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 07:37:14 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-dft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 25 -name n1bv_25 -include n1b.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 224 FP additions, 193 FP multiplications,
|
Chris@10
|
32 * (or, 43 additions, 12 multiplications, 181 fused multiply/add),
|
Chris@10
|
33 * 215 stack variables, 67 constants, and 50 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "n1b.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void n1bv_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DVK(KP792626838, +0.792626838241819413632131824093538848057784557);
|
Chris@10
|
40 DVK(KP876091699, +0.876091699473550838204498029706869638173524346);
|
Chris@10
|
41 DVK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
Chris@10
|
42 DVK(KP617882369, +0.617882369114440893914546919006756321695042882);
|
Chris@10
|
43 DVK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
Chris@10
|
44 DVK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
45 DVK(KP999544308, +0.999544308746292983948881682379742149196758193);
|
Chris@10
|
46 DVK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
47 DVK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
48 DVK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
49 DVK(KP829049696, +0.829049696159252993975487806364305442437946767);
|
Chris@10
|
50 DVK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
51 DVK(KP916574801, +0.916574801383451584742370439148878693530976769);
|
Chris@10
|
52 DVK(KP894834959, +0.894834959464455102997960030820114611498661386);
|
Chris@10
|
53 DVK(KP809385824, +0.809385824416008241660603814668679683846476688);
|
Chris@10
|
54 DVK(KP447417479, +0.447417479732227551498980015410057305749330693);
|
Chris@10
|
55 DVK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
Chris@10
|
56 DVK(KP897376177, +0.897376177523557693138608077137219684419427330);
|
Chris@10
|
57 DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
58 DVK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
Chris@10
|
59 DVK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
Chris@10
|
60 DVK(KP855719849, +0.855719849902058969314654733608091555096772472);
|
Chris@10
|
61 DVK(KP237294955, +0.237294955877110315393888866460840817927895961);
|
Chris@10
|
62 DVK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
Chris@10
|
63 DVK(KP904508497, +0.904508497187473712051146708591409529430077295);
|
Chris@10
|
64 DVK(KP997675361, +0.997675361079556513670859573984492383596555031);
|
Chris@10
|
65 DVK(KP262346850, +0.262346850930607871785420028382979691334784273);
|
Chris@10
|
66 DVK(KP763932022, +0.763932022500210303590826331268723764559381640);
|
Chris@10
|
67 DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
68 DVK(KP690983005, +0.690983005625052575897706582817180941139845410);
|
Chris@10
|
69 DVK(KP952936919, +0.952936919628306576880750665357914584765951388);
|
Chris@10
|
70 DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
71 DVK(KP570584518, +0.570584518783621657366766175430996792655723863);
|
Chris@10
|
72 DVK(KP669429328, +0.669429328479476605641803240971985825917022098);
|
Chris@10
|
73 DVK(KP923225144, +0.923225144846402650453449441572664695995209956);
|
Chris@10
|
74 DVK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
Chris@10
|
75 DVK(KP956723877, +0.956723877038460305821989399535483155872969262);
|
Chris@10
|
76 DVK(KP522616830, +0.522616830205754336872861364785224694908468440);
|
Chris@10
|
77 DVK(KP945422727, +0.945422727388575946270360266328811958657216298);
|
Chris@10
|
78 DVK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
79 DVK(KP982009705, +0.982009705009746369461829878184175962711969869);
|
Chris@10
|
80 DVK(KP921078979, +0.921078979742360627699756128143719920817673854);
|
Chris@10
|
81 DVK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
Chris@10
|
82 DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
83 DVK(KP958953096, +0.958953096729998668045963838399037225970891871);
|
Chris@10
|
84 DVK(KP867381224, +0.867381224396525206773171885031575671309956167);
|
Chris@10
|
85 DVK(KP269969613, +0.269969613759572083574752974412347470060951301);
|
Chris@10
|
86 DVK(KP244189809, +0.244189809627953270309879511234821255780225091);
|
Chris@10
|
87 DVK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
Chris@10
|
88 DVK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
Chris@10
|
89 DVK(KP132830569, +0.132830569247582714407653942074819768844536507);
|
Chris@10
|
90 DVK(KP120146378, +0.120146378570687701782758537356596213647956445);
|
Chris@10
|
91 DVK(KP987388751, +0.987388751065621252324603216482382109400433949);
|
Chris@10
|
92 DVK(KP893101515, +0.893101515366181661711202267938416198338079437);
|
Chris@10
|
93 DVK(KP786782374, +0.786782374965295178365099601674911834788448471);
|
Chris@10
|
94 DVK(KP869845200, +0.869845200362138853122720822420327157933056305);
|
Chris@10
|
95 DVK(KP447533225, +0.447533225982656890041886979663652563063114397);
|
Chris@10
|
96 DVK(KP494780565, +0.494780565770515410344588413655324772219443730);
|
Chris@10
|
97 DVK(KP578046249, +0.578046249379945007321754579646815604023525655);
|
Chris@10
|
98 DVK(KP522847744, +0.522847744331509716623755382187077770911012542);
|
Chris@10
|
99 DVK(KP059835404, +0.059835404262124915169548397419498386427871950);
|
Chris@10
|
100 DVK(KP066152395, +0.066152395967733048213034281011006031460903353);
|
Chris@10
|
101 DVK(KP603558818, +0.603558818296015001454675132653458027918768137);
|
Chris@10
|
102 DVK(KP667278218, +0.667278218140296670899089292254759909713898805);
|
Chris@10
|
103 DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
104 DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
105 DVK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
106 {
|
Chris@10
|
107 INT i;
|
Chris@10
|
108 const R *xi;
|
Chris@10
|
109 R *xo;
|
Chris@10
|
110 xi = ii;
|
Chris@10
|
111 xo = io;
|
Chris@10
|
112 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(50, is), MAKE_VOLATILE_STRIDE(50, os)) {
|
Chris@10
|
113 V T1g, T1k, T1I, T24, T2a, T1G, T1A, T1l, T1B, T1H, T1d;
|
Chris@10
|
114 {
|
Chris@10
|
115 V T2z, T1q, Ta, T9, T3n, Ty, Tl, T2O, T2W, T2l, T2s, TV, T1i, T1K, T1S;
|
Chris@10
|
116 V T3z, T3t, Tk, T3o, Tp, T2g, T2N, T2V, T2o, T2t, T1a, T1j, T1J, T1R, Tz;
|
Chris@10
|
117 V Tt, TA, Tw;
|
Chris@10
|
118 {
|
Chris@10
|
119 V T1, T5, T6, T2, T3;
|
Chris@10
|
120 T1 = LD(&(xi[0]), ivs, &(xi[0]));
|
Chris@10
|
121 T5 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
|
Chris@10
|
122 T6 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
123 T2 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
124 T3 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
|
Chris@10
|
125 {
|
Chris@10
|
126 V TH, TW, TK, TS, T10, T8, TN, TT, T17, TZ, T11;
|
Chris@10
|
127 TH = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
|
Chris@10
|
128 TW = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
129 {
|
Chris@10
|
130 V TI, TJ, TL, T7, T1p, T4, T1o, TM, TX, TY;
|
Chris@10
|
131 TI = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
132 TJ = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
|
Chris@10
|
133 TL = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
|
Chris@10
|
134 T7 = VADD(T5, T6);
|
Chris@10
|
135 T1p = VSUB(T5, T6);
|
Chris@10
|
136 T4 = VADD(T2, T3);
|
Chris@10
|
137 T1o = VSUB(T2, T3);
|
Chris@10
|
138 TM = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
139 TX = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
|
Chris@10
|
140 TK = VADD(TI, TJ);
|
Chris@10
|
141 TS = VSUB(TI, TJ);
|
Chris@10
|
142 TY = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
143 T10 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
144 T2z = VFNMS(LDK(KP618033988), T1o, T1p);
|
Chris@10
|
145 T1q = VFMA(LDK(KP618033988), T1p, T1o);
|
Chris@10
|
146 Ta = VSUB(T4, T7);
|
Chris@10
|
147 T8 = VADD(T4, T7);
|
Chris@10
|
148 TN = VADD(TL, TM);
|
Chris@10
|
149 TT = VSUB(TM, TL);
|
Chris@10
|
150 T17 = VSUB(TX, TY);
|
Chris@10
|
151 TZ = VADD(TX, TY);
|
Chris@10
|
152 T11 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
|
Chris@10
|
153 }
|
Chris@10
|
154 {
|
Chris@10
|
155 V Tc, T2m, T19, Tn, To, Tr, Tj, T16, T2n, Ts, Tu, Tv;
|
Chris@10
|
156 {
|
Chris@10
|
157 V TU, T2j, TO, TQ, T12, T18;
|
Chris@10
|
158 Tc = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
159 T9 = VFNMS(LDK(KP250000000), T8, T1);
|
Chris@10
|
160 T3n = VADD(T1, T8);
|
Chris@10
|
161 TU = VFNMS(LDK(KP618033988), TT, TS);
|
Chris@10
|
162 T2j = VFMA(LDK(KP618033988), TS, TT);
|
Chris@10
|
163 TO = VADD(TK, TN);
|
Chris@10
|
164 TQ = VSUB(TN, TK);
|
Chris@10
|
165 T12 = VADD(T10, T11);
|
Chris@10
|
166 T18 = VSUB(T10, T11);
|
Chris@10
|
167 Ty = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
|
Chris@10
|
168 {
|
Chris@10
|
169 V T3r, T15, T13, Tf, Ti, T2k, TR, TP, T3s, T14;
|
Chris@10
|
170 {
|
Chris@10
|
171 V Td, Te, Tg, Th;
|
Chris@10
|
172 Td = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
|
Chris@10
|
173 Te = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
174 Tg = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
175 Th = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
|
Chris@10
|
176 TP = VFNMS(LDK(KP250000000), TO, TH);
|
Chris@10
|
177 T3r = VADD(TH, TO);
|
Chris@10
|
178 T2m = VFNMS(LDK(KP618033988), T17, T18);
|
Chris@10
|
179 T19 = VFMA(LDK(KP618033988), T18, T17);
|
Chris@10
|
180 T15 = VSUB(T12, TZ);
|
Chris@10
|
181 T13 = VADD(TZ, T12);
|
Chris@10
|
182 Tf = VADD(Td, Te);
|
Chris@10
|
183 Tn = VSUB(Td, Te);
|
Chris@10
|
184 To = VSUB(Th, Tg);
|
Chris@10
|
185 Ti = VADD(Tg, Th);
|
Chris@10
|
186 }
|
Chris@10
|
187 T2k = VFMA(LDK(KP559016994), TQ, TP);
|
Chris@10
|
188 TR = VFNMS(LDK(KP559016994), TQ, TP);
|
Chris@10
|
189 Tr = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
|
Chris@10
|
190 T3s = VADD(TW, T13);
|
Chris@10
|
191 T14 = VFNMS(LDK(KP250000000), T13, TW);
|
Chris@10
|
192 Tj = VADD(Tf, Ti);
|
Chris@10
|
193 Tl = VSUB(Tf, Ti);
|
Chris@10
|
194 T2O = VFNMS(LDK(KP667278218), T2k, T2j);
|
Chris@10
|
195 T2W = VFMA(LDK(KP603558818), T2j, T2k);
|
Chris@10
|
196 T2l = VFMA(LDK(KP066152395), T2k, T2j);
|
Chris@10
|
197 T2s = VFNMS(LDK(KP059835404), T2j, T2k);
|
Chris@10
|
198 TV = VFNMS(LDK(KP522847744), TU, TR);
|
Chris@10
|
199 T1i = VFMA(LDK(KP578046249), TR, TU);
|
Chris@10
|
200 T1K = VFNMS(LDK(KP494780565), TR, TU);
|
Chris@10
|
201 T1S = VFMA(LDK(KP447533225), TU, TR);
|
Chris@10
|
202 T16 = VFNMS(LDK(KP559016994), T15, T14);
|
Chris@10
|
203 T2n = VFMA(LDK(KP559016994), T15, T14);
|
Chris@10
|
204 T3z = VSUB(T3r, T3s);
|
Chris@10
|
205 T3t = VADD(T3r, T3s);
|
Chris@10
|
206 Ts = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
207 Tu = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
208 Tv = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
|
Chris@10
|
209 }
|
Chris@10
|
210 }
|
Chris@10
|
211 Tk = VFNMS(LDK(KP250000000), Tj, Tc);
|
Chris@10
|
212 T3o = VADD(Tc, Tj);
|
Chris@10
|
213 Tp = VFNMS(LDK(KP618033988), To, Tn);
|
Chris@10
|
214 T2g = VFMA(LDK(KP618033988), Tn, To);
|
Chris@10
|
215 T2N = VFMA(LDK(KP066152395), T2n, T2m);
|
Chris@10
|
216 T2V = VFNMS(LDK(KP059835404), T2m, T2n);
|
Chris@10
|
217 T2o = VFMA(LDK(KP869845200), T2n, T2m);
|
Chris@10
|
218 T2t = VFNMS(LDK(KP786782374), T2m, T2n);
|
Chris@10
|
219 T1a = VFNMS(LDK(KP893101515), T19, T16);
|
Chris@10
|
220 T1j = VFMA(LDK(KP987388751), T16, T19);
|
Chris@10
|
221 T1J = VFNMS(LDK(KP120146378), T19, T16);
|
Chris@10
|
222 T1R = VFMA(LDK(KP132830569), T16, T19);
|
Chris@10
|
223 Tz = VADD(Ts, Tr);
|
Chris@10
|
224 Tt = VSUB(Tr, Ts);
|
Chris@10
|
225 TA = VADD(Tv, Tu);
|
Chris@10
|
226 Tw = VSUB(Tu, Tv);
|
Chris@10
|
227 }
|
Chris@10
|
228 }
|
Chris@10
|
229 }
|
Chris@10
|
230 {
|
Chris@10
|
231 V T2p, T2I, T2u, T2C, Tx, T2d, T2X, T34, T2P, T3b, T2b, Tb, T2Q, T2Z, T2h;
|
Chris@10
|
232 V T2w, Tq, T1e, T1M, T1U, TE, T2c, T3q, T3y;
|
Chris@10
|
233 T2p = VFNMS(LDK(KP772036680), T2o, T2l);
|
Chris@10
|
234 T2I = VFMA(LDK(KP772036680), T2o, T2l);
|
Chris@10
|
235 T2u = VFMA(LDK(KP772036680), T2t, T2s);
|
Chris@10
|
236 T2C = VFNMS(LDK(KP772036680), T2t, T2s);
|
Chris@10
|
237 {
|
Chris@10
|
238 V TD, TB, Tm, T2f, T3p, TC;
|
Chris@10
|
239 Tx = VFMA(LDK(KP618033988), Tw, Tt);
|
Chris@10
|
240 T2d = VFNMS(LDK(KP618033988), Tt, Tw);
|
Chris@10
|
241 TD = VSUB(Tz, TA);
|
Chris@10
|
242 TB = VADD(Tz, TA);
|
Chris@10
|
243 Tm = VFMA(LDK(KP559016994), Tl, Tk);
|
Chris@10
|
244 T2f = VFNMS(LDK(KP559016994), Tl, Tk);
|
Chris@10
|
245 T2X = VFMA(LDK(KP845997307), T2W, T2V);
|
Chris@10
|
246 T34 = VFNMS(LDK(KP845997307), T2W, T2V);
|
Chris@10
|
247 T2P = VFNMS(LDK(KP845997307), T2O, T2N);
|
Chris@10
|
248 T3b = VFMA(LDK(KP845997307), T2O, T2N);
|
Chris@10
|
249 T2b = VFNMS(LDK(KP559016994), Ta, T9);
|
Chris@10
|
250 Tb = VFMA(LDK(KP559016994), Ta, T9);
|
Chris@10
|
251 T3p = VADD(Ty, TB);
|
Chris@10
|
252 TC = VFMS(LDK(KP250000000), TB, Ty);
|
Chris@10
|
253 T2Q = VFNMS(LDK(KP522847744), T2g, T2f);
|
Chris@10
|
254 T2Z = VFMA(LDK(KP578046249), T2f, T2g);
|
Chris@10
|
255 T2h = VFMA(LDK(KP893101515), T2g, T2f);
|
Chris@10
|
256 T2w = VFNMS(LDK(KP987388751), T2f, T2g);
|
Chris@10
|
257 Tq = VFNMS(LDK(KP244189809), Tp, Tm);
|
Chris@10
|
258 T1e = VFMA(LDK(KP269969613), Tm, Tp);
|
Chris@10
|
259 T1M = VFMA(LDK(KP667278218), Tm, Tp);
|
Chris@10
|
260 T1U = VFNMS(LDK(KP603558818), Tp, Tm);
|
Chris@10
|
261 TE = VFNMS(LDK(KP559016994), TD, TC);
|
Chris@10
|
262 T2c = VFMA(LDK(KP559016994), TD, TC);
|
Chris@10
|
263 T3q = VADD(T3o, T3p);
|
Chris@10
|
264 T3y = VSUB(T3o, T3p);
|
Chris@10
|
265 }
|
Chris@10
|
266 {
|
Chris@10
|
267 V T1Z, T25, T1P, T22, T1X, TG, T1b, T28, T1t, T1y, T1x, T1E, T1Q, T1Y;
|
Chris@10
|
268 {
|
Chris@10
|
269 V T26, T1L, T1T, TF, T1f, T1W, T3m, T3g, T2M, T2G, T39, T3j, T21, T1O, T20;
|
Chris@10
|
270 V T27;
|
Chris@10
|
271 T26 = VFMA(LDK(KP867381224), T1K, T1J);
|
Chris@10
|
272 T1L = VFNMS(LDK(KP867381224), T1K, T1J);
|
Chris@10
|
273 T20 = VFNMS(LDK(KP958953096), T1S, T1R);
|
Chris@10
|
274 T1T = VFMA(LDK(KP958953096), T1S, T1R);
|
Chris@10
|
275 {
|
Chris@10
|
276 V T2R, T2Y, T2e, T2v, T1N, T1V;
|
Chris@10
|
277 T2R = VFNMS(LDK(KP494780565), T2c, T2d);
|
Chris@10
|
278 T2Y = VFMA(LDK(KP447533225), T2d, T2c);
|
Chris@10
|
279 T2e = VFMA(LDK(KP120146378), T2d, T2c);
|
Chris@10
|
280 T2v = VFNMS(LDK(KP132830569), T2c, T2d);
|
Chris@10
|
281 TF = VFNMS(LDK(KP667278218), TE, Tx);
|
Chris@10
|
282 T1f = VFMA(LDK(KP603558818), Tx, TE);
|
Chris@10
|
283 T1N = VFMA(LDK(KP869845200), TE, Tx);
|
Chris@10
|
284 T1V = VFNMS(LDK(KP786782374), Tx, TE);
|
Chris@10
|
285 {
|
Chris@10
|
286 V T3A, T3C, T3w, T3u;
|
Chris@10
|
287 T3A = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T3z, T3y));
|
Chris@10
|
288 T3C = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T3y, T3z));
|
Chris@10
|
289 T3w = VSUB(T3q, T3t);
|
Chris@10
|
290 T3u = VADD(T3q, T3t);
|
Chris@10
|
291 {
|
Chris@10
|
292 V T2B, T2x, T2H, T2i;
|
Chris@10
|
293 T2B = VFMA(LDK(KP734762448), T2w, T2v);
|
Chris@10
|
294 T2x = VFNMS(LDK(KP734762448), T2w, T2v);
|
Chris@10
|
295 T2H = VFNMS(LDK(KP734762448), T2h, T2e);
|
Chris@10
|
296 T2i = VFMA(LDK(KP734762448), T2h, T2e);
|
Chris@10
|
297 {
|
Chris@10
|
298 V T30, T35, T3c, T2S, T3v;
|
Chris@10
|
299 T30 = VFNMS(LDK(KP921078979), T2Z, T2Y);
|
Chris@10
|
300 T35 = VFMA(LDK(KP921078979), T2Z, T2Y);
|
Chris@10
|
301 T3c = VFMA(LDK(KP982009705), T2R, T2Q);
|
Chris@10
|
302 T2S = VFNMS(LDK(KP982009705), T2R, T2Q);
|
Chris@10
|
303 T1W = VFMA(LDK(KP912575812), T1V, T1U);
|
Chris@10
|
304 T1Z = VFNMS(LDK(KP912575812), T1V, T1U);
|
Chris@10
|
305 T1O = VFMA(LDK(KP912575812), T1N, T1M);
|
Chris@10
|
306 T25 = VFNMS(LDK(KP912575812), T1N, T1M);
|
Chris@10
|
307 ST(&(xo[0]), VADD(T3u, T3n), ovs, &(xo[0]));
|
Chris@10
|
308 T3v = VFNMS(LDK(KP250000000), T3u, T3n);
|
Chris@10
|
309 {
|
Chris@10
|
310 V T2y, T2J, T2q, T2D;
|
Chris@10
|
311 T2y = VFMA(LDK(KP945422727), T2x, T2u);
|
Chris@10
|
312 T2J = VFMA(LDK(KP522616830), T2x, T2I);
|
Chris@10
|
313 T2q = VFMA(LDK(KP956723877), T2p, T2i);
|
Chris@10
|
314 T2D = VFNMS(LDK(KP522616830), T2i, T2C);
|
Chris@10
|
315 {
|
Chris@10
|
316 V T3e, T31, T36, T2T;
|
Chris@10
|
317 T3e = VFMA(LDK(KP906616052), T30, T2X);
|
Chris@10
|
318 T31 = VFNMS(LDK(KP906616052), T30, T2X);
|
Chris@10
|
319 T36 = VFNMS(LDK(KP923225144), T2S, T2P);
|
Chris@10
|
320 T2T = VFMA(LDK(KP923225144), T2S, T2P);
|
Chris@10
|
321 {
|
Chris@10
|
322 V T3k, T3d, T3x, T3B;
|
Chris@10
|
323 T3k = VFNMS(LDK(KP669429328), T3b, T3c);
|
Chris@10
|
324 T3d = VFMA(LDK(KP570584518), T3c, T3b);
|
Chris@10
|
325 T3x = VFMA(LDK(KP559016994), T3w, T3v);
|
Chris@10
|
326 T3B = VFNMS(LDK(KP559016994), T3w, T3v);
|
Chris@10
|
327 {
|
Chris@10
|
328 V T2A, T2K, T2r, T2E;
|
Chris@10
|
329 T2A = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T2z, T2y));
|
Chris@10
|
330 T2K = VFNMS(LDK(KP690983005), T2J, T2u);
|
Chris@10
|
331 T2r = VFMA(LDK(KP992114701), T2q, T2b);
|
Chris@10
|
332 T2E = VFMA(LDK(KP763932022), T2D, T2p);
|
Chris@10
|
333 {
|
Chris@10
|
334 V T32, T3a, T37, T3h;
|
Chris@10
|
335 T32 = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T2z, T31));
|
Chris@10
|
336 T3a = VFMA(LDK(KP262346850), T31, T2z);
|
Chris@10
|
337 T37 = VFNMS(LDK(KP997675361), T36, T35);
|
Chris@10
|
338 T3h = VFNMS(LDK(KP904508497), T36, T34);
|
Chris@10
|
339 {
|
Chris@10
|
340 V T2U, T33, T3l, T3f;
|
Chris@10
|
341 T2U = VFMA(LDK(KP949179823), T2T, T2b);
|
Chris@10
|
342 T33 = VFNMS(LDK(KP237294955), T2T, T2b);
|
Chris@10
|
343 T3l = VFNMS(LDK(KP669429328), T3e, T3k);
|
Chris@10
|
344 T3f = VFMA(LDK(KP618033988), T3e, T3d);
|
Chris@10
|
345 ST(&(xo[WS(os, 20)]), VFNMSI(T3A, T3x), ovs, &(xo[0]));
|
Chris@10
|
346 ST(&(xo[WS(os, 5)]), VFMAI(T3A, T3x), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
347 ST(&(xo[WS(os, 15)]), VFMAI(T3C, T3B), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
348 ST(&(xo[WS(os, 10)]), VFNMSI(T3C, T3B), ovs, &(xo[0]));
|
Chris@10
|
349 {
|
Chris@10
|
350 V T2L, T2F, T38, T3i;
|
Chris@10
|
351 T2L = VFMA(LDK(KP855719849), T2K, T2H);
|
Chris@10
|
352 ST(&(xo[WS(os, 3)]), VFMAI(T2A, T2r), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
353 ST(&(xo[WS(os, 22)]), VFNMSI(T2A, T2r), ovs, &(xo[0]));
|
Chris@10
|
354 T2F = VFNMS(LDK(KP855719849), T2E, T2B);
|
Chris@10
|
355 T38 = VFMA(LDK(KP560319534), T37, T34);
|
Chris@10
|
356 T3i = VFNMS(LDK(KP681693190), T3h, T35);
|
Chris@10
|
357 ST(&(xo[WS(os, 2)]), VFMAI(T32, T2U), ovs, &(xo[0]));
|
Chris@10
|
358 ST(&(xo[WS(os, 23)]), VFNMSI(T32, T2U), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
359 T3m = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T3l, T3a));
|
Chris@10
|
360 T3g = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T3f, T3a));
|
Chris@10
|
361 T2M = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2L, T2z));
|
Chris@10
|
362 T2G = VFMA(LDK(KP897376177), T2F, T2b);
|
Chris@10
|
363 T39 = VFNMS(LDK(KP949179823), T38, T33);
|
Chris@10
|
364 T3j = VFNMS(LDK(KP860541664), T3i, T33);
|
Chris@10
|
365 T21 = VFMA(LDK(KP447417479), T1O, T20);
|
Chris@10
|
366 }
|
Chris@10
|
367 }
|
Chris@10
|
368 }
|
Chris@10
|
369 }
|
Chris@10
|
370 }
|
Chris@10
|
371 }
|
Chris@10
|
372 }
|
Chris@10
|
373 }
|
Chris@10
|
374 }
|
Chris@10
|
375 }
|
Chris@10
|
376 }
|
Chris@10
|
377 T1P = VFNMS(LDK(KP809385824), T1O, T1L);
|
Chris@10
|
378 ST(&(xo[WS(os, 17)]), VFNMSI(T2M, T2G), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
379 ST(&(xo[WS(os, 8)]), VFMAI(T2M, T2G), ovs, &(xo[0]));
|
Chris@10
|
380 ST(&(xo[WS(os, 13)]), VFMAI(T3g, T39), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
381 ST(&(xo[WS(os, 12)]), VFNMSI(T3g, T39), ovs, &(xo[0]));
|
Chris@10
|
382 ST(&(xo[WS(os, 18)]), VFMAI(T3m, T3j), ovs, &(xo[0]));
|
Chris@10
|
383 ST(&(xo[WS(os, 7)]), VFNMSI(T3m, T3j), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
384 T22 = VFMA(LDK(KP690983005), T21, T1L);
|
Chris@10
|
385 T27 = VFMA(LDK(KP447417479), T1W, T26);
|
Chris@10
|
386 T1X = VFMA(LDK(KP894834959), T1W, T1T);
|
Chris@10
|
387 {
|
Chris@10
|
388 V T1r, T1s, T1v, T1w;
|
Chris@10
|
389 T1r = VFNMS(LDK(KP916574801), T1f, T1e);
|
Chris@10
|
390 T1g = VFMA(LDK(KP916574801), T1f, T1e);
|
Chris@10
|
391 T1k = VFNMS(LDK(KP831864738), T1j, T1i);
|
Chris@10
|
392 T1s = VFMA(LDK(KP831864738), T1j, T1i);
|
Chris@10
|
393 T1v = VFNMS(LDK(KP829049696), TF, Tq);
|
Chris@10
|
394 TG = VFMA(LDK(KP829049696), TF, Tq);
|
Chris@10
|
395 T1b = VFMA(LDK(KP831864738), T1a, TV);
|
Chris@10
|
396 T1w = VFNMS(LDK(KP831864738), T1a, TV);
|
Chris@10
|
397 T28 = VFNMS(LDK(KP763932022), T27, T1T);
|
Chris@10
|
398 T1t = VFMA(LDK(KP904730450), T1s, T1r);
|
Chris@10
|
399 T1y = VFNMS(LDK(KP904730450), T1s, T1r);
|
Chris@10
|
400 T1x = VFMA(LDK(KP559154169), T1w, T1v);
|
Chris@10
|
401 T1E = VFNMS(LDK(KP683113946), T1v, T1w);
|
Chris@10
|
402 }
|
Chris@10
|
403 }
|
Chris@10
|
404 T1Q = VFNMS(LDK(KP992114701), T1P, Tb);
|
Chris@10
|
405 T1Y = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T1X, T1q));
|
Chris@10
|
406 {
|
Chris@10
|
407 V T1u, T1F, T1z, T1h, T1c, T23, T29;
|
Chris@10
|
408 T23 = VFNMS(LDK(KP999544308), T22, T1Z);
|
Chris@10
|
409 T29 = VFNMS(LDK(KP999544308), T28, T25);
|
Chris@10
|
410 T1I = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1t, T1q));
|
Chris@10
|
411 T1u = VFNMS(LDK(KP242145790), T1t, T1q);
|
Chris@10
|
412 T1F = VFMA(LDK(KP617882369), T1y, T1E);
|
Chris@10
|
413 T1z = VFMA(LDK(KP559016994), T1y, T1x);
|
Chris@10
|
414 T1h = VFNMS(LDK(KP904730450), T1b, TG);
|
Chris@10
|
415 T1c = VFMA(LDK(KP904730450), T1b, TG);
|
Chris@10
|
416 ST(&(xo[WS(os, 21)]), VFMAI(T1Y, T1Q), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
417 ST(&(xo[WS(os, 4)]), VFNMSI(T1Y, T1Q), ovs, &(xo[0]));
|
Chris@10
|
418 T24 = VFNMS(LDK(KP803003575), T23, Tb);
|
Chris@10
|
419 T2a = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T29, T1q));
|
Chris@10
|
420 T1G = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T1F, T1u));
|
Chris@10
|
421 T1A = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1z, T1u));
|
Chris@10
|
422 T1l = VFNMS(LDK(KP904730450), T1k, T1h);
|
Chris@10
|
423 T1B = VADD(T1g, T1h);
|
Chris@10
|
424 T1H = VFMA(LDK(KP968583161), T1c, Tb);
|
Chris@10
|
425 T1d = VFNMS(LDK(KP242145790), T1c, Tb);
|
Chris@10
|
426 }
|
Chris@10
|
427 }
|
Chris@10
|
428 }
|
Chris@10
|
429 }
|
Chris@10
|
430 ST(&(xo[WS(os, 16)]), VFMAI(T2a, T24), ovs, &(xo[0]));
|
Chris@10
|
431 ST(&(xo[WS(os, 9)]), VFNMSI(T2a, T24), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
432 {
|
Chris@10
|
433 V T1m, T1C, T1n, T1D;
|
Chris@10
|
434 T1m = VFNMS(LDK(KP618033988), T1l, T1g);
|
Chris@10
|
435 T1C = VFNMS(LDK(KP683113946), T1B, T1k);
|
Chris@10
|
436 ST(&(xo[WS(os, 24)]), VFNMSI(T1I, T1H), ovs, &(xo[0]));
|
Chris@10
|
437 ST(&(xo[WS(os, 1)]), VFMAI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
438 T1n = VFNMS(LDK(KP876091699), T1m, T1d);
|
Chris@10
|
439 T1D = VFMA(LDK(KP792626838), T1C, T1d);
|
Chris@10
|
440 ST(&(xo[WS(os, 19)]), VFNMSI(T1A, T1n), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
441 ST(&(xo[WS(os, 6)]), VFMAI(T1A, T1n), ovs, &(xo[0]));
|
Chris@10
|
442 ST(&(xo[WS(os, 14)]), VFNMSI(T1G, T1D), ovs, &(xo[0]));
|
Chris@10
|
443 ST(&(xo[WS(os, 11)]), VFMAI(T1G, T1D), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
444 }
|
Chris@10
|
445 }
|
Chris@10
|
446 }
|
Chris@10
|
447 VLEAVE();
|
Chris@10
|
448 }
|
Chris@10
|
449
|
Chris@10
|
450 static const kdft_desc desc = { 25, XSIMD_STRING("n1bv_25"), {43, 12, 181, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
451
|
Chris@10
|
452 void XSIMD(codelet_n1bv_25) (planner *p) {
|
Chris@10
|
453 X(kdft_register) (p, n1bv_25, &desc);
|
Chris@10
|
454 }
|
Chris@10
|
455
|
Chris@10
|
456 #else /* HAVE_FMA */
|
Chris@10
|
457
|
Chris@10
|
458 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 25 -name n1bv_25 -include n1b.h */
|
Chris@10
|
459
|
Chris@10
|
460 /*
|
Chris@10
|
461 * This function contains 224 FP additions, 140 FP multiplications,
|
Chris@10
|
462 * (or, 147 additions, 63 multiplications, 77 fused multiply/add),
|
Chris@10
|
463 * 115 stack variables, 40 constants, and 50 memory accesses
|
Chris@10
|
464 */
|
Chris@10
|
465 #include "n1b.h"
|
Chris@10
|
466
|
Chris@10
|
467 static void n1bv_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
468 {
|
Chris@10
|
469 DVK(KP809016994, +0.809016994374947424102293417182819058860154590);
|
Chris@10
|
470 DVK(KP309016994, +0.309016994374947424102293417182819058860154590);
|
Chris@10
|
471 DVK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
472 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344);
|
Chris@10
|
473 DVK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
474 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418);
|
Chris@10
|
475 DVK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
476 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596);
|
Chris@10
|
477 DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
478 DVK(KP250666467, +0.250666467128608490746237519633017587885836494);
|
Chris@10
|
479 DVK(KP851558583, +0.851558583130145297725004891488503407959946084);
|
Chris@10
|
480 DVK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
481 DVK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
482 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
|
Chris@10
|
483 DVK(KP497379774, +0.497379774329709576484567492012895936835134813);
|
Chris@10
|
484 DVK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
485 DVK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
486 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
|
Chris@10
|
487 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483);
|
Chris@10
|
488 DVK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
489 DVK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
490 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
|
Chris@10
|
491 DVK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
492 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691);
|
Chris@10
|
493 DVK(KP963507348, +0.963507348203430549974383005744259307057084020);
|
Chris@10
|
494 DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
495 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
|
Chris@10
|
496 DVK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
497 DVK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
498 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328);
|
Chris@10
|
499 DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
500 DVK(KP125581039, +0.125581039058626752152356449131262266244969664);
|
Chris@10
|
501 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465);
|
Chris@10
|
502 DVK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
503 DVK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
Chris@10
|
504 DVK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
Chris@10
|
505 DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
506 DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
507 DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
508 DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
509 {
|
Chris@10
|
510 INT i;
|
Chris@10
|
511 const R *xi;
|
Chris@10
|
512 R *xo;
|
Chris@10
|
513 xi = ii;
|
Chris@10
|
514 xo = io;
|
Chris@10
|
515 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(50, is), MAKE_VOLATILE_STRIDE(50, os)) {
|
Chris@10
|
516 V T1b, T2o, T1v, T1e, T2W, T2P, T2Q, T2U, T11, T27, TY, T26, T12, T2f, T1j;
|
Chris@10
|
517 V T28, TM, T24, TJ, T23, TN, T2e, T1i, T25, T2M, T2N, T2T, Tm, T1W, Tt;
|
Chris@10
|
518 V T1X, Tu, T20, Tw, T1Y, T7, T1U, Te, T1T, Tf, T21, Tx, T1V;
|
Chris@10
|
519 {
|
Chris@10
|
520 V T1c, T1a, T1t, T17, T1r;
|
Chris@10
|
521 T1c = LD(&(xi[0]), ivs, &(xi[0]));
|
Chris@10
|
522 {
|
Chris@10
|
523 V T18, T19, T15, T16;
|
Chris@10
|
524 T18 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
|
Chris@10
|
525 T19 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
526 T1a = VADD(T18, T19);
|
Chris@10
|
527 T1t = VSUB(T18, T19);
|
Chris@10
|
528 T15 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
529 T16 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
|
Chris@10
|
530 T17 = VADD(T15, T16);
|
Chris@10
|
531 T1r = VSUB(T15, T16);
|
Chris@10
|
532 }
|
Chris@10
|
533 {
|
Chris@10
|
534 V T2n, T1s, T1u, T1d;
|
Chris@10
|
535 T1b = VMUL(LDK(KP559016994), VSUB(T17, T1a));
|
Chris@10
|
536 T2n = VMUL(LDK(KP587785252), T1r);
|
Chris@10
|
537 T2o = VFNMS(LDK(KP951056516), T1t, T2n);
|
Chris@10
|
538 T1s = VMUL(LDK(KP951056516), T1r);
|
Chris@10
|
539 T1u = VMUL(LDK(KP587785252), T1t);
|
Chris@10
|
540 T1v = VADD(T1s, T1u);
|
Chris@10
|
541 T1d = VADD(T17, T1a);
|
Chris@10
|
542 T1e = VFNMS(LDK(KP250000000), T1d, T1c);
|
Chris@10
|
543 T2W = VADD(T1c, T1d);
|
Chris@10
|
544 }
|
Chris@10
|
545 }
|
Chris@10
|
546 {
|
Chris@10
|
547 V TG, TV, TF, TL, TH, TK, TU, T10, TW, TZ, TX, TI;
|
Chris@10
|
548 TG = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
|
Chris@10
|
549 TV = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
550 {
|
Chris@10
|
551 V Tz, TA, TB, TC, TD, TE;
|
Chris@10
|
552 Tz = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
553 TA = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
|
Chris@10
|
554 TB = VADD(Tz, TA);
|
Chris@10
|
555 TC = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
|
Chris@10
|
556 TD = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
557 TE = VADD(TC, TD);
|
Chris@10
|
558 TF = VMUL(LDK(KP559016994), VSUB(TB, TE));
|
Chris@10
|
559 TL = VSUB(TC, TD);
|
Chris@10
|
560 TH = VADD(TB, TE);
|
Chris@10
|
561 TK = VSUB(Tz, TA);
|
Chris@10
|
562 }
|
Chris@10
|
563 {
|
Chris@10
|
564 V TO, TP, TQ, TR, TS, TT;
|
Chris@10
|
565 TO = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
|
Chris@10
|
566 TP = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
567 TQ = VADD(TO, TP);
|
Chris@10
|
568 TR = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
569 TS = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
|
Chris@10
|
570 TT = VADD(TR, TS);
|
Chris@10
|
571 TU = VMUL(LDK(KP559016994), VSUB(TQ, TT));
|
Chris@10
|
572 T10 = VSUB(TR, TS);
|
Chris@10
|
573 TW = VADD(TQ, TT);
|
Chris@10
|
574 TZ = VSUB(TO, TP);
|
Chris@10
|
575 }
|
Chris@10
|
576 T2P = VADD(TG, TH);
|
Chris@10
|
577 T2Q = VADD(TV, TW);
|
Chris@10
|
578 T2U = VADD(T2P, T2Q);
|
Chris@10
|
579 T11 = VFMA(LDK(KP475528258), TZ, VMUL(LDK(KP293892626), T10));
|
Chris@10
|
580 T27 = VFNMS(LDK(KP475528258), T10, VMUL(LDK(KP293892626), TZ));
|
Chris@10
|
581 TX = VFNMS(LDK(KP250000000), TW, TV);
|
Chris@10
|
582 TY = VADD(TU, TX);
|
Chris@10
|
583 T26 = VSUB(TX, TU);
|
Chris@10
|
584 T12 = VFNMS(LDK(KP1_369094211), T11, VMUL(LDK(KP728968627), TY));
|
Chris@10
|
585 T2f = VFMA(LDK(KP125581039), T27, VMUL(LDK(KP998026728), T26));
|
Chris@10
|
586 T1j = VFMA(LDK(KP1_457937254), T11, VMUL(LDK(KP684547105), TY));
|
Chris@10
|
587 T28 = VFNMS(LDK(KP1_996053456), T27, VMUL(LDK(KP062790519), T26));
|
Chris@10
|
588 TM = VFMA(LDK(KP475528258), TK, VMUL(LDK(KP293892626), TL));
|
Chris@10
|
589 T24 = VFNMS(LDK(KP475528258), TL, VMUL(LDK(KP293892626), TK));
|
Chris@10
|
590 TI = VFNMS(LDK(KP250000000), TH, TG);
|
Chris@10
|
591 TJ = VADD(TF, TI);
|
Chris@10
|
592 T23 = VSUB(TI, TF);
|
Chris@10
|
593 TN = VFNMS(LDK(KP963507348), TM, VMUL(LDK(KP876306680), TJ));
|
Chris@10
|
594 T2e = VFMA(LDK(KP1_071653589), T24, VMUL(LDK(KP844327925), T23));
|
Chris@10
|
595 T1i = VFMA(LDK(KP1_752613360), TM, VMUL(LDK(KP481753674), TJ));
|
Chris@10
|
596 T25 = VFNMS(LDK(KP1_688655851), T24, VMUL(LDK(KP535826794), T23));
|
Chris@10
|
597 }
|
Chris@10
|
598 {
|
Chris@10
|
599 V Tb, Tq, T3, Tc, T6, Ta, Ti, Tr, Tl, Tp, Ts, Td;
|
Chris@10
|
600 Tb = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
|
Chris@10
|
601 Tq = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
602 {
|
Chris@10
|
603 V T1, T2, T8, T4, T5, T9;
|
Chris@10
|
604 T1 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
605 T2 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
|
Chris@10
|
606 T8 = VADD(T1, T2);
|
Chris@10
|
607 T4 = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
|
Chris@10
|
608 T5 = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
609 T9 = VADD(T4, T5);
|
Chris@10
|
610 T3 = VSUB(T1, T2);
|
Chris@10
|
611 Tc = VADD(T8, T9);
|
Chris@10
|
612 T6 = VSUB(T4, T5);
|
Chris@10
|
613 Ta = VMUL(LDK(KP559016994), VSUB(T8, T9));
|
Chris@10
|
614 }
|
Chris@10
|
615 {
|
Chris@10
|
616 V Tg, Th, Tn, Tj, Tk, To;
|
Chris@10
|
617 Tg = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
|
Chris@10
|
618 Th = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
619 Tn = VADD(Tg, Th);
|
Chris@10
|
620 Tj = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
|
Chris@10
|
621 Tk = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
|
Chris@10
|
622 To = VADD(Tj, Tk);
|
Chris@10
|
623 Ti = VSUB(Tg, Th);
|
Chris@10
|
624 Tr = VADD(Tn, To);
|
Chris@10
|
625 Tl = VSUB(Tj, Tk);
|
Chris@10
|
626 Tp = VMUL(LDK(KP559016994), VSUB(Tn, To));
|
Chris@10
|
627 }
|
Chris@10
|
628 T2M = VADD(Tq, Tr);
|
Chris@10
|
629 T2N = VADD(Tb, Tc);
|
Chris@10
|
630 T2T = VADD(T2M, T2N);
|
Chris@10
|
631 Tm = VFMA(LDK(KP475528258), Ti, VMUL(LDK(KP293892626), Tl));
|
Chris@10
|
632 T1W = VFNMS(LDK(KP475528258), Tl, VMUL(LDK(KP293892626), Ti));
|
Chris@10
|
633 Ts = VFNMS(LDK(KP250000000), Tr, Tq);
|
Chris@10
|
634 Tt = VADD(Tp, Ts);
|
Chris@10
|
635 T1X = VSUB(Ts, Tp);
|
Chris@10
|
636 Tu = VFMA(LDK(KP1_937166322), Tm, VMUL(LDK(KP248689887), Tt));
|
Chris@10
|
637 T20 = VFNMS(LDK(KP963507348), T1W, VMUL(LDK(KP876306680), T1X));
|
Chris@10
|
638 Tw = VFNMS(LDK(KP497379774), Tm, VMUL(LDK(KP968583161), Tt));
|
Chris@10
|
639 T1Y = VFMA(LDK(KP1_752613360), T1W, VMUL(LDK(KP481753674), T1X));
|
Chris@10
|
640 T7 = VFMA(LDK(KP475528258), T3, VMUL(LDK(KP293892626), T6));
|
Chris@10
|
641 T1U = VFNMS(LDK(KP475528258), T6, VMUL(LDK(KP293892626), T3));
|
Chris@10
|
642 Td = VFNMS(LDK(KP250000000), Tc, Tb);
|
Chris@10
|
643 Te = VADD(Ta, Td);
|
Chris@10
|
644 T1T = VSUB(Td, Ta);
|
Chris@10
|
645 Tf = VFMA(LDK(KP1_071653589), T7, VMUL(LDK(KP844327925), Te));
|
Chris@10
|
646 T21 = VFMA(LDK(KP1_809654104), T1U, VMUL(LDK(KP425779291), T1T));
|
Chris@10
|
647 Tx = VFNMS(LDK(KP1_688655851), T7, VMUL(LDK(KP535826794), Te));
|
Chris@10
|
648 T1V = VFNMS(LDK(KP851558583), T1U, VMUL(LDK(KP904827052), T1T));
|
Chris@10
|
649 }
|
Chris@10
|
650 {
|
Chris@10
|
651 V T2V, T2X, T2Y, T2S, T30, T2O, T2R, T31, T2Z;
|
Chris@10
|
652 T2V = VMUL(LDK(KP559016994), VSUB(T2T, T2U));
|
Chris@10
|
653 T2X = VADD(T2T, T2U);
|
Chris@10
|
654 T2Y = VFNMS(LDK(KP250000000), T2X, T2W);
|
Chris@10
|
655 T2O = VSUB(T2M, T2N);
|
Chris@10
|
656 T2R = VSUB(T2P, T2Q);
|
Chris@10
|
657 T2S = VBYI(VFMA(LDK(KP951056516), T2O, VMUL(LDK(KP587785252), T2R)));
|
Chris@10
|
658 T30 = VBYI(VFNMS(LDK(KP951056516), T2R, VMUL(LDK(KP587785252), T2O)));
|
Chris@10
|
659 ST(&(xo[0]), VADD(T2W, T2X), ovs, &(xo[0]));
|
Chris@10
|
660 T31 = VSUB(T2Y, T2V);
|
Chris@10
|
661 ST(&(xo[WS(os, 10)]), VADD(T30, T31), ovs, &(xo[0]));
|
Chris@10
|
662 ST(&(xo[WS(os, 15)]), VSUB(T31, T30), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
663 T2Z = VADD(T2V, T2Y);
|
Chris@10
|
664 ST(&(xo[WS(os, 5)]), VADD(T2S, T2Z), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
665 ST(&(xo[WS(os, 20)]), VSUB(T2Z, T2S), ovs, &(xo[0]));
|
Chris@10
|
666 }
|
Chris@10
|
667 {
|
Chris@10
|
668 V T1Z, T2i, T2j, T2g, T2w, T2x, T2y, T2G, T2H, T2I, T2D, T2E, T2F, T2z, T2A;
|
Chris@10
|
669 V T2B, T2p, T2m, T2q, T2b, T2c, T2a, T2d, T2h, T2r;
|
Chris@10
|
670 T1Z = VSUB(T1V, T1Y);
|
Chris@10
|
671 T2i = VADD(T20, T21);
|
Chris@10
|
672 T2j = VSUB(T25, T28);
|
Chris@10
|
673 T2g = VSUB(T2e, T2f);
|
Chris@10
|
674 T2w = VFMA(LDK(KP1_369094211), T1W, VMUL(LDK(KP728968627), T1X));
|
Chris@10
|
675 T2x = VFNMS(LDK(KP992114701), T1T, VMUL(LDK(KP250666467), T1U));
|
Chris@10
|
676 T2y = VADD(T2w, T2x);
|
Chris@10
|
677 T2G = VFNMS(LDK(KP125581039), T24, VMUL(LDK(KP998026728), T23));
|
Chris@10
|
678 T2H = VFMA(LDK(KP1_274847979), T27, VMUL(LDK(KP770513242), T26));
|
Chris@10
|
679 T2I = VADD(T2G, T2H);
|
Chris@10
|
680 T2D = VFNMS(LDK(KP1_457937254), T1W, VMUL(LDK(KP684547105), T1X));
|
Chris@10
|
681 T2E = VFMA(LDK(KP1_984229402), T1U, VMUL(LDK(KP125333233), T1T));
|
Chris@10
|
682 T2F = VADD(T2D, T2E);
|
Chris@10
|
683 T2z = VFMA(LDK(KP1_996053456), T24, VMUL(LDK(KP062790519), T23));
|
Chris@10
|
684 T2A = VFNMS(LDK(KP637423989), T26, VMUL(LDK(KP1_541026485), T27));
|
Chris@10
|
685 T2B = VADD(T2z, T2A);
|
Chris@10
|
686 {
|
Chris@10
|
687 V T2k, T2l, T22, T29;
|
Chris@10
|
688 T2k = VADD(T1Y, T1V);
|
Chris@10
|
689 T2l = VADD(T2e, T2f);
|
Chris@10
|
690 T2p = VADD(T2k, T2l);
|
Chris@10
|
691 T2m = VMUL(LDK(KP559016994), VSUB(T2k, T2l));
|
Chris@10
|
692 T2q = VFNMS(LDK(KP250000000), T2p, T2o);
|
Chris@10
|
693 T2b = VSUB(T1e, T1b);
|
Chris@10
|
694 T22 = VSUB(T20, T21);
|
Chris@10
|
695 T29 = VADD(T25, T28);
|
Chris@10
|
696 T2c = VADD(T22, T29);
|
Chris@10
|
697 T2a = VMUL(LDK(KP559016994), VSUB(T22, T29));
|
Chris@10
|
698 T2d = VFNMS(LDK(KP250000000), T2c, T2b);
|
Chris@10
|
699 }
|
Chris@10
|
700 {
|
Chris@10
|
701 V T2u, T2v, T2C, T2J;
|
Chris@10
|
702 T2u = VADD(T2b, T2c);
|
Chris@10
|
703 T2v = VBYI(VADD(T2o, T2p));
|
Chris@10
|
704 ST(&(xo[WS(os, 23)]), VSUB(T2u, T2v), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
705 ST(&(xo[WS(os, 2)]), VADD(T2u, T2v), ovs, &(xo[0]));
|
Chris@10
|
706 T2C = VADD(T2b, VADD(T2y, T2B));
|
Chris@10
|
707 T2J = VBYI(VSUB(VADD(T2F, T2I), T2o));
|
Chris@10
|
708 ST(&(xo[WS(os, 22)]), VSUB(T2C, T2J), ovs, &(xo[0]));
|
Chris@10
|
709 ST(&(xo[WS(os, 3)]), VADD(T2C, T2J), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
710 }
|
Chris@10
|
711 T2h = VFMA(LDK(KP951056516), T1Z, VADD(T2a, VFNMS(LDK(KP587785252), T2g, T2d)));
|
Chris@10
|
712 T2r = VBYI(VADD(VFMA(LDK(KP951056516), T2i, VMUL(LDK(KP587785252), T2j)), VADD(T2m, T2q)));
|
Chris@10
|
713 ST(&(xo[WS(os, 18)]), VSUB(T2h, T2r), ovs, &(xo[0]));
|
Chris@10
|
714 ST(&(xo[WS(os, 7)]), VADD(T2h, T2r), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
715 {
|
Chris@10
|
716 V T2s, T2t, T2K, T2L;
|
Chris@10
|
717 T2s = VFMA(LDK(KP587785252), T1Z, VFMA(LDK(KP951056516), T2g, VSUB(T2d, T2a)));
|
Chris@10
|
718 T2t = VBYI(VADD(VFNMS(LDK(KP951056516), T2j, VMUL(LDK(KP587785252), T2i)), VSUB(T2q, T2m)));
|
Chris@10
|
719 ST(&(xo[WS(os, 13)]), VSUB(T2s, T2t), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
720 ST(&(xo[WS(os, 12)]), VADD(T2s, T2t), ovs, &(xo[0]));
|
Chris@10
|
721 T2K = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T2w, T2x), VFMA(LDK(KP309016994), T2F, VFNMS(LDK(KP809016994), T2I, VMUL(LDK(KP587785252), VSUB(T2z, T2A))))), T2o));
|
Chris@10
|
722 T2L = VFMA(LDK(KP309016994), T2y, VFMA(LDK(KP951056516), VSUB(T2E, T2D), VFMA(LDK(KP587785252), VSUB(T2H, T2G), VFNMS(LDK(KP809016994), T2B, T2b))));
|
Chris@10
|
723 ST(&(xo[WS(os, 8)]), VADD(T2K, T2L), ovs, &(xo[0]));
|
Chris@10
|
724 ST(&(xo[WS(os, 17)]), VSUB(T2L, T2K), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
725 }
|
Chris@10
|
726 }
|
Chris@10
|
727 {
|
Chris@10
|
728 V Tv, T1m, T1n, T1k, T1D, T1E, T1F, T1N, T1O, T1P, T1K, T1L, T1M, T1G, T1H;
|
Chris@10
|
729 V T1I, T1w, T1q, T1x, T1f, T1g, T14, T1h, T1l, T1y;
|
Chris@10
|
730 Tv = VSUB(Tf, Tu);
|
Chris@10
|
731 T1m = VSUB(Tw, Tx);
|
Chris@10
|
732 T1n = VSUB(TN, T12);
|
Chris@10
|
733 T1k = VSUB(T1i, T1j);
|
Chris@10
|
734 T1D = VFMA(LDK(KP1_688655851), Tm, VMUL(LDK(KP535826794), Tt));
|
Chris@10
|
735 T1E = VFMA(LDK(KP1_541026485), T7, VMUL(LDK(KP637423989), Te));
|
Chris@10
|
736 T1F = VSUB(T1D, T1E);
|
Chris@10
|
737 T1N = VFMA(LDK(KP851558583), TM, VMUL(LDK(KP904827052), TJ));
|
Chris@10
|
738 T1O = VFMA(LDK(KP1_984229402), T11, VMUL(LDK(KP125333233), TY));
|
Chris@10
|
739 T1P = VADD(T1N, T1O);
|
Chris@10
|
740 T1K = VFNMS(LDK(KP1_071653589), Tm, VMUL(LDK(KP844327925), Tt));
|
Chris@10
|
741 T1L = VFNMS(LDK(KP770513242), Te, VMUL(LDK(KP1_274847979), T7));
|
Chris@10
|
742 T1M = VADD(T1K, T1L);
|
Chris@10
|
743 T1G = VFNMS(LDK(KP425779291), TJ, VMUL(LDK(KP1_809654104), TM));
|
Chris@10
|
744 T1H = VFNMS(LDK(KP992114701), TY, VMUL(LDK(KP250666467), T11));
|
Chris@10
|
745 T1I = VADD(T1G, T1H);
|
Chris@10
|
746 {
|
Chris@10
|
747 V T1o, T1p, Ty, T13;
|
Chris@10
|
748 T1o = VADD(Tu, Tf);
|
Chris@10
|
749 T1p = VADD(T1i, T1j);
|
Chris@10
|
750 T1w = VADD(T1o, T1p);
|
Chris@10
|
751 T1q = VMUL(LDK(KP559016994), VSUB(T1o, T1p));
|
Chris@10
|
752 T1x = VFNMS(LDK(KP250000000), T1w, T1v);
|
Chris@10
|
753 T1f = VADD(T1b, T1e);
|
Chris@10
|
754 Ty = VADD(Tw, Tx);
|
Chris@10
|
755 T13 = VADD(TN, T12);
|
Chris@10
|
756 T1g = VADD(Ty, T13);
|
Chris@10
|
757 T14 = VMUL(LDK(KP559016994), VSUB(Ty, T13));
|
Chris@10
|
758 T1h = VFNMS(LDK(KP250000000), T1g, T1f);
|
Chris@10
|
759 }
|
Chris@10
|
760 {
|
Chris@10
|
761 V T1B, T1C, T1J, T1Q;
|
Chris@10
|
762 T1B = VADD(T1f, T1g);
|
Chris@10
|
763 T1C = VBYI(VADD(T1v, T1w));
|
Chris@10
|
764 ST(&(xo[WS(os, 24)]), VSUB(T1B, T1C), ovs, &(xo[0]));
|
Chris@10
|
765 ST(&(xo[WS(os, 1)]), VADD(T1B, T1C), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
766 T1J = VADD(T1f, VADD(T1F, T1I));
|
Chris@10
|
767 T1Q = VBYI(VSUB(VADD(T1M, T1P), T1v));
|
Chris@10
|
768 ST(&(xo[WS(os, 21)]), VSUB(T1J, T1Q), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
769 ST(&(xo[WS(os, 4)]), VADD(T1J, T1Q), ovs, &(xo[0]));
|
Chris@10
|
770 }
|
Chris@10
|
771 T1l = VFMA(LDK(KP951056516), Tv, VADD(T14, VFNMS(LDK(KP587785252), T1k, T1h)));
|
Chris@10
|
772 T1y = VBYI(VADD(VFMA(LDK(KP951056516), T1m, VMUL(LDK(KP587785252), T1n)), VADD(T1q, T1x)));
|
Chris@10
|
773 ST(&(xo[WS(os, 19)]), VSUB(T1l, T1y), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
774 ST(&(xo[WS(os, 6)]), VADD(T1l, T1y), ovs, &(xo[0]));
|
Chris@10
|
775 {
|
Chris@10
|
776 V T1z, T1A, T1R, T1S;
|
Chris@10
|
777 T1z = VFMA(LDK(KP587785252), Tv, VFMA(LDK(KP951056516), T1k, VSUB(T1h, T14)));
|
Chris@10
|
778 T1A = VBYI(VADD(VFNMS(LDK(KP951056516), T1n, VMUL(LDK(KP587785252), T1m)), VSUB(T1x, T1q)));
|
Chris@10
|
779 ST(&(xo[WS(os, 14)]), VSUB(T1z, T1A), ovs, &(xo[0]));
|
Chris@10
|
780 ST(&(xo[WS(os, 11)]), VADD(T1z, T1A), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
781 T1R = VBYI(VSUB(VFMA(LDK(KP309016994), T1M, VFMA(LDK(KP951056516), VADD(T1D, T1E), VFNMS(LDK(KP809016994), T1P, VMUL(LDK(KP587785252), VSUB(T1G, T1H))))), T1v));
|
Chris@10
|
782 T1S = VFMA(LDK(KP951056516), VSUB(T1L, T1K), VFMA(LDK(KP309016994), T1F, VFMA(LDK(KP587785252), VSUB(T1O, T1N), VFNMS(LDK(KP809016994), T1I, T1f))));
|
Chris@10
|
783 ST(&(xo[WS(os, 9)]), VADD(T1R, T1S), ovs, &(xo[WS(os, 1)]));
|
Chris@10
|
784 ST(&(xo[WS(os, 16)]), VSUB(T1S, T1R), ovs, &(xo[0]));
|
Chris@10
|
785 }
|
Chris@10
|
786 }
|
Chris@10
|
787 }
|
Chris@10
|
788 }
|
Chris@10
|
789 VLEAVE();
|
Chris@10
|
790 }
|
Chris@10
|
791
|
Chris@10
|
792 static const kdft_desc desc = { 25, XSIMD_STRING("n1bv_25"), {147, 63, 77, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
793
|
Chris@10
|
794 void XSIMD(codelet_n1bv_25) (planner *p) {
|
Chris@10
|
795 X(kdft_register) (p, n1bv_25, &desc);
|
Chris@10
|
796 }
|
Chris@10
|
797
|
Chris@10
|
798 #endif /* HAVE_FMA */
|