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 07:41:30 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 25 -dif -name hb2_25 -include hb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 440 FP additions, 434 FP multiplications,
|
Chris@10
|
32 * (or, 84 additions, 78 multiplications, 356 fused multiply/add),
|
Chris@10
|
33 * 234 stack variables, 47 constants, and 100 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hb.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hb2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP833417178, +0.833417178328688677408962550243238843138996060);
|
Chris@10
|
40 DK(KP921177326, +0.921177326965143320250447435415066029359282231);
|
Chris@10
|
41 DK(KP541454447, +0.541454447536312777046285590082819509052033189);
|
Chris@10
|
42 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
43 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
Chris@10
|
44 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
45 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
Chris@10
|
46 DK(KP554608978, +0.554608978404018097464974850792216217022558774);
|
Chris@10
|
47 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
48 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
49 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
Chris@10
|
50 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
51 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
52 DK(KP871714437, +0.871714437527667770979999223229522602943903653);
|
Chris@10
|
53 DK(KP851038619, +0.851038619207379630836264138867114231259902550);
|
Chris@10
|
54 DK(KP943557151, +0.943557151597354104399655195398983005179443399);
|
Chris@10
|
55 DK(KP726211448, +0.726211448929902658173535992263577167607493062);
|
Chris@10
|
56 DK(KP525970792, +0.525970792408939708442463226536226366643874659);
|
Chris@10
|
57 DK(KP912018591, +0.912018591466481957908415381764119056233607330);
|
Chris@10
|
58 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
59 DK(KP994076283, +0.994076283785401014123185814696322018529298887);
|
Chris@10
|
60 DK(KP614372930, +0.614372930789563808870829930444362096004872855);
|
Chris@10
|
61 DK(KP621716863, +0.621716863012209892444754556304102309693593202);
|
Chris@10
|
62 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
Chris@10
|
63 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
Chris@10
|
64 DK(KP557913902, +0.557913902031834264187699648465567037992437152);
|
Chris@10
|
65 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
66 DK(KP249506682, +0.249506682107067890488084201715862638334226305);
|
Chris@10
|
67 DK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
Chris@10
|
68 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
Chris@10
|
69 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
Chris@10
|
70 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
Chris@10
|
71 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
Chris@10
|
72 DK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
Chris@10
|
73 DK(KP968479752, +0.968479752739016373193524836781420152702090879);
|
Chris@10
|
74 DK(KP062914667, +0.062914667253649757225485955897349402364686947);
|
Chris@10
|
75 DK(KP827271945, +0.827271945972475634034355757144307982555673741);
|
Chris@10
|
76 DK(KP470564281, +0.470564281212251493087595091036643380879947982);
|
Chris@10
|
77 DK(KP126329378, +0.126329378446108174786050455341811215027378105);
|
Chris@10
|
78 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
Chris@10
|
79 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
Chris@10
|
80 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
Chris@10
|
81 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
Chris@10
|
82 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
83 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
84 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
85 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
86 {
|
Chris@10
|
87 INT m;
|
Chris@10
|
88 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(50, rs)) {
|
Chris@10
|
89 E TN, TQ, T4e, T2y, T4i, T3U, T4u, T4o, T4G, T4C, T2F, T41, T3Q, T4q, T3a;
|
Chris@10
|
90 E T3F, T4a, T4w, T46, T44;
|
Chris@10
|
91 {
|
Chris@10
|
92 E TT, TO, TR, T23, T2d, T2x, TP, TV, T2p, T85, T4d, T25, TX;
|
Chris@10
|
93 TN = W[0];
|
Chris@10
|
94 TT = W[4];
|
Chris@10
|
95 TO = W[2];
|
Chris@10
|
96 TR = W[3];
|
Chris@10
|
97 T23 = W[6];
|
Chris@10
|
98 T2d = TN * TT;
|
Chris@10
|
99 T2x = TO * TT;
|
Chris@10
|
100 TP = TN * TO;
|
Chris@10
|
101 TV = TN * TR;
|
Chris@10
|
102 T2p = TT * T23;
|
Chris@10
|
103 T85 = TN * T23;
|
Chris@10
|
104 T4d = TO * T23;
|
Chris@10
|
105 T25 = W[7];
|
Chris@10
|
106 TQ = W[1];
|
Chris@10
|
107 TX = W[5];
|
Chris@10
|
108 {
|
Chris@10
|
109 E T86, T4n, TW, T4l, TS, T71, T2q, T4z, T2e, T8a, T2u, T76, T2k, T4B, T6E;
|
Chris@10
|
110 E T6U, T6Y, T5T, T8i, T1I, T2a, T26, TY, T8d, T8s, T8o, T5C, T5w, T7g, T7c;
|
Chris@10
|
111 E T5M, T5I, T9, T40, T1R, T3X, T6H, T7F, T5W, T7n, T4N, T68, T1S, T1k, T1T;
|
Chris@10
|
112 E T1D, T1Y, T1Z, T10, TM, T7K, T7A, T6p, T6w, T4X, T56, T3K, T2U, T7x, T7J;
|
Chris@10
|
113 E T6v, T6i, T50, T57, T3L, T39, T4Q, T59, T3O, T3E, T67, T7t, T7H, T6y, T63;
|
Chris@10
|
114 E T4T, T5a, T3N, T3p, T66, T7o;
|
Chris@10
|
115 {
|
Chris@10
|
116 E T2A, T2z, T6G, T2E, T5V, T6F;
|
Chris@10
|
117 {
|
Chris@10
|
118 E T1, T1J, T3Y, T3Z, T8, T2C, T1M, T1P, T2D, T4h, T89, T2t, T3W, T1Q, T3V;
|
Chris@10
|
119 T1 = cr[0];
|
Chris@10
|
120 T4e = FMA(TR, T25, T4d);
|
Chris@10
|
121 T4h = TO * T25;
|
Chris@10
|
122 T89 = TN * T25;
|
Chris@10
|
123 T2t = TT * T25;
|
Chris@10
|
124 T86 = FMA(TQ, T25, T85);
|
Chris@10
|
125 T4n = FNMS(TQ, TO, TV);
|
Chris@10
|
126 TW = FMA(TQ, TO, TV);
|
Chris@10
|
127 T4l = FMA(TQ, TR, TP);
|
Chris@10
|
128 TS = FNMS(TQ, TR, TP);
|
Chris@10
|
129 T71 = FNMS(TR, TX, T2x);
|
Chris@10
|
130 T2y = FMA(TR, TX, T2x);
|
Chris@10
|
131 T2q = FMA(TX, T25, T2p);
|
Chris@10
|
132 T4z = FMA(TQ, TX, T2d);
|
Chris@10
|
133 T2e = FNMS(TQ, TX, T2d);
|
Chris@10
|
134 {
|
Chris@10
|
135 E T3T, T2j, T4t, T6T;
|
Chris@10
|
136 T3T = TO * TX;
|
Chris@10
|
137 T2j = TN * TX;
|
Chris@10
|
138 T4i = FNMS(TR, T23, T4h);
|
Chris@10
|
139 T8a = FNMS(TQ, T23, T89);
|
Chris@10
|
140 T2u = FNMS(TX, T23, T2t);
|
Chris@10
|
141 T4t = T4l * TX;
|
Chris@10
|
142 T6T = T4l * T23;
|
Chris@10
|
143 {
|
Chris@10
|
144 E T6X, T4m, T1H, T29;
|
Chris@10
|
145 T6X = T4l * T25;
|
Chris@10
|
146 T4m = T4l * TT;
|
Chris@10
|
147 T1H = TS * TX;
|
Chris@10
|
148 T29 = TS * T25;
|
Chris@10
|
149 {
|
Chris@10
|
150 E T24, TU, T4F, T4A;
|
Chris@10
|
151 T24 = TS * T23;
|
Chris@10
|
152 TU = TS * TT;
|
Chris@10
|
153 T4F = T4z * T25;
|
Chris@10
|
154 T4A = T4z * T23;
|
Chris@10
|
155 {
|
Chris@10
|
156 E T8r, T8n, T5B, T5v;
|
Chris@10
|
157 T8r = T2y * T25;
|
Chris@10
|
158 T8n = T2y * T23;
|
Chris@10
|
159 T5B = T2e * T25;
|
Chris@10
|
160 T5v = T2e * T23;
|
Chris@10
|
161 T3U = FNMS(TR, TT, T3T);
|
Chris@10
|
162 T76 = FMA(TR, TT, T3T);
|
Chris@10
|
163 T2k = FMA(TQ, TT, T2j);
|
Chris@10
|
164 T4B = FNMS(TQ, TT, T2j);
|
Chris@10
|
165 T4u = FMA(T4n, TT, T4t);
|
Chris@10
|
166 T6E = FNMS(T4n, TT, T4t);
|
Chris@10
|
167 T6U = FMA(T4n, T25, T6T);
|
Chris@10
|
168 T6Y = FNMS(T4n, T23, T6X);
|
Chris@10
|
169 T5T = FMA(T4n, TX, T4m);
|
Chris@10
|
170 T4o = FNMS(T4n, TX, T4m);
|
Chris@10
|
171 T8i = FMA(TW, TT, T1H);
|
Chris@10
|
172 T1I = FNMS(TW, TT, T1H);
|
Chris@10
|
173 T2a = FNMS(TW, T23, T29);
|
Chris@10
|
174 T26 = FMA(TW, T25, T24);
|
Chris@10
|
175 TY = FMA(TW, TX, TU);
|
Chris@10
|
176 T8d = FNMS(TW, TX, TU);
|
Chris@10
|
177 T8s = FNMS(T3U, T23, T8r);
|
Chris@10
|
178 T8o = FMA(T3U, T25, T8n);
|
Chris@10
|
179 T5C = FNMS(T2k, T23, T5B);
|
Chris@10
|
180 T5w = FMA(T2k, T25, T5v);
|
Chris@10
|
181 T4G = FNMS(T4B, T23, T4F);
|
Chris@10
|
182 T4C = FMA(T4B, T25, T4A);
|
Chris@10
|
183 {
|
Chris@10
|
184 E T7f, T7b, T5L, T5H;
|
Chris@10
|
185 T7f = T5T * T25;
|
Chris@10
|
186 T7b = T5T * T23;
|
Chris@10
|
187 T5L = TY * T25;
|
Chris@10
|
188 T5H = TY * T23;
|
Chris@10
|
189 T7g = FNMS(T6E, T23, T7f);
|
Chris@10
|
190 T7c = FMA(T6E, T25, T7b);
|
Chris@10
|
191 T5M = FNMS(T1I, T23, T5L);
|
Chris@10
|
192 T5I = FMA(T1I, T25, T5H);
|
Chris@10
|
193 T1J = ci[WS(rs, 24)];
|
Chris@10
|
194 }
|
Chris@10
|
195 }
|
Chris@10
|
196 }
|
Chris@10
|
197 }
|
Chris@10
|
198 }
|
Chris@10
|
199 {
|
Chris@10
|
200 E T2, T3, T5, T6;
|
Chris@10
|
201 T2 = cr[WS(rs, 5)];
|
Chris@10
|
202 T3 = ci[WS(rs, 4)];
|
Chris@10
|
203 T5 = cr[WS(rs, 10)];
|
Chris@10
|
204 T6 = ci[WS(rs, 9)];
|
Chris@10
|
205 {
|
Chris@10
|
206 E T1K, T4, T7, T1L, T1N, T1O;
|
Chris@10
|
207 T1K = ci[WS(rs, 19)];
|
Chris@10
|
208 T3Y = T2 - T3;
|
Chris@10
|
209 T4 = T2 + T3;
|
Chris@10
|
210 T3Z = T5 - T6;
|
Chris@10
|
211 T7 = T5 + T6;
|
Chris@10
|
212 T1L = cr[WS(rs, 20)];
|
Chris@10
|
213 T1N = ci[WS(rs, 14)];
|
Chris@10
|
214 T1O = cr[WS(rs, 15)];
|
Chris@10
|
215 T8 = T4 + T7;
|
Chris@10
|
216 T2A = T4 - T7;
|
Chris@10
|
217 T2C = T1K + T1L;
|
Chris@10
|
218 T1M = T1K - T1L;
|
Chris@10
|
219 T1P = T1N - T1O;
|
Chris@10
|
220 T2D = T1N + T1O;
|
Chris@10
|
221 }
|
Chris@10
|
222 }
|
Chris@10
|
223 T2z = FNMS(KP250000000, T8, T1);
|
Chris@10
|
224 T9 = T1 + T8;
|
Chris@10
|
225 T3W = T1M - T1P;
|
Chris@10
|
226 T1Q = T1M + T1P;
|
Chris@10
|
227 T40 = FMA(KP618033988, T3Z, T3Y);
|
Chris@10
|
228 T6G = FNMS(KP618033988, T3Y, T3Z);
|
Chris@10
|
229 T2E = FMA(KP618033988, T2D, T2C);
|
Chris@10
|
230 T5V = FNMS(KP618033988, T2C, T2D);
|
Chris@10
|
231 T1R = T1J + T1Q;
|
Chris@10
|
232 T3V = FNMS(KP250000000, T1Q, T1J);
|
Chris@10
|
233 T6F = FNMS(KP559016994, T3W, T3V);
|
Chris@10
|
234 T3X = FMA(KP559016994, T3W, T3V);
|
Chris@10
|
235 }
|
Chris@10
|
236 {
|
Chris@10
|
237 E T2S, T6n, T2H, T2G, Ti, T5Y, T3C, T3r, TK, T3q, T30, T6d, T33, Tr, T32;
|
Chris@10
|
238 E T3v, T61, T3y, T1C, T3x, T2L, T6k, T2O, T1a, T2N, T6g, T37, T2W, Tt, T1j;
|
Chris@10
|
239 E T2V, Tx, T3g, T3j, Tw, T3l, T1t, T3i, Ty;
|
Chris@10
|
240 {
|
Chris@10
|
241 E T1u, T1v, T1A, T3u, T1w;
|
Chris@10
|
242 {
|
Chris@10
|
243 E TC, TI, T3B, TD, TE;
|
Chris@10
|
244 {
|
Chris@10
|
245 E Ta, Te, Tf, Tb, Tc, T5U, T2B, T2R, Tg;
|
Chris@10
|
246 Ta = cr[WS(rs, 1)];
|
Chris@10
|
247 T5U = FNMS(KP559016994, T2A, T2z);
|
Chris@10
|
248 T2B = FMA(KP559016994, T2A, T2z);
|
Chris@10
|
249 T6H = FNMS(KP951056516, T6G, T6F);
|
Chris@10
|
250 T7F = FMA(KP951056516, T6G, T6F);
|
Chris@10
|
251 Te = cr[WS(rs, 11)];
|
Chris@10
|
252 T5W = FMA(KP951056516, T5V, T5U);
|
Chris@10
|
253 T7n = FNMS(KP951056516, T5V, T5U);
|
Chris@10
|
254 T4N = FMA(KP951056516, T2E, T2B);
|
Chris@10
|
255 T2F = FNMS(KP951056516, T2E, T2B);
|
Chris@10
|
256 Tf = ci[WS(rs, 8)];
|
Chris@10
|
257 Tb = cr[WS(rs, 6)];
|
Chris@10
|
258 Tc = ci[WS(rs, 3)];
|
Chris@10
|
259 TC = cr[WS(rs, 3)];
|
Chris@10
|
260 T2R = Tf - Te;
|
Chris@10
|
261 Tg = Te + Tf;
|
Chris@10
|
262 {
|
Chris@10
|
263 E T2Q, Td, Th, TG, TH;
|
Chris@10
|
264 T2Q = Tb - Tc;
|
Chris@10
|
265 Td = Tb + Tc;
|
Chris@10
|
266 TG = ci[WS(rs, 11)];
|
Chris@10
|
267 TH = ci[WS(rs, 6)];
|
Chris@10
|
268 T2S = FNMS(KP618033988, T2R, T2Q);
|
Chris@10
|
269 T6n = FMA(KP618033988, T2Q, T2R);
|
Chris@10
|
270 Th = Td + Tg;
|
Chris@10
|
271 T2H = Td - Tg;
|
Chris@10
|
272 TI = TG + TH;
|
Chris@10
|
273 T3B = TG - TH;
|
Chris@10
|
274 T2G = FNMS(KP250000000, Th, Ta);
|
Chris@10
|
275 Ti = Ta + Th;
|
Chris@10
|
276 TD = cr[WS(rs, 8)];
|
Chris@10
|
277 TE = ci[WS(rs, 1)];
|
Chris@10
|
278 }
|
Chris@10
|
279 }
|
Chris@10
|
280 {
|
Chris@10
|
281 E Tj, Tk, Tp, T2Z, TJ, Tl;
|
Chris@10
|
282 Tj = cr[WS(rs, 4)];
|
Chris@10
|
283 {
|
Chris@10
|
284 E Tn, To, T3A, TF;
|
Chris@10
|
285 Tn = ci[WS(rs, 10)];
|
Chris@10
|
286 To = ci[WS(rs, 5)];
|
Chris@10
|
287 T3A = TD - TE;
|
Chris@10
|
288 TF = TD + TE;
|
Chris@10
|
289 Tk = cr[WS(rs, 9)];
|
Chris@10
|
290 Tp = Tn + To;
|
Chris@10
|
291 T2Z = To - Tn;
|
Chris@10
|
292 T5Y = FNMS(KP618033988, T3A, T3B);
|
Chris@10
|
293 T3C = FMA(KP618033988, T3B, T3A);
|
Chris@10
|
294 T3r = TI - TF;
|
Chris@10
|
295 TJ = TF + TI;
|
Chris@10
|
296 Tl = ci[0];
|
Chris@10
|
297 }
|
Chris@10
|
298 T1u = ci[WS(rs, 21)];
|
Chris@10
|
299 TK = TC + TJ;
|
Chris@10
|
300 T3q = FNMS(KP250000000, TJ, TC);
|
Chris@10
|
301 {
|
Chris@10
|
302 E T1y, Tm, T2Y, T1z, Tq;
|
Chris@10
|
303 T1y = cr[WS(rs, 13)];
|
Chris@10
|
304 Tm = Tk + Tl;
|
Chris@10
|
305 T2Y = Tl - Tk;
|
Chris@10
|
306 T1z = cr[WS(rs, 18)];
|
Chris@10
|
307 T1v = ci[WS(rs, 16)];
|
Chris@10
|
308 T30 = FMA(KP618033988, T2Z, T2Y);
|
Chris@10
|
309 T6d = FNMS(KP618033988, T2Y, T2Z);
|
Chris@10
|
310 T33 = Tm - Tp;
|
Chris@10
|
311 Tq = Tm + Tp;
|
Chris@10
|
312 T1A = T1y + T1z;
|
Chris@10
|
313 T3u = T1z - T1y;
|
Chris@10
|
314 Tr = Tj + Tq;
|
Chris@10
|
315 T32 = FMS(KP250000000, Tq, Tj);
|
Chris@10
|
316 T1w = cr[WS(rs, 23)];
|
Chris@10
|
317 }
|
Chris@10
|
318 }
|
Chris@10
|
319 }
|
Chris@10
|
320 {
|
Chris@10
|
321 E T1b, T1c, T1h, T36, T1d;
|
Chris@10
|
322 {
|
Chris@10
|
323 E T12, T13, T18, T2K, T1B, T14;
|
Chris@10
|
324 T12 = ci[WS(rs, 23)];
|
Chris@10
|
325 {
|
Chris@10
|
326 E T16, T17, T3t, T1x;
|
Chris@10
|
327 T16 = ci[WS(rs, 13)];
|
Chris@10
|
328 T17 = cr[WS(rs, 16)];
|
Chris@10
|
329 T3t = T1v + T1w;
|
Chris@10
|
330 T1x = T1v - T1w;
|
Chris@10
|
331 T13 = ci[WS(rs, 18)];
|
Chris@10
|
332 T18 = T16 - T17;
|
Chris@10
|
333 T2K = T16 + T17;
|
Chris@10
|
334 T3v = FMA(KP618033988, T3u, T3t);
|
Chris@10
|
335 T61 = FNMS(KP618033988, T3t, T3u);
|
Chris@10
|
336 T3y = T1x + T1A;
|
Chris@10
|
337 T1B = T1x - T1A;
|
Chris@10
|
338 T14 = cr[WS(rs, 21)];
|
Chris@10
|
339 }
|
Chris@10
|
340 T1b = ci[WS(rs, 20)];
|
Chris@10
|
341 T1C = T1u + T1B;
|
Chris@10
|
342 T3x = FMS(KP250000000, T1B, T1u);
|
Chris@10
|
343 {
|
Chris@10
|
344 E T1f, T15, T2J, T1g, T19;
|
Chris@10
|
345 T1f = cr[WS(rs, 14)];
|
Chris@10
|
346 T15 = T13 - T14;
|
Chris@10
|
347 T2J = T13 + T14;
|
Chris@10
|
348 T1g = cr[WS(rs, 19)];
|
Chris@10
|
349 T1c = ci[WS(rs, 15)];
|
Chris@10
|
350 T2L = FMA(KP618033988, T2K, T2J);
|
Chris@10
|
351 T6k = FNMS(KP618033988, T2J, T2K);
|
Chris@10
|
352 T2O = T15 - T18;
|
Chris@10
|
353 T19 = T15 + T18;
|
Chris@10
|
354 T1h = T1f + T1g;
|
Chris@10
|
355 T36 = T1g - T1f;
|
Chris@10
|
356 T1a = T12 + T19;
|
Chris@10
|
357 T2N = FNMS(KP250000000, T19, T12);
|
Chris@10
|
358 T1d = cr[WS(rs, 24)];
|
Chris@10
|
359 }
|
Chris@10
|
360 }
|
Chris@10
|
361 {
|
Chris@10
|
362 E T1l, T1p, T1o, T3e, T1i, T1q;
|
Chris@10
|
363 T1l = ci[WS(rs, 22)];
|
Chris@10
|
364 {
|
Chris@10
|
365 E T1m, T1n, T35, T1e;
|
Chris@10
|
366 T1m = ci[WS(rs, 17)];
|
Chris@10
|
367 T1n = cr[WS(rs, 22)];
|
Chris@10
|
368 T35 = T1c + T1d;
|
Chris@10
|
369 T1e = T1c - T1d;
|
Chris@10
|
370 T1p = ci[WS(rs, 12)];
|
Chris@10
|
371 T1o = T1m - T1n;
|
Chris@10
|
372 T3e = T1m + T1n;
|
Chris@10
|
373 T6g = FNMS(KP618033988, T35, T36);
|
Chris@10
|
374 T37 = FMA(KP618033988, T36, T35);
|
Chris@10
|
375 T2W = T1e + T1h;
|
Chris@10
|
376 T1i = T1e - T1h;
|
Chris@10
|
377 T1q = cr[WS(rs, 17)];
|
Chris@10
|
378 }
|
Chris@10
|
379 Tt = cr[WS(rs, 2)];
|
Chris@10
|
380 T1j = T1b + T1i;
|
Chris@10
|
381 T2V = FMS(KP250000000, T1i, T1b);
|
Chris@10
|
382 {
|
Chris@10
|
383 E Tu, T1r, T3f, Tv, T1s;
|
Chris@10
|
384 Tu = cr[WS(rs, 7)];
|
Chris@10
|
385 T1r = T1p - T1q;
|
Chris@10
|
386 T3f = T1p + T1q;
|
Chris@10
|
387 Tv = ci[WS(rs, 2)];
|
Chris@10
|
388 Tx = cr[WS(rs, 12)];
|
Chris@10
|
389 T3g = FMA(KP618033988, T3f, T3e);
|
Chris@10
|
390 T68 = FNMS(KP618033988, T3e, T3f);
|
Chris@10
|
391 T3j = T1o - T1r;
|
Chris@10
|
392 T1s = T1o + T1r;
|
Chris@10
|
393 Tw = Tu + Tv;
|
Chris@10
|
394 T3l = Tu - Tv;
|
Chris@10
|
395 T1t = T1l + T1s;
|
Chris@10
|
396 T3i = FMS(KP250000000, T1s, T1l);
|
Chris@10
|
397 Ty = ci[WS(rs, 7)];
|
Chris@10
|
398 }
|
Chris@10
|
399 }
|
Chris@10
|
400 }
|
Chris@10
|
401 }
|
Chris@10
|
402 {
|
Chris@10
|
403 E T3n, T65, T3c, T3b, T2P, T2M, T4W;
|
Chris@10
|
404 {
|
Chris@10
|
405 E TA, T3m, Tz, TB, Ts;
|
Chris@10
|
406 T3m = Ty - Tx;
|
Chris@10
|
407 Tz = Tx + Ty;
|
Chris@10
|
408 T1S = T1a + T1j;
|
Chris@10
|
409 T1k = T1a - T1j;
|
Chris@10
|
410 T3n = FNMS(KP618033988, T3m, T3l);
|
Chris@10
|
411 T65 = FMA(KP618033988, T3l, T3m);
|
Chris@10
|
412 TA = Tw + Tz;
|
Chris@10
|
413 T3c = Tz - Tw;
|
Chris@10
|
414 T3b = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
415 TB = Tt + TA;
|
Chris@10
|
416 T1T = T1t + T1C;
|
Chris@10
|
417 T1D = T1t - T1C;
|
Chris@10
|
418 T1Y = Ti - Tr;
|
Chris@10
|
419 Ts = Ti + Tr;
|
Chris@10
|
420 {
|
Chris@10
|
421 E T2I, T6j, T6m, TL;
|
Chris@10
|
422 T2I = FMA(KP559016994, T2H, T2G);
|
Chris@10
|
423 T6j = FNMS(KP559016994, T2H, T2G);
|
Chris@10
|
424 T6m = FNMS(KP559016994, T2O, T2N);
|
Chris@10
|
425 T2P = FMA(KP559016994, T2O, T2N);
|
Chris@10
|
426 TL = TB + TK;
|
Chris@10
|
427 T1Z = TB - TK;
|
Chris@10
|
428 {
|
Chris@10
|
429 E T6l, T7y, T6o, T7z;
|
Chris@10
|
430 T6l = FMA(KP951056516, T6k, T6j);
|
Chris@10
|
431 T7y = FNMS(KP951056516, T6k, T6j);
|
Chris@10
|
432 T6o = FMA(KP951056516, T6n, T6m);
|
Chris@10
|
433 T7z = FNMS(KP951056516, T6n, T6m);
|
Chris@10
|
434 T10 = Ts - TL;
|
Chris@10
|
435 TM = Ts + TL;
|
Chris@10
|
436 T2M = FNMS(KP951056516, T2L, T2I);
|
Chris@10
|
437 T4W = FMA(KP951056516, T2L, T2I);
|
Chris@10
|
438 T7K = FMA(KP939062505, T7y, T7z);
|
Chris@10
|
439 T7A = FNMS(KP939062505, T7z, T7y);
|
Chris@10
|
440 T6p = FNMS(KP549754652, T6o, T6l);
|
Chris@10
|
441 T6w = FMA(KP549754652, T6l, T6o);
|
Chris@10
|
442 }
|
Chris@10
|
443 }
|
Chris@10
|
444 }
|
Chris@10
|
445 {
|
Chris@10
|
446 E T34, T31, T4Y, T60, T3s, T3z, T5X;
|
Chris@10
|
447 {
|
Chris@10
|
448 E T2X, T6c, T6f, T4V, T2T;
|
Chris@10
|
449 T2X = FNMS(KP559016994, T2W, T2V);
|
Chris@10
|
450 T6c = FMA(KP559016994, T2W, T2V);
|
Chris@10
|
451 T6f = FMA(KP559016994, T33, T32);
|
Chris@10
|
452 T34 = FNMS(KP559016994, T33, T32);
|
Chris@10
|
453 T4V = FNMS(KP951056516, T2S, T2P);
|
Chris@10
|
454 T2T = FMA(KP951056516, T2S, T2P);
|
Chris@10
|
455 {
|
Chris@10
|
456 E T7w, T6e, T7v, T6h;
|
Chris@10
|
457 T7w = FMA(KP951056516, T6d, T6c);
|
Chris@10
|
458 T6e = FNMS(KP951056516, T6d, T6c);
|
Chris@10
|
459 T7v = FMA(KP951056516, T6g, T6f);
|
Chris@10
|
460 T6h = FNMS(KP951056516, T6g, T6f);
|
Chris@10
|
461 T4X = FNMS(KP634619297, T4W, T4V);
|
Chris@10
|
462 T56 = FMA(KP634619297, T4V, T4W);
|
Chris@10
|
463 T3K = FMA(KP256756360, T2M, T2T);
|
Chris@10
|
464 T2U = FNMS(KP256756360, T2T, T2M);
|
Chris@10
|
465 T7x = FMA(KP126329378, T7w, T7v);
|
Chris@10
|
466 T7J = FNMS(KP126329378, T7v, T7w);
|
Chris@10
|
467 T6v = FNMS(KP470564281, T6e, T6h);
|
Chris@10
|
468 T6i = FMA(KP470564281, T6h, T6e);
|
Chris@10
|
469 T31 = FMA(KP951056516, T30, T2X);
|
Chris@10
|
470 T4Y = FNMS(KP951056516, T30, T2X);
|
Chris@10
|
471 }
|
Chris@10
|
472 T60 = FMA(KP559016994, T3r, T3q);
|
Chris@10
|
473 T3s = FNMS(KP559016994, T3r, T3q);
|
Chris@10
|
474 T3z = FNMS(KP559016994, T3y, T3x);
|
Chris@10
|
475 T5X = FMA(KP559016994, T3y, T3x);
|
Chris@10
|
476 }
|
Chris@10
|
477 {
|
Chris@10
|
478 E T5Z, T7r, T4Z, T38;
|
Chris@10
|
479 T4Z = FNMS(KP951056516, T37, T34);
|
Chris@10
|
480 T38 = FMA(KP951056516, T37, T34);
|
Chris@10
|
481 {
|
Chris@10
|
482 E T4O, T3w, T4P, T3D;
|
Chris@10
|
483 T4O = FMA(KP951056516, T3v, T3s);
|
Chris@10
|
484 T3w = FNMS(KP951056516, T3v, T3s);
|
Chris@10
|
485 T4P = FMA(KP951056516, T3C, T3z);
|
Chris@10
|
486 T3D = FNMS(KP951056516, T3C, T3z);
|
Chris@10
|
487 T50 = FNMS(KP827271945, T4Z, T4Y);
|
Chris@10
|
488 T57 = FMA(KP827271945, T4Y, T4Z);
|
Chris@10
|
489 T3L = FMA(KP634619297, T31, T38);
|
Chris@10
|
490 T39 = FNMS(KP634619297, T38, T31);
|
Chris@10
|
491 T4Q = FNMS(KP126329378, T4P, T4O);
|
Chris@10
|
492 T59 = FMA(KP126329378, T4O, T4P);
|
Chris@10
|
493 T3O = FNMS(KP939062505, T3w, T3D);
|
Chris@10
|
494 T3E = FMA(KP939062505, T3D, T3w);
|
Chris@10
|
495 T5Z = FMA(KP951056516, T5Y, T5X);
|
Chris@10
|
496 T7r = FNMS(KP951056516, T5Y, T5X);
|
Chris@10
|
497 }
|
Chris@10
|
498 {
|
Chris@10
|
499 E T3d, T3k, T64, T7s, T62;
|
Chris@10
|
500 T67 = FMA(KP559016994, T3c, T3b);
|
Chris@10
|
501 T3d = FNMS(KP559016994, T3c, T3b);
|
Chris@10
|
502 T3k = FNMS(KP559016994, T3j, T3i);
|
Chris@10
|
503 T64 = FMA(KP559016994, T3j, T3i);
|
Chris@10
|
504 T7s = FNMS(KP951056516, T61, T60);
|
Chris@10
|
505 T62 = FMA(KP951056516, T61, T60);
|
Chris@10
|
506 {
|
Chris@10
|
507 E T4S, T3h, T4R, T3o;
|
Chris@10
|
508 T4S = FMA(KP951056516, T3g, T3d);
|
Chris@10
|
509 T3h = FNMS(KP951056516, T3g, T3d);
|
Chris@10
|
510 T4R = FMA(KP951056516, T3n, T3k);
|
Chris@10
|
511 T3o = FNMS(KP951056516, T3n, T3k);
|
Chris@10
|
512 T7t = FNMS(KP827271945, T7s, T7r);
|
Chris@10
|
513 T7H = FMA(KP827271945, T7r, T7s);
|
Chris@10
|
514 T6y = FNMS(KP062914667, T5Z, T62);
|
Chris@10
|
515 T63 = FMA(KP062914667, T62, T5Z);
|
Chris@10
|
516 T4T = FNMS(KP470564281, T4S, T4R);
|
Chris@10
|
517 T5a = FMA(KP470564281, T4R, T4S);
|
Chris@10
|
518 T3N = FNMS(KP549754652, T3h, T3o);
|
Chris@10
|
519 T3p = FMA(KP549754652, T3o, T3h);
|
Chris@10
|
520 T66 = FNMS(KP951056516, T65, T64);
|
Chris@10
|
521 T7o = FMA(KP951056516, T65, T64);
|
Chris@10
|
522 }
|
Chris@10
|
523 }
|
Chris@10
|
524 }
|
Chris@10
|
525 }
|
Chris@10
|
526 }
|
Chris@10
|
527 }
|
Chris@10
|
528 }
|
Chris@10
|
529 {
|
Chris@10
|
530 E T7q, T7G, T6J, T6I, T6q, T6b, T6B, T73, T6Q, T78, T6z, T6a;
|
Chris@10
|
531 cr[0] = T9 + TM;
|
Chris@10
|
532 {
|
Chris@10
|
533 E T1U, T2l, T1X, T2g, T1E, TZ, T2m, T20, T2v, T2n;
|
Chris@10
|
534 {
|
Chris@10
|
535 E T1W, T7p, T69, T1V;
|
Chris@10
|
536 T1W = T1S - T1T;
|
Chris@10
|
537 T1U = T1S + T1T;
|
Chris@10
|
538 T7p = FNMS(KP951056516, T68, T67);
|
Chris@10
|
539 T69 = FMA(KP951056516, T68, T67);
|
Chris@10
|
540 T1V = FNMS(KP250000000, T1U, T1R);
|
Chris@10
|
541 T7q = FMA(KP062914667, T7p, T7o);
|
Chris@10
|
542 T7G = FNMS(KP062914667, T7o, T7p);
|
Chris@10
|
543 T6z = FNMS(KP634619297, T66, T69);
|
Chris@10
|
544 T6a = FMA(KP634619297, T69, T66);
|
Chris@10
|
545 T2l = FNMS(KP559016994, T1W, T1V);
|
Chris@10
|
546 T1X = FMA(KP559016994, T1W, T1V);
|
Chris@10
|
547 T2g = FNMS(KP618033988, T1k, T1D);
|
Chris@10
|
548 T1E = FMA(KP618033988, T1D, T1k);
|
Chris@10
|
549 TZ = FNMS(KP250000000, TM, T9);
|
Chris@10
|
550 T2m = FNMS(KP618033988, T1Y, T1Z);
|
Chris@10
|
551 T20 = FMA(KP618033988, T1Z, T1Y);
|
Chris@10
|
552 }
|
Chris@10
|
553 ci[0] = T1R + T1U;
|
Chris@10
|
554 T2v = FMA(KP951056516, T2m, T2l);
|
Chris@10
|
555 T2n = FNMS(KP951056516, T2m, T2l);
|
Chris@10
|
556 {
|
Chris@10
|
557 E T2b, T21, T2f, T11;
|
Chris@10
|
558 T2b = FNMS(KP951056516, T20, T1X);
|
Chris@10
|
559 T21 = FMA(KP951056516, T20, T1X);
|
Chris@10
|
560 T2f = FNMS(KP559016994, T10, TZ);
|
Chris@10
|
561 T11 = FMA(KP559016994, T10, TZ);
|
Chris@10
|
562 {
|
Chris@10
|
563 E T2h, T2r, T27, T1F;
|
Chris@10
|
564 T2h = FMA(KP951056516, T2g, T2f);
|
Chris@10
|
565 T2r = FNMS(KP951056516, T2g, T2f);
|
Chris@10
|
566 T27 = FMA(KP951056516, T1E, T11);
|
Chris@10
|
567 T1F = FNMS(KP951056516, T1E, T11);
|
Chris@10
|
568 {
|
Chris@10
|
569 E T2o, T2i, T2w, T2s;
|
Chris@10
|
570 T2o = T2k * T2h;
|
Chris@10
|
571 T2i = T2e * T2h;
|
Chris@10
|
572 T2w = T2u * T2r;
|
Chris@10
|
573 T2s = T2q * T2r;
|
Chris@10
|
574 {
|
Chris@10
|
575 E T2c, T28, T22, T1G;
|
Chris@10
|
576 T2c = T2a * T27;
|
Chris@10
|
577 T28 = T26 * T27;
|
Chris@10
|
578 T22 = T1I * T1F;
|
Chris@10
|
579 T1G = TY * T1F;
|
Chris@10
|
580 ci[WS(rs, 15)] = FMA(T2q, T2v, T2w);
|
Chris@10
|
581 cr[WS(rs, 15)] = FNMS(T2u, T2v, T2s);
|
Chris@10
|
582 ci[WS(rs, 20)] = FMA(T26, T2b, T2c);
|
Chris@10
|
583 cr[WS(rs, 20)] = FNMS(T2a, T2b, T28);
|
Chris@10
|
584 ci[WS(rs, 5)] = FMA(TY, T21, T22);
|
Chris@10
|
585 cr[WS(rs, 5)] = FNMS(T1I, T21, T1G);
|
Chris@10
|
586 cr[WS(rs, 10)] = FNMS(T2k, T2n, T2i);
|
Chris@10
|
587 ci[WS(rs, 10)] = FMA(T2e, T2n, T2o);
|
Chris@10
|
588 }
|
Chris@10
|
589 }
|
Chris@10
|
590 }
|
Chris@10
|
591 }
|
Chris@10
|
592 }
|
Chris@10
|
593 {
|
Chris@10
|
594 E T6x, T6A, T6O, T6P;
|
Chris@10
|
595 T6x = FMA(KP968479752, T6w, T6v);
|
Chris@10
|
596 T6J = FNMS(KP968479752, T6w, T6v);
|
Chris@10
|
597 T6I = FMA(KP845997307, T6z, T6y);
|
Chris@10
|
598 T6A = FNMS(KP845997307, T6z, T6y);
|
Chris@10
|
599 T6O = FNMS(KP968479752, T6p, T6i);
|
Chris@10
|
600 T6q = FMA(KP968479752, T6p, T6i);
|
Chris@10
|
601 T6b = FMA(KP845997307, T6a, T63);
|
Chris@10
|
602 T6P = FNMS(KP845997307, T6a, T63);
|
Chris@10
|
603 T6B = FNMS(KP681693190, T6A, T6x);
|
Chris@10
|
604 T73 = FMA(KP560319534, T6x, T6A);
|
Chris@10
|
605 T6Q = FMA(KP681693190, T6P, T6O);
|
Chris@10
|
606 T78 = FNMS(KP560319534, T6O, T6P);
|
Chris@10
|
607 }
|
Chris@10
|
608 {
|
Chris@10
|
609 E T7U, T8f, T7B, T7u, T82, T8k, T7Y, T7M;
|
Chris@10
|
610 {
|
Chris@10
|
611 E T7L, T7I, T80, T81;
|
Chris@10
|
612 {
|
Chris@10
|
613 E T7S, T6r, T6t, T6K, T6M, T7T, T6s, T7j;
|
Chris@10
|
614 T7S = FNMS(KP734762448, T7K, T7J);
|
Chris@10
|
615 T7L = FMA(KP734762448, T7K, T7J);
|
Chris@10
|
616 T6r = FMA(KP906616052, T6q, T6b);
|
Chris@10
|
617 T6t = FNMS(KP906616052, T6q, T6b);
|
Chris@10
|
618 T6K = FNMS(KP906616052, T6J, T6I);
|
Chris@10
|
619 T6M = FMA(KP906616052, T6J, T6I);
|
Chris@10
|
620 T7I = FMA(KP772036680, T7H, T7G);
|
Chris@10
|
621 T7T = FNMS(KP772036680, T7H, T7G);
|
Chris@10
|
622 T6s = FNMS(KP249506682, T6r, T5W);
|
Chris@10
|
623 T7j = FMA(KP998026728, T6r, T5W);
|
Chris@10
|
624 {
|
Chris@10
|
625 E T6L, T7l, T72, T6u;
|
Chris@10
|
626 T6L = FNMS(KP249506682, T6K, T6H);
|
Chris@10
|
627 T7l = FMA(KP998026728, T6K, T6H);
|
Chris@10
|
628 T72 = FMA(KP557913902, T6t, T6s);
|
Chris@10
|
629 T6u = FNMS(KP557913902, T6t, T6s);
|
Chris@10
|
630 {
|
Chris@10
|
631 E T7k, T6N, T77, T7m;
|
Chris@10
|
632 T7k = T4l * T7j;
|
Chris@10
|
633 T6N = FNMS(KP557913902, T6M, T6L);
|
Chris@10
|
634 T77 = FMA(KP557913902, T6M, T6L);
|
Chris@10
|
635 T7m = T4l * T7l;
|
Chris@10
|
636 {
|
Chris@10
|
637 E T74, T7d, T6V, T6C;
|
Chris@10
|
638 T74 = FNMS(KP949179823, T73, T72);
|
Chris@10
|
639 T7d = FMA(KP949179823, T73, T72);
|
Chris@10
|
640 T6V = FMA(KP860541664, T6B, T6u);
|
Chris@10
|
641 T6C = FNMS(KP860541664, T6B, T6u);
|
Chris@10
|
642 cr[WS(rs, 2)] = FNMS(T4n, T7l, T7k);
|
Chris@10
|
643 {
|
Chris@10
|
644 E T7h, T79, T6R, T6Z;
|
Chris@10
|
645 T7h = FNMS(KP949179823, T78, T77);
|
Chris@10
|
646 T79 = FMA(KP949179823, T78, T77);
|
Chris@10
|
647 T6R = FNMS(KP860541664, T6Q, T6N);
|
Chris@10
|
648 T6Z = FMA(KP860541664, T6Q, T6N);
|
Chris@10
|
649 ci[WS(rs, 2)] = FMA(T4n, T7j, T7m);
|
Chris@10
|
650 {
|
Chris@10
|
651 E T75, T7e, T6W, T6D;
|
Chris@10
|
652 T75 = T71 * T74;
|
Chris@10
|
653 T7e = T7c * T7d;
|
Chris@10
|
654 T6W = T6U * T6V;
|
Chris@10
|
655 T6D = T5T * T6C;
|
Chris@10
|
656 {
|
Chris@10
|
657 E T7a, T7i, T70, T6S;
|
Chris@10
|
658 T7a = T71 * T79;
|
Chris@10
|
659 T7i = T7c * T7h;
|
Chris@10
|
660 T70 = T6U * T6Z;
|
Chris@10
|
661 T6S = T5T * T6R;
|
Chris@10
|
662 cr[WS(rs, 12)] = FNMS(T76, T79, T75);
|
Chris@10
|
663 cr[WS(rs, 17)] = FNMS(T7g, T7h, T7e);
|
Chris@10
|
664 cr[WS(rs, 22)] = FNMS(T6Y, T6Z, T6W);
|
Chris@10
|
665 cr[WS(rs, 7)] = FNMS(T6E, T6R, T6D);
|
Chris@10
|
666 ci[WS(rs, 12)] = FMA(T76, T74, T7a);
|
Chris@10
|
667 ci[WS(rs, 17)] = FMA(T7g, T7d, T7i);
|
Chris@10
|
668 ci[WS(rs, 22)] = FMA(T6Y, T6V, T70);
|
Chris@10
|
669 ci[WS(rs, 7)] = FMA(T6E, T6C, T6S);
|
Chris@10
|
670 T7U = FNMS(KP621716863, T7T, T7S);
|
Chris@10
|
671 T8f = FMA(KP614372930, T7S, T7T);
|
Chris@10
|
672 }
|
Chris@10
|
673 }
|
Chris@10
|
674 }
|
Chris@10
|
675 }
|
Chris@10
|
676 }
|
Chris@10
|
677 }
|
Chris@10
|
678 }
|
Chris@10
|
679 T80 = FNMS(KP734762448, T7A, T7x);
|
Chris@10
|
680 T7B = FMA(KP734762448, T7A, T7x);
|
Chris@10
|
681 T7u = FMA(KP772036680, T7t, T7q);
|
Chris@10
|
682 T81 = FNMS(KP772036680, T7t, T7q);
|
Chris@10
|
683 T82 = FNMS(KP621716863, T81, T80);
|
Chris@10
|
684 T8k = FMA(KP614372930, T80, T81);
|
Chris@10
|
685 T7Y = FNMS(KP994076283, T7L, T7I);
|
Chris@10
|
686 T7M = FMA(KP994076283, T7L, T7I);
|
Chris@10
|
687 }
|
Chris@10
|
688 {
|
Chris@10
|
689 E T5y, T5c, T51, T4U, T5f, T5E, T5o, T5i, T5k;
|
Chris@10
|
690 {
|
Chris@10
|
691 E T5h, T5g, T5m, T5n, T58, T5b;
|
Chris@10
|
692 T5h = FMA(KP912575812, T57, T56);
|
Chris@10
|
693 T58 = FNMS(KP912575812, T57, T56);
|
Chris@10
|
694 T5b = FNMS(KP912018591, T5a, T59);
|
Chris@10
|
695 T5g = FMA(KP912018591, T5a, T59);
|
Chris@10
|
696 {
|
Chris@10
|
697 E T7X, T7N, T7C, T7Q;
|
Chris@10
|
698 T7X = FNMS(KP249506682, T7M, T7F);
|
Chris@10
|
699 T7N = FMA(KP998026728, T7M, T7F);
|
Chris@10
|
700 T7C = FMA(KP994076283, T7B, T7u);
|
Chris@10
|
701 T7Q = FNMS(KP994076283, T7B, T7u);
|
Chris@10
|
702 T5y = FMA(KP525970792, T58, T5b);
|
Chris@10
|
703 T5c = FNMS(KP726211448, T5b, T58);
|
Chris@10
|
704 {
|
Chris@10
|
705 E T7Z, T8j, T7P, T7D;
|
Chris@10
|
706 T7Z = FNMS(KP557913902, T7Y, T7X);
|
Chris@10
|
707 T8j = FMA(KP557913902, T7Y, T7X);
|
Chris@10
|
708 T7P = FNMS(KP249506682, T7C, T7n);
|
Chris@10
|
709 T7D = FMA(KP998026728, T7C, T7n);
|
Chris@10
|
710 {
|
Chris@10
|
711 E T8b, T83, T8t, T8l;
|
Chris@10
|
712 T8b = FMA(KP943557151, T82, T7Z);
|
Chris@10
|
713 T83 = FNMS(KP943557151, T82, T7Z);
|
Chris@10
|
714 T8t = FMA(KP949179823, T8k, T8j);
|
Chris@10
|
715 T8l = FNMS(KP949179823, T8k, T8j);
|
Chris@10
|
716 {
|
Chris@10
|
717 E T8e, T7R, T7O, T7E;
|
Chris@10
|
718 T8e = FMA(KP557913902, T7Q, T7P);
|
Chris@10
|
719 T7R = FNMS(KP557913902, T7Q, T7P);
|
Chris@10
|
720 T7O = TR * T7D;
|
Chris@10
|
721 T7E = TO * T7D;
|
Chris@10
|
722 {
|
Chris@10
|
723 E T8g, T8p, T7V, T87;
|
Chris@10
|
724 T8g = FMA(KP949179823, T8f, T8e);
|
Chris@10
|
725 T8p = FNMS(KP949179823, T8f, T8e);
|
Chris@10
|
726 T7V = FMA(KP943557151, T7U, T7R);
|
Chris@10
|
727 T87 = FNMS(KP943557151, T7U, T7R);
|
Chris@10
|
728 ci[WS(rs, 3)] = FMA(TO, T7N, T7O);
|
Chris@10
|
729 cr[WS(rs, 3)] = FNMS(TR, T7N, T7E);
|
Chris@10
|
730 {
|
Chris@10
|
731 E T8m, T8h, T8u, T8q;
|
Chris@10
|
732 T8m = T8i * T8g;
|
Chris@10
|
733 T8h = T8d * T8g;
|
Chris@10
|
734 T8u = T8s * T8p;
|
Chris@10
|
735 T8q = T8o * T8p;
|
Chris@10
|
736 {
|
Chris@10
|
737 E T84, T7W, T8c, T88;
|
Chris@10
|
738 T84 = T4B * T7V;
|
Chris@10
|
739 T7W = T4z * T7V;
|
Chris@10
|
740 T8c = T8a * T87;
|
Chris@10
|
741 T88 = T86 * T87;
|
Chris@10
|
742 ci[WS(rs, 13)] = FMA(T8d, T8l, T8m);
|
Chris@10
|
743 cr[WS(rs, 13)] = FNMS(T8i, T8l, T8h);
|
Chris@10
|
744 ci[WS(rs, 18)] = FMA(T8o, T8t, T8u);
|
Chris@10
|
745 cr[WS(rs, 18)] = FNMS(T8s, T8t, T8q);
|
Chris@10
|
746 ci[WS(rs, 8)] = FMA(T4z, T83, T84);
|
Chris@10
|
747 cr[WS(rs, 8)] = FNMS(T4B, T83, T7W);
|
Chris@10
|
748 ci[WS(rs, 23)] = FMA(T86, T8b, T8c);
|
Chris@10
|
749 cr[WS(rs, 23)] = FNMS(T8a, T8b, T88);
|
Chris@10
|
750 }
|
Chris@10
|
751 }
|
Chris@10
|
752 }
|
Chris@10
|
753 }
|
Chris@10
|
754 }
|
Chris@10
|
755 }
|
Chris@10
|
756 }
|
Chris@10
|
757 T51 = FMA(KP912575812, T50, T4X);
|
Chris@10
|
758 T5m = FNMS(KP912575812, T50, T4X);
|
Chris@10
|
759 T5n = FMA(KP912018591, T4T, T4Q);
|
Chris@10
|
760 T4U = FNMS(KP912018591, T4T, T4Q);
|
Chris@10
|
761 T41 = FMA(KP951056516, T40, T3X);
|
Chris@10
|
762 T5f = FNMS(KP951056516, T40, T3X);
|
Chris@10
|
763 T5E = FMA(KP525970792, T5m, T5n);
|
Chris@10
|
764 T5o = FNMS(KP726211448, T5n, T5m);
|
Chris@10
|
765 T5i = FMA(KP851038619, T5h, T5g);
|
Chris@10
|
766 T5k = FNMS(KP851038619, T5h, T5g);
|
Chris@10
|
767 }
|
Chris@10
|
768 {
|
Chris@10
|
769 E T42, T43, T48, T49, T3M, T3P;
|
Chris@10
|
770 T3M = FMA(KP871714437, T3L, T3K);
|
Chris@10
|
771 T42 = FNMS(KP871714437, T3L, T3K);
|
Chris@10
|
772 T43 = FMA(KP831864738, T3O, T3N);
|
Chris@10
|
773 T3P = FNMS(KP831864738, T3O, T3N);
|
Chris@10
|
774 {
|
Chris@10
|
775 E T5R, T5j, T54, T52;
|
Chris@10
|
776 T5R = FMA(KP992114701, T5i, T5f);
|
Chris@10
|
777 T5j = FNMS(KP248028675, T5i, T5f);
|
Chris@10
|
778 T54 = FNMS(KP851038619, T51, T4U);
|
Chris@10
|
779 T52 = FMA(KP851038619, T51, T4U);
|
Chris@10
|
780 T3Q = FNMS(KP559154169, T3P, T3M);
|
Chris@10
|
781 T4q = FMA(KP683113946, T3M, T3P);
|
Chris@10
|
782 {
|
Chris@10
|
783 E T5D, T5l, T5P, T53;
|
Chris@10
|
784 T5D = FMA(KP554608978, T5k, T5j);
|
Chris@10
|
785 T5l = FNMS(KP554608978, T5k, T5j);
|
Chris@10
|
786 T5P = FNMS(KP992114701, T52, T4N);
|
Chris@10
|
787 T53 = FMA(KP248028675, T52, T4N);
|
Chris@10
|
788 {
|
Chris@10
|
789 E T5p, T5t, T5F, T5N;
|
Chris@10
|
790 T5p = FNMS(KP803003575, T5o, T5l);
|
Chris@10
|
791 T5t = FMA(KP803003575, T5o, T5l);
|
Chris@10
|
792 T5F = FNMS(KP943557151, T5E, T5D);
|
Chris@10
|
793 T5N = FMA(KP943557151, T5E, T5D);
|
Chris@10
|
794 {
|
Chris@10
|
795 E T55, T5x, T5S, T5Q;
|
Chris@10
|
796 T55 = FMA(KP554608978, T54, T53);
|
Chris@10
|
797 T5x = FNMS(KP554608978, T54, T53);
|
Chris@10
|
798 T5S = TW * T5P;
|
Chris@10
|
799 T5Q = TS * T5P;
|
Chris@10
|
800 {
|
Chris@10
|
801 E T5J, T5z, T5r, T5d;
|
Chris@10
|
802 T5J = FMA(KP943557151, T5y, T5x);
|
Chris@10
|
803 T5z = FNMS(KP943557151, T5y, T5x);
|
Chris@10
|
804 T5r = FMA(KP803003575, T5c, T55);
|
Chris@10
|
805 T5d = FNMS(KP803003575, T5c, T55);
|
Chris@10
|
806 ci[WS(rs, 4)] = FMA(TS, T5R, T5S);
|
Chris@10
|
807 cr[WS(rs, 4)] = FNMS(TW, T5R, T5Q);
|
Chris@10
|
808 {
|
Chris@10
|
809 E T5G, T5A, T5O, T5K;
|
Chris@10
|
810 T5G = T5C * T5z;
|
Chris@10
|
811 T5A = T5w * T5z;
|
Chris@10
|
812 T5O = T5M * T5J;
|
Chris@10
|
813 T5K = T5I * T5J;
|
Chris@10
|
814 {
|
Chris@10
|
815 E T5q, T5e, T5u, T5s;
|
Chris@10
|
816 T5q = TX * T5d;
|
Chris@10
|
817 T5e = TT * T5d;
|
Chris@10
|
818 T5u = T25 * T5r;
|
Chris@10
|
819 T5s = T23 * T5r;
|
Chris@10
|
820 ci[WS(rs, 14)] = FMA(T5w, T5F, T5G);
|
Chris@10
|
821 cr[WS(rs, 14)] = FNMS(T5C, T5F, T5A);
|
Chris@10
|
822 ci[WS(rs, 19)] = FMA(T5I, T5N, T5O);
|
Chris@10
|
823 cr[WS(rs, 19)] = FNMS(T5M, T5N, T5K);
|
Chris@10
|
824 ci[WS(rs, 9)] = FMA(TT, T5p, T5q);
|
Chris@10
|
825 cr[WS(rs, 9)] = FNMS(TX, T5p, T5e);
|
Chris@10
|
826 ci[WS(rs, 24)] = FMA(T23, T5t, T5u);
|
Chris@10
|
827 cr[WS(rs, 24)] = FNMS(T25, T5t, T5s);
|
Chris@10
|
828 }
|
Chris@10
|
829 }
|
Chris@10
|
830 }
|
Chris@10
|
831 }
|
Chris@10
|
832 }
|
Chris@10
|
833 }
|
Chris@10
|
834 }
|
Chris@10
|
835 T48 = FNMS(KP871714437, T39, T2U);
|
Chris@10
|
836 T3a = FMA(KP871714437, T39, T2U);
|
Chris@10
|
837 T3F = FMA(KP831864738, T3E, T3p);
|
Chris@10
|
838 T49 = FNMS(KP831864738, T3E, T3p);
|
Chris@10
|
839 T4a = FMA(KP559154169, T49, T48);
|
Chris@10
|
840 T4w = FNMS(KP683113946, T48, T49);
|
Chris@10
|
841 T46 = FMA(KP904730450, T43, T42);
|
Chris@10
|
842 T44 = FNMS(KP904730450, T43, T42);
|
Chris@10
|
843 }
|
Chris@10
|
844 }
|
Chris@10
|
845 }
|
Chris@10
|
846 }
|
Chris@10
|
847 }
|
Chris@10
|
848 }
|
Chris@10
|
849 {
|
Chris@10
|
850 E T45, T4L, T3G, T3I;
|
Chris@10
|
851 T45 = FNMS(KP242145790, T44, T41);
|
Chris@10
|
852 T4L = FMA(KP968583161, T44, T41);
|
Chris@10
|
853 T3G = FMA(KP904730450, T3F, T3a);
|
Chris@10
|
854 T3I = FNMS(KP904730450, T3F, T3a);
|
Chris@10
|
855 {
|
Chris@10
|
856 E T4v, T47, T4J, T3H;
|
Chris@10
|
857 T4v = FNMS(KP541454447, T46, T45);
|
Chris@10
|
858 T47 = FMA(KP541454447, T46, T45);
|
Chris@10
|
859 T4J = FMA(KP968583161, T3G, T2F);
|
Chris@10
|
860 T3H = FNMS(KP242145790, T3G, T2F);
|
Chris@10
|
861 {
|
Chris@10
|
862 E T4b, T4j, T4x, T4H;
|
Chris@10
|
863 T4b = FMA(KP921177326, T4a, T47);
|
Chris@10
|
864 T4j = FNMS(KP921177326, T4a, T47);
|
Chris@10
|
865 T4x = FNMS(KP833417178, T4w, T4v);
|
Chris@10
|
866 T4H = FMA(KP833417178, T4w, T4v);
|
Chris@10
|
867 {
|
Chris@10
|
868 E T3J, T4p, T4M, T4K;
|
Chris@10
|
869 T3J = FMA(KP541454447, T3I, T3H);
|
Chris@10
|
870 T4p = FNMS(KP541454447, T3I, T3H);
|
Chris@10
|
871 T4M = TQ * T4J;
|
Chris@10
|
872 T4K = TN * T4J;
|
Chris@10
|
873 {
|
Chris@10
|
874 E T4D, T4r, T4f, T3R;
|
Chris@10
|
875 T4D = FMA(KP833417178, T4q, T4p);
|
Chris@10
|
876 T4r = FNMS(KP833417178, T4q, T4p);
|
Chris@10
|
877 T4f = FMA(KP921177326, T3Q, T3J);
|
Chris@10
|
878 T3R = FNMS(KP921177326, T3Q, T3J);
|
Chris@10
|
879 ci[WS(rs, 1)] = FMA(TN, T4L, T4M);
|
Chris@10
|
880 cr[WS(rs, 1)] = FNMS(TQ, T4L, T4K);
|
Chris@10
|
881 {
|
Chris@10
|
882 E T4y, T4s, T4I, T4E;
|
Chris@10
|
883 T4y = T4u * T4r;
|
Chris@10
|
884 T4s = T4o * T4r;
|
Chris@10
|
885 T4I = T4G * T4D;
|
Chris@10
|
886 T4E = T4C * T4D;
|
Chris@10
|
887 {
|
Chris@10
|
888 E T4c, T3S, T4k, T4g;
|
Chris@10
|
889 T4c = T3U * T3R;
|
Chris@10
|
890 T3S = T2y * T3R;
|
Chris@10
|
891 T4k = T4i * T4f;
|
Chris@10
|
892 T4g = T4e * T4f;
|
Chris@10
|
893 ci[WS(rs, 11)] = FMA(T4o, T4x, T4y);
|
Chris@10
|
894 cr[WS(rs, 11)] = FNMS(T4u, T4x, T4s);
|
Chris@10
|
895 ci[WS(rs, 16)] = FMA(T4C, T4H, T4I);
|
Chris@10
|
896 cr[WS(rs, 16)] = FNMS(T4G, T4H, T4E);
|
Chris@10
|
897 ci[WS(rs, 6)] = FMA(T2y, T4b, T4c);
|
Chris@10
|
898 cr[WS(rs, 6)] = FNMS(T3U, T4b, T3S);
|
Chris@10
|
899 ci[WS(rs, 21)] = FMA(T4e, T4j, T4k);
|
Chris@10
|
900 cr[WS(rs, 21)] = FNMS(T4i, T4j, T4g);
|
Chris@10
|
901 }
|
Chris@10
|
902 }
|
Chris@10
|
903 }
|
Chris@10
|
904 }
|
Chris@10
|
905 }
|
Chris@10
|
906 }
|
Chris@10
|
907 }
|
Chris@10
|
908 }
|
Chris@10
|
909 }
|
Chris@10
|
910 }
|
Chris@10
|
911
|
Chris@10
|
912 static const tw_instr twinstr[] = {
|
Chris@10
|
913 {TW_CEXP, 1, 1},
|
Chris@10
|
914 {TW_CEXP, 1, 3},
|
Chris@10
|
915 {TW_CEXP, 1, 9},
|
Chris@10
|
916 {TW_CEXP, 1, 24},
|
Chris@10
|
917 {TW_NEXT, 1, 0}
|
Chris@10
|
918 };
|
Chris@10
|
919
|
Chris@10
|
920 static const hc2hc_desc desc = { 25, "hb2_25", twinstr, &GENUS, {84, 78, 356, 0} };
|
Chris@10
|
921
|
Chris@10
|
922 void X(codelet_hb2_25) (planner *p) {
|
Chris@10
|
923 X(khc2hc_register) (p, hb2_25, &desc);
|
Chris@10
|
924 }
|
Chris@10
|
925 #else /* HAVE_FMA */
|
Chris@10
|
926
|
Chris@10
|
927 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 25 -dif -name hb2_25 -include hb.h */
|
Chris@10
|
928
|
Chris@10
|
929 /*
|
Chris@10
|
930 * This function contains 440 FP additions, 340 FP multiplications,
|
Chris@10
|
931 * (or, 280 additions, 180 multiplications, 160 fused multiply/add),
|
Chris@10
|
932 * 155 stack variables, 20 constants, and 100 memory accesses
|
Chris@10
|
933 */
|
Chris@10
|
934 #include "hb.h"
|
Chris@10
|
935
|
Chris@10
|
936 static void hb2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
937 {
|
Chris@10
|
938 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
939 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
940 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
941 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
942 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
943 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
944 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
945 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
946 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
947 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
948 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
949 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
950 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
951 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
952 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
953 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
954 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
955 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
956 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
957 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
958 {
|
Chris@10
|
959 INT m;
|
Chris@10
|
960 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(50, rs)) {
|
Chris@10
|
961 E TN, TQ, TO, TR, TT, TY, T2t, T2r, TZ, TU, T4f, T4l, T2d, T4v, T5m;
|
Chris@10
|
962 E T2j, T5l, T4X, T2v, T11, T3R, T1L, T5d, T6x, T5h, T6t, T25, T26, T27, T29;
|
Chris@10
|
963 E T6D, T7v, T49, T7l, T7p, T7t, T2p, T2n, T4b, T4p, T5n, T6B, T5b, T5p, T6p;
|
Chris@10
|
964 E T6r, T59, T4r;
|
Chris@10
|
965 {
|
Chris@10
|
966 E T2c, T4j, T2h, T4e, T2b, T4k, T2i, T4d;
|
Chris@10
|
967 {
|
Chris@10
|
968 E TP, TX, TS, TW;
|
Chris@10
|
969 TN = W[0];
|
Chris@10
|
970 TQ = W[1];
|
Chris@10
|
971 TO = W[2];
|
Chris@10
|
972 TR = W[3];
|
Chris@10
|
973 TP = TN * TO;
|
Chris@10
|
974 TX = TQ * TO;
|
Chris@10
|
975 TS = TQ * TR;
|
Chris@10
|
976 TW = TN * TR;
|
Chris@10
|
977 TT = TP - TS;
|
Chris@10
|
978 TY = TW + TX;
|
Chris@10
|
979 T2t = TW - TX;
|
Chris@10
|
980 T2r = TP + TS;
|
Chris@10
|
981 TZ = W[5];
|
Chris@10
|
982 T2c = TQ * TZ;
|
Chris@10
|
983 T4j = TO * TZ;
|
Chris@10
|
984 T2h = TN * TZ;
|
Chris@10
|
985 T4e = TR * TZ;
|
Chris@10
|
986 TU = W[4];
|
Chris@10
|
987 T2b = TN * TU;
|
Chris@10
|
988 T4k = TR * TU;
|
Chris@10
|
989 T2i = TQ * TU;
|
Chris@10
|
990 T4d = TO * TU;
|
Chris@10
|
991 }
|
Chris@10
|
992 T4f = T4d - T4e;
|
Chris@10
|
993 T4l = T4j + T4k;
|
Chris@10
|
994 {
|
Chris@10
|
995 E T2s, T2u, TV, T10, T3P, T3Q, T1J, T1K;
|
Chris@10
|
996 T2d = T2b - T2c;
|
Chris@10
|
997 T4v = T2b + T2c;
|
Chris@10
|
998 T5m = T4j - T4k;
|
Chris@10
|
999 T2j = T2h + T2i;
|
Chris@10
|
1000 T5l = T4d + T4e;
|
Chris@10
|
1001 T4X = T2h - T2i;
|
Chris@10
|
1002 T2s = T2r * TU;
|
Chris@10
|
1003 T2u = T2t * TZ;
|
Chris@10
|
1004 T2v = T2s + T2u;
|
Chris@10
|
1005 TV = TT * TU;
|
Chris@10
|
1006 T10 = TY * TZ;
|
Chris@10
|
1007 T11 = TV + T10;
|
Chris@10
|
1008 T3P = T2r * TZ;
|
Chris@10
|
1009 T3Q = T2t * TU;
|
Chris@10
|
1010 T3R = T3P - T3Q;
|
Chris@10
|
1011 T1J = TT * TZ;
|
Chris@10
|
1012 T1K = TY * TU;
|
Chris@10
|
1013 T1L = T1J - T1K;
|
Chris@10
|
1014 T5d = TV - T10;
|
Chris@10
|
1015 T6x = T3P + T3Q;
|
Chris@10
|
1016 T5h = T1J + T1K;
|
Chris@10
|
1017 T6t = T2s - T2u;
|
Chris@10
|
1018 T25 = W[6];
|
Chris@10
|
1019 T26 = W[7];
|
Chris@10
|
1020 T27 = FMA(TT, T25, TY * T26);
|
Chris@10
|
1021 T29 = FNMS(TY, T25, TT * T26);
|
Chris@10
|
1022 T6D = FNMS(T4X, T25, T4v * T26);
|
Chris@10
|
1023 T7v = FNMS(T1L, T25, T11 * T26);
|
Chris@10
|
1024 T49 = FMA(T2r, T25, T2t * T26);
|
Chris@10
|
1025 T7l = FMA(T2d, T25, T2j * T26);
|
Chris@10
|
1026 T7p = FNMS(T2j, T25, T2d * T26);
|
Chris@10
|
1027 T7t = FMA(T11, T25, T1L * T26);
|
Chris@10
|
1028 T2p = FNMS(TZ, T25, TU * T26);
|
Chris@10
|
1029 T2n = FMA(TU, T25, TZ * T26);
|
Chris@10
|
1030 T4b = FNMS(T2t, T25, T2r * T26);
|
Chris@10
|
1031 T4p = FMA(T2v, T25, T3R * T26);
|
Chris@10
|
1032 T5n = FMA(T5l, T25, T5m * T26);
|
Chris@10
|
1033 T6B = FMA(T4v, T25, T4X * T26);
|
Chris@10
|
1034 T5b = FNMS(TQ, T25, TN * T26);
|
Chris@10
|
1035 T5p = FNMS(T5m, T25, T5l * T26);
|
Chris@10
|
1036 T6p = FMA(TO, T25, TR * T26);
|
Chris@10
|
1037 T6r = FNMS(TR, T25, TO * T26);
|
Chris@10
|
1038 T59 = FMA(TN, T25, TQ * T26);
|
Chris@10
|
1039 T4r = FNMS(T3R, T25, T2v * T26);
|
Chris@10
|
1040 }
|
Chris@10
|
1041 }
|
Chris@10
|
1042 {
|
Chris@10
|
1043 E T9, T6i, T40, T3z, T5Y, Ti, Tr, Ts, T1d, T1m, T1P, T2K, T4P, T3H, T4y;
|
Chris@10
|
1044 E T5G, T71, T65, T6N, T5z, T70, T64, T6K, T2Z, T4Q, T3I, T4B, T20, T5Z, T3C;
|
Chris@10
|
1045 E T43, T6j, TB, TK, TL, T1w, T1F, T1Q, T3f, T4S, T3K, T4F, T5V, T74, T68;
|
Chris@10
|
1046 E T6U, T5O, T73, T67, T6R, T3u, T4T, T3L, T4I;
|
Chris@10
|
1047 {
|
Chris@10
|
1048 E T1, T4, T7, T8, T3Z, T3Y, T3x, T3y;
|
Chris@10
|
1049 T1 = cr[0];
|
Chris@10
|
1050 {
|
Chris@10
|
1051 E T2, T3, T5, T6;
|
Chris@10
|
1052 T2 = cr[WS(rs, 5)];
|
Chris@10
|
1053 T3 = ci[WS(rs, 4)];
|
Chris@10
|
1054 T4 = T2 + T3;
|
Chris@10
|
1055 T5 = cr[WS(rs, 10)];
|
Chris@10
|
1056 T6 = ci[WS(rs, 9)];
|
Chris@10
|
1057 T7 = T5 + T6;
|
Chris@10
|
1058 T8 = T4 + T7;
|
Chris@10
|
1059 T3Z = T5 - T6;
|
Chris@10
|
1060 T3Y = T2 - T3;
|
Chris@10
|
1061 }
|
Chris@10
|
1062 T9 = T1 + T8;
|
Chris@10
|
1063 T6i = FMA(KP951056516, T3Y, KP587785252 * T3Z);
|
Chris@10
|
1064 T40 = FNMS(KP951056516, T3Z, KP587785252 * T3Y);
|
Chris@10
|
1065 T3x = FNMS(KP250000000, T8, T1);
|
Chris@10
|
1066 T3y = KP559016994 * (T4 - T7);
|
Chris@10
|
1067 T3z = T3x - T3y;
|
Chris@10
|
1068 T5Y = T3y + T3x;
|
Chris@10
|
1069 }
|
Chris@10
|
1070 {
|
Chris@10
|
1071 E Ta, T2x, T5w, T2F, Th, T2w, T1e, T2P, T5B, T2X, T1l, T2O, Tj, T2N, T5D;
|
Chris@10
|
1072 E T2T, Tq, T2S, T15, T2B, T5u, T2H, T1c, T2G;
|
Chris@10
|
1073 {
|
Chris@10
|
1074 E Tg, T2E, Td, T2D;
|
Chris@10
|
1075 Ta = cr[WS(rs, 1)];
|
Chris@10
|
1076 {
|
Chris@10
|
1077 E Te, Tf, Tb, Tc;
|
Chris@10
|
1078 Te = cr[WS(rs, 11)];
|
Chris@10
|
1079 Tf = ci[WS(rs, 8)];
|
Chris@10
|
1080 Tg = Te + Tf;
|
Chris@10
|
1081 T2E = Te - Tf;
|
Chris@10
|
1082 Tb = cr[WS(rs, 6)];
|
Chris@10
|
1083 Tc = ci[WS(rs, 3)];
|
Chris@10
|
1084 Td = Tb + Tc;
|
Chris@10
|
1085 T2D = Tb - Tc;
|
Chris@10
|
1086 }
|
Chris@10
|
1087 T2x = KP559016994 * (Td - Tg);
|
Chris@10
|
1088 T5w = FMA(KP951056516, T2D, KP587785252 * T2E);
|
Chris@10
|
1089 T2F = FNMS(KP951056516, T2E, KP587785252 * T2D);
|
Chris@10
|
1090 Th = Td + Tg;
|
Chris@10
|
1091 T2w = FNMS(KP250000000, Th, Ta);
|
Chris@10
|
1092 }
|
Chris@10
|
1093 {
|
Chris@10
|
1094 E T1k, T2W, T1h, T2V;
|
Chris@10
|
1095 T1e = ci[WS(rs, 20)];
|
Chris@10
|
1096 {
|
Chris@10
|
1097 E T1i, T1j, T1f, T1g;
|
Chris@10
|
1098 T1i = cr[WS(rs, 14)];
|
Chris@10
|
1099 T1j = cr[WS(rs, 19)];
|
Chris@10
|
1100 T1k = T1i + T1j;
|
Chris@10
|
1101 T2W = T1j - T1i;
|
Chris@10
|
1102 T1f = ci[WS(rs, 15)];
|
Chris@10
|
1103 T1g = cr[WS(rs, 24)];
|
Chris@10
|
1104 T1h = T1f - T1g;
|
Chris@10
|
1105 T2V = T1f + T1g;
|
Chris@10
|
1106 }
|
Chris@10
|
1107 T2P = KP559016994 * (T1h + T1k);
|
Chris@10
|
1108 T5B = FMA(KP951056516, T2V, KP587785252 * T2W);
|
Chris@10
|
1109 T2X = FNMS(KP951056516, T2W, KP587785252 * T2V);
|
Chris@10
|
1110 T1l = T1h - T1k;
|
Chris@10
|
1111 T2O = FNMS(KP250000000, T1l, T1e);
|
Chris@10
|
1112 }
|
Chris@10
|
1113 {
|
Chris@10
|
1114 E Tp, T2M, Tm, T2L;
|
Chris@10
|
1115 Tj = cr[WS(rs, 4)];
|
Chris@10
|
1116 {
|
Chris@10
|
1117 E Tn, To, Tk, Tl;
|
Chris@10
|
1118 Tn = ci[WS(rs, 10)];
|
Chris@10
|
1119 To = ci[WS(rs, 5)];
|
Chris@10
|
1120 Tp = Tn + To;
|
Chris@10
|
1121 T2M = Tn - To;
|
Chris@10
|
1122 Tk = cr[WS(rs, 9)];
|
Chris@10
|
1123 Tl = ci[0];
|
Chris@10
|
1124 Tm = Tk + Tl;
|
Chris@10
|
1125 T2L = Tk - Tl;
|
Chris@10
|
1126 }
|
Chris@10
|
1127 T2N = FNMS(KP951056516, T2M, KP587785252 * T2L);
|
Chris@10
|
1128 T5D = FMA(KP951056516, T2L, KP587785252 * T2M);
|
Chris@10
|
1129 T2T = KP559016994 * (Tm - Tp);
|
Chris@10
|
1130 Tq = Tm + Tp;
|
Chris@10
|
1131 T2S = FNMS(KP250000000, Tq, Tj);
|
Chris@10
|
1132 }
|
Chris@10
|
1133 {
|
Chris@10
|
1134 E T1b, T2A, T18, T2z;
|
Chris@10
|
1135 T15 = ci[WS(rs, 23)];
|
Chris@10
|
1136 {
|
Chris@10
|
1137 E T19, T1a, T16, T17;
|
Chris@10
|
1138 T19 = ci[WS(rs, 13)];
|
Chris@10
|
1139 T1a = cr[WS(rs, 16)];
|
Chris@10
|
1140 T1b = T19 - T1a;
|
Chris@10
|
1141 T2A = T19 + T1a;
|
Chris@10
|
1142 T16 = ci[WS(rs, 18)];
|
Chris@10
|
1143 T17 = cr[WS(rs, 21)];
|
Chris@10
|
1144 T18 = T16 - T17;
|
Chris@10
|
1145 T2z = T16 + T17;
|
Chris@10
|
1146 }
|
Chris@10
|
1147 T2B = FNMS(KP951056516, T2A, KP587785252 * T2z);
|
Chris@10
|
1148 T5u = FMA(KP951056516, T2z, KP587785252 * T2A);
|
Chris@10
|
1149 T2H = KP559016994 * (T18 - T1b);
|
Chris@10
|
1150 T1c = T18 + T1b;
|
Chris@10
|
1151 T2G = FNMS(KP250000000, T1c, T15);
|
Chris@10
|
1152 }
|
Chris@10
|
1153 Ti = Ta + Th;
|
Chris@10
|
1154 Tr = Tj + Tq;
|
Chris@10
|
1155 Ts = Ti + Tr;
|
Chris@10
|
1156 T1d = T15 + T1c;
|
Chris@10
|
1157 T1m = T1e + T1l;
|
Chris@10
|
1158 T1P = T1d + T1m;
|
Chris@10
|
1159 {
|
Chris@10
|
1160 E T2C, T4w, T2J, T4x, T2y, T2I;
|
Chris@10
|
1161 T2y = T2w - T2x;
|
Chris@10
|
1162 T2C = T2y - T2B;
|
Chris@10
|
1163 T4w = T2y + T2B;
|
Chris@10
|
1164 T2I = T2G - T2H;
|
Chris@10
|
1165 T2J = T2F + T2I;
|
Chris@10
|
1166 T4x = T2I - T2F;
|
Chris@10
|
1167 T2K = FNMS(KP481753674, T2J, KP876306680 * T2C);
|
Chris@10
|
1168 T4P = FMA(KP728968627, T4x, KP684547105 * T4w);
|
Chris@10
|
1169 T3H = FMA(KP876306680, T2J, KP481753674 * T2C);
|
Chris@10
|
1170 T4y = FNMS(KP684547105, T4x, KP728968627 * T4w);
|
Chris@10
|
1171 }
|
Chris@10
|
1172 {
|
Chris@10
|
1173 E T5C, T6M, T5F, T6L, T5A, T5E;
|
Chris@10
|
1174 T5A = T2T + T2S;
|
Chris@10
|
1175 T5C = T5A - T5B;
|
Chris@10
|
1176 T6M = T5A + T5B;
|
Chris@10
|
1177 T5E = T2O + T2P;
|
Chris@10
|
1178 T5F = T5D + T5E;
|
Chris@10
|
1179 T6L = T5E - T5D;
|
Chris@10
|
1180 T5G = FNMS(KP844327925, T5F, KP535826794 * T5C);
|
Chris@10
|
1181 T71 = FMA(KP637423989, T6L, KP770513242 * T6M);
|
Chris@10
|
1182 T65 = FMA(KP535826794, T5F, KP844327925 * T5C);
|
Chris@10
|
1183 T6N = FNMS(KP637423989, T6M, KP770513242 * T6L);
|
Chris@10
|
1184 }
|
Chris@10
|
1185 {
|
Chris@10
|
1186 E T5v, T6I, T5y, T6J, T5t, T5x;
|
Chris@10
|
1187 T5t = T2x + T2w;
|
Chris@10
|
1188 T5v = T5t - T5u;
|
Chris@10
|
1189 T6I = T5t + T5u;
|
Chris@10
|
1190 T5x = T2H + T2G;
|
Chris@10
|
1191 T5y = T5w + T5x;
|
Chris@10
|
1192 T6J = T5x - T5w;
|
Chris@10
|
1193 T5z = FNMS(KP248689887, T5y, KP968583161 * T5v);
|
Chris@10
|
1194 T70 = FMA(KP535826794, T6J, KP844327925 * T6I);
|
Chris@10
|
1195 T64 = FMA(KP968583161, T5y, KP248689887 * T5v);
|
Chris@10
|
1196 T6K = FNMS(KP844327925, T6J, KP535826794 * T6I);
|
Chris@10
|
1197 }
|
Chris@10
|
1198 {
|
Chris@10
|
1199 E T2R, T4z, T2Y, T4A, T2Q, T2U;
|
Chris@10
|
1200 T2Q = T2O - T2P;
|
Chris@10
|
1201 T2R = T2N + T2Q;
|
Chris@10
|
1202 T4z = T2Q - T2N;
|
Chris@10
|
1203 T2U = T2S - T2T;
|
Chris@10
|
1204 T2Y = T2U - T2X;
|
Chris@10
|
1205 T4A = T2U + T2X;
|
Chris@10
|
1206 T2Z = FMA(KP904827052, T2R, KP425779291 * T2Y);
|
Chris@10
|
1207 T4Q = FNMS(KP992114701, T4z, KP125333233 * T4A);
|
Chris@10
|
1208 T3I = FNMS(KP425779291, T2R, KP904827052 * T2Y);
|
Chris@10
|
1209 T4B = FMA(KP125333233, T4z, KP992114701 * T4A);
|
Chris@10
|
1210 }
|
Chris@10
|
1211 }
|
Chris@10
|
1212 {
|
Chris@10
|
1213 E T1S, T1V, T1Y, T1Z, T3B, T3A, T41, T42;
|
Chris@10
|
1214 T1S = ci[WS(rs, 24)];
|
Chris@10
|
1215 {
|
Chris@10
|
1216 E T1T, T1U, T1W, T1X;
|
Chris@10
|
1217 T1T = ci[WS(rs, 19)];
|
Chris@10
|
1218 T1U = cr[WS(rs, 20)];
|
Chris@10
|
1219 T1V = T1T - T1U;
|
Chris@10
|
1220 T1W = ci[WS(rs, 14)];
|
Chris@10
|
1221 T1X = cr[WS(rs, 15)];
|
Chris@10
|
1222 T1Y = T1W - T1X;
|
Chris@10
|
1223 T1Z = T1V + T1Y;
|
Chris@10
|
1224 T3B = T1W + T1X;
|
Chris@10
|
1225 T3A = T1T + T1U;
|
Chris@10
|
1226 }
|
Chris@10
|
1227 T20 = T1S + T1Z;
|
Chris@10
|
1228 T5Z = FMA(KP951056516, T3A, KP587785252 * T3B);
|
Chris@10
|
1229 T3C = FNMS(KP951056516, T3B, KP587785252 * T3A);
|
Chris@10
|
1230 T41 = FNMS(KP250000000, T1Z, T1S);
|
Chris@10
|
1231 T42 = KP559016994 * (T1V - T1Y);
|
Chris@10
|
1232 T43 = T41 - T42;
|
Chris@10
|
1233 T6j = T42 + T41;
|
Chris@10
|
1234 }
|
Chris@10
|
1235 {
|
Chris@10
|
1236 E Tt, T32, T5L, T3a, TA, T31, T1o, T36, T5J, T3c, T1v, T3b, TC, T3h, T5S;
|
Chris@10
|
1237 E T3p, TJ, T3g, T1x, T3l, T5Q, T3r, T1E, T3q;
|
Chris@10
|
1238 {
|
Chris@10
|
1239 E Tw, T38, Tz, T39;
|
Chris@10
|
1240 Tt = cr[WS(rs, 2)];
|
Chris@10
|
1241 {
|
Chris@10
|
1242 E Tu, Tv, Tx, Ty;
|
Chris@10
|
1243 Tu = cr[WS(rs, 7)];
|
Chris@10
|
1244 Tv = ci[WS(rs, 2)];
|
Chris@10
|
1245 Tw = Tu + Tv;
|
Chris@10
|
1246 T38 = Tu - Tv;
|
Chris@10
|
1247 Tx = cr[WS(rs, 12)];
|
Chris@10
|
1248 Ty = ci[WS(rs, 7)];
|
Chris@10
|
1249 Tz = Tx + Ty;
|
Chris@10
|
1250 T39 = Tx - Ty;
|
Chris@10
|
1251 }
|
Chris@10
|
1252 T32 = KP559016994 * (Tw - Tz);
|
Chris@10
|
1253 T5L = FMA(KP951056516, T38, KP587785252 * T39);
|
Chris@10
|
1254 T3a = FNMS(KP951056516, T39, KP587785252 * T38);
|
Chris@10
|
1255 TA = Tw + Tz;
|
Chris@10
|
1256 T31 = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
1257 }
|
Chris@10
|
1258 {
|
Chris@10
|
1259 E T1r, T34, T1u, T35;
|
Chris@10
|
1260 T1o = ci[WS(rs, 22)];
|
Chris@10
|
1261 {
|
Chris@10
|
1262 E T1p, T1q, T1s, T1t;
|
Chris@10
|
1263 T1p = ci[WS(rs, 17)];
|
Chris@10
|
1264 T1q = cr[WS(rs, 22)];
|
Chris@10
|
1265 T1r = T1p - T1q;
|
Chris@10
|
1266 T34 = T1p + T1q;
|
Chris@10
|
1267 T1s = ci[WS(rs, 12)];
|
Chris@10
|
1268 T1t = cr[WS(rs, 17)];
|
Chris@10
|
1269 T1u = T1s - T1t;
|
Chris@10
|
1270 T35 = T1s + T1t;
|
Chris@10
|
1271 }
|
Chris@10
|
1272 T36 = FNMS(KP951056516, T35, KP587785252 * T34);
|
Chris@10
|
1273 T5J = FMA(KP951056516, T34, KP587785252 * T35);
|
Chris@10
|
1274 T3c = KP559016994 * (T1r - T1u);
|
Chris@10
|
1275 T1v = T1r + T1u;
|
Chris@10
|
1276 T3b = FNMS(KP250000000, T1v, T1o);
|
Chris@10
|
1277 }
|
Chris@10
|
1278 {
|
Chris@10
|
1279 E TI, T3o, TF, T3n;
|
Chris@10
|
1280 TC = cr[WS(rs, 3)];
|
Chris@10
|
1281 {
|
Chris@10
|
1282 E TG, TH, TD, TE;
|
Chris@10
|
1283 TG = ci[WS(rs, 11)];
|
Chris@10
|
1284 TH = ci[WS(rs, 6)];
|
Chris@10
|
1285 TI = TG + TH;
|
Chris@10
|
1286 T3o = TG - TH;
|
Chris@10
|
1287 TD = cr[WS(rs, 8)];
|
Chris@10
|
1288 TE = ci[WS(rs, 1)];
|
Chris@10
|
1289 TF = TD + TE;
|
Chris@10
|
1290 T3n = TD - TE;
|
Chris@10
|
1291 }
|
Chris@10
|
1292 T3h = KP559016994 * (TF - TI);
|
Chris@10
|
1293 T5S = FMA(KP951056516, T3n, KP587785252 * T3o);
|
Chris@10
|
1294 T3p = FNMS(KP951056516, T3o, KP587785252 * T3n);
|
Chris@10
|
1295 TJ = TF + TI;
|
Chris@10
|
1296 T3g = FNMS(KP250000000, TJ, TC);
|
Chris@10
|
1297 }
|
Chris@10
|
1298 {
|
Chris@10
|
1299 E T1D, T3k, T1A, T3j;
|
Chris@10
|
1300 T1x = ci[WS(rs, 21)];
|
Chris@10
|
1301 {
|
Chris@10
|
1302 E T1B, T1C, T1y, T1z;
|
Chris@10
|
1303 T1B = cr[WS(rs, 13)];
|
Chris@10
|
1304 T1C = cr[WS(rs, 18)];
|
Chris@10
|
1305 T1D = T1B + T1C;
|
Chris@10
|
1306 T3k = T1C - T1B;
|
Chris@10
|
1307 T1y = ci[WS(rs, 16)];
|
Chris@10
|
1308 T1z = cr[WS(rs, 23)];
|
Chris@10
|
1309 T1A = T1y - T1z;
|
Chris@10
|
1310 T3j = T1y + T1z;
|
Chris@10
|
1311 }
|
Chris@10
|
1312 T3l = FNMS(KP951056516, T3k, KP587785252 * T3j);
|
Chris@10
|
1313 T5Q = FMA(KP951056516, T3j, KP587785252 * T3k);
|
Chris@10
|
1314 T3r = KP559016994 * (T1A + T1D);
|
Chris@10
|
1315 T1E = T1A - T1D;
|
Chris@10
|
1316 T3q = FNMS(KP250000000, T1E, T1x);
|
Chris@10
|
1317 }
|
Chris@10
|
1318 TB = Tt + TA;
|
Chris@10
|
1319 TK = TC + TJ;
|
Chris@10
|
1320 TL = TB + TK;
|
Chris@10
|
1321 T1w = T1o + T1v;
|
Chris@10
|
1322 T1F = T1x + T1E;
|
Chris@10
|
1323 T1Q = T1w + T1F;
|
Chris@10
|
1324 {
|
Chris@10
|
1325 E T37, T4D, T3e, T4E, T33, T3d;
|
Chris@10
|
1326 T33 = T31 - T32;
|
Chris@10
|
1327 T37 = T33 - T36;
|
Chris@10
|
1328 T4D = T33 + T36;
|
Chris@10
|
1329 T3d = T3b - T3c;
|
Chris@10
|
1330 T3e = T3a + T3d;
|
Chris@10
|
1331 T4E = T3d - T3a;
|
Chris@10
|
1332 T3f = FNMS(KP844327925, T3e, KP535826794 * T37);
|
Chris@10
|
1333 T4S = FMA(KP062790519, T4E, KP998026728 * T4D);
|
Chris@10
|
1334 T3K = FMA(KP535826794, T3e, KP844327925 * T37);
|
Chris@10
|
1335 T4F = FNMS(KP998026728, T4E, KP062790519 * T4D);
|
Chris@10
|
1336 }
|
Chris@10
|
1337 {
|
Chris@10
|
1338 E T5R, T6T, T5U, T6S, T5P, T5T;
|
Chris@10
|
1339 T5P = T3h + T3g;
|
Chris@10
|
1340 T5R = T5P - T5Q;
|
Chris@10
|
1341 T6T = T5P + T5Q;
|
Chris@10
|
1342 T5T = T3q + T3r;
|
Chris@10
|
1343 T5U = T5S + T5T;
|
Chris@10
|
1344 T6S = T5T - T5S;
|
Chris@10
|
1345 T5V = FNMS(KP684547105, T5U, KP728968627 * T5R);
|
Chris@10
|
1346 T74 = FNMS(KP992114701, T6S, KP125333233 * T6T);
|
Chris@10
|
1347 T68 = FMA(KP728968627, T5U, KP684547105 * T5R);
|
Chris@10
|
1348 T6U = FMA(KP125333233, T6S, KP992114701 * T6T);
|
Chris@10
|
1349 }
|
Chris@10
|
1350 {
|
Chris@10
|
1351 E T5K, T6Q, T5N, T6P, T5I, T5M;
|
Chris@10
|
1352 T5I = T32 + T31;
|
Chris@10
|
1353 T5K = T5I - T5J;
|
Chris@10
|
1354 T6Q = T5I + T5J;
|
Chris@10
|
1355 T5M = T3c + T3b;
|
Chris@10
|
1356 T5N = T5L + T5M;
|
Chris@10
|
1357 T6P = T5M - T5L;
|
Chris@10
|
1358 T5O = FNMS(KP481753674, T5N, KP876306680 * T5K);
|
Chris@10
|
1359 T73 = FNMS(KP425779291, T6P, KP904827052 * T6Q);
|
Chris@10
|
1360 T67 = FMA(KP876306680, T5N, KP481753674 * T5K);
|
Chris@10
|
1361 T6R = FMA(KP904827052, T6P, KP425779291 * T6Q);
|
Chris@10
|
1362 }
|
Chris@10
|
1363 {
|
Chris@10
|
1364 E T3m, T4H, T3t, T4G, T3i, T3s;
|
Chris@10
|
1365 T3i = T3g - T3h;
|
Chris@10
|
1366 T3m = T3i - T3l;
|
Chris@10
|
1367 T4H = T3i + T3l;
|
Chris@10
|
1368 T3s = T3q - T3r;
|
Chris@10
|
1369 T3t = T3p + T3s;
|
Chris@10
|
1370 T4G = T3s - T3p;
|
Chris@10
|
1371 T3u = FNMS(KP998026728, T3t, KP062790519 * T3m);
|
Chris@10
|
1372 T4T = FNMS(KP637423989, T4G, KP770513242 * T4H);
|
Chris@10
|
1373 T3L = FMA(KP062790519, T3t, KP998026728 * T3m);
|
Chris@10
|
1374 T4I = FMA(KP770513242, T4G, KP637423989 * T4H);
|
Chris@10
|
1375 }
|
Chris@10
|
1376 }
|
Chris@10
|
1377 {
|
Chris@10
|
1378 E TM, T14, T2e, T21, T23, T2l, T1H, T2f, T1O, T2k;
|
Chris@10
|
1379 {
|
Chris@10
|
1380 E T12, T13, T1R, T22;
|
Chris@10
|
1381 T12 = KP559016994 * (Ts - TL);
|
Chris@10
|
1382 TM = Ts + TL;
|
Chris@10
|
1383 T13 = FNMS(KP250000000, TM, T9);
|
Chris@10
|
1384 T14 = T12 + T13;
|
Chris@10
|
1385 T2e = T13 - T12;
|
Chris@10
|
1386 T1R = KP559016994 * (T1P - T1Q);
|
Chris@10
|
1387 T21 = T1P + T1Q;
|
Chris@10
|
1388 T22 = FNMS(KP250000000, T21, T20);
|
Chris@10
|
1389 T23 = T1R + T22;
|
Chris@10
|
1390 T2l = T22 - T1R;
|
Chris@10
|
1391 }
|
Chris@10
|
1392 {
|
Chris@10
|
1393 E T1n, T1G, T1M, T1N;
|
Chris@10
|
1394 T1n = T1d - T1m;
|
Chris@10
|
1395 T1G = T1w - T1F;
|
Chris@10
|
1396 T1H = FMA(KP951056516, T1n, KP587785252 * T1G);
|
Chris@10
|
1397 T2f = FNMS(KP951056516, T1G, KP587785252 * T1n);
|
Chris@10
|
1398 T1M = Ti - Tr;
|
Chris@10
|
1399 T1N = TB - TK;
|
Chris@10
|
1400 T1O = FMA(KP951056516, T1M, KP587785252 * T1N);
|
Chris@10
|
1401 T2k = FNMS(KP951056516, T1N, KP587785252 * T1M);
|
Chris@10
|
1402 }
|
Chris@10
|
1403 {
|
Chris@10
|
1404 E T1I, T24, T2o, T2q;
|
Chris@10
|
1405 cr[0] = T9 + TM;
|
Chris@10
|
1406 ci[0] = T20 + T21;
|
Chris@10
|
1407 T1I = T14 - T1H;
|
Chris@10
|
1408 T24 = T1O + T23;
|
Chris@10
|
1409 cr[WS(rs, 5)] = FNMS(T1L, T24, T11 * T1I);
|
Chris@10
|
1410 ci[WS(rs, 5)] = FMA(T1L, T1I, T11 * T24);
|
Chris@10
|
1411 T2o = T2e + T2f;
|
Chris@10
|
1412 T2q = T2l - T2k;
|
Chris@10
|
1413 cr[WS(rs, 15)] = FNMS(T2p, T2q, T2n * T2o);
|
Chris@10
|
1414 ci[WS(rs, 15)] = FMA(T2p, T2o, T2n * T2q);
|
Chris@10
|
1415 {
|
Chris@10
|
1416 E T2g, T2m, T28, T2a;
|
Chris@10
|
1417 T2g = T2e - T2f;
|
Chris@10
|
1418 T2m = T2k + T2l;
|
Chris@10
|
1419 cr[WS(rs, 10)] = FNMS(T2j, T2m, T2d * T2g);
|
Chris@10
|
1420 ci[WS(rs, 10)] = FMA(T2j, T2g, T2d * T2m);
|
Chris@10
|
1421 T28 = T14 + T1H;
|
Chris@10
|
1422 T2a = T23 - T1O;
|
Chris@10
|
1423 cr[WS(rs, 20)] = FNMS(T29, T2a, T27 * T28);
|
Chris@10
|
1424 ci[WS(rs, 20)] = FMA(T29, T28, T27 * T2a);
|
Chris@10
|
1425 }
|
Chris@10
|
1426 }
|
Chris@10
|
1427 }
|
Chris@10
|
1428 {
|
Chris@10
|
1429 E T76, T7n, T7a, T7q, T6H, T6W, T6X, T6Y, T7e, T7f, T7d, T7g, T7x, T7y;
|
Chris@10
|
1430 {
|
Chris@10
|
1431 E T72, T75, T78, T79;
|
Chris@10
|
1432 T72 = T70 + T71;
|
Chris@10
|
1433 T75 = T73 - T74;
|
Chris@10
|
1434 T76 = FMA(KP951056516, T72, KP587785252 * T75);
|
Chris@10
|
1435 T7n = FNMS(KP951056516, T75, KP587785252 * T72);
|
Chris@10
|
1436 T78 = T6K - T6N;
|
Chris@10
|
1437 T79 = T6U - T6R;
|
Chris@10
|
1438 T7a = FMA(KP951056516, T78, KP587785252 * T79);
|
Chris@10
|
1439 T7q = FNMS(KP951056516, T79, KP587785252 * T78);
|
Chris@10
|
1440 }
|
Chris@10
|
1441 {
|
Chris@10
|
1442 E T6O, T6V, T7b, T7c;
|
Chris@10
|
1443 T6H = T5Y + T5Z;
|
Chris@10
|
1444 T6O = T6K + T6N;
|
Chris@10
|
1445 T6V = T6R + T6U;
|
Chris@10
|
1446 T6W = T6O - T6V;
|
Chris@10
|
1447 T6X = FNMS(KP250000000, T6W, T6H);
|
Chris@10
|
1448 T6Y = KP559016994 * (T6O + T6V);
|
Chris@10
|
1449 T7e = T6j - T6i;
|
Chris@10
|
1450 T7b = T70 - T71;
|
Chris@10
|
1451 T7c = T73 + T74;
|
Chris@10
|
1452 T7f = T7b + T7c;
|
Chris@10
|
1453 T7d = KP559016994 * (T7b - T7c);
|
Chris@10
|
1454 T7g = FNMS(KP250000000, T7f, T7e);
|
Chris@10
|
1455 }
|
Chris@10
|
1456 T7x = T6H + T6W;
|
Chris@10
|
1457 T7y = T7e + T7f;
|
Chris@10
|
1458 cr[WS(rs, 4)] = FNMS(TY, T7y, TT * T7x);
|
Chris@10
|
1459 ci[WS(rs, 4)] = FMA(TY, T7x, TT * T7y);
|
Chris@10
|
1460 {
|
Chris@10
|
1461 E T7o, T7u, T7s, T7w, T7m, T7r;
|
Chris@10
|
1462 T7m = T6X - T6Y;
|
Chris@10
|
1463 T7o = T7m - T7n;
|
Chris@10
|
1464 T7u = T7m + T7n;
|
Chris@10
|
1465 T7r = T7g - T7d;
|
Chris@10
|
1466 T7s = T7q + T7r;
|
Chris@10
|
1467 T7w = T7r - T7q;
|
Chris@10
|
1468 cr[WS(rs, 14)] = FNMS(T7p, T7s, T7l * T7o);
|
Chris@10
|
1469 ci[WS(rs, 14)] = FMA(T7p, T7o, T7l * T7s);
|
Chris@10
|
1470 cr[WS(rs, 19)] = FNMS(T7v, T7w, T7t * T7u);
|
Chris@10
|
1471 ci[WS(rs, 19)] = FMA(T7v, T7u, T7t * T7w);
|
Chris@10
|
1472 }
|
Chris@10
|
1473 {
|
Chris@10
|
1474 E T77, T7j, T7i, T7k, T6Z, T7h;
|
Chris@10
|
1475 T6Z = T6X + T6Y;
|
Chris@10
|
1476 T77 = T6Z - T76;
|
Chris@10
|
1477 T7j = T6Z + T76;
|
Chris@10
|
1478 T7h = T7d + T7g;
|
Chris@10
|
1479 T7i = T7a + T7h;
|
Chris@10
|
1480 T7k = T7h - T7a;
|
Chris@10
|
1481 cr[WS(rs, 9)] = FNMS(TZ, T7i, TU * T77);
|
Chris@10
|
1482 ci[WS(rs, 9)] = FMA(TZ, T77, TU * T7i);
|
Chris@10
|
1483 cr[WS(rs, 24)] = FNMS(T26, T7k, T25 * T7j);
|
Chris@10
|
1484 ci[WS(rs, 24)] = FMA(T26, T7j, T25 * T7k);
|
Chris@10
|
1485 }
|
Chris@10
|
1486 }
|
Chris@10
|
1487 {
|
Chris@10
|
1488 E T3N, T4h, T3U, T4m, T3D, T3E, T3w, T3F, T44, T45, T3X, T46, T4t, T4u;
|
Chris@10
|
1489 {
|
Chris@10
|
1490 E T3J, T3M, T3S, T3T;
|
Chris@10
|
1491 T3J = T3H - T3I;
|
Chris@10
|
1492 T3M = T3K - T3L;
|
Chris@10
|
1493 T3N = FMA(KP951056516, T3J, KP587785252 * T3M);
|
Chris@10
|
1494 T4h = FNMS(KP951056516, T3M, KP587785252 * T3J);
|
Chris@10
|
1495 T3S = T2K + T2Z;
|
Chris@10
|
1496 T3T = T3f - T3u;
|
Chris@10
|
1497 T3U = FMA(KP951056516, T3S, KP587785252 * T3T);
|
Chris@10
|
1498 T4m = FNMS(KP951056516, T3T, KP587785252 * T3S);
|
Chris@10
|
1499 }
|
Chris@10
|
1500 {
|
Chris@10
|
1501 E T30, T3v, T3V, T3W;
|
Chris@10
|
1502 T3D = T3z - T3C;
|
Chris@10
|
1503 T30 = T2K - T2Z;
|
Chris@10
|
1504 T3v = T3f + T3u;
|
Chris@10
|
1505 T3E = T30 + T3v;
|
Chris@10
|
1506 T3w = KP559016994 * (T30 - T3v);
|
Chris@10
|
1507 T3F = FNMS(KP250000000, T3E, T3D);
|
Chris@10
|
1508 T44 = T40 + T43;
|
Chris@10
|
1509 T3V = T3H + T3I;
|
Chris@10
|
1510 T3W = T3K + T3L;
|
Chris@10
|
1511 T45 = T3V + T3W;
|
Chris@10
|
1512 T3X = KP559016994 * (T3V - T3W);
|
Chris@10
|
1513 T46 = FNMS(KP250000000, T45, T44);
|
Chris@10
|
1514 }
|
Chris@10
|
1515 T4t = T3D + T3E;
|
Chris@10
|
1516 T4u = T44 + T45;
|
Chris@10
|
1517 cr[WS(rs, 2)] = FNMS(T2t, T4u, T2r * T4t);
|
Chris@10
|
1518 ci[WS(rs, 2)] = FMA(T2t, T4t, T2r * T4u);
|
Chris@10
|
1519 {
|
Chris@10
|
1520 E T4i, T4q, T4o, T4s, T4g, T4n;
|
Chris@10
|
1521 T4g = T3F - T3w;
|
Chris@10
|
1522 T4i = T4g - T4h;
|
Chris@10
|
1523 T4q = T4g + T4h;
|
Chris@10
|
1524 T4n = T46 - T3X;
|
Chris@10
|
1525 T4o = T4m + T4n;
|
Chris@10
|
1526 T4s = T4n - T4m;
|
Chris@10
|
1527 cr[WS(rs, 12)] = FNMS(T4l, T4o, T4f * T4i);
|
Chris@10
|
1528 ci[WS(rs, 12)] = FMA(T4l, T4i, T4f * T4o);
|
Chris@10
|
1529 cr[WS(rs, 17)] = FNMS(T4r, T4s, T4p * T4q);
|
Chris@10
|
1530 ci[WS(rs, 17)] = FMA(T4r, T4q, T4p * T4s);
|
Chris@10
|
1531 }
|
Chris@10
|
1532 {
|
Chris@10
|
1533 E T3O, T4a, T48, T4c, T3G, T47;
|
Chris@10
|
1534 T3G = T3w + T3F;
|
Chris@10
|
1535 T3O = T3G - T3N;
|
Chris@10
|
1536 T4a = T3G + T3N;
|
Chris@10
|
1537 T47 = T3X + T46;
|
Chris@10
|
1538 T48 = T3U + T47;
|
Chris@10
|
1539 T4c = T47 - T3U;
|
Chris@10
|
1540 cr[WS(rs, 7)] = FNMS(T3R, T48, T2v * T3O);
|
Chris@10
|
1541 ci[WS(rs, 7)] = FMA(T3R, T3O, T2v * T48);
|
Chris@10
|
1542 cr[WS(rs, 22)] = FNMS(T4b, T4c, T49 * T4a);
|
Chris@10
|
1543 ci[WS(rs, 22)] = FMA(T4b, T4a, T49 * T4c);
|
Chris@10
|
1544 }
|
Chris@10
|
1545 }
|
Chris@10
|
1546 {
|
Chris@10
|
1547 E T4V, T5f, T50, T5i, T4L, T4M, T4K, T4N, T54, T55, T53, T56, T5r, T5s;
|
Chris@10
|
1548 {
|
Chris@10
|
1549 E T4R, T4U, T4Y, T4Z;
|
Chris@10
|
1550 T4R = T4P - T4Q;
|
Chris@10
|
1551 T4U = T4S - T4T;
|
Chris@10
|
1552 T4V = FMA(KP951056516, T4R, KP587785252 * T4U);
|
Chris@10
|
1553 T5f = FNMS(KP951056516, T4U, KP587785252 * T4R);
|
Chris@10
|
1554 T4Y = T4y + T4B;
|
Chris@10
|
1555 T4Z = T4F + T4I;
|
Chris@10
|
1556 T50 = FMA(KP951056516, T4Y, KP587785252 * T4Z);
|
Chris@10
|
1557 T5i = FNMS(KP951056516, T4Z, KP587785252 * T4Y);
|
Chris@10
|
1558 }
|
Chris@10
|
1559 {
|
Chris@10
|
1560 E T4C, T4J, T51, T52;
|
Chris@10
|
1561 T4L = T3z + T3C;
|
Chris@10
|
1562 T4C = T4y - T4B;
|
Chris@10
|
1563 T4J = T4F - T4I;
|
Chris@10
|
1564 T4M = T4C + T4J;
|
Chris@10
|
1565 T4K = KP559016994 * (T4C - T4J);
|
Chris@10
|
1566 T4N = FNMS(KP250000000, T4M, T4L);
|
Chris@10
|
1567 T54 = T43 - T40;
|
Chris@10
|
1568 T51 = T4P + T4Q;
|
Chris@10
|
1569 T52 = T4S + T4T;
|
Chris@10
|
1570 T55 = T51 + T52;
|
Chris@10
|
1571 T53 = KP559016994 * (T51 - T52);
|
Chris@10
|
1572 T56 = FNMS(KP250000000, T55, T54);
|
Chris@10
|
1573 }
|
Chris@10
|
1574 T5r = T4L + T4M;
|
Chris@10
|
1575 T5s = T54 + T55;
|
Chris@10
|
1576 cr[WS(rs, 3)] = FNMS(TR, T5s, TO * T5r);
|
Chris@10
|
1577 ci[WS(rs, 3)] = FMA(TR, T5r, TO * T5s);
|
Chris@10
|
1578 {
|
Chris@10
|
1579 E T5g, T5o, T5k, T5q, T5e, T5j;
|
Chris@10
|
1580 T5e = T4N - T4K;
|
Chris@10
|
1581 T5g = T5e - T5f;
|
Chris@10
|
1582 T5o = T5e + T5f;
|
Chris@10
|
1583 T5j = T56 - T53;
|
Chris@10
|
1584 T5k = T5i + T5j;
|
Chris@10
|
1585 T5q = T5j - T5i;
|
Chris@10
|
1586 cr[WS(rs, 13)] = FNMS(T5h, T5k, T5d * T5g);
|
Chris@10
|
1587 ci[WS(rs, 13)] = FMA(T5h, T5g, T5d * T5k);
|
Chris@10
|
1588 cr[WS(rs, 18)] = FNMS(T5p, T5q, T5n * T5o);
|
Chris@10
|
1589 ci[WS(rs, 18)] = FMA(T5p, T5o, T5n * T5q);
|
Chris@10
|
1590 }
|
Chris@10
|
1591 {
|
Chris@10
|
1592 E T4W, T5a, T58, T5c, T4O, T57;
|
Chris@10
|
1593 T4O = T4K + T4N;
|
Chris@10
|
1594 T4W = T4O - T4V;
|
Chris@10
|
1595 T5a = T4O + T4V;
|
Chris@10
|
1596 T57 = T53 + T56;
|
Chris@10
|
1597 T58 = T50 + T57;
|
Chris@10
|
1598 T5c = T57 - T50;
|
Chris@10
|
1599 cr[WS(rs, 8)] = FNMS(T4X, T58, T4v * T4W);
|
Chris@10
|
1600 ci[WS(rs, 8)] = FMA(T4X, T4W, T4v * T58);
|
Chris@10
|
1601 cr[WS(rs, 23)] = FNMS(T5b, T5c, T59 * T5a);
|
Chris@10
|
1602 ci[WS(rs, 23)] = FMA(T5b, T5a, T59 * T5c);
|
Chris@10
|
1603 }
|
Chris@10
|
1604 }
|
Chris@10
|
1605 {
|
Chris@10
|
1606 E T6a, T6v, T6e, T6y, T60, T61, T5X, T62, T6k, T6l, T6h, T6m, T6F, T6G;
|
Chris@10
|
1607 {
|
Chris@10
|
1608 E T66, T69, T6c, T6d;
|
Chris@10
|
1609 T66 = T64 - T65;
|
Chris@10
|
1610 T69 = T67 - T68;
|
Chris@10
|
1611 T6a = FMA(KP951056516, T66, KP587785252 * T69);
|
Chris@10
|
1612 T6v = FNMS(KP951056516, T69, KP587785252 * T66);
|
Chris@10
|
1613 T6c = T5z - T5G;
|
Chris@10
|
1614 T6d = T5O - T5V;
|
Chris@10
|
1615 T6e = FMA(KP951056516, T6c, KP587785252 * T6d);
|
Chris@10
|
1616 T6y = FNMS(KP951056516, T6d, KP587785252 * T6c);
|
Chris@10
|
1617 }
|
Chris@10
|
1618 {
|
Chris@10
|
1619 E T5H, T5W, T6f, T6g;
|
Chris@10
|
1620 T60 = T5Y - T5Z;
|
Chris@10
|
1621 T5H = T5z + T5G;
|
Chris@10
|
1622 T5W = T5O + T5V;
|
Chris@10
|
1623 T61 = T5H + T5W;
|
Chris@10
|
1624 T5X = KP559016994 * (T5H - T5W);
|
Chris@10
|
1625 T62 = FNMS(KP250000000, T61, T60);
|
Chris@10
|
1626 T6k = T6i + T6j;
|
Chris@10
|
1627 T6f = T64 + T65;
|
Chris@10
|
1628 T6g = T67 + T68;
|
Chris@10
|
1629 T6l = T6f + T6g;
|
Chris@10
|
1630 T6h = KP559016994 * (T6f - T6g);
|
Chris@10
|
1631 T6m = FNMS(KP250000000, T6l, T6k);
|
Chris@10
|
1632 }
|
Chris@10
|
1633 T6F = T60 + T61;
|
Chris@10
|
1634 T6G = T6k + T6l;
|
Chris@10
|
1635 cr[WS(rs, 1)] = FNMS(TQ, T6G, TN * T6F);
|
Chris@10
|
1636 ci[WS(rs, 1)] = FMA(TQ, T6F, TN * T6G);
|
Chris@10
|
1637 {
|
Chris@10
|
1638 E T6w, T6C, T6A, T6E, T6u, T6z;
|
Chris@10
|
1639 T6u = T62 - T5X;
|
Chris@10
|
1640 T6w = T6u - T6v;
|
Chris@10
|
1641 T6C = T6u + T6v;
|
Chris@10
|
1642 T6z = T6m - T6h;
|
Chris@10
|
1643 T6A = T6y + T6z;
|
Chris@10
|
1644 T6E = T6z - T6y;
|
Chris@10
|
1645 cr[WS(rs, 11)] = FNMS(T6x, T6A, T6t * T6w);
|
Chris@10
|
1646 ci[WS(rs, 11)] = FMA(T6x, T6w, T6t * T6A);
|
Chris@10
|
1647 cr[WS(rs, 16)] = FNMS(T6D, T6E, T6B * T6C);
|
Chris@10
|
1648 ci[WS(rs, 16)] = FMA(T6D, T6C, T6B * T6E);
|
Chris@10
|
1649 }
|
Chris@10
|
1650 {
|
Chris@10
|
1651 E T6b, T6q, T6o, T6s, T63, T6n;
|
Chris@10
|
1652 T63 = T5X + T62;
|
Chris@10
|
1653 T6b = T63 - T6a;
|
Chris@10
|
1654 T6q = T63 + T6a;
|
Chris@10
|
1655 T6n = T6h + T6m;
|
Chris@10
|
1656 T6o = T6e + T6n;
|
Chris@10
|
1657 T6s = T6n - T6e;
|
Chris@10
|
1658 cr[WS(rs, 6)] = FNMS(T5m, T6o, T5l * T6b);
|
Chris@10
|
1659 ci[WS(rs, 6)] = FMA(T5m, T6b, T5l * T6o);
|
Chris@10
|
1660 cr[WS(rs, 21)] = FNMS(T6r, T6s, T6p * T6q);
|
Chris@10
|
1661 ci[WS(rs, 21)] = FMA(T6r, T6q, T6p * T6s);
|
Chris@10
|
1662 }
|
Chris@10
|
1663 }
|
Chris@10
|
1664 }
|
Chris@10
|
1665 }
|
Chris@10
|
1666 }
|
Chris@10
|
1667 }
|
Chris@10
|
1668
|
Chris@10
|
1669 static const tw_instr twinstr[] = {
|
Chris@10
|
1670 {TW_CEXP, 1, 1},
|
Chris@10
|
1671 {TW_CEXP, 1, 3},
|
Chris@10
|
1672 {TW_CEXP, 1, 9},
|
Chris@10
|
1673 {TW_CEXP, 1, 24},
|
Chris@10
|
1674 {TW_NEXT, 1, 0}
|
Chris@10
|
1675 };
|
Chris@10
|
1676
|
Chris@10
|
1677 static const hc2hc_desc desc = { 25, "hb2_25", twinstr, &GENUS, {280, 180, 160, 0} };
|
Chris@10
|
1678
|
Chris@10
|
1679 void X(codelet_hb2_25) (planner *p) {
|
Chris@10
|
1680 X(khc2hc_register) (p, hb2_25, &desc);
|
Chris@10
|
1681 }
|
Chris@10
|
1682 #endif /* HAVE_FMA */
|