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 07:35:46 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-dft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_notw.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 25 -name n1_25 -include n.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 352 FP additions, 268 FP multiplications,
|
Chris@10
|
32 * (or, 84 additions, 0 multiplications, 268 fused multiply/add),
|
Chris@10
|
33 * 164 stack variables, 47 constants, and 100 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "n.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void n1_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
Chris@10
|
40 DK(KP554608978, +0.554608978404018097464974850792216217022558774);
|
Chris@10
|
41 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
Chris@10
|
42 DK(KP726211448, +0.726211448929902658173535992263577167607493062);
|
Chris@10
|
43 DK(KP525970792, +0.525970792408939708442463226536226366643874659);
|
Chris@10
|
44 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
45 DK(KP851038619, +0.851038619207379630836264138867114231259902550);
|
Chris@10
|
46 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
47 DK(KP912018591, +0.912018591466481957908415381764119056233607330);
|
Chris@10
|
48 DK(KP943557151, +0.943557151597354104399655195398983005179443399);
|
Chris@10
|
49 DK(KP614372930, +0.614372930789563808870829930444362096004872855);
|
Chris@10
|
50 DK(KP621716863, +0.621716863012209892444754556304102309693593202);
|
Chris@10
|
51 DK(KP994076283, +0.994076283785401014123185814696322018529298887);
|
Chris@10
|
52 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
Chris@10
|
53 DK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
Chris@10
|
54 DK(KP126329378, +0.126329378446108174786050455341811215027378105);
|
Chris@10
|
55 DK(KP827271945, +0.827271945972475634034355757144307982555673741);
|
Chris@10
|
56 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
Chris@10
|
57 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
Chris@10
|
58 DK(KP557913902, +0.557913902031834264187699648465567037992437152);
|
Chris@10
|
59 DK(KP249506682, +0.249506682107067890488084201715862638334226305);
|
Chris@10
|
60 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
Chris@10
|
61 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
Chris@10
|
62 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
63 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
Chris@10
|
64 DK(KP968479752, +0.968479752739016373193524836781420152702090879);
|
Chris@10
|
65 DK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
Chris@10
|
66 DK(KP470564281, +0.470564281212251493087595091036643380879947982);
|
Chris@10
|
67 DK(KP062914667, +0.062914667253649757225485955897349402364686947);
|
Chris@10
|
68 DK(KP921177326, +0.921177326965143320250447435415066029359282231);
|
Chris@10
|
69 DK(KP833417178, +0.833417178328688677408962550243238843138996060);
|
Chris@10
|
70 DK(KP541454447, +0.541454447536312777046285590082819509052033189);
|
Chris@10
|
71 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
Chris@10
|
72 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
73 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
74 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
75 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
76 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
77 DK(KP871714437, +0.871714437527667770979999223229522602943903653);
|
Chris@10
|
78 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
Chris@10
|
79 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
Chris@10
|
80 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
Chris@10
|
81 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
Chris@10
|
82 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
83 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
84 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
85 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
86 {
|
Chris@10
|
87 INT i;
|
Chris@10
|
88 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(100, is), MAKE_VOLATILE_STRIDE(100, os)) {
|
Chris@10
|
89 E T3Y, T3U, T3W, T42, T44, T3X, T3R, T3V, T3Z, T43;
|
Chris@10
|
90 {
|
Chris@10
|
91 E T4Q, T1U, T9, T3b, T45, T3e, T46, T1D, T4P, T1R, Ts, T1K, T18, T1E, T4z;
|
Chris@10
|
92 E T5f, T3z, T22, T4s, T5b, T3C, T2o, T3D, T2h, T4p, T5c, T4w, T5e, T3A, T29;
|
Chris@10
|
93 E T2z, T2y, TL, T1L, T1r, T1F, T4a, T57, T3v, T2x, T4k, T55, T3s, T2T, T2D;
|
Chris@10
|
94 E T4c, T3t, T2M, T4h, T54, T1v, T1C, T1Q;
|
Chris@10
|
95 {
|
Chris@10
|
96 E T1, T2, T3, T5, T6;
|
Chris@10
|
97 T1 = ri[0];
|
Chris@10
|
98 T2 = ri[WS(is, 5)];
|
Chris@10
|
99 T3 = ri[WS(is, 20)];
|
Chris@10
|
100 T5 = ri[WS(is, 10)];
|
Chris@10
|
101 T6 = ri[WS(is, 15)];
|
Chris@10
|
102 {
|
Chris@10
|
103 E T3a, T3c, T1y, T1z, T1A, T39, T4, T1S, T1B, T3d;
|
Chris@10
|
104 T1v = ii[0];
|
Chris@10
|
105 T4 = T2 + T3;
|
Chris@10
|
106 T1S = T2 - T3;
|
Chris@10
|
107 {
|
Chris@10
|
108 E T7, T1T, T8, T1w, T1x;
|
Chris@10
|
109 T7 = T5 + T6;
|
Chris@10
|
110 T1T = T5 - T6;
|
Chris@10
|
111 T1w = ii[WS(is, 5)];
|
Chris@10
|
112 T1x = ii[WS(is, 20)];
|
Chris@10
|
113 T4Q = FNMS(KP618033988, T1S, T1T);
|
Chris@10
|
114 T1U = FMA(KP618033988, T1T, T1S);
|
Chris@10
|
115 T8 = T4 + T7;
|
Chris@10
|
116 T3a = T4 - T7;
|
Chris@10
|
117 T3c = T1w - T1x;
|
Chris@10
|
118 T1y = T1w + T1x;
|
Chris@10
|
119 T1z = ii[WS(is, 10)];
|
Chris@10
|
120 T1A = ii[WS(is, 15)];
|
Chris@10
|
121 T39 = FNMS(KP250000000, T8, T1);
|
Chris@10
|
122 T9 = T1 + T8;
|
Chris@10
|
123 }
|
Chris@10
|
124 T1B = T1z + T1A;
|
Chris@10
|
125 T3d = T1z - T1A;
|
Chris@10
|
126 T3b = FMA(KP559016994, T3a, T39);
|
Chris@10
|
127 T45 = FNMS(KP559016994, T3a, T39);
|
Chris@10
|
128 T3e = FMA(KP618033988, T3d, T3c);
|
Chris@10
|
129 T46 = FNMS(KP618033988, T3c, T3d);
|
Chris@10
|
130 T1C = T1y + T1B;
|
Chris@10
|
131 T1Q = T1y - T1B;
|
Chris@10
|
132 }
|
Chris@10
|
133 }
|
Chris@10
|
134 {
|
Chris@10
|
135 E T24, T23, T28, T4v;
|
Chris@10
|
136 {
|
Chris@10
|
137 E Ta, TQ, Tj, TZ, T1Z, T20, Th, T26, T27, T1X, TX, T2l, T2m, Tq, T2c;
|
Chris@10
|
138 E T2e, T12, T15, T2f, T1P, TT, TW;
|
Chris@10
|
139 Ta = ri[WS(is, 1)];
|
Chris@10
|
140 T1P = FNMS(KP250000000, T1C, T1v);
|
Chris@10
|
141 T1D = T1v + T1C;
|
Chris@10
|
142 TQ = ii[WS(is, 1)];
|
Chris@10
|
143 Tj = ri[WS(is, 4)];
|
Chris@10
|
144 T4P = FNMS(KP559016994, T1Q, T1P);
|
Chris@10
|
145 T1R = FMA(KP559016994, T1Q, T1P);
|
Chris@10
|
146 TZ = ii[WS(is, 4)];
|
Chris@10
|
147 {
|
Chris@10
|
148 E Tb, Tc, Te, Tf;
|
Chris@10
|
149 Tb = ri[WS(is, 6)];
|
Chris@10
|
150 Tc = ri[WS(is, 21)];
|
Chris@10
|
151 Te = ri[WS(is, 11)];
|
Chris@10
|
152 Tf = ri[WS(is, 16)];
|
Chris@10
|
153 {
|
Chris@10
|
154 E TR, Td, Tg, TS, TU, TV;
|
Chris@10
|
155 TR = ii[WS(is, 6)];
|
Chris@10
|
156 T1Z = Tc - Tb;
|
Chris@10
|
157 Td = Tb + Tc;
|
Chris@10
|
158 T20 = Tf - Te;
|
Chris@10
|
159 Tg = Te + Tf;
|
Chris@10
|
160 TS = ii[WS(is, 21)];
|
Chris@10
|
161 TU = ii[WS(is, 11)];
|
Chris@10
|
162 TV = ii[WS(is, 16)];
|
Chris@10
|
163 Th = Td + Tg;
|
Chris@10
|
164 T24 = Td - Tg;
|
Chris@10
|
165 T26 = TR - TS;
|
Chris@10
|
166 TT = TR + TS;
|
Chris@10
|
167 TW = TU + TV;
|
Chris@10
|
168 T27 = TV - TU;
|
Chris@10
|
169 }
|
Chris@10
|
170 }
|
Chris@10
|
171 {
|
Chris@10
|
172 E Tk, Tl, Tn, To;
|
Chris@10
|
173 Tk = ri[WS(is, 9)];
|
Chris@10
|
174 T1X = TT - TW;
|
Chris@10
|
175 TX = TT + TW;
|
Chris@10
|
176 Tl = ri[WS(is, 24)];
|
Chris@10
|
177 Tn = ri[WS(is, 14)];
|
Chris@10
|
178 To = ri[WS(is, 19)];
|
Chris@10
|
179 {
|
Chris@10
|
180 E T10, Tm, Tp, T11, T13, T14;
|
Chris@10
|
181 T10 = ii[WS(is, 9)];
|
Chris@10
|
182 T2l = Tl - Tk;
|
Chris@10
|
183 Tm = Tk + Tl;
|
Chris@10
|
184 T2m = To - Tn;
|
Chris@10
|
185 Tp = Tn + To;
|
Chris@10
|
186 T11 = ii[WS(is, 24)];
|
Chris@10
|
187 T13 = ii[WS(is, 14)];
|
Chris@10
|
188 T14 = ii[WS(is, 19)];
|
Chris@10
|
189 Tq = Tm + Tp;
|
Chris@10
|
190 T2c = Tm - Tp;
|
Chris@10
|
191 T2e = T11 - T10;
|
Chris@10
|
192 T12 = T10 + T11;
|
Chris@10
|
193 T15 = T13 + T14;
|
Chris@10
|
194 T2f = T14 - T13;
|
Chris@10
|
195 }
|
Chris@10
|
196 }
|
Chris@10
|
197 {
|
Chris@10
|
198 E T2j, T2b, T1W, T21, T4y, T2i;
|
Chris@10
|
199 {
|
Chris@10
|
200 E Ti, T16, Tr, TY, T17;
|
Chris@10
|
201 T23 = FNMS(KP250000000, Th, Ta);
|
Chris@10
|
202 Ti = Ta + Th;
|
Chris@10
|
203 T2j = T15 - T12;
|
Chris@10
|
204 T16 = T12 + T15;
|
Chris@10
|
205 Tr = Tj + Tq;
|
Chris@10
|
206 T2b = FMS(KP250000000, Tq, Tj);
|
Chris@10
|
207 T1W = FNMS(KP250000000, TX, TQ);
|
Chris@10
|
208 TY = TQ + TX;
|
Chris@10
|
209 T21 = FMA(KP618033988, T20, T1Z);
|
Chris@10
|
210 T4y = FNMS(KP618033988, T1Z, T20);
|
Chris@10
|
211 T2i = FNMS(KP250000000, T16, TZ);
|
Chris@10
|
212 T17 = TZ + T16;
|
Chris@10
|
213 Ts = Ti + Tr;
|
Chris@10
|
214 T1K = Ti - Tr;
|
Chris@10
|
215 T18 = TY - T17;
|
Chris@10
|
216 T1E = TY + T17;
|
Chris@10
|
217 }
|
Chris@10
|
218 {
|
Chris@10
|
219 E T2n, T4r, T4x, T1Y;
|
Chris@10
|
220 T2n = FMA(KP618033988, T2m, T2l);
|
Chris@10
|
221 T4r = FNMS(KP618033988, T2l, T2m);
|
Chris@10
|
222 T4x = FNMS(KP559016994, T1X, T1W);
|
Chris@10
|
223 T1Y = FMA(KP559016994, T1X, T1W);
|
Chris@10
|
224 {
|
Chris@10
|
225 E T4o, T2g, T2d, T4n, T4q, T2k;
|
Chris@10
|
226 T4o = FNMS(KP618033988, T2e, T2f);
|
Chris@10
|
227 T2g = FMA(KP618033988, T2f, T2e);
|
Chris@10
|
228 T4z = FMA(KP951056516, T4y, T4x);
|
Chris@10
|
229 T5f = FNMS(KP951056516, T4y, T4x);
|
Chris@10
|
230 T3z = FNMS(KP951056516, T21, T1Y);
|
Chris@10
|
231 T22 = FMA(KP951056516, T21, T1Y);
|
Chris@10
|
232 T4q = FMA(KP559016994, T2j, T2i);
|
Chris@10
|
233 T2k = FNMS(KP559016994, T2j, T2i);
|
Chris@10
|
234 T4s = FMA(KP951056516, T4r, T4q);
|
Chris@10
|
235 T5b = FNMS(KP951056516, T4r, T4q);
|
Chris@10
|
236 T3C = FNMS(KP951056516, T2n, T2k);
|
Chris@10
|
237 T2o = FMA(KP951056516, T2n, T2k);
|
Chris@10
|
238 T2d = FNMS(KP559016994, T2c, T2b);
|
Chris@10
|
239 T4n = FMA(KP559016994, T2c, T2b);
|
Chris@10
|
240 T28 = FNMS(KP618033988, T27, T26);
|
Chris@10
|
241 T4v = FMA(KP618033988, T26, T27);
|
Chris@10
|
242 T3D = FNMS(KP951056516, T2g, T2d);
|
Chris@10
|
243 T2h = FMA(KP951056516, T2g, T2d);
|
Chris@10
|
244 T4p = FMA(KP951056516, T4o, T4n);
|
Chris@10
|
245 T5c = FNMS(KP951056516, T4o, T4n);
|
Chris@10
|
246 }
|
Chris@10
|
247 }
|
Chris@10
|
248 }
|
Chris@10
|
249 }
|
Chris@10
|
250 {
|
Chris@10
|
251 E Tt, T19, TC, T1i, T2u, T2v, TA, T2B, T2C, T2s, T1g, T2J, T2K, TJ, T2O;
|
Chris@10
|
252 E T2Q, T1l, T1o, T2R;
|
Chris@10
|
253 {
|
Chris@10
|
254 E T4u, T25, T1c, T1f;
|
Chris@10
|
255 Tt = ri[WS(is, 2)];
|
Chris@10
|
256 T19 = ii[WS(is, 2)];
|
Chris@10
|
257 TC = ri[WS(is, 3)];
|
Chris@10
|
258 T4u = FNMS(KP559016994, T24, T23);
|
Chris@10
|
259 T25 = FMA(KP559016994, T24, T23);
|
Chris@10
|
260 T1i = ii[WS(is, 3)];
|
Chris@10
|
261 {
|
Chris@10
|
262 E Tu, Tv, Tx, Ty;
|
Chris@10
|
263 Tu = ri[WS(is, 7)];
|
Chris@10
|
264 T4w = FNMS(KP951056516, T4v, T4u);
|
Chris@10
|
265 T5e = FMA(KP951056516, T4v, T4u);
|
Chris@10
|
266 T3A = FNMS(KP951056516, T28, T25);
|
Chris@10
|
267 T29 = FMA(KP951056516, T28, T25);
|
Chris@10
|
268 Tv = ri[WS(is, 22)];
|
Chris@10
|
269 Tx = ri[WS(is, 12)];
|
Chris@10
|
270 Ty = ri[WS(is, 17)];
|
Chris@10
|
271 {
|
Chris@10
|
272 E T1a, Tw, Tz, T1b, T1d, T1e;
|
Chris@10
|
273 T1a = ii[WS(is, 7)];
|
Chris@10
|
274 T2u = Tv - Tu;
|
Chris@10
|
275 Tw = Tu + Tv;
|
Chris@10
|
276 T2v = Ty - Tx;
|
Chris@10
|
277 Tz = Tx + Ty;
|
Chris@10
|
278 T1b = ii[WS(is, 22)];
|
Chris@10
|
279 T1d = ii[WS(is, 12)];
|
Chris@10
|
280 T1e = ii[WS(is, 17)];
|
Chris@10
|
281 TA = Tw + Tz;
|
Chris@10
|
282 T2z = Tz - Tw;
|
Chris@10
|
283 T2B = T1b - T1a;
|
Chris@10
|
284 T1c = T1a + T1b;
|
Chris@10
|
285 T1f = T1d + T1e;
|
Chris@10
|
286 T2C = T1d - T1e;
|
Chris@10
|
287 }
|
Chris@10
|
288 }
|
Chris@10
|
289 {
|
Chris@10
|
290 E TD, TE, TG, TH;
|
Chris@10
|
291 TD = ri[WS(is, 8)];
|
Chris@10
|
292 T2s = T1f - T1c;
|
Chris@10
|
293 T1g = T1c + T1f;
|
Chris@10
|
294 TE = ri[WS(is, 23)];
|
Chris@10
|
295 TG = ri[WS(is, 13)];
|
Chris@10
|
296 TH = ri[WS(is, 18)];
|
Chris@10
|
297 {
|
Chris@10
|
298 E T1j, TF, TI, T1k, T1m, T1n;
|
Chris@10
|
299 T1j = ii[WS(is, 8)];
|
Chris@10
|
300 T2J = TD - TE;
|
Chris@10
|
301 TF = TD + TE;
|
Chris@10
|
302 T2K = TG - TH;
|
Chris@10
|
303 TI = TG + TH;
|
Chris@10
|
304 T1k = ii[WS(is, 23)];
|
Chris@10
|
305 T1m = ii[WS(is, 13)];
|
Chris@10
|
306 T1n = ii[WS(is, 18)];
|
Chris@10
|
307 TJ = TF + TI;
|
Chris@10
|
308 T2O = TI - TF;
|
Chris@10
|
309 T2Q = T1k - T1j;
|
Chris@10
|
310 T1l = T1j + T1k;
|
Chris@10
|
311 T1o = T1m + T1n;
|
Chris@10
|
312 T2R = T1n - T1m;
|
Chris@10
|
313 }
|
Chris@10
|
314 }
|
Chris@10
|
315 }
|
Chris@10
|
316 {
|
Chris@10
|
317 E T2H, T2N, T2r, T2w, T49, T2G;
|
Chris@10
|
318 {
|
Chris@10
|
319 E TB, T1p, TK, T1h, T1q;
|
Chris@10
|
320 T2y = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
321 TB = Tt + TA;
|
Chris@10
|
322 T2H = T1o - T1l;
|
Chris@10
|
323 T1p = T1l + T1o;
|
Chris@10
|
324 TK = TC + TJ;
|
Chris@10
|
325 T2N = FNMS(KP250000000, TJ, TC);
|
Chris@10
|
326 T2r = FNMS(KP250000000, T1g, T19);
|
Chris@10
|
327 T1h = T19 + T1g;
|
Chris@10
|
328 T2w = FMA(KP618033988, T2v, T2u);
|
Chris@10
|
329 T49 = FNMS(KP618033988, T2u, T2v);
|
Chris@10
|
330 T2G = FNMS(KP250000000, T1p, T1i);
|
Chris@10
|
331 T1q = T1i + T1p;
|
Chris@10
|
332 TL = TB + TK;
|
Chris@10
|
333 T1L = TB - TK;
|
Chris@10
|
334 T1r = T1h - T1q;
|
Chris@10
|
335 T1F = T1h + T1q;
|
Chris@10
|
336 }
|
Chris@10
|
337 {
|
Chris@10
|
338 E T2S, T4j, T48, T2t;
|
Chris@10
|
339 T2S = FMA(KP618033988, T2R, T2Q);
|
Chris@10
|
340 T4j = FNMS(KP618033988, T2Q, T2R);
|
Chris@10
|
341 T48 = FMA(KP559016994, T2s, T2r);
|
Chris@10
|
342 T2t = FNMS(KP559016994, T2s, T2r);
|
Chris@10
|
343 {
|
Chris@10
|
344 E T4g, T2L, T2I, T4f, T4i, T2P;
|
Chris@10
|
345 T4g = FNMS(KP618033988, T2J, T2K);
|
Chris@10
|
346 T2L = FMA(KP618033988, T2K, T2J);
|
Chris@10
|
347 T4a = FMA(KP951056516, T49, T48);
|
Chris@10
|
348 T57 = FNMS(KP951056516, T49, T48);
|
Chris@10
|
349 T3v = FNMS(KP951056516, T2w, T2t);
|
Chris@10
|
350 T2x = FMA(KP951056516, T2w, T2t);
|
Chris@10
|
351 T4i = FMA(KP559016994, T2O, T2N);
|
Chris@10
|
352 T2P = FNMS(KP559016994, T2O, T2N);
|
Chris@10
|
353 T4k = FNMS(KP951056516, T4j, T4i);
|
Chris@10
|
354 T55 = FMA(KP951056516, T4j, T4i);
|
Chris@10
|
355 T3s = FMA(KP951056516, T2S, T2P);
|
Chris@10
|
356 T2T = FNMS(KP951056516, T2S, T2P);
|
Chris@10
|
357 T2I = FNMS(KP559016994, T2H, T2G);
|
Chris@10
|
358 T4f = FMA(KP559016994, T2H, T2G);
|
Chris@10
|
359 T2D = FNMS(KP618033988, T2C, T2B);
|
Chris@10
|
360 T4c = FMA(KP618033988, T2B, T2C);
|
Chris@10
|
361 T3t = FMA(KP951056516, T2L, T2I);
|
Chris@10
|
362 T2M = FNMS(KP951056516, T2L, T2I);
|
Chris@10
|
363 T4h = FNMS(KP951056516, T4g, T4f);
|
Chris@10
|
364 T54 = FMA(KP951056516, T4g, T4f);
|
Chris@10
|
365 }
|
Chris@10
|
366 }
|
Chris@10
|
367 }
|
Chris@10
|
368 }
|
Chris@10
|
369 }
|
Chris@10
|
370 {
|
Chris@10
|
371 E T4d, T58, T3w, T3H, T3r, T3k, T36, T38, T3o, T3q, T3j, T2Z, T37;
|
Chris@10
|
372 {
|
Chris@10
|
373 E T2E, T1s, T1u, TP, T1t;
|
Chris@10
|
374 {
|
Chris@10
|
375 E TM, TO, TN, T4b, T2A;
|
Chris@10
|
376 TM = Ts + TL;
|
Chris@10
|
377 TO = Ts - TL;
|
Chris@10
|
378 T4b = FMA(KP559016994, T2z, T2y);
|
Chris@10
|
379 T2A = FNMS(KP559016994, T2z, T2y);
|
Chris@10
|
380 TN = FNMS(KP250000000, TM, T9);
|
Chris@10
|
381 T4d = FMA(KP951056516, T4c, T4b);
|
Chris@10
|
382 T58 = FNMS(KP951056516, T4c, T4b);
|
Chris@10
|
383 T3w = FMA(KP951056516, T2D, T2A);
|
Chris@10
|
384 T2E = FNMS(KP951056516, T2D, T2A);
|
Chris@10
|
385 T1s = FMA(KP618033988, T1r, T18);
|
Chris@10
|
386 T1u = FNMS(KP618033988, T18, T1r);
|
Chris@10
|
387 ro[0] = T9 + TM;
|
Chris@10
|
388 TP = FMA(KP559016994, TO, TN);
|
Chris@10
|
389 T1t = FNMS(KP559016994, TO, TN);
|
Chris@10
|
390 }
|
Chris@10
|
391 {
|
Chris@10
|
392 E T1J, T1N, T1M, T1O, T1G, T1I, T1H;
|
Chris@10
|
393 T1G = T1E + T1F;
|
Chris@10
|
394 T1I = T1E - T1F;
|
Chris@10
|
395 ro[WS(os, 15)] = FMA(KP951056516, T1u, T1t);
|
Chris@10
|
396 ro[WS(os, 10)] = FNMS(KP951056516, T1u, T1t);
|
Chris@10
|
397 ro[WS(os, 5)] = FMA(KP951056516, T1s, TP);
|
Chris@10
|
398 ro[WS(os, 20)] = FNMS(KP951056516, T1s, TP);
|
Chris@10
|
399 T1H = FNMS(KP250000000, T1G, T1D);
|
Chris@10
|
400 io[0] = T1D + T1G;
|
Chris@10
|
401 T1J = FMA(KP559016994, T1I, T1H);
|
Chris@10
|
402 T1N = FNMS(KP559016994, T1I, T1H);
|
Chris@10
|
403 T1M = FMA(KP618033988, T1L, T1K);
|
Chris@10
|
404 T1O = FNMS(KP618033988, T1K, T1L);
|
Chris@10
|
405 {
|
Chris@10
|
406 E T1V, T3f, T3m, T3n, T2W, T2Y, T32, T3g, T3h, T35, T3i, T2X;
|
Chris@10
|
407 T3H = FMA(KP951056516, T1U, T1R);
|
Chris@10
|
408 T1V = FNMS(KP951056516, T1U, T1R);
|
Chris@10
|
409 T3f = FMA(KP951056516, T3e, T3b);
|
Chris@10
|
410 T3r = FNMS(KP951056516, T3e, T3b);
|
Chris@10
|
411 io[WS(os, 15)] = FNMS(KP951056516, T1O, T1N);
|
Chris@10
|
412 io[WS(os, 10)] = FMA(KP951056516, T1O, T1N);
|
Chris@10
|
413 io[WS(os, 20)] = FMA(KP951056516, T1M, T1J);
|
Chris@10
|
414 io[WS(os, 5)] = FNMS(KP951056516, T1M, T1J);
|
Chris@10
|
415 {
|
Chris@10
|
416 E T30, T2a, T2p, T31, T33, T2F, T2U, T34, T2q, T2V;
|
Chris@10
|
417 T30 = FMA(KP256756360, T22, T29);
|
Chris@10
|
418 T2a = FNMS(KP256756360, T29, T22);
|
Chris@10
|
419 T2p = FMA(KP634619297, T2o, T2h);
|
Chris@10
|
420 T31 = FNMS(KP634619297, T2h, T2o);
|
Chris@10
|
421 T33 = FMA(KP549754652, T2x, T2E);
|
Chris@10
|
422 T2F = FNMS(KP549754652, T2E, T2x);
|
Chris@10
|
423 T2U = FNMS(KP939062505, T2T, T2M);
|
Chris@10
|
424 T34 = FMA(KP939062505, T2M, T2T);
|
Chris@10
|
425 T3m = FNMS(KP871714437, T2p, T2a);
|
Chris@10
|
426 T2q = FMA(KP871714437, T2p, T2a);
|
Chris@10
|
427 T3n = FNMS(KP831864738, T2U, T2F);
|
Chris@10
|
428 T2V = FMA(KP831864738, T2U, T2F);
|
Chris@10
|
429 T2W = FMA(KP904730450, T2V, T2q);
|
Chris@10
|
430 T2Y = FNMS(KP904730450, T2V, T2q);
|
Chris@10
|
431 T32 = FNMS(KP871714437, T31, T30);
|
Chris@10
|
432 T3g = FMA(KP871714437, T31, T30);
|
Chris@10
|
433 T3h = FMA(KP831864738, T34, T33);
|
Chris@10
|
434 T35 = FNMS(KP831864738, T34, T33);
|
Chris@10
|
435 }
|
Chris@10
|
436 io[WS(os, 1)] = FMA(KP968583161, T2W, T1V);
|
Chris@10
|
437 T3i = FMA(KP904730450, T3h, T3g);
|
Chris@10
|
438 T3k = FNMS(KP904730450, T3h, T3g);
|
Chris@10
|
439 T36 = FMA(KP559154169, T35, T32);
|
Chris@10
|
440 T38 = FNMS(KP683113946, T32, T35);
|
Chris@10
|
441 ro[WS(os, 1)] = FMA(KP968583161, T3i, T3f);
|
Chris@10
|
442 T2X = FNMS(KP242145790, T2W, T1V);
|
Chris@10
|
443 T3o = FMA(KP559154169, T3n, T3m);
|
Chris@10
|
444 T3q = FNMS(KP683113946, T3m, T3n);
|
Chris@10
|
445 T3j = FNMS(KP242145790, T3i, T3f);
|
Chris@10
|
446 T2Z = FMA(KP541454447, T2Y, T2X);
|
Chris@10
|
447 T37 = FNMS(KP541454447, T2Y, T2X);
|
Chris@10
|
448 }
|
Chris@10
|
449 }
|
Chris@10
|
450 }
|
Chris@10
|
451 {
|
Chris@10
|
452 E T47, T4R, T5A, T5w, T5y, T5E, T5G, T5z, T5t, T5x;
|
Chris@10
|
453 {
|
Chris@10
|
454 E T53, T5j, T5u, T5v, T5i, T5D, T5m, T5p, T5C, T3p, T3l, T5s, T5q, T5r;
|
Chris@10
|
455 T47 = FMA(KP951056516, T46, T45);
|
Chris@10
|
456 T53 = FNMS(KP951056516, T46, T45);
|
Chris@10
|
457 T3p = FNMS(KP541454447, T3k, T3j);
|
Chris@10
|
458 T3l = FMA(KP541454447, T3k, T3j);
|
Chris@10
|
459 io[WS(os, 16)] = FNMS(KP833417178, T38, T37);
|
Chris@10
|
460 io[WS(os, 11)] = FMA(KP833417178, T38, T37);
|
Chris@10
|
461 io[WS(os, 21)] = FMA(KP921177326, T36, T2Z);
|
Chris@10
|
462 io[WS(os, 6)] = FNMS(KP921177326, T36, T2Z);
|
Chris@10
|
463 ro[WS(os, 11)] = FNMS(KP833417178, T3q, T3p);
|
Chris@10
|
464 ro[WS(os, 16)] = FMA(KP833417178, T3q, T3p);
|
Chris@10
|
465 ro[WS(os, 21)] = FNMS(KP921177326, T3o, T3l);
|
Chris@10
|
466 ro[WS(os, 6)] = FMA(KP921177326, T3o, T3l);
|
Chris@10
|
467 T5j = FMA(KP951056516, T4Q, T4P);
|
Chris@10
|
468 T4R = FNMS(KP951056516, T4Q, T4P);
|
Chris@10
|
469 {
|
Chris@10
|
470 E T5k, T56, T59, T5l, T5n, T5d, T5g, T5o, T5a, T5h;
|
Chris@10
|
471 T5k = FNMS(KP062914667, T54, T55);
|
Chris@10
|
472 T56 = FMA(KP062914667, T55, T54);
|
Chris@10
|
473 T59 = FMA(KP634619297, T58, T57);
|
Chris@10
|
474 T5l = FNMS(KP634619297, T57, T58);
|
Chris@10
|
475 T5n = FNMS(KP470564281, T5b, T5c);
|
Chris@10
|
476 T5d = FMA(KP470564281, T5c, T5b);
|
Chris@10
|
477 T5g = FMA(KP549754652, T5f, T5e);
|
Chris@10
|
478 T5o = FNMS(KP549754652, T5e, T5f);
|
Chris@10
|
479 T5u = FNMS(KP845997307, T59, T56);
|
Chris@10
|
480 T5a = FMA(KP845997307, T59, T56);
|
Chris@10
|
481 T5v = FNMS(KP968479752, T5g, T5d);
|
Chris@10
|
482 T5h = FMA(KP968479752, T5g, T5d);
|
Chris@10
|
483 T5i = FMA(KP906616052, T5h, T5a);
|
Chris@10
|
484 T5A = FNMS(KP906616052, T5h, T5a);
|
Chris@10
|
485 T5D = FNMS(KP845997307, T5l, T5k);
|
Chris@10
|
486 T5m = FMA(KP845997307, T5l, T5k);
|
Chris@10
|
487 T5p = FMA(KP968479752, T5o, T5n);
|
Chris@10
|
488 T5C = FNMS(KP968479752, T5o, T5n);
|
Chris@10
|
489 }
|
Chris@10
|
490 ro[WS(os, 2)] = FMA(KP998026728, T5i, T53);
|
Chris@10
|
491 T5s = FMA(KP906616052, T5p, T5m);
|
Chris@10
|
492 T5q = FNMS(KP906616052, T5p, T5m);
|
Chris@10
|
493 T5w = FNMS(KP560319534, T5v, T5u);
|
Chris@10
|
494 T5y = FMA(KP681693190, T5u, T5v);
|
Chris@10
|
495 T5E = FNMS(KP681693190, T5D, T5C);
|
Chris@10
|
496 T5G = FMA(KP560319534, T5C, T5D);
|
Chris@10
|
497 T5r = FMA(KP249506682, T5q, T5j);
|
Chris@10
|
498 io[WS(os, 2)] = FNMS(KP998026728, T5q, T5j);
|
Chris@10
|
499 T5z = FNMS(KP249506682, T5i, T53);
|
Chris@10
|
500 T5t = FNMS(KP557913902, T5s, T5r);
|
Chris@10
|
501 T5x = FMA(KP557913902, T5s, T5r);
|
Chris@10
|
502 }
|
Chris@10
|
503 {
|
Chris@10
|
504 E T4W, T4M, T4O, T50, T52, T4V, T4F, T4N;
|
Chris@10
|
505 {
|
Chris@10
|
506 E T4Y, T4Z, T4C, T4E, T4I, T4T, T4S, T4L, T5F, T5B, T4U, T4D;
|
Chris@10
|
507 T5F = FMA(KP557913902, T5A, T5z);
|
Chris@10
|
508 T5B = FNMS(KP557913902, T5A, T5z);
|
Chris@10
|
509 io[WS(os, 7)] = FMA(KP860541664, T5y, T5x);
|
Chris@10
|
510 io[WS(os, 22)] = FNMS(KP860541664, T5y, T5x);
|
Chris@10
|
511 io[WS(os, 17)] = FMA(KP949179823, T5w, T5t);
|
Chris@10
|
512 io[WS(os, 12)] = FNMS(KP949179823, T5w, T5t);
|
Chris@10
|
513 ro[WS(os, 12)] = FNMS(KP949179823, T5G, T5F);
|
Chris@10
|
514 ro[WS(os, 17)] = FMA(KP949179823, T5G, T5F);
|
Chris@10
|
515 ro[WS(os, 7)] = FNMS(KP860541664, T5E, T5B);
|
Chris@10
|
516 ro[WS(os, 22)] = FMA(KP860541664, T5E, T5B);
|
Chris@10
|
517 {
|
Chris@10
|
518 E T4J, T4e, T4l, T4K, T4G, T4t, T4A, T4H, T4m, T4B;
|
Chris@10
|
519 T4J = FNMS(KP062914667, T4a, T4d);
|
Chris@10
|
520 T4e = FMA(KP062914667, T4d, T4a);
|
Chris@10
|
521 T4l = FNMS(KP827271945, T4k, T4h);
|
Chris@10
|
522 T4K = FMA(KP827271945, T4h, T4k);
|
Chris@10
|
523 T4G = FNMS(KP126329378, T4p, T4s);
|
Chris@10
|
524 T4t = FMA(KP126329378, T4s, T4p);
|
Chris@10
|
525 T4A = FMA(KP939062505, T4z, T4w);
|
Chris@10
|
526 T4H = FNMS(KP939062505, T4w, T4z);
|
Chris@10
|
527 T4Y = FNMS(KP772036680, T4l, T4e);
|
Chris@10
|
528 T4m = FMA(KP772036680, T4l, T4e);
|
Chris@10
|
529 T4Z = FNMS(KP734762448, T4A, T4t);
|
Chris@10
|
530 T4B = FMA(KP734762448, T4A, T4t);
|
Chris@10
|
531 T4C = FMA(KP994076283, T4B, T4m);
|
Chris@10
|
532 T4E = FNMS(KP994076283, T4B, T4m);
|
Chris@10
|
533 T4I = FMA(KP734762448, T4H, T4G);
|
Chris@10
|
534 T4T = FNMS(KP734762448, T4H, T4G);
|
Chris@10
|
535 T4S = FMA(KP772036680, T4K, T4J);
|
Chris@10
|
536 T4L = FNMS(KP772036680, T4K, T4J);
|
Chris@10
|
537 }
|
Chris@10
|
538 ro[WS(os, 3)] = FMA(KP998026728, T4C, T47);
|
Chris@10
|
539 T4U = FMA(KP994076283, T4T, T4S);
|
Chris@10
|
540 T4W = FNMS(KP994076283, T4T, T4S);
|
Chris@10
|
541 T4M = FNMS(KP621716863, T4L, T4I);
|
Chris@10
|
542 T4O = FMA(KP614372930, T4I, T4L);
|
Chris@10
|
543 io[WS(os, 3)] = FNMS(KP998026728, T4U, T4R);
|
Chris@10
|
544 T4D = FNMS(KP249506682, T4C, T47);
|
Chris@10
|
545 T50 = FMA(KP614372930, T4Z, T4Y);
|
Chris@10
|
546 T52 = FNMS(KP621716863, T4Y, T4Z);
|
Chris@10
|
547 T4V = FMA(KP249506682, T4U, T4R);
|
Chris@10
|
548 T4F = FNMS(KP557913902, T4E, T4D);
|
Chris@10
|
549 T4N = FMA(KP557913902, T4E, T4D);
|
Chris@10
|
550 }
|
Chris@10
|
551 {
|
Chris@10
|
552 E T3S, T3T, T3G, T41, T3K, T3N, T40, T51, T4X, T3Q, T3O, T3P;
|
Chris@10
|
553 T51 = FMA(KP557913902, T4W, T4V);
|
Chris@10
|
554 T4X = FNMS(KP557913902, T4W, T4V);
|
Chris@10
|
555 ro[WS(os, 18)] = FNMS(KP949179823, T4O, T4N);
|
Chris@10
|
556 ro[WS(os, 13)] = FMA(KP949179823, T4O, T4N);
|
Chris@10
|
557 ro[WS(os, 8)] = FMA(KP943557151, T4M, T4F);
|
Chris@10
|
558 ro[WS(os, 23)] = FNMS(KP943557151, T4M, T4F);
|
Chris@10
|
559 io[WS(os, 8)] = FMA(KP943557151, T52, T51);
|
Chris@10
|
560 io[WS(os, 23)] = FNMS(KP943557151, T52, T51);
|
Chris@10
|
561 io[WS(os, 18)] = FNMS(KP949179823, T50, T4X);
|
Chris@10
|
562 io[WS(os, 13)] = FMA(KP949179823, T50, T4X);
|
Chris@10
|
563 {
|
Chris@10
|
564 E T3I, T3u, T3x, T3J, T3L, T3B, T3E, T3M, T3y, T3F;
|
Chris@10
|
565 T3I = FMA(KP126329378, T3s, T3t);
|
Chris@10
|
566 T3u = FNMS(KP126329378, T3t, T3s);
|
Chris@10
|
567 T3x = FNMS(KP470564281, T3w, T3v);
|
Chris@10
|
568 T3J = FMA(KP470564281, T3v, T3w);
|
Chris@10
|
569 T3L = FNMS(KP634619297, T3z, T3A);
|
Chris@10
|
570 T3B = FMA(KP634619297, T3A, T3z);
|
Chris@10
|
571 T3E = FNMS(KP827271945, T3D, T3C);
|
Chris@10
|
572 T3M = FMA(KP827271945, T3C, T3D);
|
Chris@10
|
573 T3S = FMA(KP912018591, T3x, T3u);
|
Chris@10
|
574 T3y = FNMS(KP912018591, T3x, T3u);
|
Chris@10
|
575 T3T = FMA(KP912575812, T3E, T3B);
|
Chris@10
|
576 T3F = FNMS(KP912575812, T3E, T3B);
|
Chris@10
|
577 T3G = FNMS(KP851038619, T3F, T3y);
|
Chris@10
|
578 T3Y = FMA(KP851038619, T3F, T3y);
|
Chris@10
|
579 T41 = FNMS(KP912018591, T3J, T3I);
|
Chris@10
|
580 T3K = FMA(KP912018591, T3J, T3I);
|
Chris@10
|
581 T3N = FMA(KP912575812, T3M, T3L);
|
Chris@10
|
582 T40 = FNMS(KP912575812, T3M, T3L);
|
Chris@10
|
583 }
|
Chris@10
|
584 ro[WS(os, 4)] = FNMS(KP992114701, T3G, T3r);
|
Chris@10
|
585 T3Q = FNMS(KP851038619, T3N, T3K);
|
Chris@10
|
586 T3O = FMA(KP851038619, T3N, T3K);
|
Chris@10
|
587 T3U = FNMS(KP525970792, T3T, T3S);
|
Chris@10
|
588 T3W = FMA(KP726211448, T3S, T3T);
|
Chris@10
|
589 T42 = FNMS(KP726211448, T41, T40);
|
Chris@10
|
590 T44 = FMA(KP525970792, T40, T41);
|
Chris@10
|
591 T3P = FMA(KP248028675, T3O, T3H);
|
Chris@10
|
592 io[WS(os, 4)] = FNMS(KP992114701, T3O, T3H);
|
Chris@10
|
593 T3X = FMA(KP248028675, T3G, T3r);
|
Chris@10
|
594 T3R = FNMS(KP554608978, T3Q, T3P);
|
Chris@10
|
595 T3V = FMA(KP554608978, T3Q, T3P);
|
Chris@10
|
596 }
|
Chris@10
|
597 }
|
Chris@10
|
598 }
|
Chris@10
|
599 }
|
Chris@10
|
600 }
|
Chris@10
|
601 T3Z = FMA(KP554608978, T3Y, T3X);
|
Chris@10
|
602 T43 = FNMS(KP554608978, T3Y, T3X);
|
Chris@10
|
603 io[WS(os, 9)] = FNMS(KP803003575, T3W, T3V);
|
Chris@10
|
604 io[WS(os, 24)] = FMA(KP803003575, T3W, T3V);
|
Chris@10
|
605 io[WS(os, 19)] = FNMS(KP943557151, T3U, T3R);
|
Chris@10
|
606 io[WS(os, 14)] = FMA(KP943557151, T3U, T3R);
|
Chris@10
|
607 ro[WS(os, 14)] = FNMS(KP943557151, T44, T43);
|
Chris@10
|
608 ro[WS(os, 19)] = FMA(KP943557151, T44, T43);
|
Chris@10
|
609 ro[WS(os, 24)] = FMA(KP803003575, T42, T3Z);
|
Chris@10
|
610 ro[WS(os, 9)] = FNMS(KP803003575, T42, T3Z);
|
Chris@10
|
611 }
|
Chris@10
|
612 }
|
Chris@10
|
613 }
|
Chris@10
|
614
|
Chris@10
|
615 static const kdft_desc desc = { 25, "n1_25", {84, 0, 268, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
616
|
Chris@10
|
617 void X(codelet_n1_25) (planner *p) {
|
Chris@10
|
618 X(kdft_register) (p, n1_25, &desc);
|
Chris@10
|
619 }
|
Chris@10
|
620
|
Chris@10
|
621 #else /* HAVE_FMA */
|
Chris@10
|
622
|
Chris@10
|
623 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 25 -name n1_25 -include n.h */
|
Chris@10
|
624
|
Chris@10
|
625 /*
|
Chris@10
|
626 * This function contains 352 FP additions, 184 FP multiplications,
|
Chris@10
|
627 * (or, 260 additions, 92 multiplications, 92 fused multiply/add),
|
Chris@10
|
628 * 101 stack variables, 20 constants, and 100 memory accesses
|
Chris@10
|
629 */
|
Chris@10
|
630 #include "n.h"
|
Chris@10
|
631
|
Chris@10
|
632 static void n1_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
633 {
|
Chris@10
|
634 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
635 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
636 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
637 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
638 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
639 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
640 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
641 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
642 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
643 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
644 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
645 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
646 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
647 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
648 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
649 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
650 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
651 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
652 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
653 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
654 {
|
Chris@10
|
655 INT i;
|
Chris@10
|
656 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(100, is), MAKE_VOLATILE_STRIDE(100, os)) {
|
Chris@10
|
657 E T9, T4u, T2T, TP, T3H, TW, T5y, T3I, T2Q, T4v, Ti, Tr, Ts, T5m, T5n;
|
Chris@10
|
658 E T5v, T18, T4G, T34, T3M, T1G, T4J, T38, T3T, T1v, T4K, T37, T3W, T1j, T4H;
|
Chris@10
|
659 E T35, T3P, TB, TK, TL, T5p, T5q, T5w, T1T, T4N, T3c, T41, T2r, T4Q, T3e;
|
Chris@10
|
660 E T4b, T2g, T4R, T3f, T48, T24, T4O, T3b, T44;
|
Chris@10
|
661 {
|
Chris@10
|
662 E T1, T4, T7, T8, T2S, T2R, TN, TO;
|
Chris@10
|
663 T1 = ri[0];
|
Chris@10
|
664 {
|
Chris@10
|
665 E T2, T3, T5, T6;
|
Chris@10
|
666 T2 = ri[WS(is, 5)];
|
Chris@10
|
667 T3 = ri[WS(is, 20)];
|
Chris@10
|
668 T4 = T2 + T3;
|
Chris@10
|
669 T5 = ri[WS(is, 10)];
|
Chris@10
|
670 T6 = ri[WS(is, 15)];
|
Chris@10
|
671 T7 = T5 + T6;
|
Chris@10
|
672 T8 = T4 + T7;
|
Chris@10
|
673 T2S = T5 - T6;
|
Chris@10
|
674 T2R = T2 - T3;
|
Chris@10
|
675 }
|
Chris@10
|
676 T9 = T1 + T8;
|
Chris@10
|
677 T4u = FNMS(KP587785252, T2R, KP951056516 * T2S);
|
Chris@10
|
678 T2T = FMA(KP951056516, T2R, KP587785252 * T2S);
|
Chris@10
|
679 TN = KP559016994 * (T4 - T7);
|
Chris@10
|
680 TO = FNMS(KP250000000, T8, T1);
|
Chris@10
|
681 TP = TN + TO;
|
Chris@10
|
682 T3H = TO - TN;
|
Chris@10
|
683 }
|
Chris@10
|
684 {
|
Chris@10
|
685 E T2N, T2K, T2L, TS, T2O, TV, T2M, T2P;
|
Chris@10
|
686 T2N = ii[0];
|
Chris@10
|
687 {
|
Chris@10
|
688 E TQ, TR, TT, TU;
|
Chris@10
|
689 TQ = ii[WS(is, 5)];
|
Chris@10
|
690 TR = ii[WS(is, 20)];
|
Chris@10
|
691 T2K = TQ + TR;
|
Chris@10
|
692 TT = ii[WS(is, 10)];
|
Chris@10
|
693 TU = ii[WS(is, 15)];
|
Chris@10
|
694 T2L = TT + TU;
|
Chris@10
|
695 TS = TQ - TR;
|
Chris@10
|
696 T2O = T2K + T2L;
|
Chris@10
|
697 TV = TT - TU;
|
Chris@10
|
698 }
|
Chris@10
|
699 TW = FMA(KP951056516, TS, KP587785252 * TV);
|
Chris@10
|
700 T5y = T2N + T2O;
|
Chris@10
|
701 T3I = FNMS(KP587785252, TS, KP951056516 * TV);
|
Chris@10
|
702 T2M = KP559016994 * (T2K - T2L);
|
Chris@10
|
703 T2P = FNMS(KP250000000, T2O, T2N);
|
Chris@10
|
704 T2Q = T2M + T2P;
|
Chris@10
|
705 T4v = T2P - T2M;
|
Chris@10
|
706 }
|
Chris@10
|
707 {
|
Chris@10
|
708 E Ta, T1c, Tj, T1z, Th, T1h, TY, T1g, T13, T1d, T16, T1b, Tq, T1E, T1l;
|
Chris@10
|
709 E T1D, T1q, T1A, T1t, T1y;
|
Chris@10
|
710 Ta = ri[WS(is, 1)];
|
Chris@10
|
711 T1c = ii[WS(is, 1)];
|
Chris@10
|
712 Tj = ri[WS(is, 4)];
|
Chris@10
|
713 T1z = ii[WS(is, 4)];
|
Chris@10
|
714 {
|
Chris@10
|
715 E Tb, Tc, Td, Te, Tf, Tg;
|
Chris@10
|
716 Tb = ri[WS(is, 6)];
|
Chris@10
|
717 Tc = ri[WS(is, 21)];
|
Chris@10
|
718 Td = Tb + Tc;
|
Chris@10
|
719 Te = ri[WS(is, 11)];
|
Chris@10
|
720 Tf = ri[WS(is, 16)];
|
Chris@10
|
721 Tg = Te + Tf;
|
Chris@10
|
722 Th = Td + Tg;
|
Chris@10
|
723 T1h = Te - Tf;
|
Chris@10
|
724 TY = KP559016994 * (Td - Tg);
|
Chris@10
|
725 T1g = Tb - Tc;
|
Chris@10
|
726 }
|
Chris@10
|
727 {
|
Chris@10
|
728 E T11, T12, T19, T14, T15, T1a;
|
Chris@10
|
729 T11 = ii[WS(is, 6)];
|
Chris@10
|
730 T12 = ii[WS(is, 21)];
|
Chris@10
|
731 T19 = T11 + T12;
|
Chris@10
|
732 T14 = ii[WS(is, 11)];
|
Chris@10
|
733 T15 = ii[WS(is, 16)];
|
Chris@10
|
734 T1a = T14 + T15;
|
Chris@10
|
735 T13 = T11 - T12;
|
Chris@10
|
736 T1d = T19 + T1a;
|
Chris@10
|
737 T16 = T14 - T15;
|
Chris@10
|
738 T1b = KP559016994 * (T19 - T1a);
|
Chris@10
|
739 }
|
Chris@10
|
740 {
|
Chris@10
|
741 E Tk, Tl, Tm, Tn, To, Tp;
|
Chris@10
|
742 Tk = ri[WS(is, 9)];
|
Chris@10
|
743 Tl = ri[WS(is, 24)];
|
Chris@10
|
744 Tm = Tk + Tl;
|
Chris@10
|
745 Tn = ri[WS(is, 14)];
|
Chris@10
|
746 To = ri[WS(is, 19)];
|
Chris@10
|
747 Tp = Tn + To;
|
Chris@10
|
748 Tq = Tm + Tp;
|
Chris@10
|
749 T1E = Tn - To;
|
Chris@10
|
750 T1l = KP559016994 * (Tm - Tp);
|
Chris@10
|
751 T1D = Tk - Tl;
|
Chris@10
|
752 }
|
Chris@10
|
753 {
|
Chris@10
|
754 E T1o, T1p, T1w, T1r, T1s, T1x;
|
Chris@10
|
755 T1o = ii[WS(is, 9)];
|
Chris@10
|
756 T1p = ii[WS(is, 24)];
|
Chris@10
|
757 T1w = T1o + T1p;
|
Chris@10
|
758 T1r = ii[WS(is, 14)];
|
Chris@10
|
759 T1s = ii[WS(is, 19)];
|
Chris@10
|
760 T1x = T1r + T1s;
|
Chris@10
|
761 T1q = T1o - T1p;
|
Chris@10
|
762 T1A = T1w + T1x;
|
Chris@10
|
763 T1t = T1r - T1s;
|
Chris@10
|
764 T1y = KP559016994 * (T1w - T1x);
|
Chris@10
|
765 }
|
Chris@10
|
766 Ti = Ta + Th;
|
Chris@10
|
767 Tr = Tj + Tq;
|
Chris@10
|
768 Ts = Ti + Tr;
|
Chris@10
|
769 T5m = T1c + T1d;
|
Chris@10
|
770 T5n = T1z + T1A;
|
Chris@10
|
771 T5v = T5m + T5n;
|
Chris@10
|
772 {
|
Chris@10
|
773 E T17, T3L, T10, T3K, TZ;
|
Chris@10
|
774 T17 = FMA(KP951056516, T13, KP587785252 * T16);
|
Chris@10
|
775 T3L = FNMS(KP587785252, T13, KP951056516 * T16);
|
Chris@10
|
776 TZ = FNMS(KP250000000, Th, Ta);
|
Chris@10
|
777 T10 = TY + TZ;
|
Chris@10
|
778 T3K = TZ - TY;
|
Chris@10
|
779 T18 = T10 + T17;
|
Chris@10
|
780 T4G = T3K + T3L;
|
Chris@10
|
781 T34 = T10 - T17;
|
Chris@10
|
782 T3M = T3K - T3L;
|
Chris@10
|
783 }
|
Chris@10
|
784 {
|
Chris@10
|
785 E T1F, T3R, T1C, T3S, T1B;
|
Chris@10
|
786 T1F = FMA(KP951056516, T1D, KP587785252 * T1E);
|
Chris@10
|
787 T3R = FNMS(KP587785252, T1D, KP951056516 * T1E);
|
Chris@10
|
788 T1B = FNMS(KP250000000, T1A, T1z);
|
Chris@10
|
789 T1C = T1y + T1B;
|
Chris@10
|
790 T3S = T1B - T1y;
|
Chris@10
|
791 T1G = T1C - T1F;
|
Chris@10
|
792 T4J = T3S - T3R;
|
Chris@10
|
793 T38 = T1F + T1C;
|
Chris@10
|
794 T3T = T3R + T3S;
|
Chris@10
|
795 }
|
Chris@10
|
796 {
|
Chris@10
|
797 E T1u, T3V, T1n, T3U, T1m;
|
Chris@10
|
798 T1u = FMA(KP951056516, T1q, KP587785252 * T1t);
|
Chris@10
|
799 T3V = FNMS(KP587785252, T1q, KP951056516 * T1t);
|
Chris@10
|
800 T1m = FNMS(KP250000000, Tq, Tj);
|
Chris@10
|
801 T1n = T1l + T1m;
|
Chris@10
|
802 T3U = T1m - T1l;
|
Chris@10
|
803 T1v = T1n + T1u;
|
Chris@10
|
804 T4K = T3U + T3V;
|
Chris@10
|
805 T37 = T1n - T1u;
|
Chris@10
|
806 T3W = T3U - T3V;
|
Chris@10
|
807 }
|
Chris@10
|
808 {
|
Chris@10
|
809 E T1i, T3N, T1f, T3O, T1e;
|
Chris@10
|
810 T1i = FMA(KP951056516, T1g, KP587785252 * T1h);
|
Chris@10
|
811 T3N = FNMS(KP587785252, T1g, KP951056516 * T1h);
|
Chris@10
|
812 T1e = FNMS(KP250000000, T1d, T1c);
|
Chris@10
|
813 T1f = T1b + T1e;
|
Chris@10
|
814 T3O = T1e - T1b;
|
Chris@10
|
815 T1j = T1f - T1i;
|
Chris@10
|
816 T4H = T3O - T3N;
|
Chris@10
|
817 T35 = T1i + T1f;
|
Chris@10
|
818 T3P = T3N + T3O;
|
Chris@10
|
819 }
|
Chris@10
|
820 }
|
Chris@10
|
821 {
|
Chris@10
|
822 E Tt, T1X, TC, T2k, TA, T22, T1J, T21, T1O, T1Y, T1R, T1W, TJ, T2p, T26;
|
Chris@10
|
823 E T2o, T2b, T2l, T2e, T2j;
|
Chris@10
|
824 Tt = ri[WS(is, 2)];
|
Chris@10
|
825 T1X = ii[WS(is, 2)];
|
Chris@10
|
826 TC = ri[WS(is, 3)];
|
Chris@10
|
827 T2k = ii[WS(is, 3)];
|
Chris@10
|
828 {
|
Chris@10
|
829 E Tu, Tv, Tw, Tx, Ty, Tz;
|
Chris@10
|
830 Tu = ri[WS(is, 7)];
|
Chris@10
|
831 Tv = ri[WS(is, 22)];
|
Chris@10
|
832 Tw = Tu + Tv;
|
Chris@10
|
833 Tx = ri[WS(is, 12)];
|
Chris@10
|
834 Ty = ri[WS(is, 17)];
|
Chris@10
|
835 Tz = Tx + Ty;
|
Chris@10
|
836 TA = Tw + Tz;
|
Chris@10
|
837 T22 = Tx - Ty;
|
Chris@10
|
838 T1J = KP559016994 * (Tw - Tz);
|
Chris@10
|
839 T21 = Tu - Tv;
|
Chris@10
|
840 }
|
Chris@10
|
841 {
|
Chris@10
|
842 E T1M, T1N, T1U, T1P, T1Q, T1V;
|
Chris@10
|
843 T1M = ii[WS(is, 7)];
|
Chris@10
|
844 T1N = ii[WS(is, 22)];
|
Chris@10
|
845 T1U = T1M + T1N;
|
Chris@10
|
846 T1P = ii[WS(is, 12)];
|
Chris@10
|
847 T1Q = ii[WS(is, 17)];
|
Chris@10
|
848 T1V = T1P + T1Q;
|
Chris@10
|
849 T1O = T1M - T1N;
|
Chris@10
|
850 T1Y = T1U + T1V;
|
Chris@10
|
851 T1R = T1P - T1Q;
|
Chris@10
|
852 T1W = KP559016994 * (T1U - T1V);
|
Chris@10
|
853 }
|
Chris@10
|
854 {
|
Chris@10
|
855 E TD, TE, TF, TG, TH, TI;
|
Chris@10
|
856 TD = ri[WS(is, 8)];
|
Chris@10
|
857 TE = ri[WS(is, 23)];
|
Chris@10
|
858 TF = TD + TE;
|
Chris@10
|
859 TG = ri[WS(is, 13)];
|
Chris@10
|
860 TH = ri[WS(is, 18)];
|
Chris@10
|
861 TI = TG + TH;
|
Chris@10
|
862 TJ = TF + TI;
|
Chris@10
|
863 T2p = TG - TH;
|
Chris@10
|
864 T26 = KP559016994 * (TF - TI);
|
Chris@10
|
865 T2o = TD - TE;
|
Chris@10
|
866 }
|
Chris@10
|
867 {
|
Chris@10
|
868 E T29, T2a, T2h, T2c, T2d, T2i;
|
Chris@10
|
869 T29 = ii[WS(is, 8)];
|
Chris@10
|
870 T2a = ii[WS(is, 23)];
|
Chris@10
|
871 T2h = T29 + T2a;
|
Chris@10
|
872 T2c = ii[WS(is, 13)];
|
Chris@10
|
873 T2d = ii[WS(is, 18)];
|
Chris@10
|
874 T2i = T2c + T2d;
|
Chris@10
|
875 T2b = T29 - T2a;
|
Chris@10
|
876 T2l = T2h + T2i;
|
Chris@10
|
877 T2e = T2c - T2d;
|
Chris@10
|
878 T2j = KP559016994 * (T2h - T2i);
|
Chris@10
|
879 }
|
Chris@10
|
880 TB = Tt + TA;
|
Chris@10
|
881 TK = TC + TJ;
|
Chris@10
|
882 TL = TB + TK;
|
Chris@10
|
883 T5p = T1X + T1Y;
|
Chris@10
|
884 T5q = T2k + T2l;
|
Chris@10
|
885 T5w = T5p + T5q;
|
Chris@10
|
886 {
|
Chris@10
|
887 E T1S, T40, T1L, T3Z, T1K;
|
Chris@10
|
888 T1S = FMA(KP951056516, T1O, KP587785252 * T1R);
|
Chris@10
|
889 T40 = FNMS(KP587785252, T1O, KP951056516 * T1R);
|
Chris@10
|
890 T1K = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
891 T1L = T1J + T1K;
|
Chris@10
|
892 T3Z = T1K - T1J;
|
Chris@10
|
893 T1T = T1L + T1S;
|
Chris@10
|
894 T4N = T3Z + T40;
|
Chris@10
|
895 T3c = T1L - T1S;
|
Chris@10
|
896 T41 = T3Z - T40;
|
Chris@10
|
897 }
|
Chris@10
|
898 {
|
Chris@10
|
899 E T2q, T49, T2n, T4a, T2m;
|
Chris@10
|
900 T2q = FMA(KP951056516, T2o, KP587785252 * T2p);
|
Chris@10
|
901 T49 = FNMS(KP587785252, T2o, KP951056516 * T2p);
|
Chris@10
|
902 T2m = FNMS(KP250000000, T2l, T2k);
|
Chris@10
|
903 T2n = T2j + T2m;
|
Chris@10
|
904 T4a = T2m - T2j;
|
Chris@10
|
905 T2r = T2n - T2q;
|
Chris@10
|
906 T4Q = T4a - T49;
|
Chris@10
|
907 T3e = T2q + T2n;
|
Chris@10
|
908 T4b = T49 + T4a;
|
Chris@10
|
909 }
|
Chris@10
|
910 {
|
Chris@10
|
911 E T2f, T47, T28, T46, T27;
|
Chris@10
|
912 T2f = FMA(KP951056516, T2b, KP587785252 * T2e);
|
Chris@10
|
913 T47 = FNMS(KP587785252, T2b, KP951056516 * T2e);
|
Chris@10
|
914 T27 = FNMS(KP250000000, TJ, TC);
|
Chris@10
|
915 T28 = T26 + T27;
|
Chris@10
|
916 T46 = T27 - T26;
|
Chris@10
|
917 T2g = T28 + T2f;
|
Chris@10
|
918 T4R = T46 + T47;
|
Chris@10
|
919 T3f = T28 - T2f;
|
Chris@10
|
920 T48 = T46 - T47;
|
Chris@10
|
921 }
|
Chris@10
|
922 {
|
Chris@10
|
923 E T23, T42, T20, T43, T1Z;
|
Chris@10
|
924 T23 = FMA(KP951056516, T21, KP587785252 * T22);
|
Chris@10
|
925 T42 = FNMS(KP587785252, T21, KP951056516 * T22);
|
Chris@10
|
926 T1Z = FNMS(KP250000000, T1Y, T1X);
|
Chris@10
|
927 T20 = T1W + T1Z;
|
Chris@10
|
928 T43 = T1Z - T1W;
|
Chris@10
|
929 T24 = T20 - T23;
|
Chris@10
|
930 T4O = T43 - T42;
|
Chris@10
|
931 T3b = T23 + T20;
|
Chris@10
|
932 T44 = T42 + T43;
|
Chris@10
|
933 }
|
Chris@10
|
934 }
|
Chris@10
|
935 {
|
Chris@10
|
936 E T5j, TM, T5k, T5s, T5u, T5o, T5r, T5t, T5l;
|
Chris@10
|
937 T5j = KP559016994 * (Ts - TL);
|
Chris@10
|
938 TM = Ts + TL;
|
Chris@10
|
939 T5k = FNMS(KP250000000, TM, T9);
|
Chris@10
|
940 T5o = T5m - T5n;
|
Chris@10
|
941 T5r = T5p - T5q;
|
Chris@10
|
942 T5s = FMA(KP951056516, T5o, KP587785252 * T5r);
|
Chris@10
|
943 T5u = FNMS(KP587785252, T5o, KP951056516 * T5r);
|
Chris@10
|
944 ro[0] = T9 + TM;
|
Chris@10
|
945 T5t = T5k - T5j;
|
Chris@10
|
946 ro[WS(os, 10)] = T5t - T5u;
|
Chris@10
|
947 ro[WS(os, 15)] = T5t + T5u;
|
Chris@10
|
948 T5l = T5j + T5k;
|
Chris@10
|
949 ro[WS(os, 20)] = T5l - T5s;
|
Chris@10
|
950 ro[WS(os, 5)] = T5l + T5s;
|
Chris@10
|
951 }
|
Chris@10
|
952 {
|
Chris@10
|
953 E T5x, T5z, T5A, T5E, T5F, T5C, T5D, T5G, T5B;
|
Chris@10
|
954 T5x = KP559016994 * (T5v - T5w);
|
Chris@10
|
955 T5z = T5v + T5w;
|
Chris@10
|
956 T5A = FNMS(KP250000000, T5z, T5y);
|
Chris@10
|
957 T5C = Ti - Tr;
|
Chris@10
|
958 T5D = TB - TK;
|
Chris@10
|
959 T5E = FMA(KP951056516, T5C, KP587785252 * T5D);
|
Chris@10
|
960 T5F = FNMS(KP587785252, T5C, KP951056516 * T5D);
|
Chris@10
|
961 io[0] = T5y + T5z;
|
Chris@10
|
962 T5G = T5A - T5x;
|
Chris@10
|
963 io[WS(os, 10)] = T5F + T5G;
|
Chris@10
|
964 io[WS(os, 15)] = T5G - T5F;
|
Chris@10
|
965 T5B = T5x + T5A;
|
Chris@10
|
966 io[WS(os, 5)] = T5B - T5E;
|
Chris@10
|
967 io[WS(os, 20)] = T5E + T5B;
|
Chris@10
|
968 }
|
Chris@10
|
969 {
|
Chris@10
|
970 E TX, T2U, T2u, T2Z, T2v, T2Y, T2A, T2V, T2D, T2J;
|
Chris@10
|
971 TX = TP + TW;
|
Chris@10
|
972 T2U = T2Q - T2T;
|
Chris@10
|
973 {
|
Chris@10
|
974 E T1k, T1H, T1I, T25, T2s, T2t;
|
Chris@10
|
975 T1k = FMA(KP968583161, T18, KP248689887 * T1j);
|
Chris@10
|
976 T1H = FMA(KP535826794, T1v, KP844327925 * T1G);
|
Chris@10
|
977 T1I = T1k + T1H;
|
Chris@10
|
978 T25 = FMA(KP876306680, T1T, KP481753674 * T24);
|
Chris@10
|
979 T2s = FMA(KP728968627, T2g, KP684547105 * T2r);
|
Chris@10
|
980 T2t = T25 + T2s;
|
Chris@10
|
981 T2u = T1I + T2t;
|
Chris@10
|
982 T2Z = T25 - T2s;
|
Chris@10
|
983 T2v = KP559016994 * (T1I - T2t);
|
Chris@10
|
984 T2Y = T1k - T1H;
|
Chris@10
|
985 }
|
Chris@10
|
986 {
|
Chris@10
|
987 E T2y, T2z, T2H, T2B, T2C, T2I;
|
Chris@10
|
988 T2y = FNMS(KP248689887, T18, KP968583161 * T1j);
|
Chris@10
|
989 T2z = FNMS(KP844327925, T1v, KP535826794 * T1G);
|
Chris@10
|
990 T2H = T2y + T2z;
|
Chris@10
|
991 T2B = FNMS(KP481753674, T1T, KP876306680 * T24);
|
Chris@10
|
992 T2C = FNMS(KP684547105, T2g, KP728968627 * T2r);
|
Chris@10
|
993 T2I = T2B + T2C;
|
Chris@10
|
994 T2A = T2y - T2z;
|
Chris@10
|
995 T2V = T2H + T2I;
|
Chris@10
|
996 T2D = T2B - T2C;
|
Chris@10
|
997 T2J = KP559016994 * (T2H - T2I);
|
Chris@10
|
998 }
|
Chris@10
|
999 ro[WS(os, 1)] = TX + T2u;
|
Chris@10
|
1000 io[WS(os, 1)] = T2U + T2V;
|
Chris@10
|
1001 {
|
Chris@10
|
1002 E T2E, T2G, T2x, T2F, T2w;
|
Chris@10
|
1003 T2E = FMA(KP951056516, T2A, KP587785252 * T2D);
|
Chris@10
|
1004 T2G = FNMS(KP587785252, T2A, KP951056516 * T2D);
|
Chris@10
|
1005 T2w = FNMS(KP250000000, T2u, TX);
|
Chris@10
|
1006 T2x = T2v + T2w;
|
Chris@10
|
1007 T2F = T2w - T2v;
|
Chris@10
|
1008 ro[WS(os, 21)] = T2x - T2E;
|
Chris@10
|
1009 ro[WS(os, 16)] = T2F + T2G;
|
Chris@10
|
1010 ro[WS(os, 6)] = T2x + T2E;
|
Chris@10
|
1011 ro[WS(os, 11)] = T2F - T2G;
|
Chris@10
|
1012 }
|
Chris@10
|
1013 {
|
Chris@10
|
1014 E T30, T31, T2X, T32, T2W;
|
Chris@10
|
1015 T30 = FMA(KP951056516, T2Y, KP587785252 * T2Z);
|
Chris@10
|
1016 T31 = FNMS(KP587785252, T2Y, KP951056516 * T2Z);
|
Chris@10
|
1017 T2W = FNMS(KP250000000, T2V, T2U);
|
Chris@10
|
1018 T2X = T2J + T2W;
|
Chris@10
|
1019 T32 = T2W - T2J;
|
Chris@10
|
1020 io[WS(os, 6)] = T2X - T30;
|
Chris@10
|
1021 io[WS(os, 16)] = T32 - T31;
|
Chris@10
|
1022 io[WS(os, 21)] = T30 + T2X;
|
Chris@10
|
1023 io[WS(os, 11)] = T31 + T32;
|
Chris@10
|
1024 }
|
Chris@10
|
1025 }
|
Chris@10
|
1026 {
|
Chris@10
|
1027 E T4F, T52, T4U, T5b, T56, T57, T51, T5f, T53, T5e;
|
Chris@10
|
1028 T4F = T3H + T3I;
|
Chris@10
|
1029 T52 = T4v - T4u;
|
Chris@10
|
1030 {
|
Chris@10
|
1031 E T4I, T4L, T4M, T4P, T4S, T4T;
|
Chris@10
|
1032 T4I = FMA(KP728968627, T4G, KP684547105 * T4H);
|
Chris@10
|
1033 T4L = FNMS(KP992114701, T4K, KP125333233 * T4J);
|
Chris@10
|
1034 T4M = T4I + T4L;
|
Chris@10
|
1035 T4P = FMA(KP062790519, T4N, KP998026728 * T4O);
|
Chris@10
|
1036 T4S = FNMS(KP637423989, T4R, KP770513242 * T4Q);
|
Chris@10
|
1037 T4T = T4P + T4S;
|
Chris@10
|
1038 T4U = T4M + T4T;
|
Chris@10
|
1039 T5b = KP559016994 * (T4M - T4T);
|
Chris@10
|
1040 T56 = T4I - T4L;
|
Chris@10
|
1041 T57 = T4P - T4S;
|
Chris@10
|
1042 }
|
Chris@10
|
1043 {
|
Chris@10
|
1044 E T4V, T4W, T4X, T4Y, T4Z, T50;
|
Chris@10
|
1045 T4V = FNMS(KP684547105, T4G, KP728968627 * T4H);
|
Chris@10
|
1046 T4W = FMA(KP125333233, T4K, KP992114701 * T4J);
|
Chris@10
|
1047 T4X = T4V - T4W;
|
Chris@10
|
1048 T4Y = FNMS(KP998026728, T4N, KP062790519 * T4O);
|
Chris@10
|
1049 T4Z = FMA(KP770513242, T4R, KP637423989 * T4Q);
|
Chris@10
|
1050 T50 = T4Y - T4Z;
|
Chris@10
|
1051 T51 = KP559016994 * (T4X - T50);
|
Chris@10
|
1052 T5f = T4Y + T4Z;
|
Chris@10
|
1053 T53 = T4X + T50;
|
Chris@10
|
1054 T5e = T4V + T4W;
|
Chris@10
|
1055 }
|
Chris@10
|
1056 ro[WS(os, 3)] = T4F + T4U;
|
Chris@10
|
1057 io[WS(os, 3)] = T52 + T53;
|
Chris@10
|
1058 {
|
Chris@10
|
1059 E T58, T59, T55, T5a, T54;
|
Chris@10
|
1060 T58 = FMA(KP951056516, T56, KP587785252 * T57);
|
Chris@10
|
1061 T59 = FNMS(KP587785252, T56, KP951056516 * T57);
|
Chris@10
|
1062 T54 = FNMS(KP250000000, T53, T52);
|
Chris@10
|
1063 T55 = T51 + T54;
|
Chris@10
|
1064 T5a = T54 - T51;
|
Chris@10
|
1065 io[WS(os, 8)] = T55 - T58;
|
Chris@10
|
1066 io[WS(os, 18)] = T5a - T59;
|
Chris@10
|
1067 io[WS(os, 23)] = T58 + T55;
|
Chris@10
|
1068 io[WS(os, 13)] = T59 + T5a;
|
Chris@10
|
1069 }
|
Chris@10
|
1070 {
|
Chris@10
|
1071 E T5g, T5i, T5d, T5h, T5c;
|
Chris@10
|
1072 T5g = FMA(KP951056516, T5e, KP587785252 * T5f);
|
Chris@10
|
1073 T5i = FNMS(KP587785252, T5e, KP951056516 * T5f);
|
Chris@10
|
1074 T5c = FNMS(KP250000000, T4U, T4F);
|
Chris@10
|
1075 T5d = T5b + T5c;
|
Chris@10
|
1076 T5h = T5c - T5b;
|
Chris@10
|
1077 ro[WS(os, 23)] = T5d - T5g;
|
Chris@10
|
1078 ro[WS(os, 18)] = T5h + T5i;
|
Chris@10
|
1079 ro[WS(os, 8)] = T5d + T5g;
|
Chris@10
|
1080 ro[WS(os, 13)] = T5h - T5i;
|
Chris@10
|
1081 }
|
Chris@10
|
1082 }
|
Chris@10
|
1083 {
|
Chris@10
|
1084 E T3J, T4w, T4e, T4B, T4f, T4A, T4k, T4x, T4n, T4t;
|
Chris@10
|
1085 T3J = T3H - T3I;
|
Chris@10
|
1086 T4w = T4u + T4v;
|
Chris@10
|
1087 {
|
Chris@10
|
1088 E T3Q, T3X, T3Y, T45, T4c, T4d;
|
Chris@10
|
1089 T3Q = FMA(KP876306680, T3M, KP481753674 * T3P);
|
Chris@10
|
1090 T3X = FNMS(KP425779291, T3W, KP904827052 * T3T);
|
Chris@10
|
1091 T3Y = T3Q + T3X;
|
Chris@10
|
1092 T45 = FMA(KP535826794, T41, KP844327925 * T44);
|
Chris@10
|
1093 T4c = FMA(KP062790519, T48, KP998026728 * T4b);
|
Chris@10
|
1094 T4d = T45 + T4c;
|
Chris@10
|
1095 T4e = T3Y + T4d;
|
Chris@10
|
1096 T4B = T45 - T4c;
|
Chris@10
|
1097 T4f = KP559016994 * (T3Y - T4d);
|
Chris@10
|
1098 T4A = T3Q - T3X;
|
Chris@10
|
1099 }
|
Chris@10
|
1100 {
|
Chris@10
|
1101 E T4i, T4j, T4r, T4l, T4m, T4s;
|
Chris@10
|
1102 T4i = FNMS(KP481753674, T3M, KP876306680 * T3P);
|
Chris@10
|
1103 T4j = FMA(KP904827052, T3W, KP425779291 * T3T);
|
Chris@10
|
1104 T4r = T4i - T4j;
|
Chris@10
|
1105 T4l = FNMS(KP844327925, T41, KP535826794 * T44);
|
Chris@10
|
1106 T4m = FNMS(KP998026728, T48, KP062790519 * T4b);
|
Chris@10
|
1107 T4s = T4l + T4m;
|
Chris@10
|
1108 T4k = T4i + T4j;
|
Chris@10
|
1109 T4x = T4r + T4s;
|
Chris@10
|
1110 T4n = T4l - T4m;
|
Chris@10
|
1111 T4t = KP559016994 * (T4r - T4s);
|
Chris@10
|
1112 }
|
Chris@10
|
1113 ro[WS(os, 2)] = T3J + T4e;
|
Chris@10
|
1114 io[WS(os, 2)] = T4w + T4x;
|
Chris@10
|
1115 {
|
Chris@10
|
1116 E T4o, T4q, T4h, T4p, T4g;
|
Chris@10
|
1117 T4o = FMA(KP951056516, T4k, KP587785252 * T4n);
|
Chris@10
|
1118 T4q = FNMS(KP587785252, T4k, KP951056516 * T4n);
|
Chris@10
|
1119 T4g = FNMS(KP250000000, T4e, T3J);
|
Chris@10
|
1120 T4h = T4f + T4g;
|
Chris@10
|
1121 T4p = T4g - T4f;
|
Chris@10
|
1122 ro[WS(os, 22)] = T4h - T4o;
|
Chris@10
|
1123 ro[WS(os, 17)] = T4p + T4q;
|
Chris@10
|
1124 ro[WS(os, 7)] = T4h + T4o;
|
Chris@10
|
1125 ro[WS(os, 12)] = T4p - T4q;
|
Chris@10
|
1126 }
|
Chris@10
|
1127 {
|
Chris@10
|
1128 E T4C, T4D, T4z, T4E, T4y;
|
Chris@10
|
1129 T4C = FMA(KP951056516, T4A, KP587785252 * T4B);
|
Chris@10
|
1130 T4D = FNMS(KP587785252, T4A, KP951056516 * T4B);
|
Chris@10
|
1131 T4y = FNMS(KP250000000, T4x, T4w);
|
Chris@10
|
1132 T4z = T4t + T4y;
|
Chris@10
|
1133 T4E = T4y - T4t;
|
Chris@10
|
1134 io[WS(os, 7)] = T4z - T4C;
|
Chris@10
|
1135 io[WS(os, 17)] = T4E - T4D;
|
Chris@10
|
1136 io[WS(os, 22)] = T4C + T4z;
|
Chris@10
|
1137 io[WS(os, 12)] = T4D + T4E;
|
Chris@10
|
1138 }
|
Chris@10
|
1139 }
|
Chris@10
|
1140 {
|
Chris@10
|
1141 E T33, T3j, T3i, T3z, T3r, T3s, T3q, T3D, T3v, T3C;
|
Chris@10
|
1142 T33 = TP - TW;
|
Chris@10
|
1143 T3j = T2T + T2Q;
|
Chris@10
|
1144 {
|
Chris@10
|
1145 E T36, T39, T3a, T3d, T3g, T3h;
|
Chris@10
|
1146 T36 = FMA(KP535826794, T34, KP844327925 * T35);
|
Chris@10
|
1147 T39 = FMA(KP637423989, T37, KP770513242 * T38);
|
Chris@10
|
1148 T3a = T36 - T39;
|
Chris@10
|
1149 T3d = FNMS(KP425779291, T3c, KP904827052 * T3b);
|
Chris@10
|
1150 T3g = FNMS(KP992114701, T3f, KP125333233 * T3e);
|
Chris@10
|
1151 T3h = T3d + T3g;
|
Chris@10
|
1152 T3i = T3a + T3h;
|
Chris@10
|
1153 T3z = KP559016994 * (T3a - T3h);
|
Chris@10
|
1154 T3r = T3d - T3g;
|
Chris@10
|
1155 T3s = T36 + T39;
|
Chris@10
|
1156 }
|
Chris@10
|
1157 {
|
Chris@10
|
1158 E T3k, T3l, T3m, T3n, T3o, T3p;
|
Chris@10
|
1159 T3k = FNMS(KP844327925, T34, KP535826794 * T35);
|
Chris@10
|
1160 T3l = FNMS(KP637423989, T38, KP770513242 * T37);
|
Chris@10
|
1161 T3m = T3k + T3l;
|
Chris@10
|
1162 T3n = FMA(KP904827052, T3c, KP425779291 * T3b);
|
Chris@10
|
1163 T3o = FMA(KP125333233, T3f, KP992114701 * T3e);
|
Chris@10
|
1164 T3p = T3n + T3o;
|
Chris@10
|
1165 T3q = T3m - T3p;
|
Chris@10
|
1166 T3D = T3o - T3n;
|
Chris@10
|
1167 T3v = KP559016994 * (T3m + T3p);
|
Chris@10
|
1168 T3C = T3k - T3l;
|
Chris@10
|
1169 }
|
Chris@10
|
1170 ro[WS(os, 4)] = T33 + T3i;
|
Chris@10
|
1171 io[WS(os, 4)] = T3j + T3q;
|
Chris@10
|
1172 {
|
Chris@10
|
1173 E T3t, T3y, T3w, T3x, T3u;
|
Chris@10
|
1174 T3t = FNMS(KP587785252, T3s, KP951056516 * T3r);
|
Chris@10
|
1175 T3y = FMA(KP951056516, T3s, KP587785252 * T3r);
|
Chris@10
|
1176 T3u = FNMS(KP250000000, T3q, T3j);
|
Chris@10
|
1177 T3w = T3u - T3v;
|
Chris@10
|
1178 T3x = T3u + T3v;
|
Chris@10
|
1179 io[WS(os, 14)] = T3t + T3w;
|
Chris@10
|
1180 io[WS(os, 24)] = T3y + T3x;
|
Chris@10
|
1181 io[WS(os, 19)] = T3w - T3t;
|
Chris@10
|
1182 io[WS(os, 9)] = T3x - T3y;
|
Chris@10
|
1183 }
|
Chris@10
|
1184 {
|
Chris@10
|
1185 E T3E, T3G, T3B, T3F, T3A;
|
Chris@10
|
1186 T3E = FMA(KP951056516, T3C, KP587785252 * T3D);
|
Chris@10
|
1187 T3G = FNMS(KP587785252, T3C, KP951056516 * T3D);
|
Chris@10
|
1188 T3A = FNMS(KP250000000, T3i, T33);
|
Chris@10
|
1189 T3B = T3z + T3A;
|
Chris@10
|
1190 T3F = T3A - T3z;
|
Chris@10
|
1191 ro[WS(os, 24)] = T3B - T3E;
|
Chris@10
|
1192 ro[WS(os, 19)] = T3F + T3G;
|
Chris@10
|
1193 ro[WS(os, 9)] = T3B + T3E;
|
Chris@10
|
1194 ro[WS(os, 14)] = T3F - T3G;
|
Chris@10
|
1195 }
|
Chris@10
|
1196 }
|
Chris@10
|
1197 }
|
Chris@10
|
1198 }
|
Chris@10
|
1199 }
|
Chris@10
|
1200
|
Chris@10
|
1201 static const kdft_desc desc = { 25, "n1_25", {260, 92, 92, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
1202
|
Chris@10
|
1203 void X(codelet_n1_25) (planner *p) {
|
Chris@10
|
1204 X(kdft_register) (p, n1_25, &desc);
|
Chris@10
|
1205 }
|
Chris@10
|
1206
|
Chris@10
|
1207 #endif /* HAVE_FMA */
|