Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:40:24 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cfII_25 -dft-II -include r2cfII.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 212 FP additions, 177 FP multiplications,
|
Chris@10
|
32 * (or, 47 additions, 12 multiplications, 165 fused multiply/add),
|
Chris@10
|
33 * 163 stack variables, 67 constants, and 50 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "r2cfII.h"
|
Chris@10
|
36
|
Chris@10
|
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)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP876091699, +0.876091699473550838204498029706869638173524346);
|
Chris@10
|
40 DK(KP792626838, +0.792626838241819413632131824093538848057784557);
|
Chris@10
|
41 DK(KP690668130, +0.690668130712929053565177988380887884042527623);
|
Chris@10
|
42 DK(KP809385824, +0.809385824416008241660603814668679683846476688);
|
Chris@10
|
43 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
Chris@10
|
44 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
Chris@10
|
45 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
Chris@10
|
46 DK(KP237294955, +0.237294955877110315393888866460840817927895961);
|
Chris@10
|
47 DK(KP897376177, +0.897376177523557693138608077137219684419427330);
|
Chris@10
|
48 DK(KP997675361, +0.997675361079556513670859573984492383596555031);
|
Chris@10
|
49 DK(KP584303379, +0.584303379262766050358567120694562180043261496);
|
Chris@10
|
50 DK(KP653711795, +0.653711795629256296299985401753308353544378892);
|
Chris@10
|
51 DK(KP591287873, +0.591287873858343558732323717242372865934480959);
|
Chris@10
|
52 DK(KP645989928, +0.645989928319777763844272876603899665178054552);
|
Chris@10
|
53 DK(KP956723877, +0.956723877038460305821989399535483155872969262);
|
Chris@10
|
54 DK(KP952936919, +0.952936919628306576880750665357914584765951388);
|
Chris@10
|
55 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
56 DK(KP945422727, +0.945422727388575946270360266328811958657216298);
|
Chris@10
|
57 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
58 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
59 DK(KP999754674, +0.999754674276473633366203429228112409535557487);
|
Chris@10
|
60 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
61 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
Chris@10
|
62 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
Chris@10
|
63 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
64 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
65 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
Chris@10
|
66 DK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
Chris@10
|
67 DK(KP669429328, +0.669429328479476605641803240971985825917022098);
|
Chris@10
|
68 DK(KP916574801, +0.916574801383451584742370439148878693530976769);
|
Chris@10
|
69 DK(KP829049696, +0.829049696159252993975487806364305442437946767);
|
Chris@10
|
70 DK(KP923225144, +0.923225144846402650453449441572664695995209956);
|
Chris@10
|
71 DK(KP262346850, +0.262346850930607871785420028382979691334784273);
|
Chris@10
|
72 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
73 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
Chris@10
|
74 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
Chris@10
|
75 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
76 DK(KP763583905, +0.763583905359130246362948588764067237776594106);
|
Chris@10
|
77 DK(KP921078979, +0.921078979742360627699756128143719920817673854);
|
Chris@10
|
78 DK(KP904508497, +0.904508497187473712051146708591409529430077295);
|
Chris@10
|
79 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
Chris@10
|
80 DK(KP894834959, +0.894834959464455102997960030820114611498661386);
|
Chris@10
|
81 DK(KP982009705, +0.982009705009746369461829878184175962711969869);
|
Chris@10
|
82 DK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
Chris@10
|
83 DK(KP958953096, +0.958953096729998668045963838399037225970891871);
|
Chris@10
|
84 DK(KP867381224, +0.867381224396525206773171885031575671309956167);
|
Chris@10
|
85 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
86 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
87 DK(KP869845200, +0.869845200362138853122720822420327157933056305);
|
Chris@10
|
88 DK(KP786782374, +0.786782374965295178365099601674911834788448471);
|
Chris@10
|
89 DK(KP120146378, +0.120146378570687701782758537356596213647956445);
|
Chris@10
|
90 DK(KP132830569, +0.132830569247582714407653942074819768844536507);
|
Chris@10
|
91 DK(KP269969613, +0.269969613759572083574752974412347470060951301);
|
Chris@10
|
92 DK(KP244189809, +0.244189809627953270309879511234821255780225091);
|
Chris@10
|
93 DK(KP987388751, +0.987388751065621252324603216482382109400433949);
|
Chris@10
|
94 DK(KP893101515, +0.893101515366181661711202267938416198338079437);
|
Chris@10
|
95 DK(KP494780565, +0.494780565770515410344588413655324772219443730);
|
Chris@10
|
96 DK(KP447533225, +0.447533225982656890041886979663652563063114397);
|
Chris@10
|
97 DK(KP522847744, +0.522847744331509716623755382187077770911012542);
|
Chris@10
|
98 DK(KP578046249, +0.578046249379945007321754579646815604023525655);
|
Chris@10
|
99 DK(KP066152395, +0.066152395967733048213034281011006031460903353);
|
Chris@10
|
100 DK(KP059835404, +0.059835404262124915169548397419498386427871950);
|
Chris@10
|
101 DK(KP667278218, +0.667278218140296670899089292254759909713898805);
|
Chris@10
|
102 DK(KP603558818, +0.603558818296015001454675132653458027918768137);
|
Chris@10
|
103 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
104 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
105 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
106 {
|
Chris@10
|
107 INT i;
|
Chris@10
|
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)) {
|
Chris@10
|
109 E T2R, T2T, T2D, T2C, T2H, T2G, T2B, T2P, T2S;
|
Chris@10
|
110 {
|
Chris@10
|
111 E T2A, TJ, T1K, T3l, T2z, TB, T2d, T2l, T1N, T21, T15, T1g, T1s, T1D, T9;
|
Chris@10
|
112 E T25, T1X, T2o, T2g, T1z, T1u, T1j, TQ, Ti, T1a, T2f, T2p, T1U, T24, TX;
|
Chris@10
|
113 E T1k, T1v, T1A, T19, Ts, T18, T1P;
|
Chris@10
|
114 {
|
Chris@10
|
115 E Tt, Tw, TZ, Tx, Ty;
|
Chris@10
|
116 {
|
Chris@10
|
117 E T2v, TG, TH, TD, TE, TI, T2x;
|
Chris@10
|
118 T2v = R0[0];
|
Chris@10
|
119 TG = R0[WS(rs, 10)];
|
Chris@10
|
120 TH = R1[WS(rs, 2)];
|
Chris@10
|
121 TD = R0[WS(rs, 5)];
|
Chris@10
|
122 TE = R1[WS(rs, 7)];
|
Chris@10
|
123 Tt = R0[WS(rs, 2)];
|
Chris@10
|
124 TI = TG + TH;
|
Chris@10
|
125 T2x = TG - TH;
|
Chris@10
|
126 {
|
Chris@10
|
127 E TF, T2w, Tu, Tv, T2y;
|
Chris@10
|
128 TF = TD + TE;
|
Chris@10
|
129 T2w = TD - TE;
|
Chris@10
|
130 Tu = R0[WS(rs, 7)];
|
Chris@10
|
131 Tv = R1[WS(rs, 9)];
|
Chris@10
|
132 T2A = T2w - T2x;
|
Chris@10
|
133 T2y = T2w + T2x;
|
Chris@10
|
134 TJ = FMA(KP618033988, TI, TF);
|
Chris@10
|
135 T1K = FNMS(KP618033988, TF, TI);
|
Chris@10
|
136 T3l = T2v + T2y;
|
Chris@10
|
137 T2z = FNMS(KP250000000, T2y, T2v);
|
Chris@10
|
138 Tw = Tu - Tv;
|
Chris@10
|
139 TZ = Tu + Tv;
|
Chris@10
|
140 Tx = R0[WS(rs, 12)];
|
Chris@10
|
141 Ty = R1[WS(rs, 4)];
|
Chris@10
|
142 }
|
Chris@10
|
143 }
|
Chris@10
|
144 {
|
Chris@10
|
145 E TO, TN, TM, T1V;
|
Chris@10
|
146 {
|
Chris@10
|
147 E T1, T1M, T11, T13, T4, TK, T12, TL, T7, T5, TA, T6, T14, T1L, T8;
|
Chris@10
|
148 T1 = R0[WS(rs, 1)];
|
Chris@10
|
149 {
|
Chris@10
|
150 E T2, T10, Tz, T3;
|
Chris@10
|
151 T2 = R0[WS(rs, 6)];
|
Chris@10
|
152 T10 = Tx + Ty;
|
Chris@10
|
153 Tz = Tx - Ty;
|
Chris@10
|
154 T3 = R1[WS(rs, 8)];
|
Chris@10
|
155 T5 = R0[WS(rs, 11)];
|
Chris@10
|
156 T1M = FNMS(KP618033988, TZ, T10);
|
Chris@10
|
157 T11 = FMA(KP618033988, T10, TZ);
|
Chris@10
|
158 T13 = Tz - Tw;
|
Chris@10
|
159 TA = Tw + Tz;
|
Chris@10
|
160 T4 = T2 - T3;
|
Chris@10
|
161 TK = T2 + T3;
|
Chris@10
|
162 T6 = R1[WS(rs, 3)];
|
Chris@10
|
163 }
|
Chris@10
|
164 TB = Tt + TA;
|
Chris@10
|
165 T12 = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
166 TL = T5 + T6;
|
Chris@10
|
167 T7 = T5 - T6;
|
Chris@10
|
168 T14 = FNMS(KP559016994, T13, T12);
|
Chris@10
|
169 T1L = FMA(KP559016994, T13, T12);
|
Chris@10
|
170 T8 = T4 + T7;
|
Chris@10
|
171 TO = T4 - T7;
|
Chris@10
|
172 T2d = FNMS(KP603558818, T1M, T1L);
|
Chris@10
|
173 T2l = FMA(KP667278218, T1L, T1M);
|
Chris@10
|
174 T1N = FMA(KP059835404, T1M, T1L);
|
Chris@10
|
175 T21 = FNMS(KP066152395, T1L, T1M);
|
Chris@10
|
176 T15 = FMA(KP578046249, T14, T11);
|
Chris@10
|
177 T1g = FNMS(KP522847744, T11, T14);
|
Chris@10
|
178 T1s = FMA(KP447533225, T11, T14);
|
Chris@10
|
179 T1D = FNMS(KP494780565, T14, T11);
|
Chris@10
|
180 TN = FNMS(KP250000000, T8, T1);
|
Chris@10
|
181 T9 = T1 + T8;
|
Chris@10
|
182 TM = FMA(KP618033988, TL, TK);
|
Chris@10
|
183 T1V = FNMS(KP618033988, TK, TL);
|
Chris@10
|
184 }
|
Chris@10
|
185 {
|
Chris@10
|
186 E Th, Td, TU, Tc, Te;
|
Chris@10
|
187 Th = R0[WS(rs, 4)];
|
Chris@10
|
188 {
|
Chris@10
|
189 E Ta, Tb, T1W, TP;
|
Chris@10
|
190 Ta = R0[WS(rs, 9)];
|
Chris@10
|
191 Tb = R1[WS(rs, 11)];
|
Chris@10
|
192 T1W = FNMS(KP559016994, TO, TN);
|
Chris@10
|
193 TP = FMA(KP559016994, TO, TN);
|
Chris@10
|
194 Td = R1[WS(rs, 6)];
|
Chris@10
|
195 TU = Ta + Tb;
|
Chris@10
|
196 Tc = Ta - Tb;
|
Chris@10
|
197 T25 = FNMS(KP893101515, T1V, T1W);
|
Chris@10
|
198 T1X = FMA(KP987388751, T1W, T1V);
|
Chris@10
|
199 T2o = FMA(KP522847744, T1V, T1W);
|
Chris@10
|
200 T2g = FNMS(KP578046249, T1W, T1V);
|
Chris@10
|
201 T1z = FMA(KP667278218, TP, TM);
|
Chris@10
|
202 T1u = FNMS(KP603558818, TM, TP);
|
Chris@10
|
203 T1j = FNMS(KP244189809, TM, TP);
|
Chris@10
|
204 TQ = FMA(KP269969613, TP, TM);
|
Chris@10
|
205 Te = R1[WS(rs, 1)];
|
Chris@10
|
206 }
|
Chris@10
|
207 {
|
Chris@10
|
208 E Tk, T1S, TW, TS, Tn, T16, TR, T17, Tq, To, Tg, Tp, TT, T1T, Tr;
|
Chris@10
|
209 Tk = R0[WS(rs, 3)];
|
Chris@10
|
210 {
|
Chris@10
|
211 E Tl, TV, Tf, Tm;
|
Chris@10
|
212 Tl = R0[WS(rs, 8)];
|
Chris@10
|
213 TV = Te - Td;
|
Chris@10
|
214 Tf = Td + Te;
|
Chris@10
|
215 Tm = R1[WS(rs, 10)];
|
Chris@10
|
216 To = R1[0];
|
Chris@10
|
217 T1S = FMA(KP618033988, TU, TV);
|
Chris@10
|
218 TW = FNMS(KP618033988, TV, TU);
|
Chris@10
|
219 TS = Tc + Tf;
|
Chris@10
|
220 Tg = Tc - Tf;
|
Chris@10
|
221 Tn = Tl - Tm;
|
Chris@10
|
222 T16 = Tl + Tm;
|
Chris@10
|
223 Tp = R1[WS(rs, 5)];
|
Chris@10
|
224 }
|
Chris@10
|
225 Ti = Tg + Th;
|
Chris@10
|
226 TR = FNMS(KP250000000, Tg, Th);
|
Chris@10
|
227 T17 = Tp - To;
|
Chris@10
|
228 Tq = To + Tp;
|
Chris@10
|
229 TT = FMA(KP559016994, TS, TR);
|
Chris@10
|
230 T1T = FNMS(KP559016994, TS, TR);
|
Chris@10
|
231 Tr = Tn - Tq;
|
Chris@10
|
232 T1a = Tn + Tq;
|
Chris@10
|
233 T2f = FNMS(KP447533225, T1S, T1T);
|
Chris@10
|
234 T2p = FMA(KP494780565, T1T, T1S);
|
Chris@10
|
235 T1U = FMA(KP132830569, T1T, T1S);
|
Chris@10
|
236 T24 = FNMS(KP120146378, T1S, T1T);
|
Chris@10
|
237 TX = FMA(KP603558818, TW, TT);
|
Chris@10
|
238 T1k = FNMS(KP667278218, TT, TW);
|
Chris@10
|
239 T1v = FNMS(KP786782374, TW, TT);
|
Chris@10
|
240 T1A = FMA(KP869845200, TT, TW);
|
Chris@10
|
241 T19 = FNMS(KP250000000, Tr, Tk);
|
Chris@10
|
242 Ts = Tk + Tr;
|
Chris@10
|
243 T18 = FMA(KP618033988, T17, T16);
|
Chris@10
|
244 T1P = FNMS(KP618033988, T16, T17);
|
Chris@10
|
245 }
|
Chris@10
|
246 }
|
Chris@10
|
247 }
|
Chris@10
|
248 }
|
Chris@10
|
249 {
|
Chris@10
|
250 E T22, T1Q, T1h, T1c, T2O, T2N, T2m, T3a, T3b, T2q, T1y, T3f, T2e, T2h, T3e;
|
Chris@10
|
251 E T1H, T1J;
|
Chris@10
|
252 {
|
Chris@10
|
253 E T3m, T3n, T2k, T2c, T1C, T1r;
|
Chris@10
|
254 {
|
Chris@10
|
255 E Tj, TC, T1O, T1b;
|
Chris@10
|
256 T3m = T9 + Ti;
|
Chris@10
|
257 Tj = T9 - Ti;
|
Chris@10
|
258 TC = Ts - TB;
|
Chris@10
|
259 T3n = TB + Ts;
|
Chris@10
|
260 T1O = FNMS(KP559016994, T1a, T19);
|
Chris@10
|
261 T1b = FMA(KP559016994, T1a, T19);
|
Chris@10
|
262 Ci[WS(csi, 7)] = KP951056516 * (FMA(KP618033988, Tj, TC));
|
Chris@10
|
263 Ci[WS(csi, 2)] = -(KP951056516 * (FNMS(KP618033988, TC, Tj)));
|
Chris@10
|
264 T22 = FMA(KP869845200, T1O, T1P);
|
Chris@10
|
265 T1Q = FNMS(KP786782374, T1P, T1O);
|
Chris@10
|
266 T2k = FMA(KP066152395, T1O, T1P);
|
Chris@10
|
267 T2c = FNMS(KP059835404, T1P, T1O);
|
Chris@10
|
268 T1C = FNMS(KP120146378, T18, T1b);
|
Chris@10
|
269 T1r = FMA(KP132830569, T1b, T18);
|
Chris@10
|
270 T1h = FNMS(KP893101515, T18, T1b);
|
Chris@10
|
271 T1c = FMA(KP987388751, T1b, T18);
|
Chris@10
|
272 }
|
Chris@10
|
273 {
|
Chris@10
|
274 E T1B, T1E, T1t, T3o, T3q, T1w, T3p;
|
Chris@10
|
275 T1B = FMA(KP912575812, T1A, T1z);
|
Chris@10
|
276 T2O = FNMS(KP912575812, T1A, T1z);
|
Chris@10
|
277 T2N = FNMS(KP867381224, T1D, T1C);
|
Chris@10
|
278 T1E = FMA(KP867381224, T1D, T1C);
|
Chris@10
|
279 T1t = FMA(KP958953096, T1s, T1r);
|
Chris@10
|
280 T2R = FNMS(KP958953096, T1s, T1r);
|
Chris@10
|
281 T3o = T3m + T3n;
|
Chris@10
|
282 T3q = T3m - T3n;
|
Chris@10
|
283 T2T = FMA(KP912575812, T1v, T1u);
|
Chris@10
|
284 T1w = FNMS(KP912575812, T1v, T1u);
|
Chris@10
|
285 T2m = FNMS(KP845997307, T2l, T2k);
|
Chris@10
|
286 T3a = FMA(KP845997307, T2l, T2k);
|
Chris@10
|
287 T3b = FNMS(KP982009705, T2p, T2o);
|
Chris@10
|
288 T2q = FMA(KP982009705, T2p, T2o);
|
Chris@10
|
289 T3p = FNMS(KP250000000, T3o, T3l);
|
Chris@10
|
290 Cr[WS(csr, 12)] = T3o + T3l;
|
Chris@10
|
291 {
|
Chris@10
|
292 E T1x, T1F, T1G, T1I;
|
Chris@10
|
293 T1x = FMA(KP894834959, T1w, T1t);
|
Chris@10
|
294 T1F = FNMS(KP894834959, T1w, T1t);
|
Chris@10
|
295 Cr[WS(csr, 7)] = FNMS(KP559016994, T3q, T3p);
|
Chris@10
|
296 Cr[WS(csr, 2)] = FMA(KP559016994, T3q, T3p);
|
Chris@10
|
297 T1y = FMA(KP248028675, T1x, TJ);
|
Chris@10
|
298 T1G = FNMS(KP904508497, T1F, T1E);
|
Chris@10
|
299 T1I = FNMS(KP894834959, T1B, T1F);
|
Chris@10
|
300 T3f = FNMS(KP845997307, T2d, T2c);
|
Chris@10
|
301 T2e = FMA(KP845997307, T2d, T2c);
|
Chris@10
|
302 T2h = FNMS(KP921078979, T2g, T2f);
|
Chris@10
|
303 T3e = FMA(KP921078979, T2g, T2f);
|
Chris@10
|
304 T1H = FMA(KP763583905, T1G, T1B);
|
Chris@10
|
305 T1J = FMA(KP559016994, T1I, T1E);
|
Chris@10
|
306 }
|
Chris@10
|
307 }
|
Chris@10
|
308 }
|
Chris@10
|
309 {
|
Chris@10
|
310 E T1i, T1l, T23, T30, T2Z, T26, T1R, T33, T1f, T1n, T1p, T34, T1Y, T3d, T3k;
|
Chris@10
|
311 E T3i;
|
Chris@10
|
312 {
|
Chris@10
|
313 E T2j, TY, T2s, T2u, T1d, T1m, T1e;
|
Chris@10
|
314 T2D = FMA(KP831864738, T1h, T1g);
|
Chris@10
|
315 T1i = FNMS(KP831864738, T1h, T1g);
|
Chris@10
|
316 {
|
Chris@10
|
317 E T2i, T2n, T2r, T2t;
|
Chris@10
|
318 T2i = FMA(KP906616052, T2h, T2e);
|
Chris@10
|
319 T2n = FNMS(KP906616052, T2h, T2e);
|
Chris@10
|
320 Ci[WS(csi, 4)] = KP951056516 * (FNMS(KP803003575, T1H, T1y));
|
Chris@10
|
321 Ci[WS(csi, 9)] = KP951056516 * (FNMS(KP992114701, T1J, T1y));
|
Chris@10
|
322 T2j = FMA(KP262346850, T2i, T1K);
|
Chris@10
|
323 T2r = FNMS(KP923225144, T2q, T2n);
|
Chris@10
|
324 T2t = T2m + T2n;
|
Chris@10
|
325 T2C = FNMS(KP829049696, T1k, T1j);
|
Chris@10
|
326 T1l = FMA(KP829049696, T1k, T1j);
|
Chris@10
|
327 TY = FMA(KP916574801, TX, TQ);
|
Chris@10
|
328 T2H = FNMS(KP916574801, TX, TQ);
|
Chris@10
|
329 T2s = FNMS(KP618033988, T2r, T2m);
|
Chris@10
|
330 T2u = FNMS(KP669429328, T2t, T2q);
|
Chris@10
|
331 T2G = FNMS(KP831864738, T1c, T15);
|
Chris@10
|
332 T1d = FMA(KP831864738, T1c, T15);
|
Chris@10
|
333 }
|
Chris@10
|
334 T23 = FNMS(KP772036680, T22, T21);
|
Chris@10
|
335 T30 = FMA(KP772036680, T22, T21);
|
Chris@10
|
336 Ci[WS(csi, 8)] = KP951056516 * (FMA(KP949179823, T2s, T2j));
|
Chris@10
|
337 Ci[WS(csi, 3)] = KP951056516 * (FNMS(KP876306680, T2u, T2j));
|
Chris@10
|
338 T1m = FNMS(KP904730450, T1d, TY);
|
Chris@10
|
339 T1e = FMA(KP904730450, T1d, TY);
|
Chris@10
|
340 T2Z = FNMS(KP734762448, T25, T24);
|
Chris@10
|
341 T26 = FMA(KP734762448, T25, T24);
|
Chris@10
|
342 T1R = FMA(KP772036680, T1Q, T1N);
|
Chris@10
|
343 T33 = FNMS(KP772036680, T1Q, T1N);
|
Chris@10
|
344 T1f = FNMS(KP242145790, T1e, TJ);
|
Chris@10
|
345 Ci[0] = -(KP951056516 * (FMA(KP968583161, T1e, TJ)));
|
Chris@10
|
346 T1n = FNMS(KP904508497, T1m, T1l);
|
Chris@10
|
347 T1p = FNMS(KP999754674, T1m, T1i);
|
Chris@10
|
348 T34 = FNMS(KP734762448, T1X, T1U);
|
Chris@10
|
349 T1Y = FMA(KP734762448, T1X, T1U);
|
Chris@10
|
350 }
|
Chris@10
|
351 {
|
Chris@10
|
352 E T2Y, T31, T38, T36, T3c, T3g;
|
Chris@10
|
353 {
|
Chris@10
|
354 E T20, T28, T2a, T29, T2b, T35;
|
Chris@10
|
355 T2Y = FNMS(KP559016994, T2A, T2z);
|
Chris@10
|
356 T2B = FMA(KP559016994, T2A, T2z);
|
Chris@10
|
357 {
|
Chris@10
|
358 E T1o, T1q, T27, T1Z;
|
Chris@10
|
359 T1o = FNMS(KP683113946, T1n, T1i);
|
Chris@10
|
360 T1q = FMA(KP559154169, T1p, T1l);
|
Chris@10
|
361 T27 = FNMS(KP945422727, T1Y, T1R);
|
Chris@10
|
362 T1Z = FMA(KP945422727, T1Y, T1R);
|
Chris@10
|
363 Ci[WS(csi, 5)] = -(KP951056516 * (FNMS(KP876306680, T1o, T1f)));
|
Chris@10
|
364 Ci[WS(csi, 10)] = -(KP951056516 * (FNMS(KP968583161, T1q, T1f)));
|
Chris@10
|
365 T20 = FNMS(KP262346850, T1Z, T1K);
|
Chris@10
|
366 Ci[WS(csi, 1)] = -(KP998026728 * (FMA(KP952936919, T1K, T1Z)));
|
Chris@10
|
367 T28 = FMA(KP956723877, T27, T26);
|
Chris@10
|
368 T2a = T27 - T23;
|
Chris@10
|
369 }
|
Chris@10
|
370 T29 = FMA(KP645989928, T28, T23);
|
Chris@10
|
371 T2b = FMA(KP591287873, T2a, T26);
|
Chris@10
|
372 Ci[WS(csi, 6)] = -(KP951056516 * (FMA(KP949179823, T29, T20)));
|
Chris@10
|
373 Ci[WS(csi, 11)] = -(KP951056516 * (FNMS(KP992114701, T2b, T20)));
|
Chris@10
|
374 T31 = FMA(KP956723877, T30, T2Z);
|
Chris@10
|
375 T35 = FNMS(KP956723877, T30, T2Z);
|
Chris@10
|
376 T38 = FMA(KP618033988, T35, T34);
|
Chris@10
|
377 T36 = T34 + T35;
|
Chris@10
|
378 }
|
Chris@10
|
379 Cr[WS(csr, 1)] = FNMS(KP992114701, T31, T2Y);
|
Chris@10
|
380 T3c = FMA(KP923225144, T3b, T3a);
|
Chris@10
|
381 T3g = FNMS(KP923225144, T3b, T3a);
|
Chris@10
|
382 {
|
Chris@10
|
383 E T32, T37, T3h, T3j, T39;
|
Chris@10
|
384 T32 = FMA(KP248028675, T31, T2Y);
|
Chris@10
|
385 T39 = FNMS(KP653711795, T33, T38);
|
Chris@10
|
386 T37 = FMA(KP584303379, T36, T33);
|
Chris@10
|
387 T3h = FNMS(KP904508497, T3g, T3f);
|
Chris@10
|
388 T3j = FNMS(KP997675361, T3g, T3e);
|
Chris@10
|
389 Cr[WS(csr, 11)] = FNMS(KP897376177, T39, T32);
|
Chris@10
|
390 Cr[WS(csr, 6)] = FMA(KP949179823, T37, T32);
|
Chris@10
|
391 T3d = FNMS(KP237294955, T3c, T2Y);
|
Chris@10
|
392 T3k = FNMS(KP560319534, T3j, T3f);
|
Chris@10
|
393 T3i = FMA(KP681693190, T3h, T3e);
|
Chris@10
|
394 }
|
Chris@10
|
395 }
|
Chris@10
|
396 Cr[WS(csr, 8)] = FMA(KP949179823, T3k, T3d);
|
Chris@10
|
397 Cr[WS(csr, 3)] = FMA(KP860541664, T3i, T3d);
|
Chris@10
|
398 T2P = FNMS(KP809385824, T2O, T2N);
|
Chris@10
|
399 T2S = FMA(KP809385824, T2O, T2N);
|
Chris@10
|
400 }
|
Chris@10
|
401 }
|
Chris@10
|
402 }
|
Chris@10
|
403 {
|
Chris@10
|
404 E T2F, T2K, T2M, T2Q;
|
Chris@10
|
405 T2Q = FMA(KP248028675, T2P, T2B);
|
Chris@10
|
406 {
|
Chris@10
|
407 E T2U, T2W, T2E, T2I;
|
Chris@10
|
408 T2U = FNMS(KP894834959, T2T, T2S);
|
Chris@10
|
409 T2W = T2R + T2S;
|
Chris@10
|
410 T2E = FMA(KP904730450, T2D, T2C);
|
Chris@10
|
411 T2I = FNMS(KP904730450, T2D, T2C);
|
Chris@10
|
412 {
|
Chris@10
|
413 E T2V, T2X, T2J, T2L;
|
Chris@10
|
414 T2V = FNMS(KP618033988, T2U, T2R);
|
Chris@10
|
415 T2X = FNMS(KP690668130, T2W, T2T);
|
Chris@10
|
416 T2F = FNMS(KP242145790, T2E, T2B);
|
Chris@10
|
417 Cr[0] = FMA(KP968583161, T2E, T2B);
|
Chris@10
|
418 T2J = T2H + T2I;
|
Chris@10
|
419 T2L = FMA(KP904730450, T2G, T2I);
|
Chris@10
|
420 Cr[WS(csr, 9)] = FMA(KP897376177, T2V, T2Q);
|
Chris@10
|
421 Cr[WS(csr, 4)] = FNMS(KP803003575, T2X, T2Q);
|
Chris@10
|
422 T2K = FNMS(KP683113946, T2J, T2G);
|
Chris@10
|
423 T2M = FMA(KP618033988, T2L, T2H);
|
Chris@10
|
424 }
|
Chris@10
|
425 }
|
Chris@10
|
426 Cr[WS(csr, 5)] = FMA(KP792626838, T2K, T2F);
|
Chris@10
|
427 Cr[WS(csr, 10)] = FMA(KP876091699, T2M, T2F);
|
Chris@10
|
428 }
|
Chris@10
|
429 }
|
Chris@10
|
430 }
|
Chris@10
|
431 }
|
Chris@10
|
432
|
Chris@10
|
433 static const kr2c_desc desc = { 25, "r2cfII_25", {47, 12, 165, 0}, &GENUS };
|
Chris@10
|
434
|
Chris@10
|
435 void X(codelet_r2cfII_25) (planner *p) {
|
Chris@10
|
436 X(kr2c_register) (p, r2cfII_25, &desc);
|
Chris@10
|
437 }
|
Chris@10
|
438
|
Chris@10
|
439 #else /* HAVE_FMA */
|
Chris@10
|
440
|
Chris@10
|
441 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cfII_25 -dft-II -include r2cfII.h */
|
Chris@10
|
442
|
Chris@10
|
443 /*
|
Chris@10
|
444 * This function contains 213 FP additions, 148 FP multiplications,
|
Chris@10
|
445 * (or, 126 additions, 61 multiplications, 87 fused multiply/add),
|
Chris@10
|
446 * 94 stack variables, 38 constants, and 50 memory accesses
|
Chris@10
|
447 */
|
Chris@10
|
448 #include "r2cfII.h"
|
Chris@10
|
449
|
Chris@10
|
450 static void r2cfII_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
451 {
|
Chris@10
|
452 DK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
|
Chris@10
|
453 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
454 DK(KP125581039, +0.125581039058626752152356449131262266244969664);
|
Chris@10
|
455 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
456 DK(KP1_369094211, +1.369094211857377347464566715242418539779038465);
|
Chris@10
|
457 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
458 DK(KP963507348, +0.963507348203430549974383005744259307057084020);
|
Chris@10
|
459 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
460 DK(KP497379774, +0.497379774329709576484567492012895936835134813);
|
Chris@10
|
461 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
462 DK(KP1_457937254, +1.457937254842823046293460638110518222745143328);
|
Chris@10
|
463 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
464 DK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
|
Chris@10
|
465 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
466 DK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
|
Chris@10
|
467 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
468 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
469 DK(KP250666467, +0.250666467128608490746237519633017587885836494);
|
Chris@10
|
470 DK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
|
Chris@10
|
471 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
472 DK(KP1_541026485, +1.541026485551578461606019272792355694543335344);
|
Chris@10
|
473 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
474 DK(KP1_688655851, +1.688655851004030157097116127933363010763318483);
|
Chris@10
|
475 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
476 DK(KP851558583, +0.851558583130145297725004891488503407959946084);
|
Chris@10
|
477 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
478 DK(KP1_984229402, +1.984229402628955662099586085571557042906073418);
|
Chris@10
|
479 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
480 DK(KP1_274847979, +1.274847979497379420353425623352032390869834596);
|
Chris@10
|
481 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
482 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
483 DK(KP1_071653589, +1.071653589957993236542617535735279956127150691);
|
Chris@10
|
484 DK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
Chris@10
|
485 DK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
Chris@10
|
486 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
487 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
488 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
489 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
490 {
|
Chris@10
|
491 INT i;
|
Chris@10
|
492 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)) {
|
Chris@10
|
493 E TE, TR, T2i, T1z, TL, TS, TB, T2d, T1l, T1i, T2c, T9, T23, TZ, TW;
|
Chris@10
|
494 E T22, Ti, T26, T16, T13, T25, Ts, T2a, T1e, T1b, T29, TP, TQ;
|
Chris@10
|
495 {
|
Chris@10
|
496 E TK, T1y, TH, T1x;
|
Chris@10
|
497 TE = R0[0];
|
Chris@10
|
498 {
|
Chris@10
|
499 E TI, TJ, TF, TG;
|
Chris@10
|
500 TI = R0[WS(rs, 10)];
|
Chris@10
|
501 TJ = R1[WS(rs, 2)];
|
Chris@10
|
502 TK = TI - TJ;
|
Chris@10
|
503 T1y = TI + TJ;
|
Chris@10
|
504 TF = R0[WS(rs, 5)];
|
Chris@10
|
505 TG = R1[WS(rs, 7)];
|
Chris@10
|
506 TH = TF - TG;
|
Chris@10
|
507 T1x = TF + TG;
|
Chris@10
|
508 }
|
Chris@10
|
509 TR = KP559016994 * (TH - TK);
|
Chris@10
|
510 T2i = FNMS(KP587785252, T1x, KP951056516 * T1y);
|
Chris@10
|
511 T1z = FMA(KP951056516, T1x, KP587785252 * T1y);
|
Chris@10
|
512 TL = TH + TK;
|
Chris@10
|
513 TS = FNMS(KP250000000, TL, TE);
|
Chris@10
|
514 }
|
Chris@10
|
515 {
|
Chris@10
|
516 E Tt, Tw, Tz, TA, T1k, T1j, T1g, T1h;
|
Chris@10
|
517 Tt = R0[WS(rs, 3)];
|
Chris@10
|
518 {
|
Chris@10
|
519 E Tu, Tv, Tx, Ty;
|
Chris@10
|
520 Tu = R0[WS(rs, 8)];
|
Chris@10
|
521 Tv = R1[WS(rs, 10)];
|
Chris@10
|
522 Tw = Tu - Tv;
|
Chris@10
|
523 Tx = R1[0];
|
Chris@10
|
524 Ty = R1[WS(rs, 5)];
|
Chris@10
|
525 Tz = Tx + Ty;
|
Chris@10
|
526 TA = Tw - Tz;
|
Chris@10
|
527 T1k = Ty - Tx;
|
Chris@10
|
528 T1j = Tu + Tv;
|
Chris@10
|
529 }
|
Chris@10
|
530 TB = Tt + TA;
|
Chris@10
|
531 T2d = FNMS(KP293892626, T1j, KP475528258 * T1k);
|
Chris@10
|
532 T1l = FMA(KP475528258, T1j, KP293892626 * T1k);
|
Chris@10
|
533 T1g = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
534 T1h = KP559016994 * (Tw + Tz);
|
Chris@10
|
535 T1i = T1g + T1h;
|
Chris@10
|
536 T2c = T1g - T1h;
|
Chris@10
|
537 }
|
Chris@10
|
538 {
|
Chris@10
|
539 E T1, T4, T7, T8, TY, TX, TU, TV;
|
Chris@10
|
540 T1 = R0[WS(rs, 1)];
|
Chris@10
|
541 {
|
Chris@10
|
542 E T2, T3, T5, T6;
|
Chris@10
|
543 T2 = R0[WS(rs, 6)];
|
Chris@10
|
544 T3 = R1[WS(rs, 8)];
|
Chris@10
|
545 T4 = T2 - T3;
|
Chris@10
|
546 T5 = R0[WS(rs, 11)];
|
Chris@10
|
547 T6 = R1[WS(rs, 3)];
|
Chris@10
|
548 T7 = T5 - T6;
|
Chris@10
|
549 T8 = T4 + T7;
|
Chris@10
|
550 TY = T5 + T6;
|
Chris@10
|
551 TX = T2 + T3;
|
Chris@10
|
552 }
|
Chris@10
|
553 T9 = T1 + T8;
|
Chris@10
|
554 T23 = FNMS(KP293892626, TX, KP475528258 * TY);
|
Chris@10
|
555 TZ = FMA(KP475528258, TX, KP293892626 * TY);
|
Chris@10
|
556 TU = KP559016994 * (T4 - T7);
|
Chris@10
|
557 TV = FNMS(KP250000000, T8, T1);
|
Chris@10
|
558 TW = TU + TV;
|
Chris@10
|
559 T22 = TV - TU;
|
Chris@10
|
560 }
|
Chris@10
|
561 {
|
Chris@10
|
562 E Ta, Td, Tg, Th, T15, T14, T11, T12;
|
Chris@10
|
563 Ta = R0[WS(rs, 4)];
|
Chris@10
|
564 {
|
Chris@10
|
565 E Tb, Tc, Te, Tf;
|
Chris@10
|
566 Tb = R0[WS(rs, 9)];
|
Chris@10
|
567 Tc = R1[WS(rs, 11)];
|
Chris@10
|
568 Td = Tb - Tc;
|
Chris@10
|
569 Te = R1[WS(rs, 1)];
|
Chris@10
|
570 Tf = R1[WS(rs, 6)];
|
Chris@10
|
571 Tg = Te + Tf;
|
Chris@10
|
572 Th = Td - Tg;
|
Chris@10
|
573 T15 = Tf - Te;
|
Chris@10
|
574 T14 = Tb + Tc;
|
Chris@10
|
575 }
|
Chris@10
|
576 Ti = Ta + Th;
|
Chris@10
|
577 T26 = FNMS(KP293892626, T14, KP475528258 * T15);
|
Chris@10
|
578 T16 = FMA(KP475528258, T14, KP293892626 * T15);
|
Chris@10
|
579 T11 = FNMS(KP250000000, Th, Ta);
|
Chris@10
|
580 T12 = KP559016994 * (Td + Tg);
|
Chris@10
|
581 T13 = T11 + T12;
|
Chris@10
|
582 T25 = T11 - T12;
|
Chris@10
|
583 }
|
Chris@10
|
584 {
|
Chris@10
|
585 E Tk, Tn, Tq, Tr, T1d, T1c, T19, T1a;
|
Chris@10
|
586 Tk = R0[WS(rs, 2)];
|
Chris@10
|
587 {
|
Chris@10
|
588 E Tl, Tm, To, Tp;
|
Chris@10
|
589 Tl = R0[WS(rs, 7)];
|
Chris@10
|
590 Tm = R1[WS(rs, 9)];
|
Chris@10
|
591 Tn = Tl - Tm;
|
Chris@10
|
592 To = R0[WS(rs, 12)];
|
Chris@10
|
593 Tp = R1[WS(rs, 4)];
|
Chris@10
|
594 Tq = To - Tp;
|
Chris@10
|
595 Tr = Tn + Tq;
|
Chris@10
|
596 T1d = To + Tp;
|
Chris@10
|
597 T1c = Tl + Tm;
|
Chris@10
|
598 }
|
Chris@10
|
599 Ts = Tk + Tr;
|
Chris@10
|
600 T2a = FNMS(KP293892626, T1c, KP475528258 * T1d);
|
Chris@10
|
601 T1e = FMA(KP475528258, T1c, KP293892626 * T1d);
|
Chris@10
|
602 T19 = KP559016994 * (Tn - Tq);
|
Chris@10
|
603 T1a = FNMS(KP250000000, Tr, Tk);
|
Chris@10
|
604 T1b = T19 + T1a;
|
Chris@10
|
605 T29 = T1a - T19;
|
Chris@10
|
606 }
|
Chris@10
|
607 TP = TB - Ts;
|
Chris@10
|
608 TQ = T9 - Ti;
|
Chris@10
|
609 Ci[WS(csi, 2)] = FNMS(KP951056516, TQ, KP587785252 * TP);
|
Chris@10
|
610 Ci[WS(csi, 7)] = FMA(KP587785252, TQ, KP951056516 * TP);
|
Chris@10
|
611 {
|
Chris@10
|
612 E TM, TD, TN, Tj, TC, TO;
|
Chris@10
|
613 TM = TE + TL;
|
Chris@10
|
614 Tj = T9 + Ti;
|
Chris@10
|
615 TC = Ts + TB;
|
Chris@10
|
616 TD = KP559016994 * (Tj - TC);
|
Chris@10
|
617 TN = Tj + TC;
|
Chris@10
|
618 Cr[WS(csr, 12)] = TM + TN;
|
Chris@10
|
619 TO = FNMS(KP250000000, TN, TM);
|
Chris@10
|
620 Cr[WS(csr, 2)] = TD + TO;
|
Chris@10
|
621 Cr[WS(csr, 7)] = TO - TD;
|
Chris@10
|
622 }
|
Chris@10
|
623 {
|
Chris@10
|
624 E TT, T1J, T1Y, T1U, T1X, T1P, T1V, T1M, T1W, T1A, T1B, T1r, T1C, T1v, T18;
|
Chris@10
|
625 E T1n, T1o, T1G, T1D;
|
Chris@10
|
626 TT = TR + TS;
|
Chris@10
|
627 {
|
Chris@10
|
628 E T1H, T1I, T1S, T1T;
|
Chris@10
|
629 T1H = FNMS(KP844327925, TW, KP1_071653589 * TZ);
|
Chris@10
|
630 T1I = FNMS(KP1_274847979, T16, KP770513242 * T13);
|
Chris@10
|
631 T1J = T1H - T1I;
|
Chris@10
|
632 T1Y = T1H + T1I;
|
Chris@10
|
633 T1S = FMA(KP125333233, T1i, KP1_984229402 * T1l);
|
Chris@10
|
634 T1T = FMA(KP904827052, T1b, KP851558583 * T1e);
|
Chris@10
|
635 T1U = T1S - T1T;
|
Chris@10
|
636 T1X = T1T + T1S;
|
Chris@10
|
637 }
|
Chris@10
|
638 {
|
Chris@10
|
639 E T1N, T1O, T1K, T1L;
|
Chris@10
|
640 T1N = FMA(KP535826794, TW, KP1_688655851 * TZ);
|
Chris@10
|
641 T1O = FMA(KP637423989, T13, KP1_541026485 * T16);
|
Chris@10
|
642 T1P = T1N - T1O;
|
Chris@10
|
643 T1V = T1N + T1O;
|
Chris@10
|
644 T1K = FNMS(KP1_809654104, T1e, KP425779291 * T1b);
|
Chris@10
|
645 T1L = FNMS(KP992114701, T1i, KP250666467 * T1l);
|
Chris@10
|
646 T1M = T1K - T1L;
|
Chris@10
|
647 T1W = T1K + T1L;
|
Chris@10
|
648 }
|
Chris@10
|
649 {
|
Chris@10
|
650 E T1p, T1q, T1t, T1u;
|
Chris@10
|
651 T1p = FMA(KP844327925, T13, KP1_071653589 * T16);
|
Chris@10
|
652 T1q = FMA(KP248689887, TW, KP1_937166322 * TZ);
|
Chris@10
|
653 T1A = T1q + T1p;
|
Chris@10
|
654 T1t = FMA(KP481753674, T1b, KP1_752613360 * T1e);
|
Chris@10
|
655 T1u = FMA(KP684547105, T1i, KP1_457937254 * T1l);
|
Chris@10
|
656 T1B = T1t + T1u;
|
Chris@10
|
657 T1r = T1p - T1q;
|
Chris@10
|
658 T1C = T1A + T1B;
|
Chris@10
|
659 T1v = T1t - T1u;
|
Chris@10
|
660 }
|
Chris@10
|
661 {
|
Chris@10
|
662 E T10, T17, T1f, T1m;
|
Chris@10
|
663 T10 = FNMS(KP497379774, TZ, KP968583161 * TW);
|
Chris@10
|
664 T17 = FNMS(KP1_688655851, T16, KP535826794 * T13);
|
Chris@10
|
665 T18 = T10 + T17;
|
Chris@10
|
666 T1f = FNMS(KP963507348, T1e, KP876306680 * T1b);
|
Chris@10
|
667 T1m = FNMS(KP1_369094211, T1l, KP728968627 * T1i);
|
Chris@10
|
668 T1n = T1f + T1m;
|
Chris@10
|
669 T1o = T18 + T1n;
|
Chris@10
|
670 T1G = T10 - T17;
|
Chris@10
|
671 T1D = T1f - T1m;
|
Chris@10
|
672 }
|
Chris@10
|
673 {
|
Chris@10
|
674 E T1R, T1Q, T20, T1Z;
|
Chris@10
|
675 Cr[0] = TT + T1o;
|
Chris@10
|
676 Ci[0] = -(T1z + T1C);
|
Chris@10
|
677 T1R = KP559016994 * (T1P + T1M);
|
Chris@10
|
678 T1Q = FMA(KP250000000, T1M - T1P, TT);
|
Chris@10
|
679 Cr[WS(csr, 4)] = FMA(KP951056516, T1J, T1Q) + FMA(KP587785252, T1U, T1R);
|
Chris@10
|
680 Cr[WS(csr, 9)] = FMA(KP951056516, T1U, T1Q) + FNMA(KP587785252, T1J, T1R);
|
Chris@10
|
681 T20 = KP559016994 * (T1Y + T1X);
|
Chris@10
|
682 T1Z = FMA(KP250000000, T1X - T1Y, T1z);
|
Chris@10
|
683 Ci[WS(csi, 9)] = FMA(KP587785252, T1V, KP951056516 * T1W) + T1Z - T20;
|
Chris@10
|
684 Ci[WS(csi, 4)] = FMA(KP587785252, T1W, T1Z) + FNMS(KP951056516, T1V, T20);
|
Chris@10
|
685 {
|
Chris@10
|
686 E T1E, T1F, T1s, T1w;
|
Chris@10
|
687 T1E = FMS(KP250000000, T1C, T1z);
|
Chris@10
|
688 T1F = KP559016994 * (T1B - T1A);
|
Chris@10
|
689 Ci[WS(csi, 5)] = FMA(KP951056516, T1D, T1E) + FNMA(KP587785252, T1G, T1F);
|
Chris@10
|
690 Ci[WS(csi, 10)] = FMA(KP951056516, T1G, KP587785252 * T1D) + T1E + T1F;
|
Chris@10
|
691 T1s = FNMS(KP250000000, T1o, TT);
|
Chris@10
|
692 T1w = KP559016994 * (T18 - T1n);
|
Chris@10
|
693 Cr[WS(csr, 5)] = FMA(KP587785252, T1r, T1s) + FMS(KP951056516, T1v, T1w);
|
Chris@10
|
694 Cr[WS(csr, 10)] = T1w + FMA(KP587785252, T1v, T1s) - (KP951056516 * T1r);
|
Chris@10
|
695 }
|
Chris@10
|
696 }
|
Chris@10
|
697 }
|
Chris@10
|
698 {
|
Chris@10
|
699 E T21, T2z, T2L, T2K, T2M, T2F, T2P, T2C, T2Q, T2l, T2o, T2p, T2w, T2u, T28;
|
Chris@10
|
700 E T2f, T2g, T2s, T2h;
|
Chris@10
|
701 T21 = TS - TR;
|
Chris@10
|
702 {
|
Chris@10
|
703 E T2x, T2y, T2I, T2J;
|
Chris@10
|
704 T2x = FNMS(KP844327925, T29, KP1_071653589 * T2a);
|
Chris@10
|
705 T2y = FNMS(KP125581039, T2d, KP998026728 * T2c);
|
Chris@10
|
706 T2z = T2x + T2y;
|
Chris@10
|
707 T2L = T2y - T2x;
|
Chris@10
|
708 T2I = FNMS(KP481753674, T22, KP1_752613360 * T23);
|
Chris@10
|
709 T2J = FMA(KP904827052, T25, KP851558583 * T26);
|
Chris@10
|
710 T2K = T2I + T2J;
|
Chris@10
|
711 T2M = T2I - T2J;
|
Chris@10
|
712 }
|
Chris@10
|
713 {
|
Chris@10
|
714 E T2D, T2E, T2A, T2B;
|
Chris@10
|
715 T2D = FMA(KP535826794, T29, KP1_688655851 * T2a);
|
Chris@10
|
716 T2E = FMA(KP062790519, T2c, KP1_996053456 * T2d);
|
Chris@10
|
717 T2F = T2D + T2E;
|
Chris@10
|
718 T2P = T2E - T2D;
|
Chris@10
|
719 T2A = FMA(KP876306680, T22, KP963507348 * T23);
|
Chris@10
|
720 T2B = FNMS(KP425779291, T25, KP1_809654104 * T26);
|
Chris@10
|
721 T2C = T2A + T2B;
|
Chris@10
|
722 T2Q = T2A - T2B;
|
Chris@10
|
723 }
|
Chris@10
|
724 {
|
Chris@10
|
725 E T2j, T2k, T2m, T2n;
|
Chris@10
|
726 T2j = FNMS(KP125333233, T25, KP1_984229402 * T26);
|
Chris@10
|
727 T2k = FMA(KP684547105, T22, KP1_457937254 * T23);
|
Chris@10
|
728 T2l = T2j - T2k;
|
Chris@10
|
729 T2m = FNMS(KP770513242, T2c, KP1_274847979 * T2d);
|
Chris@10
|
730 T2n = FMA(KP998026728, T29, KP125581039 * T2a);
|
Chris@10
|
731 T2o = T2m - T2n;
|
Chris@10
|
732 T2p = T2l + T2o;
|
Chris@10
|
733 T2w = T2k + T2j;
|
Chris@10
|
734 T2u = T2n + T2m;
|
Chris@10
|
735 }
|
Chris@10
|
736 {
|
Chris@10
|
737 E T24, T27, T2b, T2e;
|
Chris@10
|
738 T24 = FNMS(KP1_369094211, T23, KP728968627 * T22);
|
Chris@10
|
739 T27 = FMA(KP992114701, T25, KP250666467 * T26);
|
Chris@10
|
740 T28 = T24 - T27;
|
Chris@10
|
741 T2b = FNMS(KP1_996053456, T2a, KP062790519 * T29);
|
Chris@10
|
742 T2e = FMA(KP637423989, T2c, KP1_541026485 * T2d);
|
Chris@10
|
743 T2f = T2b - T2e;
|
Chris@10
|
744 T2g = T28 + T2f;
|
Chris@10
|
745 T2s = T24 + T27;
|
Chris@10
|
746 T2h = T2b + T2e;
|
Chris@10
|
747 }
|
Chris@10
|
748 {
|
Chris@10
|
749 E T2H, T2G, T2O, T2N;
|
Chris@10
|
750 Cr[WS(csr, 1)] = T21 + T2g;
|
Chris@10
|
751 Ci[WS(csi, 1)] = T2p - T2i;
|
Chris@10
|
752 T2H = KP559016994 * (T2C - T2F);
|
Chris@10
|
753 T2G = FNMS(KP250000000, T2C + T2F, T21);
|
Chris@10
|
754 Cr[WS(csr, 8)] = FMA(KP951056516, T2z, T2G) + FNMA(KP587785252, T2K, T2H);
|
Chris@10
|
755 Cr[WS(csr, 3)] = FMA(KP951056516, T2K, KP587785252 * T2z) + T2G + T2H;
|
Chris@10
|
756 T2O = KP559016994 * (T2M + T2L);
|
Chris@10
|
757 T2N = FMA(KP250000000, T2L - T2M, T2i);
|
Chris@10
|
758 Ci[WS(csi, 3)] = T2N + FMA(KP587785252, T2P, T2O) - (KP951056516 * T2Q);
|
Chris@10
|
759 Ci[WS(csi, 8)] = FMA(KP587785252, T2Q, T2N) + FMS(KP951056516, T2P, T2O);
|
Chris@10
|
760 {
|
Chris@10
|
761 E T2t, T2v, T2q, T2r;
|
Chris@10
|
762 T2t = FNMS(KP250000000, T2g, T21);
|
Chris@10
|
763 T2v = KP559016994 * (T28 - T2f);
|
Chris@10
|
764 Cr[WS(csr, 6)] = FMA(KP951056516, T2u, T2t) + FNMA(KP587785252, T2w, T2v);
|
Chris@10
|
765 Cr[WS(csr, 11)] = FMA(KP951056516, T2w, T2v) + FMA(KP587785252, T2u, T2t);
|
Chris@10
|
766 T2q = KP250000000 * T2p;
|
Chris@10
|
767 T2r = KP559016994 * (T2l - T2o);
|
Chris@10
|
768 Ci[WS(csi, 6)] = FMS(KP951056516, T2h, T2i + T2q) + FNMA(KP587785252, T2s, T2r);
|
Chris@10
|
769 Ci[WS(csi, 11)] = FMA(KP951056516, T2s, KP587785252 * T2h) + T2r - (T2i + T2q);
|
Chris@10
|
770 }
|
Chris@10
|
771 }
|
Chris@10
|
772 }
|
Chris@10
|
773 }
|
Chris@10
|
774 }
|
Chris@10
|
775 }
|
Chris@10
|
776
|
Chris@10
|
777 static const kr2c_desc desc = { 25, "r2cfII_25", {126, 61, 87, 0}, &GENUS };
|
Chris@10
|
778
|
Chris@10
|
779 void X(codelet_r2cfII_25) (planner *p) {
|
Chris@10
|
780 X(kr2c_register) (p, r2cfII_25, &desc);
|
Chris@10
|
781 }
|
Chris@10
|
782
|
Chris@10
|
783 #endif /* HAVE_FMA */
|