
|
609 TG = Tq - TF;
|
cannam@127
|
610 T1E = FMA(KP851558583, TN, KP904827052 * TU);
|
cannam@127
|
611 T1F = FMA(KP1_984229402, T12, KP125333233 * T19);
|
cannam@127
|
612 T1G = T1E + T1F;
|
cannam@127
|
613 T1B = FNMS(KP844327925, Tp, KP1_071653589 * Ti);
|
cannam@127
|
614 T1C = FNMS(KP1_274847979, Tx, KP770513242 * TE);
|
cannam@127
|
615 T1D = T1B + T1C;
|
cannam@127
|
616 TV = FNMS(KP425779291, TU, KP1_809654104 * TN);
|
cannam@127
|
617 T1a = FNMS(KP992114701, T19, KP250666467 * T12);
|
cannam@127
|
618 T1b = TV + T1a;
|
cannam@127
|
619 {
|
cannam@127
|
620 E T1m, T1n, T1p, T1q;
|
cannam@127
|
621 T1m = FMA(KP1_937166322, Ti, KP248689887 * Tp);
|
cannam@127
|
622 T1n = FMA(KP1_071653589, Tx, KP844327925 * TE);
|
cannam@127
|
623 T1o = T1m + T1n;
|
cannam@127
|
624 T1p = FMA(KP1_752613360, TN, KP481753674 * TU);
|
cannam@127
|
625 T1q = FMA(KP1_457937254, T12, KP684547105 * T19);
|
cannam@127
|
626 T1r = T1p + T1q;
|
cannam@127
|
627 T1s = T1o + T1r;
|
cannam@127
|
628 T1z = T1q - T1p;
|
cannam@127
|
629 T1x = T1n - T1m;
|
cannam@127
|
630 }
|
cannam@127
|
631 {
|
cannam@127
|
632 E T1c, T1d, T1f, T1g;
|
cannam@127
|
633 T1c = FNMS(KP497379774, Ti, KP968583161 * Tp);
|
cannam@127
|
634 T1d = FNMS(KP1_688655851, Tx, KP535826794 * TE);
|
cannam@127
|
635 T1e = T1c + T1d;
|
cannam@127
|
636 T1f = FNMS(KP963507348, TN, KP876306680 * TU);
|
cannam@127
|
637 T1g = FNMS(KP1_369094211, T12, KP728968627 * T19);
|
cannam@127
|
638 T1h = T1f + T1g;
|
cannam@127
|
639 T1i = T1e + T1h;
|
cannam@127
|
640 T1u = T1f - T1g;
|
cannam@127
|
641 T1t = T1d - T1c;
|
cannam@127
|
642 }
|
cannam@127
|
643 Cr[WS(csr, 1)] = Tb + T1i;
|
cannam@127
|
644 Ci[WS(csi, 1)] = -(T1l + T1s);
|
cannam@127
|
645 Cr[WS(csr, 4)] = Tb + TG + T1b;
|
cannam@127
|
646 Ci[WS(csi, 4)] = T1l + T1D - T1G;
|
cannam@127
|
647 Ci[WS(csi, 9)] = FMA(KP309016994, T1D, T1l) + FMA(KP587785252, T1a - TV, KP809016994 * T1G) - (KP951056516 * (Tq + TF));
|
cannam@127
|
648 Cr[WS(csr, 9)] = FMA(KP309016994, TG, Tb) + FMA(KP951056516, T1B - T1C, KP587785252 * (T1F - T1E)) - (KP809016994 * T1b);
|
cannam@127
|
649 {
|
cannam@127
|
650 E T1v, T1w, T1y, T1A;
|
cannam@127
|
651 T1v = FMS(KP250000000, T1s, T1l);
|
cannam@127
|
652 T1w = KP559016994 * (T1r - T1o);
|
cannam@127
|
653 Ci[WS(csi, 11)] = FMA(KP587785252, T1t, KP951056516 * T1u) + T1v - T1w;
|
cannam@127
|
654 Ci[WS(csi, 6)] = FMA(KP951056516, T1t, T1v) + FNMS(KP587785252, T1u, T1w);
|
cannam@127
|
655 T1y = FNMS(KP250000000, T1i, Tb);
|
cannam@127
|
656 T1A = KP559016994 * (T1e - T1h);
|
cannam@127
|
657 Cr[WS(csr, 11)] = FMA(KP587785252, T1x, T1y) + FNMA(KP951056516, T1z, T1A);
|
cannam@127
|
658 Cr[WS(csr, 6)] = FMA(KP951056516, T1x, T1A) + FMA(KP587785252, T1z, T1y);
|
cannam@127
|
659 }
|
cannam@127
|
660 }
|
cannam@127
|
661 {
|
cannam@127
|
662 E T1W, T1X, T1J, T1M, T1N, T21, T22, T23, T1Q, T1T, T1U, T1Y, T1Z, T20, T26;
|
cannam@127
|
663 E T29, T2a, T2k, T2j, T2l, T2m, T2d, T2o, T2i;
|
cannam@127
|
664 T1W = FNMS(KP587785252, T1j, T1V);
|
cannam@127
|
665 T1X = Ta - T7;
|
cannam@127
|
666 T1J = FNMS(KP125333233, T1I, KP1_984229402 * T1H);
|
cannam@127
|
667 T1M = FMA(KP1_457937254, T1K, KP684547105 * T1L);
|
cannam@127
|
668 T1N = T1J - T1M;
|
cannam@127
|
669 T21 = FNMS(KP1_996053456, T1R, KP062790519 * T1S);
|
cannam@127
|
670 T22 = FMA(KP1_541026485, T1O, KP637423989 * T1P);
|
cannam@127
|
671 T23 = T21 - T22;
|
cannam@127
|
672 T1Q = FNMS(KP770513242, T1P, KP1_274847979 * T1O);
|
cannam@127
|
673 T1T = FMA(KP125581039, T1R, KP998026728 * T1S);
|
cannam@127
|
674 T1U = T1Q - T1T;
|
cannam@127
|
675 T1Y = FNMS(KP1_369094211, T1K, KP728968627 * T1L);
|
cannam@127
|
676 T1Z = FMA(KP250666467, T1H, KP992114701 * T1I);
|
cannam@127
|
677 T20 = T1Y - T1Z;
|
cannam@127
|
678 {
|
cannam@127
|
679 E T24, T25, T27, T28;
|
cannam@127
|
680 T24 = FNMS(KP481753674, T1L, KP1_752613360 * T1K);
|
cannam@127
|
681 T25 = FMA(KP851558583, T1H, KP904827052 * T1I);
|
cannam@127
|
682 T26 = T24 - T25;
|
cannam@127
|
683 T27 = FNMS(KP844327925, T1S, KP1_071653589 * T1R);
|
cannam@127
|
684 T28 = FNMS(KP998026728, T1P, KP125581039 * T1O);
|
cannam@127
|
685 T29 = T27 + T28;
|
cannam@127
|
686 T2a = T26 + T29;
|
cannam@127
|
687 T2k = T27 - T28;
|
cannam@127
|
688 T2j = T24 + T25;
|
cannam@127
|
689 }
|
cannam@127
|
690 {
|
cannam@127
|
691 E T2b, T2c, T2g, T2h;
|
cannam@127
|
692 T2b = FNMS(KP425779291, T1I, KP1_809654104 * T1H);
|
cannam@127
|
693 T2c = FMA(KP963507348, T1K, KP876306680 * T1L);
|
cannam@127
|
694 T2l = T2c + T2b;
|
cannam@127
|
695 T2g = FMA(KP1_688655851, T1R, KP535826794 * T1S);
|
cannam@127
|
696 T2h = FMA(KP1_996053456, T1O, KP062790519 * T1P);
|
cannam@127
|
697 T2m = T2g + T2h;
|
cannam@127
|
698 T2d = T2b - T2c;
|
cannam@127
|
699 T2o = T2l + T2m;
|
cannam@127
|
700 T2i = T2g - T2h;
|
cannam@127
|
701 }
|
cannam@127
|
702 Ci[WS(csi, 2)] = T1W + T2a;
|
cannam@127
|
703 Cr[WS(csr, 2)] = T1X + T2o;
|
cannam@127
|
704 Ci[WS(csi, 3)] = T1N + T1U - T1W;
|
cannam@127
|
705 Cr[WS(csr, 3)] = T1X + T20 + T23;
|
cannam@127
|
706 Cr[WS(csr, 8)] = FMA(KP309016994, T20, T1X) + FNMA(KP809016994, T23, KP587785252 * (T1T + T1Q)) - (KP951056516 * (T1M + T1J));
|
cannam@127
|
707 Ci[WS(csi, 8)] = FNMS(KP587785252, T21 + T22, KP309016994 * T1N) + FNMA(KP809016994, T1U, KP951056516 * (T1Y + T1Z)) - T1W;
|
cannam@127
|
708 {
|
cannam@127
|
709 E T2e, T2f, T2n, T2p;
|
cannam@127
|
710 T2e = KP559016994 * (T26 - T29);
|
cannam@127
|
711 T2f = FNMS(KP250000000, T2a, T1W);
|
cannam@127
|
712 Ci[WS(csi, 7)] = FMA(KP951056516, T2d, T2e) + FNMS(KP587785252, T2i, T2f);
|
cannam@127
|
713 Ci[WS(csi, 12)] = FMA(KP587785252, T2d, T2f) + FMS(KP951056516, T2i, T2e);
|
cannam@127
|
714 T2n = KP559016994 * (T2l - T2m);
|
cannam@127
|
715 T2p = FNMS(KP250000000, T2o, T1X);
|
cannam@127
|
716 Cr[WS(csr, 7)] = FMA(KP951056516, T2j, KP587785252 * T2k) + T2n + T2p;
|
cannam@127
|
717 Cr[WS(csr, 12)] = FMA(KP587785252, T2j, T2p) + FNMA(KP951056516, T2k, T2n);
|
cannam@127
|
718 }
|
cannam@127
|
719 }
|
cannam@127
|
720 }
|
cannam@127
|
721 }
|
cannam@127
|
722 }
|
cannam@127
|
723
|
cannam@127
|
724 static const kr2c_desc desc = { 25, "r2cf_25", {117, 57, 83, 0}, &GENUS };
|
cannam@127
|
725
|
cannam@127
|
726 void X(codelet_r2cf_25) (planner *p) {
|
cannam@127
|
727 X(kr2c_register) (p, r2cf_25, &desc);
|
cannam@127
|
728 }
|
cannam@127
|
729
|
cannam@127
|
730 #endif /* HAVE_FMA */
|