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 07:39:23 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_twiddle_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 25 -name t3bv_25 -include t3b.h -sign 1 */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 268 FP additions, 281 FP multiplications,
|
Chris@10
|
32 * (or, 87 additions, 100 multiplications, 181 fused multiply/add),
|
Chris@10
|
33 * 223 stack variables, 67 constants, and 50 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "t3b.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void t3bv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DVK(KP792626838, +0.792626838241819413632131824093538848057784557);
|
Chris@10
|
40 DVK(KP876091699, +0.876091699473550838204498029706869638173524346);
|
Chris@10
|
41 DVK(KP617882369, +0.617882369114440893914546919006756321695042882);
|
Chris@10
|
42 DVK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
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(KP916574801, +0.916574801383451584742370439148878693530976769);
|
Chris@10
|
47 DVK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
48 DVK(KP809385824, +0.809385824416008241660603814668679683846476688);
|
Chris@10
|
49 DVK(KP447417479, +0.447417479732227551498980015410057305749330693);
|
Chris@10
|
50 DVK(KP894834959, +0.894834959464455102997960030820114611498661386);
|
Chris@10
|
51 DVK(KP867381224, +0.867381224396525206773171885031575671309956167);
|
Chris@10
|
52 DVK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
53 DVK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
54 DVK(KP958953096, +0.958953096729998668045963838399037225970891871);
|
Chris@10
|
55 DVK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
56 DVK(KP829049696, +0.829049696159252993975487806364305442437946767);
|
Chris@10
|
57 DVK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
Chris@10
|
58 DVK(KP897376177, +0.897376177523557693138608077137219684419427330);
|
Chris@10
|
59 DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
60 DVK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
Chris@10
|
61 DVK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
Chris@10
|
62 DVK(KP855719849, +0.855719849902058969314654733608091555096772472);
|
Chris@10
|
63 DVK(KP237294955, +0.237294955877110315393888866460840817927895961);
|
Chris@10
|
64 DVK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
Chris@10
|
65 DVK(KP904508497, +0.904508497187473712051146708591409529430077295);
|
Chris@10
|
66 DVK(KP997675361, +0.997675361079556513670859573984492383596555031);
|
Chris@10
|
67 DVK(KP763932022, +0.763932022500210303590826331268723764559381640);
|
Chris@10
|
68 DVK(KP690983005, +0.690983005625052575897706582817180941139845410);
|
Chris@10
|
69 DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
70 DVK(KP952936919, +0.952936919628306576880750665357914584765951388);
|
Chris@10
|
71 DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
72 DVK(KP262346850, +0.262346850930607871785420028382979691334784273);
|
Chris@10
|
73 DVK(KP570584518, +0.570584518783621657366766175430996792655723863);
|
Chris@10
|
74 DVK(KP669429328, +0.669429328479476605641803240971985825917022098);
|
Chris@10
|
75 DVK(KP923225144, +0.923225144846402650453449441572664695995209956);
|
Chris@10
|
76 DVK(KP945422727, +0.945422727388575946270360266328811958657216298);
|
Chris@10
|
77 DVK(KP522616830, +0.522616830205754336872861364785224694908468440);
|
Chris@10
|
78 DVK(KP956723877, +0.956723877038460305821989399535483155872969262);
|
Chris@10
|
79 DVK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
Chris@10
|
80 DVK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
Chris@10
|
81 DVK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
Chris@10
|
82 DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
83 DVK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
84 DVK(KP921078979, +0.921078979742360627699756128143719920817673854);
|
Chris@10
|
85 DVK(KP982009705, +0.982009705009746369461829878184175962711969869);
|
Chris@10
|
86 DVK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
Chris@10
|
87 DVK(KP494780565, +0.494780565770515410344588413655324772219443730);
|
Chris@10
|
88 DVK(KP447533225, +0.447533225982656890041886979663652563063114397);
|
Chris@10
|
89 DVK(KP603558818, +0.603558818296015001454675132653458027918768137);
|
Chris@10
|
90 DVK(KP667278218, +0.667278218140296670899089292254759909713898805);
|
Chris@10
|
91 DVK(KP244189809, +0.244189809627953270309879511234821255780225091);
|
Chris@10
|
92 DVK(KP269969613, +0.269969613759572083574752974412347470060951301);
|
Chris@10
|
93 DVK(KP578046249, +0.578046249379945007321754579646815604023525655);
|
Chris@10
|
94 DVK(KP522847744, +0.522847744331509716623755382187077770911012542);
|
Chris@10
|
95 DVK(KP132830569, +0.132830569247582714407653942074819768844536507);
|
Chris@10
|
96 DVK(KP120146378, +0.120146378570687701782758537356596213647956445);
|
Chris@10
|
97 DVK(KP893101515, +0.893101515366181661711202267938416198338079437);
|
Chris@10
|
98 DVK(KP987388751, +0.987388751065621252324603216482382109400433949);
|
Chris@10
|
99 DVK(KP059835404, +0.059835404262124915169548397419498386427871950);
|
Chris@10
|
100 DVK(KP066152395, +0.066152395967733048213034281011006031460903353);
|
Chris@10
|
101 DVK(KP786782374, +0.786782374965295178365099601674911834788448471);
|
Chris@10
|
102 DVK(KP869845200, +0.869845200362138853122720822420327157933056305);
|
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 m;
|
Chris@10
|
108 R *x;
|
Chris@10
|
109 x = ii;
|
Chris@10
|
110 for (m = mb, W = W + (mb * ((TWVL / VL) * 8)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(25, rs)) {
|
Chris@10
|
111 V T2t, T1Z, T2W, T28, T2Q, T2r, T2g, T2u, T2o, T2l;
|
Chris@10
|
112 {
|
Chris@10
|
113 V T2, T5, T3, T9;
|
Chris@10
|
114 T2 = LDW(&(W[0]));
|
Chris@10
|
115 T5 = LDW(&(W[TWVL * 4]));
|
Chris@10
|
116 T3 = LDW(&(W[TWVL * 2]));
|
Chris@10
|
117 T9 = LDW(&(W[TWVL * 6]));
|
Chris@10
|
118 {
|
Chris@10
|
119 V T2c, T3l, Tn, T49, Tm, T4e, TN, T32, T1d, T3a, T3f, T3z, T3H, T25, T1W;
|
Chris@10
|
120 V T2v, T2D, T4a, T1g, T18, T2Z, T11, T31, TK, T1q, T1j, T1n, T4b, T17;
|
Chris@10
|
121 {
|
Chris@10
|
122 V T1, T1l, Tr, T4, Ty, T1E, Tu, TX, TD, T1h, Tz, T1e, T1I, T1o, TU;
|
Chris@10
|
123 V Tk, T2b, T1B, T1D, T1N, T1F, Td, T2a, T1J;
|
Chris@10
|
124 {
|
Chris@10
|
125 V T7, Tb, TC, Tg, T1L, Ta, T6, Tj, T1A;
|
Chris@10
|
126 T1 = LD(&(x[0]), ms, &(x[0]));
|
Chris@10
|
127 {
|
Chris@10
|
128 V Tf, Ti, Te, Th;
|
Chris@10
|
129 Tf = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
|
Chris@10
|
130 Ti = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
131 T7 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
132 Tb = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
|
Chris@10
|
133 Te = VZMUL(T2, T5);
|
Chris@10
|
134 TC = VZMULJ(T2, T5);
|
Chris@10
|
135 T1l = VZMUL(T3, T5);
|
Chris@10
|
136 Tr = VZMULJ(T3, T5);
|
Chris@10
|
137 T4 = VZMUL(T2, T3);
|
Chris@10
|
138 Ty = VZMULJ(T2, T3);
|
Chris@10
|
139 T1E = VZMULJ(T2, T9);
|
Chris@10
|
140 Th = VZMULJ(T5, T9);
|
Chris@10
|
141 Tu = VZMULJ(T3, T9);
|
Chris@10
|
142 Tg = VZMUL(Te, Tf);
|
Chris@10
|
143 TX = VZMULJ(Te, T9);
|
Chris@10
|
144 TD = VZMULJ(TC, T9);
|
Chris@10
|
145 T1h = VZMULJ(Ty, T9);
|
Chris@10
|
146 Tz = VZMUL(Ty, T5);
|
Chris@10
|
147 T1e = VZMULJ(Ty, T5);
|
Chris@10
|
148 T1L = VZMULJ(Tr, T9);
|
Chris@10
|
149 Ta = VZMULJ(T4, T9);
|
Chris@10
|
150 T1I = VZMUL(T4, T5);
|
Chris@10
|
151 T6 = VZMULJ(T4, T5);
|
Chris@10
|
152 Tj = VZMUL(Th, Ti);
|
Chris@10
|
153 }
|
Chris@10
|
154 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
155 T1o = VZMULJ(T1e, T9);
|
Chris@10
|
156 {
|
Chris@10
|
157 V Tc, T8, T1C, T1M;
|
Chris@10
|
158 T1C = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
|
Chris@10
|
159 T1M = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
|
Chris@10
|
160 Tc = VZMUL(Ta, Tb);
|
Chris@10
|
161 T8 = VZMUL(T6, T7);
|
Chris@10
|
162 TU = VZMULJ(T6, T9);
|
Chris@10
|
163 Tk = VADD(Tg, Tj);
|
Chris@10
|
164 T2b = VSUB(Tg, Tj);
|
Chris@10
|
165 T1B = VZMUL(T3, T1A);
|
Chris@10
|
166 T1D = VZMUL(TC, T1C);
|
Chris@10
|
167 T1N = VZMUL(T1L, T1M);
|
Chris@10
|
168 T1F = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
169 Td = VADD(T8, Tc);
|
Chris@10
|
170 T2a = VSUB(T8, Tc);
|
Chris@10
|
171 T1J = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
172 }
|
Chris@10
|
173 }
|
Chris@10
|
174 {
|
Chris@10
|
175 V Tq, Tt, TF, T1T, T1H, Tw, T1U, T1O, TA, Tp, Ts, TE;
|
Chris@10
|
176 Tp = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
177 Ts = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
|
Chris@10
|
178 TE = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
|
Chris@10
|
179 {
|
Chris@10
|
180 V T1K, Tv, T1G, Tl;
|
Chris@10
|
181 Tv = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
182 T1G = VZMUL(T1E, T1F);
|
Chris@10
|
183 T2c = VFMA(LDK(KP618033988), T2b, T2a);
|
Chris@10
|
184 T3l = VFNMS(LDK(KP618033988), T2a, T2b);
|
Chris@10
|
185 Tn = VSUB(Td, Tk);
|
Chris@10
|
186 Tl = VADD(Td, Tk);
|
Chris@10
|
187 T1K = VZMUL(T1I, T1J);
|
Chris@10
|
188 Tq = VZMUL(T2, Tp);
|
Chris@10
|
189 Tt = VZMUL(Tr, Ts);
|
Chris@10
|
190 TF = VZMUL(TD, TE);
|
Chris@10
|
191 T1T = VSUB(T1D, T1G);
|
Chris@10
|
192 T1H = VADD(T1D, T1G);
|
Chris@10
|
193 T49 = VADD(T1, Tl);
|
Chris@10
|
194 Tm = VFNMS(LDK(KP250000000), Tl, T1);
|
Chris@10
|
195 Tw = VZMUL(Tu, Tv);
|
Chris@10
|
196 T1U = VSUB(T1K, T1N);
|
Chris@10
|
197 T1O = VADD(T1K, T1N);
|
Chris@10
|
198 TA = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
199 }
|
Chris@10
|
200 {
|
Chris@10
|
201 V Tx, TL, T1R, T38, T1V, T13, TQ, TZ, TS, T1Q, TV, TG, TM, T12, T1c;
|
Chris@10
|
202 V T16;
|
Chris@10
|
203 T12 = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
|
Chris@10
|
204 {
|
Chris@10
|
205 V TP, TY, T1P, TB, TR;
|
Chris@10
|
206 TP = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
|
Chris@10
|
207 TY = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
|
Chris@10
|
208 TR = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
209 Tx = VADD(Tt, Tw);
|
Chris@10
|
210 TL = VSUB(Tt, Tw);
|
Chris@10
|
211 T1R = VSUB(T1O, T1H);
|
Chris@10
|
212 T1P = VADD(T1H, T1O);
|
Chris@10
|
213 T38 = VFNMS(LDK(KP618033988), T1T, T1U);
|
Chris@10
|
214 T1V = VFMA(LDK(KP618033988), T1U, T1T);
|
Chris@10
|
215 TB = VZMUL(Tz, TA);
|
Chris@10
|
216 T13 = VZMUL(T4, T12);
|
Chris@10
|
217 TQ = VZMUL(T9, TP);
|
Chris@10
|
218 TZ = VZMUL(TX, TY);
|
Chris@10
|
219 TS = VZMUL(T5, TR);
|
Chris@10
|
220 T4e = VADD(T1B, T1P);
|
Chris@10
|
221 T1Q = VFNMS(LDK(KP250000000), T1P, T1B);
|
Chris@10
|
222 TV = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
223 TG = VADD(TB, TF);
|
Chris@10
|
224 TM = VSUB(TF, TB);
|
Chris@10
|
225 }
|
Chris@10
|
226 T1c = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
|
Chris@10
|
227 {
|
Chris@10
|
228 V T14, TT, TJ, T15, T10, TI, T1p, T1f, T1i, T1m;
|
Chris@10
|
229 T1f = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
230 T14 = VADD(TS, TQ);
|
Chris@10
|
231 TT = VSUB(TQ, TS);
|
Chris@10
|
232 {
|
Chris@10
|
233 V T39, T1S, TW, TH;
|
Chris@10
|
234 T39 = VFMA(LDK(KP559016994), T1R, T1Q);
|
Chris@10
|
235 T1S = VFNMS(LDK(KP559016994), T1R, T1Q);
|
Chris@10
|
236 TW = VZMUL(TU, TV);
|
Chris@10
|
237 TH = VADD(Tx, TG);
|
Chris@10
|
238 TJ = VSUB(Tx, TG);
|
Chris@10
|
239 TN = VFNMS(LDK(KP618033988), TM, TL);
|
Chris@10
|
240 T32 = VFMA(LDK(KP618033988), TL, TM);
|
Chris@10
|
241 T1d = VZMUL(Ty, T1c);
|
Chris@10
|
242 T3a = VFMA(LDK(KP869845200), T39, T38);
|
Chris@10
|
243 T3f = VFNMS(LDK(KP786782374), T38, T39);
|
Chris@10
|
244 T3z = VFMA(LDK(KP066152395), T39, T38);
|
Chris@10
|
245 T3H = VFNMS(LDK(KP059835404), T38, T39);
|
Chris@10
|
246 T25 = VFMA(LDK(KP987388751), T1S, T1V);
|
Chris@10
|
247 T1W = VFNMS(LDK(KP893101515), T1V, T1S);
|
Chris@10
|
248 T2v = VFNMS(LDK(KP120146378), T1V, T1S);
|
Chris@10
|
249 T2D = VFMA(LDK(KP132830569), T1S, T1V);
|
Chris@10
|
250 T15 = VADD(TZ, TW);
|
Chris@10
|
251 T10 = VSUB(TW, TZ);
|
Chris@10
|
252 TI = VFNMS(LDK(KP250000000), TH, Tq);
|
Chris@10
|
253 T4a = VADD(Tq, TH);
|
Chris@10
|
254 T1g = VZMUL(T1e, T1f);
|
Chris@10
|
255 }
|
Chris@10
|
256 T1p = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
257 T1i = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
|
Chris@10
|
258 T1m = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
|
Chris@10
|
259 T18 = VSUB(T14, T15);
|
Chris@10
|
260 T16 = VADD(T14, T15);
|
Chris@10
|
261 T2Z = VFNMS(LDK(KP618033988), TT, T10);
|
Chris@10
|
262 T11 = VFMA(LDK(KP618033988), T10, TT);
|
Chris@10
|
263 T31 = VFNMS(LDK(KP559016994), TJ, TI);
|
Chris@10
|
264 TK = VFMA(LDK(KP559016994), TJ, TI);
|
Chris@10
|
265 T1q = VZMUL(T1o, T1p);
|
Chris@10
|
266 T1j = VZMUL(T1h, T1i);
|
Chris@10
|
267 T1n = VZMUL(T1l, T1m);
|
Chris@10
|
268 }
|
Chris@10
|
269 T4b = VADD(T13, T16);
|
Chris@10
|
270 T17 = VFMS(LDK(KP250000000), T16, T13);
|
Chris@10
|
271 }
|
Chris@10
|
272 }
|
Chris@10
|
273 }
|
Chris@10
|
274 {
|
Chris@10
|
275 V T33, T3i, T3C, T3L, T20, TO, T2y, T2G, T1k, T1w, T1r, T1x, T2Y, T19, T4k;
|
Chris@10
|
276 V T4c;
|
Chris@10
|
277 T33 = VFMA(LDK(KP893101515), T32, T31);
|
Chris@10
|
278 T3i = VFNMS(LDK(KP987388751), T31, T32);
|
Chris@10
|
279 T3C = VFNMS(LDK(KP522847744), T32, T31);
|
Chris@10
|
280 T3L = VFMA(LDK(KP578046249), T31, T32);
|
Chris@10
|
281 T20 = VFMA(LDK(KP269969613), TK, TN);
|
Chris@10
|
282 TO = VFNMS(LDK(KP244189809), TN, TK);
|
Chris@10
|
283 T2y = VFMA(LDK(KP667278218), TK, TN);
|
Chris@10
|
284 T2G = VFNMS(LDK(KP603558818), TN, TK);
|
Chris@10
|
285 T1k = VADD(T1g, T1j);
|
Chris@10
|
286 T1w = VSUB(T1g, T1j);
|
Chris@10
|
287 T1r = VADD(T1n, T1q);
|
Chris@10
|
288 T1x = VSUB(T1q, T1n);
|
Chris@10
|
289 T2Y = VFMA(LDK(KP559016994), T18, T17);
|
Chris@10
|
290 T19 = VFNMS(LDK(KP559016994), T18, T17);
|
Chris@10
|
291 T4k = VSUB(T4a, T4b);
|
Chris@10
|
292 T4c = VADD(T4a, T4b);
|
Chris@10
|
293 {
|
Chris@10
|
294 V T2X, To, T35, T1y, T2H, T2z, T1a, T21, T3t, T34, T3n, T3j, T3E, T3Y, T3M;
|
Chris@10
|
295 V T3R, T1v, T36, T4l, T4f, T1u, T1s;
|
Chris@10
|
296 T2X = VFNMS(LDK(KP559016994), Tn, Tm);
|
Chris@10
|
297 To = VFMA(LDK(KP559016994), Tn, Tm);
|
Chris@10
|
298 T1u = VSUB(T1r, T1k);
|
Chris@10
|
299 T1s = VADD(T1k, T1r);
|
Chris@10
|
300 T35 = VFMA(LDK(KP618033988), T1w, T1x);
|
Chris@10
|
301 T1y = VFNMS(LDK(KP618033988), T1x, T1w);
|
Chris@10
|
302 {
|
Chris@10
|
303 V T3K, T30, T3h, T3D, T4d, T1t;
|
Chris@10
|
304 T3K = VFMA(LDK(KP447533225), T2Z, T2Y);
|
Chris@10
|
305 T30 = VFMA(LDK(KP120146378), T2Z, T2Y);
|
Chris@10
|
306 T3h = VFNMS(LDK(KP132830569), T2Y, T2Z);
|
Chris@10
|
307 T3D = VFNMS(LDK(KP494780565), T2Y, T2Z);
|
Chris@10
|
308 T2H = VFNMS(LDK(KP786782374), T11, T19);
|
Chris@10
|
309 T2z = VFMA(LDK(KP869845200), T19, T11);
|
Chris@10
|
310 T1a = VFNMS(LDK(KP667278218), T19, T11);
|
Chris@10
|
311 T21 = VFMA(LDK(KP603558818), T11, T19);
|
Chris@10
|
312 T4d = VADD(T1d, T1s);
|
Chris@10
|
313 T1t = VFNMS(LDK(KP250000000), T1s, T1d);
|
Chris@10
|
314 T3t = VFNMS(LDK(KP734762448), T33, T30);
|
Chris@10
|
315 T34 = VFMA(LDK(KP734762448), T33, T30);
|
Chris@10
|
316 T3n = VFMA(LDK(KP734762448), T3i, T3h);
|
Chris@10
|
317 T3j = VFNMS(LDK(KP734762448), T3i, T3h);
|
Chris@10
|
318 T3E = VFNMS(LDK(KP982009705), T3D, T3C);
|
Chris@10
|
319 T3Y = VFMA(LDK(KP982009705), T3D, T3C);
|
Chris@10
|
320 T3M = VFNMS(LDK(KP921078979), T3L, T3K);
|
Chris@10
|
321 T3R = VFMA(LDK(KP921078979), T3L, T3K);
|
Chris@10
|
322 T1v = VFNMS(LDK(KP559016994), T1u, T1t);
|
Chris@10
|
323 T36 = VFMA(LDK(KP559016994), T1u, T1t);
|
Chris@10
|
324 T4l = VSUB(T4d, T4e);
|
Chris@10
|
325 T4f = VADD(T4d, T4e);
|
Chris@10
|
326 }
|
Chris@10
|
327 {
|
Chris@10
|
328 V T2L, T2R, T2j, T2q, T2J, T2B, T2e, T26, T2U, T1Y, T23, T2O;
|
Chris@10
|
329 {
|
Chris@10
|
330 V T2I, T24, T2w, T2E, T48, T42, T3y, T3s, T3V, T45, T2A, T1b, T2h, T2i, T1X;
|
Chris@10
|
331 T2L = VFNMS(LDK(KP912575812), T2H, T2G);
|
Chris@10
|
332 T2I = VFMA(LDK(KP912575812), T2H, T2G);
|
Chris@10
|
333 {
|
Chris@10
|
334 V T3A, T3e, T37, T3I, T1z;
|
Chris@10
|
335 T3A = VFNMS(LDK(KP667278218), T36, T35);
|
Chris@10
|
336 T3e = VFNMS(LDK(KP059835404), T35, T36);
|
Chris@10
|
337 T37 = VFMA(LDK(KP066152395), T36, T35);
|
Chris@10
|
338 T3I = VFMA(LDK(KP603558818), T35, T36);
|
Chris@10
|
339 T24 = VFMA(LDK(KP578046249), T1v, T1y);
|
Chris@10
|
340 T1z = VFNMS(LDK(KP522847744), T1y, T1v);
|
Chris@10
|
341 T2w = VFNMS(LDK(KP494780565), T1v, T1y);
|
Chris@10
|
342 T2E = VFMA(LDK(KP447533225), T1y, T1v);
|
Chris@10
|
343 {
|
Chris@10
|
344 V T4i, T4g, T4o, T4m;
|
Chris@10
|
345 T4i = VSUB(T4c, T4f);
|
Chris@10
|
346 T4g = VADD(T4c, T4f);
|
Chris@10
|
347 T4o = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T4k, T4l));
|
Chris@10
|
348 T4m = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T4l, T4k));
|
Chris@10
|
349 {
|
Chris@10
|
350 V T3Q, T3J, T3b, T3u;
|
Chris@10
|
351 T3Q = VFNMS(LDK(KP845997307), T3I, T3H);
|
Chris@10
|
352 T3J = VFMA(LDK(KP845997307), T3I, T3H);
|
Chris@10
|
353 T3b = VFNMS(LDK(KP772036680), T3a, T37);
|
Chris@10
|
354 T3u = VFMA(LDK(KP772036680), T3a, T37);
|
Chris@10
|
355 {
|
Chris@10
|
356 V T3o, T3g, T3B, T3X, T4h;
|
Chris@10
|
357 T3o = VFNMS(LDK(KP772036680), T3f, T3e);
|
Chris@10
|
358 T3g = VFMA(LDK(KP772036680), T3f, T3e);
|
Chris@10
|
359 T3B = VFNMS(LDK(KP845997307), T3A, T3z);
|
Chris@10
|
360 T3X = VFMA(LDK(KP845997307), T3A, T3z);
|
Chris@10
|
361 ST(&(x[0]), VADD(T4g, T49), ms, &(x[0]));
|
Chris@10
|
362 T4h = VFNMS(LDK(KP250000000), T4g, T49);
|
Chris@10
|
363 {
|
Chris@10
|
364 V T40, T3N, T3c, T3v;
|
Chris@10
|
365 T40 = VFMA(LDK(KP906616052), T3M, T3J);
|
Chris@10
|
366 T3N = VFNMS(LDK(KP906616052), T3M, T3J);
|
Chris@10
|
367 T3c = VFMA(LDK(KP956723877), T3b, T34);
|
Chris@10
|
368 T3v = VFMA(LDK(KP522616830), T3j, T3u);
|
Chris@10
|
369 {
|
Chris@10
|
370 V T3p, T3k, T3S, T3F;
|
Chris@10
|
371 T3p = VFNMS(LDK(KP522616830), T34, T3o);
|
Chris@10
|
372 T3k = VFMA(LDK(KP945422727), T3j, T3g);
|
Chris@10
|
373 T3S = VFNMS(LDK(KP923225144), T3E, T3B);
|
Chris@10
|
374 T3F = VFMA(LDK(KP923225144), T3E, T3B);
|
Chris@10
|
375 {
|
Chris@10
|
376 V T46, T3Z, T4j, T4n;
|
Chris@10
|
377 T46 = VFNMS(LDK(KP669429328), T3X, T3Y);
|
Chris@10
|
378 T3Z = VFMA(LDK(KP570584518), T3Y, T3X);
|
Chris@10
|
379 T4j = VFMA(LDK(KP559016994), T4i, T4h);
|
Chris@10
|
380 T4n = VFNMS(LDK(KP559016994), T4i, T4h);
|
Chris@10
|
381 {
|
Chris@10
|
382 V T3W, T3O, T3d, T3w;
|
Chris@10
|
383 T3W = VFMA(LDK(KP262346850), T3N, T3l);
|
Chris@10
|
384 T3O = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T3l, T3N));
|
Chris@10
|
385 T3d = VFMA(LDK(KP992114701), T3c, T2X);
|
Chris@10
|
386 T3w = VFNMS(LDK(KP690983005), T3v, T3g);
|
Chris@10
|
387 {
|
Chris@10
|
388 V T3q, T3m, T3T, T43;
|
Chris@10
|
389 T3q = VFMA(LDK(KP763932022), T3p, T3b);
|
Chris@10
|
390 T3m = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T3l, T3k));
|
Chris@10
|
391 T3T = VFNMS(LDK(KP997675361), T3S, T3R);
|
Chris@10
|
392 T43 = VFNMS(LDK(KP904508497), T3S, T3Q);
|
Chris@10
|
393 {
|
Chris@10
|
394 V T3G, T3P, T47, T41;
|
Chris@10
|
395 T3G = VFMA(LDK(KP949179823), T3F, T2X);
|
Chris@10
|
396 T3P = VFNMS(LDK(KP237294955), T3F, T2X);
|
Chris@10
|
397 T47 = VFNMS(LDK(KP669429328), T40, T46);
|
Chris@10
|
398 T41 = VFMA(LDK(KP618033988), T40, T3Z);
|
Chris@10
|
399 ST(&(x[WS(rs, 20)]), VFNMSI(T4m, T4j), ms, &(x[0]));
|
Chris@10
|
400 ST(&(x[WS(rs, 5)]), VFMAI(T4m, T4j), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
401 ST(&(x[WS(rs, 15)]), VFMAI(T4o, T4n), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
402 ST(&(x[WS(rs, 10)]), VFNMSI(T4o, T4n), ms, &(x[0]));
|
Chris@10
|
403 {
|
Chris@10
|
404 V T3x, T3r, T3U, T44;
|
Chris@10
|
405 T3x = VFMA(LDK(KP855719849), T3w, T3t);
|
Chris@10
|
406 T3r = VFNMS(LDK(KP855719849), T3q, T3n);
|
Chris@10
|
407 ST(&(x[WS(rs, 3)]), VFMAI(T3m, T3d), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
408 ST(&(x[WS(rs, 22)]), VFNMSI(T3m, T3d), ms, &(x[0]));
|
Chris@10
|
409 T3U = VFMA(LDK(KP560319534), T3T, T3Q);
|
Chris@10
|
410 T44 = VFNMS(LDK(KP681693190), T43, T3R);
|
Chris@10
|
411 ST(&(x[WS(rs, 2)]), VFMAI(T3O, T3G), ms, &(x[0]));
|
Chris@10
|
412 ST(&(x[WS(rs, 23)]), VFNMSI(T3O, T3G), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
413 T48 = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T47, T3W));
|
Chris@10
|
414 T42 = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T41, T3W));
|
Chris@10
|
415 T3y = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T3x, T3l));
|
Chris@10
|
416 T3s = VFMA(LDK(KP897376177), T3r, T2X);
|
Chris@10
|
417 T3V = VFNMS(LDK(KP949179823), T3U, T3P);
|
Chris@10
|
418 T45 = VFNMS(LDK(KP860541664), T44, T3P);
|
Chris@10
|
419 T2R = VFNMS(LDK(KP912575812), T2z, T2y);
|
Chris@10
|
420 T2A = VFMA(LDK(KP912575812), T2z, T2y);
|
Chris@10
|
421 T1b = VFMA(LDK(KP829049696), T1a, TO);
|
Chris@10
|
422 T2h = VFNMS(LDK(KP829049696), T1a, TO);
|
Chris@10
|
423 T2i = VFNMS(LDK(KP831864738), T1W, T1z);
|
Chris@10
|
424 T1X = VFMA(LDK(KP831864738), T1W, T1z);
|
Chris@10
|
425 }
|
Chris@10
|
426 }
|
Chris@10
|
427 }
|
Chris@10
|
428 }
|
Chris@10
|
429 }
|
Chris@10
|
430 }
|
Chris@10
|
431 }
|
Chris@10
|
432 }
|
Chris@10
|
433 }
|
Chris@10
|
434 }
|
Chris@10
|
435 }
|
Chris@10
|
436 {
|
Chris@10
|
437 V T2M, T2F, T2x, T2S, T2T, T2N;
|
Chris@10
|
438 T2M = VFNMS(LDK(KP958953096), T2E, T2D);
|
Chris@10
|
439 T2F = VFMA(LDK(KP958953096), T2E, T2D);
|
Chris@10
|
440 ST(&(x[WS(rs, 17)]), VFNMSI(T3y, T3s), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
441 ST(&(x[WS(rs, 8)]), VFMAI(T3y, T3s), ms, &(x[0]));
|
Chris@10
|
442 ST(&(x[WS(rs, 13)]), VFMAI(T42, T3V), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
443 ST(&(x[WS(rs, 12)]), VFNMSI(T42, T3V), ms, &(x[0]));
|
Chris@10
|
444 ST(&(x[WS(rs, 7)]), VFNMSI(T48, T45), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
445 ST(&(x[WS(rs, 18)]), VFMAI(T48, T45), ms, &(x[0]));
|
Chris@10
|
446 T2j = VFMA(LDK(KP559154169), T2i, T2h);
|
Chris@10
|
447 T2q = VFNMS(LDK(KP683113946), T2h, T2i);
|
Chris@10
|
448 T2x = VFNMS(LDK(KP867381224), T2w, T2v);
|
Chris@10
|
449 T2S = VFMA(LDK(KP867381224), T2w, T2v);
|
Chris@10
|
450 T2J = VFMA(LDK(KP894834959), T2I, T2F);
|
Chris@10
|
451 T2T = VFMA(LDK(KP447417479), T2I, T2S);
|
Chris@10
|
452 T2B = VFNMS(LDK(KP809385824), T2A, T2x);
|
Chris@10
|
453 T2N = VFMA(LDK(KP447417479), T2A, T2M);
|
Chris@10
|
454 T2e = VFMA(LDK(KP831864738), T25, T24);
|
Chris@10
|
455 T26 = VFNMS(LDK(KP831864738), T25, T24);
|
Chris@10
|
456 T2U = VFNMS(LDK(KP763932022), T2T, T2F);
|
Chris@10
|
457 T1Y = VFMA(LDK(KP904730450), T1X, T1b);
|
Chris@10
|
458 T23 = VFNMS(LDK(KP904730450), T1X, T1b);
|
Chris@10
|
459 T2O = VFMA(LDK(KP690983005), T2N, T2x);
|
Chris@10
|
460 }
|
Chris@10
|
461 }
|
Chris@10
|
462 {
|
Chris@10
|
463 V T2C, T22, T2d, T2K;
|
Chris@10
|
464 T2C = VFNMS(LDK(KP992114701), T2B, To);
|
Chris@10
|
465 T22 = VFMA(LDK(KP916574801), T21, T20);
|
Chris@10
|
466 T2d = VFNMS(LDK(KP916574801), T21, T20);
|
Chris@10
|
467 T2K = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2J, T2c));
|
Chris@10
|
468 {
|
Chris@10
|
469 V T27, T2P, T2f, T2k, T2n, T2V;
|
Chris@10
|
470 T2V = VFNMS(LDK(KP999544308), T2U, T2R);
|
Chris@10
|
471 T27 = VFNMS(LDK(KP904730450), T26, T23);
|
Chris@10
|
472 T2t = VFMA(LDK(KP968583161), T1Y, To);
|
Chris@10
|
473 T1Z = VFNMS(LDK(KP242145790), T1Y, To);
|
Chris@10
|
474 T2P = VFNMS(LDK(KP999544308), T2O, T2L);
|
Chris@10
|
475 T2f = VFMA(LDK(KP904730450), T2e, T2d);
|
Chris@10
|
476 T2k = VFNMS(LDK(KP904730450), T2e, T2d);
|
Chris@10
|
477 T2n = VADD(T22, T23);
|
Chris@10
|
478 ST(&(x[WS(rs, 21)]), VFMAI(T2K, T2C), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
479 ST(&(x[WS(rs, 4)]), VFNMSI(T2K, T2C), ms, &(x[0]));
|
Chris@10
|
480 T2W = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T2V, T2c));
|
Chris@10
|
481 T28 = VFNMS(LDK(KP618033988), T27, T22);
|
Chris@10
|
482 T2Q = VFNMS(LDK(KP803003575), T2P, To);
|
Chris@10
|
483 T2r = VFMA(LDK(KP617882369), T2k, T2q);
|
Chris@10
|
484 T2g = VFNMS(LDK(KP242145790), T2f, T2c);
|
Chris@10
|
485 T2u = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T2f, T2c));
|
Chris@10
|
486 T2o = VFNMS(LDK(KP683113946), T2n, T26);
|
Chris@10
|
487 T2l = VFMA(LDK(KP559016994), T2k, T2j);
|
Chris@10
|
488 }
|
Chris@10
|
489 }
|
Chris@10
|
490 }
|
Chris@10
|
491 }
|
Chris@10
|
492 }
|
Chris@10
|
493 }
|
Chris@10
|
494 }
|
Chris@10
|
495 {
|
Chris@10
|
496 V T29, T2s, T2p, T2m;
|
Chris@10
|
497 T29 = VFNMS(LDK(KP876091699), T28, T1Z);
|
Chris@10
|
498 ST(&(x[WS(rs, 16)]), VFMAI(T2W, T2Q), ms, &(x[0]));
|
Chris@10
|
499 ST(&(x[WS(rs, 9)]), VFNMSI(T2W, T2Q), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
500 T2s = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T2r, T2g));
|
Chris@10
|
501 ST(&(x[WS(rs, 24)]), VFNMSI(T2u, T2t), ms, &(x[0]));
|
Chris@10
|
502 ST(&(x[WS(rs, 1)]), VFMAI(T2u, T2t), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
503 T2p = VFMA(LDK(KP792626838), T2o, T1Z);
|
Chris@10
|
504 T2m = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T2l, T2g));
|
Chris@10
|
505 ST(&(x[WS(rs, 11)]), VFMAI(T2s, T2p), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
506 ST(&(x[WS(rs, 14)]), VFNMSI(T2s, T2p), ms, &(x[0]));
|
Chris@10
|
507 ST(&(x[WS(rs, 19)]), VFNMSI(T2m, T29), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
508 ST(&(x[WS(rs, 6)]), VFMAI(T2m, T29), ms, &(x[0]));
|
Chris@10
|
509 }
|
Chris@10
|
510 }
|
Chris@10
|
511 }
|
Chris@10
|
512 VLEAVE();
|
Chris@10
|
513 }
|
Chris@10
|
514
|
Chris@10
|
515 static const tw_instr twinstr[] = {
|
Chris@10
|
516 VTW(0, 1),
|
Chris@10
|
517 VTW(0, 3),
|
Chris@10
|
518 VTW(0, 9),
|
Chris@10
|
519 VTW(0, 24),
|
Chris@10
|
520 {TW_NEXT, VL, 0}
|
Chris@10
|
521 };
|
Chris@10
|
522
|
Chris@10
|
523 static const ct_desc desc = { 25, XSIMD_STRING("t3bv_25"), twinstr, &GENUS, {87, 100, 181, 0}, 0, 0, 0 };
|
Chris@10
|
524
|
Chris@10
|
525 void XSIMD(codelet_t3bv_25) (planner *p) {
|
Chris@10
|
526 X(kdft_dit_register) (p, t3bv_25, &desc);
|
Chris@10
|
527 }
|
Chris@10
|
528 #else /* HAVE_FMA */
|
Chris@10
|
529
|
Chris@10
|
530 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 25 -name t3bv_25 -include t3b.h -sign 1 */
|
Chris@10
|
531
|
Chris@10
|
532 /*
|
Chris@10
|
533 * This function contains 268 FP additions, 228 FP multiplications,
|
Chris@10
|
534 * (or, 191 additions, 151 multiplications, 77 fused multiply/add),
|
Chris@10
|
535 * 124 stack variables, 40 constants, and 50 memory accesses
|
Chris@10
|
536 */
|
Chris@10
|
537 #include "t3b.h"
|
Chris@10
|
538
|
Chris@10
|
539 static void t3bv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
540 {
|
Chris@10
|
541 DVK(KP497379774, +0.497379774329709576484567492012895936835134813);
|
Chris@10
|
542 DVK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
543 DVK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
544 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
|
Chris@10
|
545 DVK(KP809016994, +0.809016994374947424102293417182819058860154590);
|
Chris@10
|
546 DVK(KP309016994, +0.309016994374947424102293417182819058860154590);
|
Chris@10
|
547 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483);
|
Chris@10
|
548 DVK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
549 DVK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
550 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
|
Chris@10
|
551 DVK(KP963507348, +0.963507348203430549974383005744259307057084020);
|
Chris@10
|
552 DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
553 DVK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
554 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691);
|
Chris@10
|
555 DVK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
556 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
|
Chris@10
|
557 DVK(KP851558583, +0.851558583130145297725004891488503407959946084);
|
Chris@10
|
558 DVK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
559 DVK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
560 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418);
|
Chris@10
|
561 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328);
|
Chris@10
|
562 DVK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
563 DVK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
564 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344);
|
Chris@10
|
565 DVK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
566 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
|
Chris@10
|
567 DVK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
568 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596);
|
Chris@10
|
569 DVK(KP125581039, +0.125581039058626752152356449131262266244969664);
|
Chris@10
|
570 DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
571 DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
572 DVK(KP250666467, +0.250666467128608490746237519633017587885836494);
|
Chris@10
|
573 DVK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
574 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465);
|
Chris@10
|
575 DVK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
Chris@10
|
576 DVK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
Chris@10
|
577 DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
578 DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
579 DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
580 DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
581 {
|
Chris@10
|
582 INT m;
|
Chris@10
|
583 R *x;
|
Chris@10
|
584 x = ii;
|
Chris@10
|
585 for (m = mb, W = W + (mb * ((TWVL / VL) * 8)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(25, rs)) {
|
Chris@10
|
586 V T1, Td, T8, T9, TF, Te, Tu, TB, TC, T1s, T15, Tf, TY, T4, Ta;
|
Chris@10
|
587 V Tx, T1T, Tg, T1N, T1v, T18, TG, T1o, T11;
|
Chris@10
|
588 T1 = LDW(&(W[TWVL * 4]));
|
Chris@10
|
589 Td = LDW(&(W[TWVL * 2]));
|
Chris@10
|
590 T8 = LDW(&(W[0]));
|
Chris@10
|
591 T9 = VZMUL(T8, T1);
|
Chris@10
|
592 TF = VZMULJ(T8, T1);
|
Chris@10
|
593 Te = VZMUL(T8, Td);
|
Chris@10
|
594 Tu = VZMULJ(Td, T1);
|
Chris@10
|
595 TB = VZMULJ(T8, Td);
|
Chris@10
|
596 TC = VZMUL(TB, T1);
|
Chris@10
|
597 T1s = VZMUL(Te, T1);
|
Chris@10
|
598 T15 = VZMUL(Td, T1);
|
Chris@10
|
599 Tf = VZMULJ(Te, T1);
|
Chris@10
|
600 TY = VZMULJ(TB, T1);
|
Chris@10
|
601 T4 = LDW(&(W[TWVL * 6]));
|
Chris@10
|
602 Ta = VZMULJ(T9, T4);
|
Chris@10
|
603 Tx = VZMULJ(Td, T4);
|
Chris@10
|
604 T1T = VZMULJ(T1, T4);
|
Chris@10
|
605 Tg = VZMULJ(Tf, T4);
|
Chris@10
|
606 T1N = VZMULJ(Te, T4);
|
Chris@10
|
607 T1v = VZMULJ(Tu, T4);
|
Chris@10
|
608 T18 = VZMULJ(TY, T4);
|
Chris@10
|
609 TG = VZMULJ(TF, T4);
|
Chris@10
|
610 T1o = VZMULJ(T8, T4);
|
Chris@10
|
611 T11 = VZMULJ(TB, T4);
|
Chris@10
|
612 {
|
Chris@10
|
613 V T1Y, T1X, T2f, T2g, T1Z, T20, T2e, T39, T1H, T2T, T1E, T3C, T2S, Tk, T2G;
|
Chris@10
|
614 V Ts, T3z, T2F, TK, T2I, TS, T3y, T2J, T1k, T2Q, T1h, T3B, T2P;
|
Chris@10
|
615 {
|
Chris@10
|
616 V T1S, T1V, T1W, T1M, T1P, T1Q, T2d;
|
Chris@10
|
617 T1Y = LD(&(x[0]), ms, &(x[0]));
|
Chris@10
|
618 {
|
Chris@10
|
619 V T1R, T1U, T1L, T1O;
|
Chris@10
|
620 T1R = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
|
Chris@10
|
621 T1S = VZMUL(T9, T1R);
|
Chris@10
|
622 T1U = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
623 T1V = VZMUL(T1T, T1U);
|
Chris@10
|
624 T1W = VADD(T1S, T1V);
|
Chris@10
|
625 T1L = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
626 T1M = VZMUL(Tf, T1L);
|
Chris@10
|
627 T1O = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
|
Chris@10
|
628 T1P = VZMUL(T1N, T1O);
|
Chris@10
|
629 T1Q = VADD(T1M, T1P);
|
Chris@10
|
630 }
|
Chris@10
|
631 T1X = VMUL(LDK(KP559016994), VSUB(T1Q, T1W));
|
Chris@10
|
632 T2f = VSUB(T1S, T1V);
|
Chris@10
|
633 T2g = VMUL(LDK(KP587785252), T2f);
|
Chris@10
|
634 T1Z = VADD(T1Q, T1W);
|
Chris@10
|
635 T20 = VFNMS(LDK(KP250000000), T1Z, T1Y);
|
Chris@10
|
636 T2d = VSUB(T1M, T1P);
|
Chris@10
|
637 T2e = VMUL(LDK(KP951056516), T2d);
|
Chris@10
|
638 T39 = VMUL(LDK(KP587785252), T2d);
|
Chris@10
|
639 }
|
Chris@10
|
640 {
|
Chris@10
|
641 V T1B, T1u, T1x, T1y, T1n, T1q, T1r, T1A;
|
Chris@10
|
642 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
643 T1B = VZMUL(Td, T1A);
|
Chris@10
|
644 {
|
Chris@10
|
645 V T1t, T1w, T1m, T1p;
|
Chris@10
|
646 T1t = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
647 T1u = VZMUL(T1s, T1t);
|
Chris@10
|
648 T1w = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
|
Chris@10
|
649 T1x = VZMUL(T1v, T1w);
|
Chris@10
|
650 T1y = VADD(T1u, T1x);
|
Chris@10
|
651 T1m = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
|
Chris@10
|
652 T1n = VZMUL(TF, T1m);
|
Chris@10
|
653 T1p = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
654 T1q = VZMUL(T1o, T1p);
|
Chris@10
|
655 T1r = VADD(T1n, T1q);
|
Chris@10
|
656 }
|
Chris@10
|
657 {
|
Chris@10
|
658 V T1F, T1G, T1z, T1C, T1D;
|
Chris@10
|
659 T1F = VSUB(T1n, T1q);
|
Chris@10
|
660 T1G = VSUB(T1u, T1x);
|
Chris@10
|
661 T1H = VFMA(LDK(KP475528258), T1F, VMUL(LDK(KP293892626), T1G));
|
Chris@10
|
662 T2T = VFNMS(LDK(KP475528258), T1G, VMUL(LDK(KP293892626), T1F));
|
Chris@10
|
663 T1z = VMUL(LDK(KP559016994), VSUB(T1r, T1y));
|
Chris@10
|
664 T1C = VADD(T1r, T1y);
|
Chris@10
|
665 T1D = VFNMS(LDK(KP250000000), T1C, T1B);
|
Chris@10
|
666 T1E = VADD(T1z, T1D);
|
Chris@10
|
667 T3C = VADD(T1B, T1C);
|
Chris@10
|
668 T2S = VSUB(T1D, T1z);
|
Chris@10
|
669 }
|
Chris@10
|
670 }
|
Chris@10
|
671 {
|
Chris@10
|
672 V Tp, Tc, Ti, Tm, T3, T6, Tl, To;
|
Chris@10
|
673 To = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
|
Chris@10
|
674 Tp = VZMUL(Te, To);
|
Chris@10
|
675 {
|
Chris@10
|
676 V Tb, Th, T2, T5;
|
Chris@10
|
677 Tb = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
|
Chris@10
|
678 Tc = VZMUL(Ta, Tb);
|
Chris@10
|
679 Th = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
680 Ti = VZMUL(Tg, Th);
|
Chris@10
|
681 Tm = VADD(Tc, Ti);
|
Chris@10
|
682 T2 = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
683 T3 = VZMUL(T1, T2);
|
Chris@10
|
684 T5 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
|
Chris@10
|
685 T6 = VZMUL(T4, T5);
|
Chris@10
|
686 Tl = VADD(T3, T6);
|
Chris@10
|
687 }
|
Chris@10
|
688 {
|
Chris@10
|
689 V T7, Tj, Tn, Tq, Tr;
|
Chris@10
|
690 T7 = VSUB(T3, T6);
|
Chris@10
|
691 Tj = VSUB(Tc, Ti);
|
Chris@10
|
692 Tk = VFMA(LDK(KP475528258), T7, VMUL(LDK(KP293892626), Tj));
|
Chris@10
|
693 T2G = VFNMS(LDK(KP475528258), Tj, VMUL(LDK(KP293892626), T7));
|
Chris@10
|
694 Tn = VMUL(LDK(KP559016994), VSUB(Tl, Tm));
|
Chris@10
|
695 Tq = VADD(Tl, Tm);
|
Chris@10
|
696 Tr = VFNMS(LDK(KP250000000), Tq, Tp);
|
Chris@10
|
697 Ts = VADD(Tn, Tr);
|
Chris@10
|
698 T3z = VADD(Tp, Tq);
|
Chris@10
|
699 T2F = VSUB(Tr, Tn);
|
Chris@10
|
700 }
|
Chris@10
|
701 }
|
Chris@10
|
702 {
|
Chris@10
|
703 V TP, TE, TI, TM, Tw, Tz, TL, TO;
|
Chris@10
|
704 TO = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
705 TP = VZMUL(T8, TO);
|
Chris@10
|
706 {
|
Chris@10
|
707 V TD, TH, Tv, Ty;
|
Chris@10
|
708 TD = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
709 TE = VZMUL(TC, TD);
|
Chris@10
|
710 TH = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
|
Chris@10
|
711 TI = VZMUL(TG, TH);
|
Chris@10
|
712 TM = VADD(TE, TI);
|
Chris@10
|
713 Tv = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
|
Chris@10
|
714 Tw = VZMUL(Tu, Tv);
|
Chris@10
|
715 Ty = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
716 Tz = VZMUL(Tx, Ty);
|
Chris@10
|
717 TL = VADD(Tw, Tz);
|
Chris@10
|
718 }
|
Chris@10
|
719 {
|
Chris@10
|
720 V TA, TJ, TN, TQ, TR;
|
Chris@10
|
721 TA = VSUB(Tw, Tz);
|
Chris@10
|
722 TJ = VSUB(TE, TI);
|
Chris@10
|
723 TK = VFMA(LDK(KP475528258), TA, VMUL(LDK(KP293892626), TJ));
|
Chris@10
|
724 T2I = VFNMS(LDK(KP475528258), TJ, VMUL(LDK(KP293892626), TA));
|
Chris@10
|
725 TN = VMUL(LDK(KP559016994), VSUB(TL, TM));
|
Chris@10
|
726 TQ = VADD(TL, TM);
|
Chris@10
|
727 TR = VFNMS(LDK(KP250000000), TQ, TP);
|
Chris@10
|
728 TS = VADD(TN, TR);
|
Chris@10
|
729 T3y = VADD(TP, TQ);
|
Chris@10
|
730 T2J = VSUB(TR, TN);
|
Chris@10
|
731 }
|
Chris@10
|
732 }
|
Chris@10
|
733 {
|
Chris@10
|
734 V T1e, T17, T1a, T1b, T10, T13, T14, T1d;
|
Chris@10
|
735 T1d = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
|
Chris@10
|
736 T1e = VZMUL(TB, T1d);
|
Chris@10
|
737 {
|
Chris@10
|
738 V T16, T19, TZ, T12;
|
Chris@10
|
739 T16 = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
|
Chris@10
|
740 T17 = VZMUL(T15, T16);
|
Chris@10
|
741 T19 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
742 T1a = VZMUL(T18, T19);
|
Chris@10
|
743 T1b = VADD(T17, T1a);
|
Chris@10
|
744 TZ = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
745 T10 = VZMUL(TY, TZ);
|
Chris@10
|
746 T12 = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
|
Chris@10
|
747 T13 = VZMUL(T11, T12);
|
Chris@10
|
748 T14 = VADD(T10, T13);
|
Chris@10
|
749 }
|
Chris@10
|
750 {
|
Chris@10
|
751 V T1i, T1j, T1c, T1f, T1g;
|
Chris@10
|
752 T1i = VSUB(T10, T13);
|
Chris@10
|
753 T1j = VSUB(T17, T1a);
|
Chris@10
|
754 T1k = VFMA(LDK(KP475528258), T1i, VMUL(LDK(KP293892626), T1j));
|
Chris@10
|
755 T2Q = VFNMS(LDK(KP475528258), T1j, VMUL(LDK(KP293892626), T1i));
|
Chris@10
|
756 T1c = VMUL(LDK(KP559016994), VSUB(T14, T1b));
|
Chris@10
|
757 T1f = VADD(T14, T1b);
|
Chris@10
|
758 T1g = VFNMS(LDK(KP250000000), T1f, T1e);
|
Chris@10
|
759 T1h = VADD(T1c, T1g);
|
Chris@10
|
760 T3B = VADD(T1e, T1f);
|
Chris@10
|
761 T2P = VSUB(T1g, T1c);
|
Chris@10
|
762 }
|
Chris@10
|
763 }
|
Chris@10
|
764 {
|
Chris@10
|
765 V T3E, T3M, T3I, T3J, T3H, T3K, T3N, T3L;
|
Chris@10
|
766 {
|
Chris@10
|
767 V T3A, T3D, T3F, T3G;
|
Chris@10
|
768 T3A = VSUB(T3y, T3z);
|
Chris@10
|
769 T3D = VSUB(T3B, T3C);
|
Chris@10
|
770 T3E = VBYI(VFMA(LDK(KP951056516), T3A, VMUL(LDK(KP587785252), T3D)));
|
Chris@10
|
771 T3M = VBYI(VFNMS(LDK(KP951056516), T3D, VMUL(LDK(KP587785252), T3A)));
|
Chris@10
|
772 T3I = VADD(T1Y, T1Z);
|
Chris@10
|
773 T3F = VADD(T3y, T3z);
|
Chris@10
|
774 T3G = VADD(T3B, T3C);
|
Chris@10
|
775 T3J = VADD(T3F, T3G);
|
Chris@10
|
776 T3H = VMUL(LDK(KP559016994), VSUB(T3F, T3G));
|
Chris@10
|
777 T3K = VFNMS(LDK(KP250000000), T3J, T3I);
|
Chris@10
|
778 }
|
Chris@10
|
779 ST(&(x[0]), VADD(T3I, T3J), ms, &(x[0]));
|
Chris@10
|
780 T3N = VSUB(T3K, T3H);
|
Chris@10
|
781 ST(&(x[WS(rs, 10)]), VADD(T3M, T3N), ms, &(x[0]));
|
Chris@10
|
782 ST(&(x[WS(rs, 15)]), VSUB(T3N, T3M), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
783 T3L = VADD(T3H, T3K);
|
Chris@10
|
784 ST(&(x[WS(rs, 5)]), VADD(T3E, T3L), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
785 ST(&(x[WS(rs, 20)]), VSUB(T3L, T3E), ms, &(x[0]));
|
Chris@10
|
786 }
|
Chris@10
|
787 {
|
Chris@10
|
788 V T2X, T3a, T3i, T3j, T3k, T3s, T3t, T3u, T3l, T3m, T3n, T3p, T3q, T3r, T2L;
|
Chris@10
|
789 V T3b, T32, T38, T2W, T35, T2Y, T34, T3w, T3x;
|
Chris@10
|
790 T2X = VSUB(T20, T1X);
|
Chris@10
|
791 T3a = VFNMS(LDK(KP951056516), T2f, T39);
|
Chris@10
|
792 T3i = VFMA(LDK(KP1_369094211), T2I, VMUL(LDK(KP728968627), T2J));
|
Chris@10
|
793 T3j = VFNMS(LDK(KP992114701), T2F, VMUL(LDK(KP250666467), T2G));
|
Chris@10
|
794 T3k = VADD(T3i, T3j);
|
Chris@10
|
795 T3s = VFNMS(LDK(KP125581039), T2Q, VMUL(LDK(KP998026728), T2P));
|
Chris@10
|
796 T3t = VFMA(LDK(KP1_274847979), T2T, VMUL(LDK(KP770513242), T2S));
|
Chris@10
|
797 T3u = VADD(T3s, T3t);
|
Chris@10
|
798 T3l = VFMA(LDK(KP1_996053456), T2Q, VMUL(LDK(KP062790519), T2P));
|
Chris@10
|
799 T3m = VFNMS(LDK(KP637423989), T2S, VMUL(LDK(KP1_541026485), T2T));
|
Chris@10
|
800 T3n = VADD(T3l, T3m);
|
Chris@10
|
801 T3p = VFNMS(LDK(KP1_457937254), T2I, VMUL(LDK(KP684547105), T2J));
|
Chris@10
|
802 T3q = VFMA(LDK(KP1_984229402), T2G, VMUL(LDK(KP125333233), T2F));
|
Chris@10
|
803 T3r = VADD(T3p, T3q);
|
Chris@10
|
804 {
|
Chris@10
|
805 V T2H, T2K, T36, T30, T31, T37;
|
Chris@10
|
806 T2H = VFNMS(LDK(KP851558583), T2G, VMUL(LDK(KP904827052), T2F));
|
Chris@10
|
807 T2K = VFMA(LDK(KP1_752613360), T2I, VMUL(LDK(KP481753674), T2J));
|
Chris@10
|
808 T36 = VADD(T2K, T2H);
|
Chris@10
|
809 T30 = VFMA(LDK(KP1_071653589), T2Q, VMUL(LDK(KP844327925), T2P));
|
Chris@10
|
810 T31 = VFMA(LDK(KP125581039), T2T, VMUL(LDK(KP998026728), T2S));
|
Chris@10
|
811 T37 = VADD(T30, T31);
|
Chris@10
|
812 T2L = VSUB(T2H, T2K);
|
Chris@10
|
813 T3b = VADD(T36, T37);
|
Chris@10
|
814 T32 = VSUB(T30, T31);
|
Chris@10
|
815 T38 = VMUL(LDK(KP559016994), VSUB(T36, T37));
|
Chris@10
|
816 }
|
Chris@10
|
817 {
|
Chris@10
|
818 V T2M, T2N, T2O, T2R, T2U, T2V;
|
Chris@10
|
819 T2M = VFNMS(LDK(KP963507348), T2I, VMUL(LDK(KP876306680), T2J));
|
Chris@10
|
820 T2N = VFMA(LDK(KP1_809654104), T2G, VMUL(LDK(KP425779291), T2F));
|
Chris@10
|
821 T2O = VSUB(T2M, T2N);
|
Chris@10
|
822 T2R = VFNMS(LDK(KP1_688655851), T2Q, VMUL(LDK(KP535826794), T2P));
|
Chris@10
|
823 T2U = VFNMS(LDK(KP1_996053456), T2T, VMUL(LDK(KP062790519), T2S));
|
Chris@10
|
824 T2V = VADD(T2R, T2U);
|
Chris@10
|
825 T2W = VMUL(LDK(KP559016994), VSUB(T2O, T2V));
|
Chris@10
|
826 T35 = VSUB(T2R, T2U);
|
Chris@10
|
827 T2Y = VADD(T2O, T2V);
|
Chris@10
|
828 T34 = VADD(T2M, T2N);
|
Chris@10
|
829 }
|
Chris@10
|
830 {
|
Chris@10
|
831 V T3g, T3h, T3o, T3v;
|
Chris@10
|
832 T3g = VADD(T2X, T2Y);
|
Chris@10
|
833 T3h = VBYI(VADD(T3a, T3b));
|
Chris@10
|
834 ST(&(x[WS(rs, 23)]), VSUB(T3g, T3h), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
835 ST(&(x[WS(rs, 2)]), VADD(T3g, T3h), ms, &(x[0]));
|
Chris@10
|
836 T3o = VADD(T2X, VADD(T3k, T3n));
|
Chris@10
|
837 T3v = VBYI(VSUB(VADD(T3r, T3u), T3a));
|
Chris@10
|
838 ST(&(x[WS(rs, 22)]), VSUB(T3o, T3v), ms, &(x[0]));
|
Chris@10
|
839 ST(&(x[WS(rs, 3)]), VADD(T3o, T3v), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
840 }
|
Chris@10
|
841 T3w = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T3i, T3j), VFMA(LDK(KP309016994), T3r, VFNMS(LDK(KP809016994), T3u, VMUL(LDK(KP587785252), VSUB(T3l, T3m))))), T3a));
|
Chris@10
|
842 T3x = VFMA(LDK(KP309016994), T3k, VFMA(LDK(KP951056516), VSUB(T3q, T3p), VFMA(LDK(KP587785252), VSUB(T3t, T3s), VFNMS(LDK(KP809016994), T3n, T2X))));
|
Chris@10
|
843 ST(&(x[WS(rs, 8)]), VADD(T3w, T3x), ms, &(x[0]));
|
Chris@10
|
844 ST(&(x[WS(rs, 17)]), VSUB(T3x, T3w), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
845 {
|
Chris@10
|
846 V T33, T3e, T3d, T3f, T2Z, T3c;
|
Chris@10
|
847 T2Z = VFNMS(LDK(KP250000000), T2Y, T2X);
|
Chris@10
|
848 T33 = VFMA(LDK(KP951056516), T2L, VADD(T2W, VFNMS(LDK(KP587785252), T32, T2Z)));
|
Chris@10
|
849 T3e = VFMA(LDK(KP587785252), T2L, VFMA(LDK(KP951056516), T32, VSUB(T2Z, T2W)));
|
Chris@10
|
850 T3c = VFNMS(LDK(KP250000000), T3b, T3a);
|
Chris@10
|
851 T3d = VBYI(VADD(VFMA(LDK(KP951056516), T34, VMUL(LDK(KP587785252), T35)), VADD(T38, T3c)));
|
Chris@10
|
852 T3f = VBYI(VADD(VFNMS(LDK(KP951056516), T35, VMUL(LDK(KP587785252), T34)), VSUB(T3c, T38)));
|
Chris@10
|
853 ST(&(x[WS(rs, 18)]), VSUB(T33, T3d), ms, &(x[0]));
|
Chris@10
|
854 ST(&(x[WS(rs, 12)]), VADD(T3e, T3f), ms, &(x[0]));
|
Chris@10
|
855 ST(&(x[WS(rs, 7)]), VADD(T33, T3d), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
856 ST(&(x[WS(rs, 13)]), VSUB(T3e, T3f), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
857 }
|
Chris@10
|
858 }
|
Chris@10
|
859 {
|
Chris@10
|
860 V T21, T2h, T2p, T2q, T2r, T2z, T2A, T2B, T2s, T2t, T2u, T2w, T2x, T2y, TU;
|
Chris@10
|
861 V T2i, T26, T2c, T1K, T29, T22, T28, T2D, T2E;
|
Chris@10
|
862 T21 = VADD(T1X, T20);
|
Chris@10
|
863 T2h = VADD(T2e, T2g);
|
Chris@10
|
864 T2p = VFMA(LDK(KP1_688655851), TK, VMUL(LDK(KP535826794), TS));
|
Chris@10
|
865 T2q = VFMA(LDK(KP1_541026485), Tk, VMUL(LDK(KP637423989), Ts));
|
Chris@10
|
866 T2r = VSUB(T2p, T2q);
|
Chris@10
|
867 T2z = VFMA(LDK(KP851558583), T1k, VMUL(LDK(KP904827052), T1h));
|
Chris@10
|
868 T2A = VFMA(LDK(KP1_984229402), T1H, VMUL(LDK(KP125333233), T1E));
|
Chris@10
|
869 T2B = VADD(T2z, T2A);
|
Chris@10
|
870 T2s = VFNMS(LDK(KP425779291), T1h, VMUL(LDK(KP1_809654104), T1k));
|
Chris@10
|
871 T2t = VFNMS(LDK(KP992114701), T1E, VMUL(LDK(KP250666467), T1H));
|
Chris@10
|
872 T2u = VADD(T2s, T2t);
|
Chris@10
|
873 T2w = VFNMS(LDK(KP1_071653589), TK, VMUL(LDK(KP844327925), TS));
|
Chris@10
|
874 T2x = VFNMS(LDK(KP770513242), Ts, VMUL(LDK(KP1_274847979), Tk));
|
Chris@10
|
875 T2y = VADD(T2w, T2x);
|
Chris@10
|
876 {
|
Chris@10
|
877 V Tt, TT, T2a, T24, T25, T2b;
|
Chris@10
|
878 Tt = VFMA(LDK(KP1_071653589), Tk, VMUL(LDK(KP844327925), Ts));
|
Chris@10
|
879 TT = VFMA(LDK(KP1_937166322), TK, VMUL(LDK(KP248689887), TS));
|
Chris@10
|
880 T2a = VADD(TT, Tt);
|
Chris@10
|
881 T24 = VFMA(LDK(KP1_752613360), T1k, VMUL(LDK(KP481753674), T1h));
|
Chris@10
|
882 T25 = VFMA(LDK(KP1_457937254), T1H, VMUL(LDK(KP684547105), T1E));
|
Chris@10
|
883 T2b = VADD(T24, T25);
|
Chris@10
|
884 TU = VSUB(Tt, TT);
|
Chris@10
|
885 T2i = VADD(T2a, T2b);
|
Chris@10
|
886 T26 = VSUB(T24, T25);
|
Chris@10
|
887 T2c = VMUL(LDK(KP559016994), VSUB(T2a, T2b));
|
Chris@10
|
888 }
|
Chris@10
|
889 {
|
Chris@10
|
890 V TV, TW, TX, T1l, T1I, T1J;
|
Chris@10
|
891 TV = VFNMS(LDK(KP497379774), TK, VMUL(LDK(KP968583161), TS));
|
Chris@10
|
892 TW = VFNMS(LDK(KP1_688655851), Tk, VMUL(LDK(KP535826794), Ts));
|
Chris@10
|
893 TX = VADD(TV, TW);
|
Chris@10
|
894 T1l = VFNMS(LDK(KP963507348), T1k, VMUL(LDK(KP876306680), T1h));
|
Chris@10
|
895 T1I = VFNMS(LDK(KP1_369094211), T1H, VMUL(LDK(KP728968627), T1E));
|
Chris@10
|
896 T1J = VADD(T1l, T1I);
|
Chris@10
|
897 T1K = VMUL(LDK(KP559016994), VSUB(TX, T1J));
|
Chris@10
|
898 T29 = VSUB(T1l, T1I);
|
Chris@10
|
899 T22 = VADD(TX, T1J);
|
Chris@10
|
900 T28 = VSUB(TV, TW);
|
Chris@10
|
901 }
|
Chris@10
|
902 {
|
Chris@10
|
903 V T2n, T2o, T2v, T2C;
|
Chris@10
|
904 T2n = VADD(T21, T22);
|
Chris@10
|
905 T2o = VBYI(VADD(T2h, T2i));
|
Chris@10
|
906 ST(&(x[WS(rs, 24)]), VSUB(T2n, T2o), ms, &(x[0]));
|
Chris@10
|
907 ST(&(x[WS(rs, 1)]), VADD(T2n, T2o), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
908 T2v = VADD(T21, VADD(T2r, T2u));
|
Chris@10
|
909 T2C = VBYI(VSUB(VADD(T2y, T2B), T2h));
|
Chris@10
|
910 ST(&(x[WS(rs, 21)]), VSUB(T2v, T2C), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
911 ST(&(x[WS(rs, 4)]), VADD(T2v, T2C), ms, &(x[0]));
|
Chris@10
|
912 }
|
Chris@10
|
913 T2D = VBYI(VSUB(VFMA(LDK(KP309016994), T2y, VFMA(LDK(KP951056516), VADD(T2p, T2q), VFNMS(LDK(KP809016994), T2B, VMUL(LDK(KP587785252), VSUB(T2s, T2t))))), T2h));
|
Chris@10
|
914 T2E = VFMA(LDK(KP951056516), VSUB(T2x, T2w), VFMA(LDK(KP309016994), T2r, VFMA(LDK(KP587785252), VSUB(T2A, T2z), VFNMS(LDK(KP809016994), T2u, T21))));
|
Chris@10
|
915 ST(&(x[WS(rs, 9)]), VADD(T2D, T2E), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
916 ST(&(x[WS(rs, 16)]), VSUB(T2E, T2D), ms, &(x[0]));
|
Chris@10
|
917 {
|
Chris@10
|
918 V T27, T2l, T2k, T2m, T23, T2j;
|
Chris@10
|
919 T23 = VFNMS(LDK(KP250000000), T22, T21);
|
Chris@10
|
920 T27 = VFMA(LDK(KP951056516), TU, VADD(T1K, VFNMS(LDK(KP587785252), T26, T23)));
|
Chris@10
|
921 T2l = VFMA(LDK(KP587785252), TU, VFMA(LDK(KP951056516), T26, VSUB(T23, T1K)));
|
Chris@10
|
922 T2j = VFNMS(LDK(KP250000000), T2i, T2h);
|
Chris@10
|
923 T2k = VBYI(VADD(VFMA(LDK(KP951056516), T28, VMUL(LDK(KP587785252), T29)), VADD(T2c, T2j)));
|
Chris@10
|
924 T2m = VBYI(VADD(VFNMS(LDK(KP951056516), T29, VMUL(LDK(KP587785252), T28)), VSUB(T2j, T2c)));
|
Chris@10
|
925 ST(&(x[WS(rs, 19)]), VSUB(T27, T2k), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
926 ST(&(x[WS(rs, 11)]), VADD(T2l, T2m), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
927 ST(&(x[WS(rs, 6)]), VADD(T27, T2k), ms, &(x[0]));
|
Chris@10
|
928 ST(&(x[WS(rs, 14)]), VSUB(T2l, T2m), ms, &(x[0]));
|
Chris@10
|
929 }
|
Chris@10
|
930 }
|
Chris@10
|
931 }
|
Chris@10
|
932 }
|
Chris@10
|
933 }
|
Chris@10
|
934 VLEAVE();
|
Chris@10
|
935 }
|
Chris@10
|
936
|
Chris@10
|
937 static const tw_instr twinstr[] = {
|
Chris@10
|
938 VTW(0, 1),
|
Chris@10
|
939 VTW(0, 3),
|
Chris@10
|
940 VTW(0, 9),
|
Chris@10
|
941 VTW(0, 24),
|
Chris@10
|
942 {TW_NEXT, VL, 0}
|
Chris@10
|
943 };
|
Chris@10
|
944
|
Chris@10
|
945 static const ct_desc desc = { 25, XSIMD_STRING("t3bv_25"), twinstr, &GENUS, {191, 151, 77, 0}, 0, 0, 0 };
|
Chris@10
|
946
|
Chris@10
|
947 void XSIMD(codelet_t3bv_25) (planner *p) {
|
Chris@10
|
948 X(kdft_dit_register) (p, t3bv_25, &desc);
|
Chris@10
|
949 }
|
Chris@10
|
950 #endif /* HAVE_FMA */
|