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 07:39:06 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 -n 25 -name t1bv_25 -include t1b.h -sign 1 */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 248 FP additions, 241 FP multiplications,
|
Chris@10
|
32 * (or, 67 additions, 60 multiplications, 181 fused multiply/add),
|
Chris@10
|
33 * 208 stack variables, 67 constants, and 50 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "t1b.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void t1bv_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(KP921078979, +0.921078979742360627699756128143719920817673854);
|
Chris@10
|
84 DVK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
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(KP269969613, +0.269969613759572083574752974412347470060951301);
|
Chris@10
|
90 DVK(KP244189809, +0.244189809627953270309879511234821255780225091);
|
Chris@10
|
91 DVK(KP667278218, +0.667278218140296670899089292254759909713898805);
|
Chris@10
|
92 DVK(KP603558818, +0.603558818296015001454675132653458027918768137);
|
Chris@10
|
93 DVK(KP522847744, +0.522847744331509716623755382187077770911012542);
|
Chris@10
|
94 DVK(KP578046249, +0.578046249379945007321754579646815604023525655);
|
Chris@10
|
95 DVK(KP987388751, +0.987388751065621252324603216482382109400433949);
|
Chris@10
|
96 DVK(KP893101515, +0.893101515366181661711202267938416198338079437);
|
Chris@10
|
97 DVK(KP120146378, +0.120146378570687701782758537356596213647956445);
|
Chris@10
|
98 DVK(KP132830569, +0.132830569247582714407653942074819768844536507);
|
Chris@10
|
99 DVK(KP869845200, +0.869845200362138853122720822420327157933056305);
|
Chris@10
|
100 DVK(KP786782374, +0.786782374965295178365099601674911834788448471);
|
Chris@10
|
101 DVK(KP066152395, +0.066152395967733048213034281011006031460903353);
|
Chris@10
|
102 DVK(KP059835404, +0.059835404262124915169548397419498386427871950);
|
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) * 48)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 48), MAKE_VOLATILE_STRIDE(25, rs)) {
|
Chris@10
|
111 V T25, T1B, T2y, T1K, T2s, T23, T1S, T26, T20, T1X;
|
Chris@10
|
112 {
|
Chris@10
|
113 V T1O, T2X, Te, T3L, Td, T3Q, T3j, T3b, T2R, T2M, T2f, T27, T1y, T1H, T3M;
|
Chris@10
|
114 V TW, TR, TK, T2B, T3n, T3e, T2U, T2F, T2i, T2a, Tz, T1C, T3N, TQ, T11;
|
Chris@10
|
115 V T1b, T1c, T16;
|
Chris@10
|
116 {
|
Chris@10
|
117 V T1, T1g, T1i, T1p, T1k, T1m, Tb, T1N, T6, T1M;
|
Chris@10
|
118 {
|
Chris@10
|
119 V T7, T9, T2, T4, T1f, T1h, T1o;
|
Chris@10
|
120 T1 = LD(&(x[0]), ms, &(x[0]));
|
Chris@10
|
121 T7 = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
|
Chris@10
|
122 T9 = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
123 T2 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
124 T4 = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
|
Chris@10
|
125 T1f = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
126 T1h = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
|
Chris@10
|
127 T1o = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
|
Chris@10
|
128 {
|
Chris@10
|
129 V T8, Ta, T3, T5, T1j;
|
Chris@10
|
130 T1j = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
131 T8 = BYTW(&(W[TWVL * 18]), T7);
|
Chris@10
|
132 Ta = BYTW(&(W[TWVL * 28]), T9);
|
Chris@10
|
133 T3 = BYTW(&(W[TWVL * 8]), T2);
|
Chris@10
|
134 T5 = BYTW(&(W[TWVL * 38]), T4);
|
Chris@10
|
135 T1g = BYTW(&(W[TWVL * 4]), T1f);
|
Chris@10
|
136 T1i = BYTW(&(W[TWVL * 14]), T1h);
|
Chris@10
|
137 T1p = BYTW(&(W[TWVL * 34]), T1o);
|
Chris@10
|
138 T1k = BYTW(&(W[TWVL * 44]), T1j);
|
Chris@10
|
139 T1m = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
140 Tb = VADD(T8, Ta);
|
Chris@10
|
141 T1N = VSUB(T8, Ta);
|
Chris@10
|
142 T6 = VADD(T3, T5);
|
Chris@10
|
143 T1M = VSUB(T3, T5);
|
Chris@10
|
144 }
|
Chris@10
|
145 }
|
Chris@10
|
146 {
|
Chris@10
|
147 V T1v, T1l, Th, Tj, T1w, T1q, Tq, Tk, Tn, Tg;
|
Chris@10
|
148 Tg = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
149 {
|
Chris@10
|
150 V Tc, Ti, T1n, Tp;
|
Chris@10
|
151 Ti = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
|
Chris@10
|
152 T1v = VSUB(T1i, T1k);
|
Chris@10
|
153 T1l = VADD(T1i, T1k);
|
Chris@10
|
154 T1n = BYTW(&(W[TWVL * 24]), T1m);
|
Chris@10
|
155 Tp = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
|
Chris@10
|
156 T1O = VFMA(LDK(KP618033988), T1N, T1M);
|
Chris@10
|
157 T2X = VFNMS(LDK(KP618033988), T1M, T1N);
|
Chris@10
|
158 Te = VSUB(T6, Tb);
|
Chris@10
|
159 Tc = VADD(T6, Tb);
|
Chris@10
|
160 Th = BYTW(&(W[0]), Tg);
|
Chris@10
|
161 Tj = BYTW(&(W[TWVL * 10]), Ti);
|
Chris@10
|
162 T1w = VSUB(T1n, T1p);
|
Chris@10
|
163 T1q = VADD(T1n, T1p);
|
Chris@10
|
164 Tq = BYTW(&(W[TWVL * 30]), Tp);
|
Chris@10
|
165 Tk = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
166 T3L = VADD(T1, Tc);
|
Chris@10
|
167 Td = VFNMS(LDK(KP250000000), Tc, T1);
|
Chris@10
|
168 Tn = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
169 }
|
Chris@10
|
170 {
|
Chris@10
|
171 V T1x, T2K, TM, TB, Tw, Tm, Tx, Tr, TI, T2L, T1u, TD, TF, TL;
|
Chris@10
|
172 TL = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
|
Chris@10
|
173 {
|
Chris@10
|
174 V T1t, Tl, To, TH, T1s, T1r, TA, TC;
|
Chris@10
|
175 TA = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
|
Chris@10
|
176 T1r = VADD(T1l, T1q);
|
Chris@10
|
177 T1t = VSUB(T1q, T1l);
|
Chris@10
|
178 T1x = VFMA(LDK(KP618033988), T1w, T1v);
|
Chris@10
|
179 T2K = VFNMS(LDK(KP618033988), T1v, T1w);
|
Chris@10
|
180 Tl = BYTW(&(W[TWVL * 40]), Tk);
|
Chris@10
|
181 To = BYTW(&(W[TWVL * 20]), Tn);
|
Chris@10
|
182 TM = BYTW(&(W[TWVL * 6]), TL);
|
Chris@10
|
183 TB = BYTW(&(W[TWVL * 46]), TA);
|
Chris@10
|
184 TH = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
|
Chris@10
|
185 T1s = VFNMS(LDK(KP250000000), T1r, T1g);
|
Chris@10
|
186 T3Q = VADD(T1g, T1r);
|
Chris@10
|
187 TC = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
188 Tw = VSUB(Tj, Tl);
|
Chris@10
|
189 Tm = VADD(Tj, Tl);
|
Chris@10
|
190 Tx = VSUB(Tq, To);
|
Chris@10
|
191 Tr = VADD(To, Tq);
|
Chris@10
|
192 TI = BYTW(&(W[TWVL * 26]), TH);
|
Chris@10
|
193 T2L = VFMA(LDK(KP559016994), T1t, T1s);
|
Chris@10
|
194 T1u = VFNMS(LDK(KP559016994), T1t, T1s);
|
Chris@10
|
195 TD = BYTW(&(W[TWVL * 16]), TC);
|
Chris@10
|
196 TF = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
197 }
|
Chris@10
|
198 {
|
Chris@10
|
199 V Tu, Ty, T2E, TE, TN, TG, Tt, TV, Ts;
|
Chris@10
|
200 TV = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
|
Chris@10
|
201 Ts = VADD(Tm, Tr);
|
Chris@10
|
202 Tu = VSUB(Tm, Tr);
|
Chris@10
|
203 Ty = VFNMS(LDK(KP618033988), Tx, Tw);
|
Chris@10
|
204 T2E = VFMA(LDK(KP618033988), Tw, Tx);
|
Chris@10
|
205 T3j = VFNMS(LDK(KP059835404), T2K, T2L);
|
Chris@10
|
206 T3b = VFMA(LDK(KP066152395), T2L, T2K);
|
Chris@10
|
207 T2R = VFNMS(LDK(KP786782374), T2K, T2L);
|
Chris@10
|
208 T2M = VFMA(LDK(KP869845200), T2L, T2K);
|
Chris@10
|
209 T2f = VFMA(LDK(KP132830569), T1u, T1x);
|
Chris@10
|
210 T27 = VFNMS(LDK(KP120146378), T1x, T1u);
|
Chris@10
|
211 T1y = VFNMS(LDK(KP893101515), T1x, T1u);
|
Chris@10
|
212 T1H = VFMA(LDK(KP987388751), T1u, T1x);
|
Chris@10
|
213 TE = VSUB(TB, TD);
|
Chris@10
|
214 TN = VADD(TD, TB);
|
Chris@10
|
215 TG = BYTW(&(W[TWVL * 36]), TF);
|
Chris@10
|
216 Tt = VFNMS(LDK(KP250000000), Ts, Th);
|
Chris@10
|
217 T3M = VADD(Th, Ts);
|
Chris@10
|
218 TW = BYTW(&(W[TWVL * 2]), TV);
|
Chris@10
|
219 {
|
Chris@10
|
220 V TJ, TO, Tv, T2D, TY, T15, T10, T13, TP;
|
Chris@10
|
221 {
|
Chris@10
|
222 V TX, T14, TZ, T12;
|
Chris@10
|
223 TX = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
224 T14 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
225 TZ = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
|
Chris@10
|
226 T12 = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
|
Chris@10
|
227 TJ = VSUB(TG, TI);
|
Chris@10
|
228 TO = VADD(TI, TG);
|
Chris@10
|
229 Tv = VFMA(LDK(KP559016994), Tu, Tt);
|
Chris@10
|
230 T2D = VFNMS(LDK(KP559016994), Tu, Tt);
|
Chris@10
|
231 TY = BYTW(&(W[TWVL * 12]), TX);
|
Chris@10
|
232 T15 = BYTW(&(W[TWVL * 32]), T14);
|
Chris@10
|
233 T10 = BYTW(&(W[TWVL * 42]), TZ);
|
Chris@10
|
234 T13 = BYTW(&(W[TWVL * 22]), T12);
|
Chris@10
|
235 }
|
Chris@10
|
236 TP = VADD(TN, TO);
|
Chris@10
|
237 TR = VSUB(TN, TO);
|
Chris@10
|
238 TK = VFMA(LDK(KP618033988), TJ, TE);
|
Chris@10
|
239 T2B = VFNMS(LDK(KP618033988), TE, TJ);
|
Chris@10
|
240 T3n = VFMA(LDK(KP578046249), T2D, T2E);
|
Chris@10
|
241 T3e = VFNMS(LDK(KP522847744), T2E, T2D);
|
Chris@10
|
242 T2U = VFNMS(LDK(KP987388751), T2D, T2E);
|
Chris@10
|
243 T2F = VFMA(LDK(KP893101515), T2E, T2D);
|
Chris@10
|
244 T2i = VFNMS(LDK(KP603558818), Ty, Tv);
|
Chris@10
|
245 T2a = VFMA(LDK(KP667278218), Tv, Ty);
|
Chris@10
|
246 Tz = VFNMS(LDK(KP244189809), Ty, Tv);
|
Chris@10
|
247 T1C = VFMA(LDK(KP269969613), Tv, Ty);
|
Chris@10
|
248 T3N = VADD(TM, TP);
|
Chris@10
|
249 TQ = VFMS(LDK(KP250000000), TP, TM);
|
Chris@10
|
250 T11 = VADD(TY, T10);
|
Chris@10
|
251 T1b = VSUB(TY, T10);
|
Chris@10
|
252 T1c = VSUB(T15, T13);
|
Chris@10
|
253 T16 = VADD(T13, T15);
|
Chris@10
|
254 }
|
Chris@10
|
255 }
|
Chris@10
|
256 }
|
Chris@10
|
257 }
|
Chris@10
|
258 }
|
Chris@10
|
259 {
|
Chris@10
|
260 V T2z, Tf, T3W, T3O, T1d, T2H, T3m, T2j, T2b, TT, T1D, T2G, T35, T2V, T2Z;
|
Chris@10
|
261 V T3A, T3g, T2I, T1a, T3R, T3X;
|
Chris@10
|
262 T2z = VFNMS(LDK(KP559016994), Te, Td);
|
Chris@10
|
263 Tf = VFMA(LDK(KP559016994), Te, Td);
|
Chris@10
|
264 {
|
Chris@10
|
265 V TS, T2A, T17, T19;
|
Chris@10
|
266 TS = VFNMS(LDK(KP559016994), TR, TQ);
|
Chris@10
|
267 T2A = VFMA(LDK(KP559016994), TR, TQ);
|
Chris@10
|
268 T3W = VSUB(T3M, T3N);
|
Chris@10
|
269 T3O = VADD(T3M, T3N);
|
Chris@10
|
270 T1d = VFNMS(LDK(KP618033988), T1c, T1b);
|
Chris@10
|
271 T2H = VFMA(LDK(KP618033988), T1b, T1c);
|
Chris@10
|
272 T17 = VADD(T11, T16);
|
Chris@10
|
273 T19 = VSUB(T16, T11);
|
Chris@10
|
274 {
|
Chris@10
|
275 V T3f, T2T, T2C, T18, T3P;
|
Chris@10
|
276 T3m = VFMA(LDK(KP447533225), T2B, T2A);
|
Chris@10
|
277 T3f = VFNMS(LDK(KP494780565), T2A, T2B);
|
Chris@10
|
278 T2T = VFNMS(LDK(KP132830569), T2A, T2B);
|
Chris@10
|
279 T2C = VFMA(LDK(KP120146378), T2B, T2A);
|
Chris@10
|
280 T2j = VFNMS(LDK(KP786782374), TK, TS);
|
Chris@10
|
281 T2b = VFMA(LDK(KP869845200), TS, TK);
|
Chris@10
|
282 TT = VFNMS(LDK(KP667278218), TS, TK);
|
Chris@10
|
283 T1D = VFMA(LDK(KP603558818), TK, TS);
|
Chris@10
|
284 T18 = VFNMS(LDK(KP250000000), T17, TW);
|
Chris@10
|
285 T3P = VADD(TW, T17);
|
Chris@10
|
286 T2G = VFMA(LDK(KP734762448), T2F, T2C);
|
Chris@10
|
287 T35 = VFNMS(LDK(KP734762448), T2F, T2C);
|
Chris@10
|
288 T2V = VFNMS(LDK(KP734762448), T2U, T2T);
|
Chris@10
|
289 T2Z = VFMA(LDK(KP734762448), T2U, T2T);
|
Chris@10
|
290 T3A = VFMA(LDK(KP982009705), T3f, T3e);
|
Chris@10
|
291 T3g = VFNMS(LDK(KP982009705), T3f, T3e);
|
Chris@10
|
292 T2I = VFMA(LDK(KP559016994), T19, T18);
|
Chris@10
|
293 T1a = VFNMS(LDK(KP559016994), T19, T18);
|
Chris@10
|
294 T3R = VADD(T3P, T3Q);
|
Chris@10
|
295 T3X = VSUB(T3P, T3Q);
|
Chris@10
|
296 }
|
Chris@10
|
297 }
|
Chris@10
|
298 {
|
Chris@10
|
299 V T2n, T2t, T1V, T22, T2l, T2d, T1Q, T1I, T2w, T1A, T1F, T2q;
|
Chris@10
|
300 {
|
Chris@10
|
301 V T2k, T1G, T28, T2g, T3K, T3E, T3a, T34, T3x, T3H, T2c, TU, T1T, T1U, T1z;
|
Chris@10
|
302 V T3o, T3t;
|
Chris@10
|
303 T2n = VFNMS(LDK(KP912575812), T2j, T2i);
|
Chris@10
|
304 T2k = VFMA(LDK(KP912575812), T2j, T2i);
|
Chris@10
|
305 T3o = VFNMS(LDK(KP921078979), T3n, T3m);
|
Chris@10
|
306 T3t = VFMA(LDK(KP921078979), T3n, T3m);
|
Chris@10
|
307 {
|
Chris@10
|
308 V T3c, T2Q, T2J, T3k, T1e;
|
Chris@10
|
309 T3c = VFNMS(LDK(KP667278218), T2I, T2H);
|
Chris@10
|
310 T2Q = VFNMS(LDK(KP059835404), T2H, T2I);
|
Chris@10
|
311 T2J = VFMA(LDK(KP066152395), T2I, T2H);
|
Chris@10
|
312 T3k = VFMA(LDK(KP603558818), T2H, T2I);
|
Chris@10
|
313 T1G = VFMA(LDK(KP578046249), T1a, T1d);
|
Chris@10
|
314 T1e = VFNMS(LDK(KP522847744), T1d, T1a);
|
Chris@10
|
315 T28 = VFNMS(LDK(KP494780565), T1a, T1d);
|
Chris@10
|
316 T2g = VFMA(LDK(KP447533225), T1d, T1a);
|
Chris@10
|
317 {
|
Chris@10
|
318 V T3U, T3S, T40, T3Y;
|
Chris@10
|
319 T3U = VSUB(T3O, T3R);
|
Chris@10
|
320 T3S = VADD(T3O, T3R);
|
Chris@10
|
321 T40 = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T3W, T3X));
|
Chris@10
|
322 T3Y = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T3X, T3W));
|
Chris@10
|
323 {
|
Chris@10
|
324 V T3s, T3l, T2N, T36;
|
Chris@10
|
325 T3s = VFNMS(LDK(KP845997307), T3k, T3j);
|
Chris@10
|
326 T3l = VFMA(LDK(KP845997307), T3k, T3j);
|
Chris@10
|
327 T2N = VFNMS(LDK(KP772036680), T2M, T2J);
|
Chris@10
|
328 T36 = VFMA(LDK(KP772036680), T2M, T2J);
|
Chris@10
|
329 {
|
Chris@10
|
330 V T30, T2S, T3d, T3z, T3T;
|
Chris@10
|
331 T30 = VFNMS(LDK(KP772036680), T2R, T2Q);
|
Chris@10
|
332 T2S = VFMA(LDK(KP772036680), T2R, T2Q);
|
Chris@10
|
333 T3d = VFNMS(LDK(KP845997307), T3c, T3b);
|
Chris@10
|
334 T3z = VFMA(LDK(KP845997307), T3c, T3b);
|
Chris@10
|
335 ST(&(x[0]), VADD(T3S, T3L), ms, &(x[0]));
|
Chris@10
|
336 T3T = VFNMS(LDK(KP250000000), T3S, T3L);
|
Chris@10
|
337 {
|
Chris@10
|
338 V T3C, T3p, T2O, T37;
|
Chris@10
|
339 T3C = VFMA(LDK(KP906616052), T3o, T3l);
|
Chris@10
|
340 T3p = VFNMS(LDK(KP906616052), T3o, T3l);
|
Chris@10
|
341 T2O = VFMA(LDK(KP956723877), T2N, T2G);
|
Chris@10
|
342 T37 = VFMA(LDK(KP522616830), T2V, T36);
|
Chris@10
|
343 {
|
Chris@10
|
344 V T31, T2W, T3u, T3h;
|
Chris@10
|
345 T31 = VFNMS(LDK(KP522616830), T2G, T30);
|
Chris@10
|
346 T2W = VFMA(LDK(KP945422727), T2V, T2S);
|
Chris@10
|
347 T3u = VFNMS(LDK(KP923225144), T3g, T3d);
|
Chris@10
|
348 T3h = VFMA(LDK(KP923225144), T3g, T3d);
|
Chris@10
|
349 {
|
Chris@10
|
350 V T3I, T3B, T3V, T3Z;
|
Chris@10
|
351 T3I = VFNMS(LDK(KP669429328), T3z, T3A);
|
Chris@10
|
352 T3B = VFMA(LDK(KP570584518), T3A, T3z);
|
Chris@10
|
353 T3V = VFMA(LDK(KP559016994), T3U, T3T);
|
Chris@10
|
354 T3Z = VFNMS(LDK(KP559016994), T3U, T3T);
|
Chris@10
|
355 {
|
Chris@10
|
356 V T3y, T3q, T2P, T38;
|
Chris@10
|
357 T3y = VFMA(LDK(KP262346850), T3p, T2X);
|
Chris@10
|
358 T3q = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T2X, T3p));
|
Chris@10
|
359 T2P = VFMA(LDK(KP992114701), T2O, T2z);
|
Chris@10
|
360 T38 = VFNMS(LDK(KP690983005), T37, T2S);
|
Chris@10
|
361 {
|
Chris@10
|
362 V T32, T2Y, T3v, T3F;
|
Chris@10
|
363 T32 = VFMA(LDK(KP763932022), T31, T2N);
|
Chris@10
|
364 T2Y = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T2X, T2W));
|
Chris@10
|
365 T3v = VFNMS(LDK(KP997675361), T3u, T3t);
|
Chris@10
|
366 T3F = VFNMS(LDK(KP904508497), T3u, T3s);
|
Chris@10
|
367 {
|
Chris@10
|
368 V T3i, T3r, T3J, T3D;
|
Chris@10
|
369 T3i = VFMA(LDK(KP949179823), T3h, T2z);
|
Chris@10
|
370 T3r = VFNMS(LDK(KP237294955), T3h, T2z);
|
Chris@10
|
371 T3J = VFNMS(LDK(KP669429328), T3C, T3I);
|
Chris@10
|
372 T3D = VFMA(LDK(KP618033988), T3C, T3B);
|
Chris@10
|
373 ST(&(x[WS(rs, 20)]), VFNMSI(T3Y, T3V), ms, &(x[0]));
|
Chris@10
|
374 ST(&(x[WS(rs, 5)]), VFMAI(T3Y, T3V), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
375 ST(&(x[WS(rs, 15)]), VFMAI(T40, T3Z), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
376 ST(&(x[WS(rs, 10)]), VFNMSI(T40, T3Z), ms, &(x[0]));
|
Chris@10
|
377 {
|
Chris@10
|
378 V T39, T33, T3w, T3G;
|
Chris@10
|
379 T39 = VFMA(LDK(KP855719849), T38, T35);
|
Chris@10
|
380 T33 = VFNMS(LDK(KP855719849), T32, T2Z);
|
Chris@10
|
381 ST(&(x[WS(rs, 3)]), VFMAI(T2Y, T2P), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
382 ST(&(x[WS(rs, 22)]), VFNMSI(T2Y, T2P), ms, &(x[0]));
|
Chris@10
|
383 T3w = VFMA(LDK(KP560319534), T3v, T3s);
|
Chris@10
|
384 T3G = VFNMS(LDK(KP681693190), T3F, T3t);
|
Chris@10
|
385 ST(&(x[WS(rs, 2)]), VFMAI(T3q, T3i), ms, &(x[0]));
|
Chris@10
|
386 ST(&(x[WS(rs, 23)]), VFNMSI(T3q, T3i), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
387 T3K = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T3J, T3y));
|
Chris@10
|
388 T3E = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T3D, T3y));
|
Chris@10
|
389 T3a = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T39, T2X));
|
Chris@10
|
390 T34 = VFMA(LDK(KP897376177), T33, T2z);
|
Chris@10
|
391 T3x = VFNMS(LDK(KP949179823), T3w, T3r);
|
Chris@10
|
392 T3H = VFNMS(LDK(KP860541664), T3G, T3r);
|
Chris@10
|
393 T2t = VFNMS(LDK(KP912575812), T2b, T2a);
|
Chris@10
|
394 T2c = VFMA(LDK(KP912575812), T2b, T2a);
|
Chris@10
|
395 TU = VFMA(LDK(KP829049696), TT, Tz);
|
Chris@10
|
396 T1T = VFNMS(LDK(KP829049696), TT, Tz);
|
Chris@10
|
397 T1U = VFNMS(LDK(KP831864738), T1y, T1e);
|
Chris@10
|
398 T1z = VFMA(LDK(KP831864738), T1y, T1e);
|
Chris@10
|
399 }
|
Chris@10
|
400 }
|
Chris@10
|
401 }
|
Chris@10
|
402 }
|
Chris@10
|
403 }
|
Chris@10
|
404 }
|
Chris@10
|
405 }
|
Chris@10
|
406 }
|
Chris@10
|
407 }
|
Chris@10
|
408 }
|
Chris@10
|
409 }
|
Chris@10
|
410 {
|
Chris@10
|
411 V T2o, T2h, T29, T2u, T2v, T2p;
|
Chris@10
|
412 T2o = VFNMS(LDK(KP958953096), T2g, T2f);
|
Chris@10
|
413 T2h = VFMA(LDK(KP958953096), T2g, T2f);
|
Chris@10
|
414 ST(&(x[WS(rs, 17)]), VFNMSI(T3a, T34), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
415 ST(&(x[WS(rs, 8)]), VFMAI(T3a, T34), ms, &(x[0]));
|
Chris@10
|
416 ST(&(x[WS(rs, 13)]), VFMAI(T3E, T3x), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
417 ST(&(x[WS(rs, 12)]), VFNMSI(T3E, T3x), ms, &(x[0]));
|
Chris@10
|
418 ST(&(x[WS(rs, 7)]), VFNMSI(T3K, T3H), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
419 ST(&(x[WS(rs, 18)]), VFMAI(T3K, T3H), ms, &(x[0]));
|
Chris@10
|
420 T1V = VFMA(LDK(KP559154169), T1U, T1T);
|
Chris@10
|
421 T22 = VFNMS(LDK(KP683113946), T1T, T1U);
|
Chris@10
|
422 T29 = VFNMS(LDK(KP867381224), T28, T27);
|
Chris@10
|
423 T2u = VFMA(LDK(KP867381224), T28, T27);
|
Chris@10
|
424 T2l = VFMA(LDK(KP894834959), T2k, T2h);
|
Chris@10
|
425 T2v = VFMA(LDK(KP447417479), T2k, T2u);
|
Chris@10
|
426 T2d = VFNMS(LDK(KP809385824), T2c, T29);
|
Chris@10
|
427 T2p = VFMA(LDK(KP447417479), T2c, T2o);
|
Chris@10
|
428 T1Q = VFMA(LDK(KP831864738), T1H, T1G);
|
Chris@10
|
429 T1I = VFNMS(LDK(KP831864738), T1H, T1G);
|
Chris@10
|
430 T2w = VFNMS(LDK(KP763932022), T2v, T2h);
|
Chris@10
|
431 T1A = VFMA(LDK(KP904730450), T1z, TU);
|
Chris@10
|
432 T1F = VFNMS(LDK(KP904730450), T1z, TU);
|
Chris@10
|
433 T2q = VFMA(LDK(KP690983005), T2p, T29);
|
Chris@10
|
434 }
|
Chris@10
|
435 }
|
Chris@10
|
436 {
|
Chris@10
|
437 V T2e, T1E, T1P, T2m;
|
Chris@10
|
438 T2e = VFNMS(LDK(KP992114701), T2d, Tf);
|
Chris@10
|
439 T1E = VFMA(LDK(KP916574801), T1D, T1C);
|
Chris@10
|
440 T1P = VFNMS(LDK(KP916574801), T1D, T1C);
|
Chris@10
|
441 T2m = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2l, T1O));
|
Chris@10
|
442 {
|
Chris@10
|
443 V T1J, T2r, T1R, T1W, T1Z, T2x;
|
Chris@10
|
444 T2x = VFNMS(LDK(KP999544308), T2w, T2t);
|
Chris@10
|
445 T1J = VFNMS(LDK(KP904730450), T1I, T1F);
|
Chris@10
|
446 T25 = VFMA(LDK(KP968583161), T1A, Tf);
|
Chris@10
|
447 T1B = VFNMS(LDK(KP242145790), T1A, Tf);
|
Chris@10
|
448 T2r = VFNMS(LDK(KP999544308), T2q, T2n);
|
Chris@10
|
449 T1R = VFMA(LDK(KP904730450), T1Q, T1P);
|
Chris@10
|
450 T1W = VFNMS(LDK(KP904730450), T1Q, T1P);
|
Chris@10
|
451 T1Z = VADD(T1E, T1F);
|
Chris@10
|
452 ST(&(x[WS(rs, 21)]), VFMAI(T2m, T2e), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
453 ST(&(x[WS(rs, 4)]), VFNMSI(T2m, T2e), ms, &(x[0]));
|
Chris@10
|
454 T2y = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T2x, T1O));
|
Chris@10
|
455 T1K = VFNMS(LDK(KP618033988), T1J, T1E);
|
Chris@10
|
456 T2s = VFNMS(LDK(KP803003575), T2r, Tf);
|
Chris@10
|
457 T23 = VFMA(LDK(KP617882369), T1W, T22);
|
Chris@10
|
458 T1S = VFNMS(LDK(KP242145790), T1R, T1O);
|
Chris@10
|
459 T26 = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1R, T1O));
|
Chris@10
|
460 T20 = VFNMS(LDK(KP683113946), T1Z, T1I);
|
Chris@10
|
461 T1X = VFMA(LDK(KP559016994), T1W, T1V);
|
Chris@10
|
462 }
|
Chris@10
|
463 }
|
Chris@10
|
464 }
|
Chris@10
|
465 }
|
Chris@10
|
466 }
|
Chris@10
|
467 {
|
Chris@10
|
468 V T1L, T24, T21, T1Y;
|
Chris@10
|
469 T1L = VFNMS(LDK(KP876091699), T1K, T1B);
|
Chris@10
|
470 ST(&(x[WS(rs, 16)]), VFMAI(T2y, T2s), ms, &(x[0]));
|
Chris@10
|
471 ST(&(x[WS(rs, 9)]), VFNMSI(T2y, T2s), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
472 T24 = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T23, T1S));
|
Chris@10
|
473 ST(&(x[WS(rs, 24)]), VFNMSI(T26, T25), ms, &(x[0]));
|
Chris@10
|
474 ST(&(x[WS(rs, 1)]), VFMAI(T26, T25), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
475 T21 = VFMA(LDK(KP792626838), T20, T1B);
|
Chris@10
|
476 T1Y = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1X, T1S));
|
Chris@10
|
477 ST(&(x[WS(rs, 11)]), VFMAI(T24, T21), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
478 ST(&(x[WS(rs, 14)]), VFNMSI(T24, T21), ms, &(x[0]));
|
Chris@10
|
479 ST(&(x[WS(rs, 19)]), VFNMSI(T1Y, T1L), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
480 ST(&(x[WS(rs, 6)]), VFMAI(T1Y, T1L), ms, &(x[0]));
|
Chris@10
|
481 }
|
Chris@10
|
482 }
|
Chris@10
|
483 }
|
Chris@10
|
484 VLEAVE();
|
Chris@10
|
485 }
|
Chris@10
|
486
|
Chris@10
|
487 static const tw_instr twinstr[] = {
|
Chris@10
|
488 VTW(0, 1),
|
Chris@10
|
489 VTW(0, 2),
|
Chris@10
|
490 VTW(0, 3),
|
Chris@10
|
491 VTW(0, 4),
|
Chris@10
|
492 VTW(0, 5),
|
Chris@10
|
493 VTW(0, 6),
|
Chris@10
|
494 VTW(0, 7),
|
Chris@10
|
495 VTW(0, 8),
|
Chris@10
|
496 VTW(0, 9),
|
Chris@10
|
497 VTW(0, 10),
|
Chris@10
|
498 VTW(0, 11),
|
Chris@10
|
499 VTW(0, 12),
|
Chris@10
|
500 VTW(0, 13),
|
Chris@10
|
501 VTW(0, 14),
|
Chris@10
|
502 VTW(0, 15),
|
Chris@10
|
503 VTW(0, 16),
|
Chris@10
|
504 VTW(0, 17),
|
Chris@10
|
505 VTW(0, 18),
|
Chris@10
|
506 VTW(0, 19),
|
Chris@10
|
507 VTW(0, 20),
|
Chris@10
|
508 VTW(0, 21),
|
Chris@10
|
509 VTW(0, 22),
|
Chris@10
|
510 VTW(0, 23),
|
Chris@10
|
511 VTW(0, 24),
|
Chris@10
|
512 {TW_NEXT, VL, 0}
|
Chris@10
|
513 };
|
Chris@10
|
514
|
Chris@10
|
515 static const ct_desc desc = { 25, XSIMD_STRING("t1bv_25"), twinstr, &GENUS, {67, 60, 181, 0}, 0, 0, 0 };
|
Chris@10
|
516
|
Chris@10
|
517 void XSIMD(codelet_t1bv_25) (planner *p) {
|
Chris@10
|
518 X(kdft_dit_register) (p, t1bv_25, &desc);
|
Chris@10
|
519 }
|
Chris@10
|
520 #else /* HAVE_FMA */
|
Chris@10
|
521
|
Chris@10
|
522 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 25 -name t1bv_25 -include t1b.h -sign 1 */
|
Chris@10
|
523
|
Chris@10
|
524 /*
|
Chris@10
|
525 * This function contains 248 FP additions, 188 FP multiplications,
|
Chris@10
|
526 * (or, 171 additions, 111 multiplications, 77 fused multiply/add),
|
Chris@10
|
527 * 100 stack variables, 40 constants, and 50 memory accesses
|
Chris@10
|
528 */
|
Chris@10
|
529 #include "t1b.h"
|
Chris@10
|
530
|
Chris@10
|
531 static void t1bv_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
532 {
|
Chris@10
|
533 DVK(KP497379774, +0.497379774329709576484567492012895936835134813);
|
Chris@10
|
534 DVK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
535 DVK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
536 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
|
Chris@10
|
537 DVK(KP809016994, +0.809016994374947424102293417182819058860154590);
|
Chris@10
|
538 DVK(KP309016994, +0.309016994374947424102293417182819058860154590);
|
Chris@10
|
539 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483);
|
Chris@10
|
540 DVK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
541 DVK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
542 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
|
Chris@10
|
543 DVK(KP963507348, +0.963507348203430549974383005744259307057084020);
|
Chris@10
|
544 DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
545 DVK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
546 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691);
|
Chris@10
|
547 DVK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
548 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
|
Chris@10
|
549 DVK(KP851558583, +0.851558583130145297725004891488503407959946084);
|
Chris@10
|
550 DVK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
551 DVK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
552 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418);
|
Chris@10
|
553 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328);
|
Chris@10
|
554 DVK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
555 DVK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
556 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344);
|
Chris@10
|
557 DVK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
558 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
|
Chris@10
|
559 DVK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
560 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596);
|
Chris@10
|
561 DVK(KP125581039, +0.125581039058626752152356449131262266244969664);
|
Chris@10
|
562 DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
563 DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
564 DVK(KP250666467, +0.250666467128608490746237519633017587885836494);
|
Chris@10
|
565 DVK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
566 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465);
|
Chris@10
|
567 DVK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
Chris@10
|
568 DVK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
Chris@10
|
569 DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
570 DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
571 DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
572 DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
573 {
|
Chris@10
|
574 INT m;
|
Chris@10
|
575 R *x;
|
Chris@10
|
576 x = ii;
|
Chris@10
|
577 for (m = mb, W = W + (mb * ((TWVL / VL) * 48)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 48), MAKE_VOLATILE_STRIDE(25, rs)) {
|
Chris@10
|
578 V T1A, T1z, T1R, T1S, T1B, T1C, T1Q, T2L, T1l, T2v, T1i, T3e, T2u, Tb, T2i;
|
Chris@10
|
579 V Tj, T3b, T2h, Tv, T2k, TD, T3a, T2l, T11, T2s, TY, T3d, T2r;
|
Chris@10
|
580 {
|
Chris@10
|
581 V T1v, T1x, T1y, T1q, T1s, T1t, T1P;
|
Chris@10
|
582 T1A = LD(&(x[0]), ms, &(x[0]));
|
Chris@10
|
583 {
|
Chris@10
|
584 V T1u, T1w, T1p, T1r;
|
Chris@10
|
585 T1u = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
|
Chris@10
|
586 T1v = BYTW(&(W[TWVL * 18]), T1u);
|
Chris@10
|
587 T1w = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
588 T1x = BYTW(&(W[TWVL * 28]), T1w);
|
Chris@10
|
589 T1y = VADD(T1v, T1x);
|
Chris@10
|
590 T1p = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
591 T1q = BYTW(&(W[TWVL * 8]), T1p);
|
Chris@10
|
592 T1r = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
|
Chris@10
|
593 T1s = BYTW(&(W[TWVL * 38]), T1r);
|
Chris@10
|
594 T1t = VADD(T1q, T1s);
|
Chris@10
|
595 }
|
Chris@10
|
596 T1z = VMUL(LDK(KP559016994), VSUB(T1t, T1y));
|
Chris@10
|
597 T1R = VSUB(T1v, T1x);
|
Chris@10
|
598 T1S = VMUL(LDK(KP587785252), T1R);
|
Chris@10
|
599 T1B = VADD(T1t, T1y);
|
Chris@10
|
600 T1C = VFNMS(LDK(KP250000000), T1B, T1A);
|
Chris@10
|
601 T1P = VSUB(T1q, T1s);
|
Chris@10
|
602 T1Q = VMUL(LDK(KP951056516), T1P);
|
Chris@10
|
603 T2L = VMUL(LDK(KP587785252), T1P);
|
Chris@10
|
604 }
|
Chris@10
|
605 {
|
Chris@10
|
606 V T1f, T19, T1b, T1c, T14, T16, T17, T1e;
|
Chris@10
|
607 T1e = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
608 T1f = BYTW(&(W[TWVL * 4]), T1e);
|
Chris@10
|
609 {
|
Chris@10
|
610 V T18, T1a, T13, T15;
|
Chris@10
|
611 T18 = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
612 T19 = BYTW(&(W[TWVL * 24]), T18);
|
Chris@10
|
613 T1a = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
|
Chris@10
|
614 T1b = BYTW(&(W[TWVL * 34]), T1a);
|
Chris@10
|
615 T1c = VADD(T19, T1b);
|
Chris@10
|
616 T13 = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
|
Chris@10
|
617 T14 = BYTW(&(W[TWVL * 14]), T13);
|
Chris@10
|
618 T15 = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
619 T16 = BYTW(&(W[TWVL * 44]), T15);
|
Chris@10
|
620 T17 = VADD(T14, T16);
|
Chris@10
|
621 }
|
Chris@10
|
622 {
|
Chris@10
|
623 V T1j, T1k, T1d, T1g, T1h;
|
Chris@10
|
624 T1j = VSUB(T14, T16);
|
Chris@10
|
625 T1k = VSUB(T19, T1b);
|
Chris@10
|
626 T1l = VFMA(LDK(KP475528258), T1j, VMUL(LDK(KP293892626), T1k));
|
Chris@10
|
627 T2v = VFNMS(LDK(KP475528258), T1k, VMUL(LDK(KP293892626), T1j));
|
Chris@10
|
628 T1d = VMUL(LDK(KP559016994), VSUB(T17, T1c));
|
Chris@10
|
629 T1g = VADD(T17, T1c);
|
Chris@10
|
630 T1h = VFNMS(LDK(KP250000000), T1g, T1f);
|
Chris@10
|
631 T1i = VADD(T1d, T1h);
|
Chris@10
|
632 T3e = VADD(T1f, T1g);
|
Chris@10
|
633 T2u = VSUB(T1h, T1d);
|
Chris@10
|
634 }
|
Chris@10
|
635 }
|
Chris@10
|
636 {
|
Chris@10
|
637 V Tg, T7, T9, Td, T2, T4, Tc, Tf;
|
Chris@10
|
638 Tf = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
|
Chris@10
|
639 Tg = BYTW(&(W[TWVL * 6]), Tf);
|
Chris@10
|
640 {
|
Chris@10
|
641 V T6, T8, T1, T3;
|
Chris@10
|
642 T6 = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
|
Chris@10
|
643 T7 = BYTW(&(W[TWVL * 26]), T6);
|
Chris@10
|
644 T8 = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
645 T9 = BYTW(&(W[TWVL * 36]), T8);
|
Chris@10
|
646 Td = VADD(T7, T9);
|
Chris@10
|
647 T1 = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
648 T2 = BYTW(&(W[TWVL * 16]), T1);
|
Chris@10
|
649 T3 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
|
Chris@10
|
650 T4 = BYTW(&(W[TWVL * 46]), T3);
|
Chris@10
|
651 Tc = VADD(T2, T4);
|
Chris@10
|
652 }
|
Chris@10
|
653 {
|
Chris@10
|
654 V T5, Ta, Te, Th, Ti;
|
Chris@10
|
655 T5 = VSUB(T2, T4);
|
Chris@10
|
656 Ta = VSUB(T7, T9);
|
Chris@10
|
657 Tb = VFMA(LDK(KP475528258), T5, VMUL(LDK(KP293892626), Ta));
|
Chris@10
|
658 T2i = VFNMS(LDK(KP475528258), Ta, VMUL(LDK(KP293892626), T5));
|
Chris@10
|
659 Te = VMUL(LDK(KP559016994), VSUB(Tc, Td));
|
Chris@10
|
660 Th = VADD(Tc, Td);
|
Chris@10
|
661 Ti = VFNMS(LDK(KP250000000), Th, Tg);
|
Chris@10
|
662 Tj = VADD(Te, Ti);
|
Chris@10
|
663 T3b = VADD(Tg, Th);
|
Chris@10
|
664 T2h = VSUB(Ti, Te);
|
Chris@10
|
665 }
|
Chris@10
|
666 }
|
Chris@10
|
667 {
|
Chris@10
|
668 V TA, Tr, Tt, Tx, Tm, To, Tw, Tz;
|
Chris@10
|
669 Tz = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
670 TA = BYTW(&(W[0]), Tz);
|
Chris@10
|
671 {
|
Chris@10
|
672 V Tq, Ts, Tl, Tn;
|
Chris@10
|
673 Tq = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
674 Tr = BYTW(&(W[TWVL * 20]), Tq);
|
Chris@10
|
675 Ts = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
|
Chris@10
|
676 Tt = BYTW(&(W[TWVL * 30]), Ts);
|
Chris@10
|
677 Tx = VADD(Tr, Tt);
|
Chris@10
|
678 Tl = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
|
Chris@10
|
679 Tm = BYTW(&(W[TWVL * 10]), Tl);
|
Chris@10
|
680 Tn = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
681 To = BYTW(&(W[TWVL * 40]), Tn);
|
Chris@10
|
682 Tw = VADD(Tm, To);
|
Chris@10
|
683 }
|
Chris@10
|
684 {
|
Chris@10
|
685 V Tp, Tu, Ty, TB, TC;
|
Chris@10
|
686 Tp = VSUB(Tm, To);
|
Chris@10
|
687 Tu = VSUB(Tr, Tt);
|
Chris@10
|
688 Tv = VFMA(LDK(KP475528258), Tp, VMUL(LDK(KP293892626), Tu));
|
Chris@10
|
689 T2k = VFNMS(LDK(KP475528258), Tu, VMUL(LDK(KP293892626), Tp));
|
Chris@10
|
690 Ty = VMUL(LDK(KP559016994), VSUB(Tw, Tx));
|
Chris@10
|
691 TB = VADD(Tw, Tx);
|
Chris@10
|
692 TC = VFNMS(LDK(KP250000000), TB, TA);
|
Chris@10
|
693 TD = VADD(Ty, TC);
|
Chris@10
|
694 T3a = VADD(TA, TB);
|
Chris@10
|
695 T2l = VSUB(TC, Ty);
|
Chris@10
|
696 }
|
Chris@10
|
697 }
|
Chris@10
|
698 {
|
Chris@10
|
699 V TV, TP, TR, TS, TK, TM, TN, TU;
|
Chris@10
|
700 TU = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
|
Chris@10
|
701 TV = BYTW(&(W[TWVL * 2]), TU);
|
Chris@10
|
702 {
|
Chris@10
|
703 V TO, TQ, TJ, TL;
|
Chris@10
|
704 TO = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
|
Chris@10
|
705 TP = BYTW(&(W[TWVL * 22]), TO);
|
Chris@10
|
706 TQ = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
707 TR = BYTW(&(W[TWVL * 32]), TQ);
|
Chris@10
|
708 TS = VADD(TP, TR);
|
Chris@10
|
709 TJ = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
710 TK = BYTW(&(W[TWVL * 12]), TJ);
|
Chris@10
|
711 TL = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
|
Chris@10
|
712 TM = BYTW(&(W[TWVL * 42]), TL);
|
Chris@10
|
713 TN = VADD(TK, TM);
|
Chris@10
|
714 }
|
Chris@10
|
715 {
|
Chris@10
|
716 V TZ, T10, TT, TW, TX;
|
Chris@10
|
717 TZ = VSUB(TK, TM);
|
Chris@10
|
718 T10 = VSUB(TP, TR);
|
Chris@10
|
719 T11 = VFMA(LDK(KP475528258), TZ, VMUL(LDK(KP293892626), T10));
|
Chris@10
|
720 T2s = VFNMS(LDK(KP475528258), T10, VMUL(LDK(KP293892626), TZ));
|
Chris@10
|
721 TT = VMUL(LDK(KP559016994), VSUB(TN, TS));
|
Chris@10
|
722 TW = VADD(TN, TS);
|
Chris@10
|
723 TX = VFNMS(LDK(KP250000000), TW, TV);
|
Chris@10
|
724 TY = VADD(TT, TX);
|
Chris@10
|
725 T3d = VADD(TV, TW);
|
Chris@10
|
726 T2r = VSUB(TX, TT);
|
Chris@10
|
727 }
|
Chris@10
|
728 }
|
Chris@10
|
729 {
|
Chris@10
|
730 V T3g, T3o, T3k, T3l, T3j, T3m, T3p, T3n;
|
Chris@10
|
731 {
|
Chris@10
|
732 V T3c, T3f, T3h, T3i;
|
Chris@10
|
733 T3c = VSUB(T3a, T3b);
|
Chris@10
|
734 T3f = VSUB(T3d, T3e);
|
Chris@10
|
735 T3g = VBYI(VFMA(LDK(KP951056516), T3c, VMUL(LDK(KP587785252), T3f)));
|
Chris@10
|
736 T3o = VBYI(VFNMS(LDK(KP951056516), T3f, VMUL(LDK(KP587785252), T3c)));
|
Chris@10
|
737 T3k = VADD(T1A, T1B);
|
Chris@10
|
738 T3h = VADD(T3a, T3b);
|
Chris@10
|
739 T3i = VADD(T3d, T3e);
|
Chris@10
|
740 T3l = VADD(T3h, T3i);
|
Chris@10
|
741 T3j = VMUL(LDK(KP559016994), VSUB(T3h, T3i));
|
Chris@10
|
742 T3m = VFNMS(LDK(KP250000000), T3l, T3k);
|
Chris@10
|
743 }
|
Chris@10
|
744 ST(&(x[0]), VADD(T3k, T3l), ms, &(x[0]));
|
Chris@10
|
745 T3p = VSUB(T3m, T3j);
|
Chris@10
|
746 ST(&(x[WS(rs, 10)]), VADD(T3o, T3p), ms, &(x[0]));
|
Chris@10
|
747 ST(&(x[WS(rs, 15)]), VSUB(T3p, T3o), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
748 T3n = VADD(T3j, T3m);
|
Chris@10
|
749 ST(&(x[WS(rs, 5)]), VADD(T3g, T3n), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
750 ST(&(x[WS(rs, 20)]), VSUB(T3n, T3g), ms, &(x[0]));
|
Chris@10
|
751 }
|
Chris@10
|
752 {
|
Chris@10
|
753 V T2z, T2M, T2U, T2V, T2W, T34, T35, T36, T2X, T2Y, T2Z, T31, T32, T33, T2n;
|
Chris@10
|
754 V T2N, T2E, T2K, T2y, T2H, T2A, T2G, T38, T39;
|
Chris@10
|
755 T2z = VSUB(T1C, T1z);
|
Chris@10
|
756 T2M = VFNMS(LDK(KP951056516), T1R, T2L);
|
Chris@10
|
757 T2U = VFMA(LDK(KP1_369094211), T2k, VMUL(LDK(KP728968627), T2l));
|
Chris@10
|
758 T2V = VFNMS(LDK(KP992114701), T2h, VMUL(LDK(KP250666467), T2i));
|
Chris@10
|
759 T2W = VADD(T2U, T2V);
|
Chris@10
|
760 T34 = VFNMS(LDK(KP125581039), T2s, VMUL(LDK(KP998026728), T2r));
|
Chris@10
|
761 T35 = VFMA(LDK(KP1_274847979), T2v, VMUL(LDK(KP770513242), T2u));
|
Chris@10
|
762 T36 = VADD(T34, T35);
|
Chris@10
|
763 T2X = VFMA(LDK(KP1_996053456), T2s, VMUL(LDK(KP062790519), T2r));
|
Chris@10
|
764 T2Y = VFNMS(LDK(KP637423989), T2u, VMUL(LDK(KP1_541026485), T2v));
|
Chris@10
|
765 T2Z = VADD(T2X, T2Y);
|
Chris@10
|
766 T31 = VFNMS(LDK(KP1_457937254), T2k, VMUL(LDK(KP684547105), T2l));
|
Chris@10
|
767 T32 = VFMA(LDK(KP1_984229402), T2i, VMUL(LDK(KP125333233), T2h));
|
Chris@10
|
768 T33 = VADD(T31, T32);
|
Chris@10
|
769 {
|
Chris@10
|
770 V T2j, T2m, T2I, T2C, T2D, T2J;
|
Chris@10
|
771 T2j = VFNMS(LDK(KP851558583), T2i, VMUL(LDK(KP904827052), T2h));
|
Chris@10
|
772 T2m = VFMA(LDK(KP1_752613360), T2k, VMUL(LDK(KP481753674), T2l));
|
Chris@10
|
773 T2I = VADD(T2m, T2j);
|
Chris@10
|
774 T2C = VFMA(LDK(KP1_071653589), T2s, VMUL(LDK(KP844327925), T2r));
|
Chris@10
|
775 T2D = VFMA(LDK(KP125581039), T2v, VMUL(LDK(KP998026728), T2u));
|
Chris@10
|
776 T2J = VADD(T2C, T2D);
|
Chris@10
|
777 T2n = VSUB(T2j, T2m);
|
Chris@10
|
778 T2N = VADD(T2I, T2J);
|
Chris@10
|
779 T2E = VSUB(T2C, T2D);
|
Chris@10
|
780 T2K = VMUL(LDK(KP559016994), VSUB(T2I, T2J));
|
Chris@10
|
781 }
|
Chris@10
|
782 {
|
Chris@10
|
783 V T2o, T2p, T2q, T2t, T2w, T2x;
|
Chris@10
|
784 T2o = VFNMS(LDK(KP963507348), T2k, VMUL(LDK(KP876306680), T2l));
|
Chris@10
|
785 T2p = VFMA(LDK(KP1_809654104), T2i, VMUL(LDK(KP425779291), T2h));
|
Chris@10
|
786 T2q = VSUB(T2o, T2p);
|
Chris@10
|
787 T2t = VFNMS(LDK(KP1_688655851), T2s, VMUL(LDK(KP535826794), T2r));
|
Chris@10
|
788 T2w = VFNMS(LDK(KP1_996053456), T2v, VMUL(LDK(KP062790519), T2u));
|
Chris@10
|
789 T2x = VADD(T2t, T2w);
|
Chris@10
|
790 T2y = VMUL(LDK(KP559016994), VSUB(T2q, T2x));
|
Chris@10
|
791 T2H = VSUB(T2t, T2w);
|
Chris@10
|
792 T2A = VADD(T2q, T2x);
|
Chris@10
|
793 T2G = VADD(T2o, T2p);
|
Chris@10
|
794 }
|
Chris@10
|
795 {
|
Chris@10
|
796 V T2S, T2T, T30, T37;
|
Chris@10
|
797 T2S = VADD(T2z, T2A);
|
Chris@10
|
798 T2T = VBYI(VADD(T2M, T2N));
|
Chris@10
|
799 ST(&(x[WS(rs, 23)]), VSUB(T2S, T2T), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
800 ST(&(x[WS(rs, 2)]), VADD(T2S, T2T), ms, &(x[0]));
|
Chris@10
|
801 T30 = VADD(T2z, VADD(T2W, T2Z));
|
Chris@10
|
802 T37 = VBYI(VSUB(VADD(T33, T36), T2M));
|
Chris@10
|
803 ST(&(x[WS(rs, 22)]), VSUB(T30, T37), ms, &(x[0]));
|
Chris@10
|
804 ST(&(x[WS(rs, 3)]), VADD(T30, T37), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
805 }
|
Chris@10
|
806 T38 = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T2U, T2V), VFMA(LDK(KP309016994), T33, VFNMS(LDK(KP809016994), T36, VMUL(LDK(KP587785252), VSUB(T2X, T2Y))))), T2M));
|
Chris@10
|
807 T39 = VFMA(LDK(KP309016994), T2W, VFMA(LDK(KP951056516), VSUB(T32, T31), VFMA(LDK(KP587785252), VSUB(T35, T34), VFNMS(LDK(KP809016994), T2Z, T2z))));
|
Chris@10
|
808 ST(&(x[WS(rs, 8)]), VADD(T38, T39), ms, &(x[0]));
|
Chris@10
|
809 ST(&(x[WS(rs, 17)]), VSUB(T39, T38), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
810 {
|
Chris@10
|
811 V T2F, T2Q, T2P, T2R, T2B, T2O;
|
Chris@10
|
812 T2B = VFNMS(LDK(KP250000000), T2A, T2z);
|
Chris@10
|
813 T2F = VFMA(LDK(KP951056516), T2n, VADD(T2y, VFNMS(LDK(KP587785252), T2E, T2B)));
|
Chris@10
|
814 T2Q = VFMA(LDK(KP587785252), T2n, VFMA(LDK(KP951056516), T2E, VSUB(T2B, T2y)));
|
Chris@10
|
815 T2O = VFNMS(LDK(KP250000000), T2N, T2M);
|
Chris@10
|
816 T2P = VBYI(VADD(VFMA(LDK(KP951056516), T2G, VMUL(LDK(KP587785252), T2H)), VADD(T2K, T2O)));
|
Chris@10
|
817 T2R = VBYI(VADD(VFNMS(LDK(KP951056516), T2H, VMUL(LDK(KP587785252), T2G)), VSUB(T2O, T2K)));
|
Chris@10
|
818 ST(&(x[WS(rs, 18)]), VSUB(T2F, T2P), ms, &(x[0]));
|
Chris@10
|
819 ST(&(x[WS(rs, 12)]), VADD(T2Q, T2R), ms, &(x[0]));
|
Chris@10
|
820 ST(&(x[WS(rs, 7)]), VADD(T2F, T2P), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
821 ST(&(x[WS(rs, 13)]), VSUB(T2Q, T2R), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
822 }
|
Chris@10
|
823 }
|
Chris@10
|
824 {
|
Chris@10
|
825 V T1D, T1T, T21, T22, T23, T2b, T2c, T2d, T24, T25, T26, T28, T29, T2a, TF;
|
Chris@10
|
826 V T1U, T1I, T1O, T1o, T1L, T1E, T1K, T2f, T2g;
|
Chris@10
|
827 T1D = VADD(T1z, T1C);
|
Chris@10
|
828 T1T = VADD(T1Q, T1S);
|
Chris@10
|
829 T21 = VFMA(LDK(KP1_688655851), Tv, VMUL(LDK(KP535826794), TD));
|
Chris@10
|
830 T22 = VFMA(LDK(KP1_541026485), Tb, VMUL(LDK(KP637423989), Tj));
|
Chris@10
|
831 T23 = VSUB(T21, T22);
|
Chris@10
|
832 T2b = VFMA(LDK(KP851558583), T11, VMUL(LDK(KP904827052), TY));
|
Chris@10
|
833 T2c = VFMA(LDK(KP1_984229402), T1l, VMUL(LDK(KP125333233), T1i));
|
Chris@10
|
834 T2d = VADD(T2b, T2c);
|
Chris@10
|
835 T24 = VFNMS(LDK(KP425779291), TY, VMUL(LDK(KP1_809654104), T11));
|
Chris@10
|
836 T25 = VFNMS(LDK(KP992114701), T1i, VMUL(LDK(KP250666467), T1l));
|
Chris@10
|
837 T26 = VADD(T24, T25);
|
Chris@10
|
838 T28 = VFNMS(LDK(KP1_071653589), Tv, VMUL(LDK(KP844327925), TD));
|
Chris@10
|
839 T29 = VFNMS(LDK(KP770513242), Tj, VMUL(LDK(KP1_274847979), Tb));
|
Chris@10
|
840 T2a = VADD(T28, T29);
|
Chris@10
|
841 {
|
Chris@10
|
842 V Tk, TE, T1M, T1G, T1H, T1N;
|
Chris@10
|
843 Tk = VFMA(LDK(KP1_071653589), Tb, VMUL(LDK(KP844327925), Tj));
|
Chris@10
|
844 TE = VFMA(LDK(KP1_937166322), Tv, VMUL(LDK(KP248689887), TD));
|
Chris@10
|
845 T1M = VADD(TE, Tk);
|
Chris@10
|
846 T1G = VFMA(LDK(KP1_752613360), T11, VMUL(LDK(KP481753674), TY));
|
Chris@10
|
847 T1H = VFMA(LDK(KP1_457937254), T1l, VMUL(LDK(KP684547105), T1i));
|
Chris@10
|
848 T1N = VADD(T1G, T1H);
|
Chris@10
|
849 TF = VSUB(Tk, TE);
|
Chris@10
|
850 T1U = VADD(T1M, T1N);
|
Chris@10
|
851 T1I = VSUB(T1G, T1H);
|
Chris@10
|
852 T1O = VMUL(LDK(KP559016994), VSUB(T1M, T1N));
|
Chris@10
|
853 }
|
Chris@10
|
854 {
|
Chris@10
|
855 V TG, TH, TI, T12, T1m, T1n;
|
Chris@10
|
856 TG = VFNMS(LDK(KP497379774), Tv, VMUL(LDK(KP968583161), TD));
|
Chris@10
|
857 TH = VFNMS(LDK(KP1_688655851), Tb, VMUL(LDK(KP535826794), Tj));
|
Chris@10
|
858 TI = VADD(TG, TH);
|
Chris@10
|
859 T12 = VFNMS(LDK(KP963507348), T11, VMUL(LDK(KP876306680), TY));
|
Chris@10
|
860 T1m = VFNMS(LDK(KP1_369094211), T1l, VMUL(LDK(KP728968627), T1i));
|
Chris@10
|
861 T1n = VADD(T12, T1m);
|
Chris@10
|
862 T1o = VMUL(LDK(KP559016994), VSUB(TI, T1n));
|
Chris@10
|
863 T1L = VSUB(T12, T1m);
|
Chris@10
|
864 T1E = VADD(TI, T1n);
|
Chris@10
|
865 T1K = VSUB(TG, TH);
|
Chris@10
|
866 }
|
Chris@10
|
867 {
|
Chris@10
|
868 V T1Z, T20, T27, T2e;
|
Chris@10
|
869 T1Z = VADD(T1D, T1E);
|
Chris@10
|
870 T20 = VBYI(VADD(T1T, T1U));
|
Chris@10
|
871 ST(&(x[WS(rs, 24)]), VSUB(T1Z, T20), ms, &(x[0]));
|
Chris@10
|
872 ST(&(x[WS(rs, 1)]), VADD(T1Z, T20), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
873 T27 = VADD(T1D, VADD(T23, T26));
|
Chris@10
|
874 T2e = VBYI(VSUB(VADD(T2a, T2d), T1T));
|
Chris@10
|
875 ST(&(x[WS(rs, 21)]), VSUB(T27, T2e), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
876 ST(&(x[WS(rs, 4)]), VADD(T27, T2e), ms, &(x[0]));
|
Chris@10
|
877 }
|
Chris@10
|
878 T2f = VBYI(VSUB(VFMA(LDK(KP309016994), T2a, VFMA(LDK(KP951056516), VADD(T21, T22), VFNMS(LDK(KP809016994), T2d, VMUL(LDK(KP587785252), VSUB(T24, T25))))), T1T));
|
Chris@10
|
879 T2g = VFMA(LDK(KP951056516), VSUB(T29, T28), VFMA(LDK(KP309016994), T23, VFMA(LDK(KP587785252), VSUB(T2c, T2b), VFNMS(LDK(KP809016994), T26, T1D))));
|
Chris@10
|
880 ST(&(x[WS(rs, 9)]), VADD(T2f, T2g), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
881 ST(&(x[WS(rs, 16)]), VSUB(T2g, T2f), ms, &(x[0]));
|
Chris@10
|
882 {
|
Chris@10
|
883 V T1J, T1X, T1W, T1Y, T1F, T1V;
|
Chris@10
|
884 T1F = VFNMS(LDK(KP250000000), T1E, T1D);
|
Chris@10
|
885 T1J = VFMA(LDK(KP951056516), TF, VADD(T1o, VFNMS(LDK(KP587785252), T1I, T1F)));
|
Chris@10
|
886 T1X = VFMA(LDK(KP587785252), TF, VFMA(LDK(KP951056516), T1I, VSUB(T1F, T1o)));
|
Chris@10
|
887 T1V = VFNMS(LDK(KP250000000), T1U, T1T);
|
Chris@10
|
888 T1W = VBYI(VADD(VFMA(LDK(KP951056516), T1K, VMUL(LDK(KP587785252), T1L)), VADD(T1O, T1V)));
|
Chris@10
|
889 T1Y = VBYI(VADD(VFNMS(LDK(KP951056516), T1L, VMUL(LDK(KP587785252), T1K)), VSUB(T1V, T1O)));
|
Chris@10
|
890 ST(&(x[WS(rs, 19)]), VSUB(T1J, T1W), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
891 ST(&(x[WS(rs, 11)]), VADD(T1X, T1Y), ms, &(x[WS(rs, 1)]));
|
Chris@10
|
892 ST(&(x[WS(rs, 6)]), VADD(T1J, T1W), ms, &(x[0]));
|
Chris@10
|
893 ST(&(x[WS(rs, 14)]), VSUB(T1X, T1Y), ms, &(x[0]));
|
Chris@10
|
894 }
|
Chris@10
|
895 }
|
Chris@10
|
896 }
|
Chris@10
|
897 }
|
Chris@10
|
898 VLEAVE();
|
Chris@10
|
899 }
|
Chris@10
|
900
|
Chris@10
|
901 static const tw_instr twinstr[] = {
|
Chris@10
|
902 VTW(0, 1),
|
Chris@10
|
903 VTW(0, 2),
|
Chris@10
|
904 VTW(0, 3),
|
Chris@10
|
905 VTW(0, 4),
|
Chris@10
|
906 VTW(0, 5),
|
Chris@10
|
907 VTW(0, 6),
|
Chris@10
|
908 VTW(0, 7),
|
Chris@10
|
909 VTW(0, 8),
|
Chris@10
|
910 VTW(0, 9),
|
Chris@10
|
911 VTW(0, 10),
|
Chris@10
|
912 VTW(0, 11),
|
Chris@10
|
913 VTW(0, 12),
|
Chris@10
|
914 VTW(0, 13),
|
Chris@10
|
915 VTW(0, 14),
|
Chris@10
|
916 VTW(0, 15),
|
Chris@10
|
917 VTW(0, 16),
|
Chris@10
|
918 VTW(0, 17),
|
Chris@10
|
919 VTW(0, 18),
|
Chris@10
|
920 VTW(0, 19),
|
Chris@10
|
921 VTW(0, 20),
|
Chris@10
|
922 VTW(0, 21),
|
Chris@10
|
923 VTW(0, 22),
|
Chris@10
|
924 VTW(0, 23),
|
Chris@10
|
925 VTW(0, 24),
|
Chris@10
|
926 {TW_NEXT, VL, 0}
|
Chris@10
|
927 };
|
Chris@10
|
928
|
Chris@10
|
929 static const ct_desc desc = { 25, XSIMD_STRING("t1bv_25"), twinstr, &GENUS, {171, 111, 77, 0}, 0, 0, 0 };
|
Chris@10
|
930
|
Chris@10
|
931 void XSIMD(codelet_t1bv_25) (planner *p) {
|
Chris@10
|
932 X(kdft_dit_register) (p, t1bv_25, &desc);
|
Chris@10
|
933 }
|
Chris@10
|
934 #endif /* HAVE_FMA */
|