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 07:41:23 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 -n 25 -dif -name hb_25 -include hb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 400 FP additions, 364 FP multiplications,
|
Chris@10
|
32 * (or, 84 additions, 48 multiplications, 316 fused multiply/add),
|
Chris@10
|
33 * 176 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 hb_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP921177326, +0.921177326965143320250447435415066029359282231);
|
Chris@10
|
40 DK(KP833417178, +0.833417178328688677408962550243238843138996060);
|
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(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
46 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
47 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
48 DK(KP871714437, +0.871714437527667770979999223229522602943903653);
|
Chris@10
|
49 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
Chris@10
|
50 DK(KP554608978, +0.554608978404018097464974850792216217022558774);
|
Chris@10
|
51 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
Chris@10
|
52 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
53 DK(KP851038619, +0.851038619207379630836264138867114231259902550);
|
Chris@10
|
54 DK(KP525970792, +0.525970792408939708442463226536226366643874659);
|
Chris@10
|
55 DK(KP726211448, +0.726211448929902658173535992263577167607493062);
|
Chris@10
|
56 DK(KP912018591, +0.912018591466481957908415381764119056233607330);
|
Chris@10
|
57 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
58 DK(KP943557151, +0.943557151597354104399655195398983005179443399);
|
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(KP772036680, +0.772036680810363904029489473607579825330539880);
|
Chris@10
|
63 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
Chris@10
|
64 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
Chris@10
|
65 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
Chris@10
|
66 DK(KP557913902, +0.557913902031834264187699648465567037992437152);
|
Chris@10
|
67 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
68 DK(KP249506682, +0.249506682107067890488084201715862638334226305);
|
Chris@10
|
69 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
Chris@10
|
70 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
Chris@10
|
71 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
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) * 48); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 48, MAKE_VOLATILE_STRIDE(50, rs)) {
|
Chris@10
|
89 E T3w, T3P, T2d, T3y, T3x, T3Q;
|
Chris@10
|
90 {
|
Chris@10
|
91 E T9, T3E, T1F, T3B, T6f, T7d, T5u, T6U, T4k, T2k, T5G, T1G, T19, T1H, T1s;
|
Chris@10
|
92 E T1M, T1N, TP, TM, T7i, T77, T5X, T64, T4u, T4D, T3p, T2z, T74, T7h, T63;
|
Chris@10
|
93 E T5Q, T4x, T4E, T3q, T2O, T4n, T4G, T3t, T3j, T5F, T70, T7f, T66, T5B, T4q;
|
Chris@10
|
94 E T4H, T3s, T34, T5E, T6V;
|
Chris@10
|
95 {
|
Chris@10
|
96 E T2f, T2e, T6e, T2j, T5t, T6d;
|
Chris@10
|
97 {
|
Chris@10
|
98 E T1, T1x, T3C, T3D, T8, T2h, T1A, T1D, T2i, T3A, T1E, T3z;
|
Chris@10
|
99 T1 = cr[0];
|
Chris@10
|
100 T1x = ci[WS(rs, 24)];
|
Chris@10
|
101 {
|
Chris@10
|
102 E T2, T3, T5, T6;
|
Chris@10
|
103 T2 = cr[WS(rs, 5)];
|
Chris@10
|
104 T3 = ci[WS(rs, 4)];
|
Chris@10
|
105 T5 = cr[WS(rs, 10)];
|
Chris@10
|
106 T6 = ci[WS(rs, 9)];
|
Chris@10
|
107 {
|
Chris@10
|
108 E T1y, T4, T7, T1z, T1B, T1C;
|
Chris@10
|
109 T1y = ci[WS(rs, 19)];
|
Chris@10
|
110 T3C = T2 - T3;
|
Chris@10
|
111 T4 = T2 + T3;
|
Chris@10
|
112 T3D = T5 - T6;
|
Chris@10
|
113 T7 = T5 + T6;
|
Chris@10
|
114 T1z = cr[WS(rs, 20)];
|
Chris@10
|
115 T1B = ci[WS(rs, 14)];
|
Chris@10
|
116 T1C = cr[WS(rs, 15)];
|
Chris@10
|
117 T8 = T4 + T7;
|
Chris@10
|
118 T2f = T4 - T7;
|
Chris@10
|
119 T2h = T1y + T1z;
|
Chris@10
|
120 T1A = T1y - T1z;
|
Chris@10
|
121 T1D = T1B - T1C;
|
Chris@10
|
122 T2i = T1B + T1C;
|
Chris@10
|
123 }
|
Chris@10
|
124 }
|
Chris@10
|
125 T2e = FNMS(KP250000000, T8, T1);
|
Chris@10
|
126 T9 = T1 + T8;
|
Chris@10
|
127 T3A = T1A - T1D;
|
Chris@10
|
128 T1E = T1A + T1D;
|
Chris@10
|
129 T3E = FMA(KP618033988, T3D, T3C);
|
Chris@10
|
130 T6e = FNMS(KP618033988, T3C, T3D);
|
Chris@10
|
131 T2j = FMA(KP618033988, T2i, T2h);
|
Chris@10
|
132 T5t = FNMS(KP618033988, T2h, T2i);
|
Chris@10
|
133 T1F = T1x + T1E;
|
Chris@10
|
134 T3z = FNMS(KP250000000, T1E, T1x);
|
Chris@10
|
135 T6d = FNMS(KP559016994, T3A, T3z);
|
Chris@10
|
136 T3B = FMA(KP559016994, T3A, T3z);
|
Chris@10
|
137 }
|
Chris@10
|
138 {
|
Chris@10
|
139 E T2x, T5V, T2m, T2l, Ti, T5w, T3h, T36, TK, T35, T2F, T5L, T2I, Tr, T2H;
|
Chris@10
|
140 E T3a, T5z, T3d, T1r, T3c, T2q, T5S, T2t, TZ, T2s, T5O, T2M, T2B, Tt, T18;
|
Chris@10
|
141 E T2A, Tx, T2V, T2Y, Tw, T30, T1i, T2X, Ty;
|
Chris@10
|
142 {
|
Chris@10
|
143 E T1j, T1k, T1p, T39, T1l;
|
Chris@10
|
144 {
|
Chris@10
|
145 E TC, TI, T3g, TD, TE;
|
Chris@10
|
146 {
|
Chris@10
|
147 E Ta, Te, Tf, Tb, Tc, T5s, T2g, T2w, Tg;
|
Chris@10
|
148 Ta = cr[WS(rs, 1)];
|
Chris@10
|
149 T5s = FNMS(KP559016994, T2f, T2e);
|
Chris@10
|
150 T2g = FMA(KP559016994, T2f, T2e);
|
Chris@10
|
151 T6f = FNMS(KP951056516, T6e, T6d);
|
Chris@10
|
152 T7d = FMA(KP951056516, T6e, T6d);
|
Chris@10
|
153 Te = cr[WS(rs, 11)];
|
Chris@10
|
154 T5u = FMA(KP951056516, T5t, T5s);
|
Chris@10
|
155 T6U = FNMS(KP951056516, T5t, T5s);
|
Chris@10
|
156 T4k = FMA(KP951056516, T2j, T2g);
|
Chris@10
|
157 T2k = FNMS(KP951056516, T2j, T2g);
|
Chris@10
|
158 Tf = ci[WS(rs, 8)];
|
Chris@10
|
159 Tb = cr[WS(rs, 6)];
|
Chris@10
|
160 Tc = ci[WS(rs, 3)];
|
Chris@10
|
161 TC = cr[WS(rs, 3)];
|
Chris@10
|
162 T2w = Tf - Te;
|
Chris@10
|
163 Tg = Te + Tf;
|
Chris@10
|
164 {
|
Chris@10
|
165 E T2v, Td, Th, TG, TH;
|
Chris@10
|
166 T2v = Tb - Tc;
|
Chris@10
|
167 Td = Tb + Tc;
|
Chris@10
|
168 TG = ci[WS(rs, 11)];
|
Chris@10
|
169 TH = ci[WS(rs, 6)];
|
Chris@10
|
170 T2x = FNMS(KP618033988, T2w, T2v);
|
Chris@10
|
171 T5V = FMA(KP618033988, T2v, T2w);
|
Chris@10
|
172 Th = Td + Tg;
|
Chris@10
|
173 T2m = Td - Tg;
|
Chris@10
|
174 TI = TG + TH;
|
Chris@10
|
175 T3g = TG - TH;
|
Chris@10
|
176 T2l = FNMS(KP250000000, Th, Ta);
|
Chris@10
|
177 Ti = Ta + Th;
|
Chris@10
|
178 TD = cr[WS(rs, 8)];
|
Chris@10
|
179 TE = ci[WS(rs, 1)];
|
Chris@10
|
180 }
|
Chris@10
|
181 }
|
Chris@10
|
182 {
|
Chris@10
|
183 E Tj, Tk, Tp, T2E, TJ, Tl;
|
Chris@10
|
184 Tj = cr[WS(rs, 4)];
|
Chris@10
|
185 {
|
Chris@10
|
186 E Tn, To, T3f, TF;
|
Chris@10
|
187 Tn = ci[WS(rs, 10)];
|
Chris@10
|
188 To = ci[WS(rs, 5)];
|
Chris@10
|
189 T3f = TD - TE;
|
Chris@10
|
190 TF = TD + TE;
|
Chris@10
|
191 Tk = cr[WS(rs, 9)];
|
Chris@10
|
192 Tp = Tn + To;
|
Chris@10
|
193 T2E = To - Tn;
|
Chris@10
|
194 T5w = FNMS(KP618033988, T3f, T3g);
|
Chris@10
|
195 T3h = FMA(KP618033988, T3g, T3f);
|
Chris@10
|
196 T36 = TI - TF;
|
Chris@10
|
197 TJ = TF + TI;
|
Chris@10
|
198 Tl = ci[0];
|
Chris@10
|
199 }
|
Chris@10
|
200 T1j = ci[WS(rs, 21)];
|
Chris@10
|
201 TK = TC + TJ;
|
Chris@10
|
202 T35 = FNMS(KP250000000, TJ, TC);
|
Chris@10
|
203 {
|
Chris@10
|
204 E T1n, Tm, T2D, T1o, Tq;
|
Chris@10
|
205 T1n = cr[WS(rs, 13)];
|
Chris@10
|
206 Tm = Tk + Tl;
|
Chris@10
|
207 T2D = Tl - Tk;
|
Chris@10
|
208 T1o = cr[WS(rs, 18)];
|
Chris@10
|
209 T1k = ci[WS(rs, 16)];
|
Chris@10
|
210 T2F = FMA(KP618033988, T2E, T2D);
|
Chris@10
|
211 T5L = FNMS(KP618033988, T2D, T2E);
|
Chris@10
|
212 T2I = Tm - Tp;
|
Chris@10
|
213 Tq = Tm + Tp;
|
Chris@10
|
214 T1p = T1n + T1o;
|
Chris@10
|
215 T39 = T1o - T1n;
|
Chris@10
|
216 Tr = Tj + Tq;
|
Chris@10
|
217 T2H = FMS(KP250000000, Tq, Tj);
|
Chris@10
|
218 T1l = cr[WS(rs, 23)];
|
Chris@10
|
219 }
|
Chris@10
|
220 }
|
Chris@10
|
221 }
|
Chris@10
|
222 {
|
Chris@10
|
223 E T10, T11, T16, T2L, T12;
|
Chris@10
|
224 {
|
Chris@10
|
225 E TR, TS, TX, T2p, T1q, TT;
|
Chris@10
|
226 TR = ci[WS(rs, 23)];
|
Chris@10
|
227 {
|
Chris@10
|
228 E TV, TW, T38, T1m;
|
Chris@10
|
229 TV = ci[WS(rs, 13)];
|
Chris@10
|
230 TW = cr[WS(rs, 16)];
|
Chris@10
|
231 T38 = T1k + T1l;
|
Chris@10
|
232 T1m = T1k - T1l;
|
Chris@10
|
233 TS = ci[WS(rs, 18)];
|
Chris@10
|
234 TX = TV - TW;
|
Chris@10
|
235 T2p = TV + TW;
|
Chris@10
|
236 T3a = FMA(KP618033988, T39, T38);
|
Chris@10
|
237 T5z = FNMS(KP618033988, T38, T39);
|
Chris@10
|
238 T3d = T1m + T1p;
|
Chris@10
|
239 T1q = T1m - T1p;
|
Chris@10
|
240 TT = cr[WS(rs, 21)];
|
Chris@10
|
241 }
|
Chris@10
|
242 T10 = ci[WS(rs, 20)];
|
Chris@10
|
243 T1r = T1j + T1q;
|
Chris@10
|
244 T3c = FMS(KP250000000, T1q, T1j);
|
Chris@10
|
245 {
|
Chris@10
|
246 E T14, TU, T2o, T15, TY;
|
Chris@10
|
247 T14 = cr[WS(rs, 14)];
|
Chris@10
|
248 TU = TS - TT;
|
Chris@10
|
249 T2o = TS + TT;
|
Chris@10
|
250 T15 = cr[WS(rs, 19)];
|
Chris@10
|
251 T11 = ci[WS(rs, 15)];
|
Chris@10
|
252 T2q = FMA(KP618033988, T2p, T2o);
|
Chris@10
|
253 T5S = FNMS(KP618033988, T2o, T2p);
|
Chris@10
|
254 T2t = TU - TX;
|
Chris@10
|
255 TY = TU + TX;
|
Chris@10
|
256 T16 = T14 + T15;
|
Chris@10
|
257 T2L = T15 - T14;
|
Chris@10
|
258 TZ = TR + TY;
|
Chris@10
|
259 T2s = FNMS(KP250000000, TY, TR);
|
Chris@10
|
260 T12 = cr[WS(rs, 24)];
|
Chris@10
|
261 }
|
Chris@10
|
262 }
|
Chris@10
|
263 {
|
Chris@10
|
264 E T1a, T1e, T1d, T2T, T17, T1f;
|
Chris@10
|
265 T1a = ci[WS(rs, 22)];
|
Chris@10
|
266 {
|
Chris@10
|
267 E T1b, T1c, T2K, T13;
|
Chris@10
|
268 T1b = ci[WS(rs, 17)];
|
Chris@10
|
269 T1c = cr[WS(rs, 22)];
|
Chris@10
|
270 T2K = T11 + T12;
|
Chris@10
|
271 T13 = T11 - T12;
|
Chris@10
|
272 T1e = ci[WS(rs, 12)];
|
Chris@10
|
273 T1d = T1b - T1c;
|
Chris@10
|
274 T2T = T1b + T1c;
|
Chris@10
|
275 T5O = FNMS(KP618033988, T2K, T2L);
|
Chris@10
|
276 T2M = FMA(KP618033988, T2L, T2K);
|
Chris@10
|
277 T2B = T13 + T16;
|
Chris@10
|
278 T17 = T13 - T16;
|
Chris@10
|
279 T1f = cr[WS(rs, 17)];
|
Chris@10
|
280 }
|
Chris@10
|
281 Tt = cr[WS(rs, 2)];
|
Chris@10
|
282 T18 = T10 + T17;
|
Chris@10
|
283 T2A = FMS(KP250000000, T17, T10);
|
Chris@10
|
284 {
|
Chris@10
|
285 E Tu, T1g, T2U, Tv, T1h;
|
Chris@10
|
286 Tu = cr[WS(rs, 7)];
|
Chris@10
|
287 T1g = T1e - T1f;
|
Chris@10
|
288 T2U = T1e + T1f;
|
Chris@10
|
289 Tv = ci[WS(rs, 2)];
|
Chris@10
|
290 Tx = cr[WS(rs, 12)];
|
Chris@10
|
291 T2V = FMA(KP618033988, T2U, T2T);
|
Chris@10
|
292 T5G = FNMS(KP618033988, T2T, T2U);
|
Chris@10
|
293 T2Y = T1d - T1g;
|
Chris@10
|
294 T1h = T1d + T1g;
|
Chris@10
|
295 Tw = Tu + Tv;
|
Chris@10
|
296 T30 = Tu - Tv;
|
Chris@10
|
297 T1i = T1a + T1h;
|
Chris@10
|
298 T2X = FMS(KP250000000, T1h, T1a);
|
Chris@10
|
299 Ty = ci[WS(rs, 7)];
|
Chris@10
|
300 }
|
Chris@10
|
301 }
|
Chris@10
|
302 }
|
Chris@10
|
303 }
|
Chris@10
|
304 {
|
Chris@10
|
305 E T32, T5D, T2R, T2Q, T2u, T2r, T4t;
|
Chris@10
|
306 {
|
Chris@10
|
307 E TA, T31, Tz, TB, Ts;
|
Chris@10
|
308 T31 = Ty - Tx;
|
Chris@10
|
309 Tz = Tx + Ty;
|
Chris@10
|
310 T1G = TZ + T18;
|
Chris@10
|
311 T19 = TZ - T18;
|
Chris@10
|
312 T32 = FNMS(KP618033988, T31, T30);
|
Chris@10
|
313 T5D = FMA(KP618033988, T30, T31);
|
Chris@10
|
314 TA = Tw + Tz;
|
Chris@10
|
315 T2R = Tz - Tw;
|
Chris@10
|
316 T2Q = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
317 TB = Tt + TA;
|
Chris@10
|
318 T1H = T1i + T1r;
|
Chris@10
|
319 T1s = T1i - T1r;
|
Chris@10
|
320 T1M = Ti - Tr;
|
Chris@10
|
321 Ts = Ti + Tr;
|
Chris@10
|
322 {
|
Chris@10
|
323 E T2n, T5R, T5U, TL;
|
Chris@10
|
324 T2n = FMA(KP559016994, T2m, T2l);
|
Chris@10
|
325 T5R = FNMS(KP559016994, T2m, T2l);
|
Chris@10
|
326 T5U = FNMS(KP559016994, T2t, T2s);
|
Chris@10
|
327 T2u = FMA(KP559016994, T2t, T2s);
|
Chris@10
|
328 TL = TB + TK;
|
Chris@10
|
329 T1N = TB - TK;
|
Chris@10
|
330 {
|
Chris@10
|
331 E T5T, T75, T5W, T76;
|
Chris@10
|
332 T5T = FMA(KP951056516, T5S, T5R);
|
Chris@10
|
333 T75 = FNMS(KP951056516, T5S, T5R);
|
Chris@10
|
334 T5W = FMA(KP951056516, T5V, T5U);
|
Chris@10
|
335 T76 = FNMS(KP951056516, T5V, T5U);
|
Chris@10
|
336 TP = Ts - TL;
|
Chris@10
|
337 TM = Ts + TL;
|
Chris@10
|
338 T2r = FNMS(KP951056516, T2q, T2n);
|
Chris@10
|
339 T4t = FMA(KP951056516, T2q, T2n);
|
Chris@10
|
340 T7i = FMA(KP939062505, T75, T76);
|
Chris@10
|
341 T77 = FNMS(KP939062505, T76, T75);
|
Chris@10
|
342 T5X = FNMS(KP549754652, T5W, T5T);
|
Chris@10
|
343 T64 = FMA(KP549754652, T5T, T5W);
|
Chris@10
|
344 }
|
Chris@10
|
345 }
|
Chris@10
|
346 }
|
Chris@10
|
347 {
|
Chris@10
|
348 E T2J, T2G, T4v, T5y, T37, T3e, T5v;
|
Chris@10
|
349 {
|
Chris@10
|
350 E T2C, T5K, T5N, T4s, T2y;
|
Chris@10
|
351 T2C = FNMS(KP559016994, T2B, T2A);
|
Chris@10
|
352 T5K = FMA(KP559016994, T2B, T2A);
|
Chris@10
|
353 T5N = FMA(KP559016994, T2I, T2H);
|
Chris@10
|
354 T2J = FNMS(KP559016994, T2I, T2H);
|
Chris@10
|
355 T4s = FNMS(KP951056516, T2x, T2u);
|
Chris@10
|
356 T2y = FMA(KP951056516, T2x, T2u);
|
Chris@10
|
357 {
|
Chris@10
|
358 E T73, T5M, T72, T5P;
|
Chris@10
|
359 T73 = FMA(KP951056516, T5L, T5K);
|
Chris@10
|
360 T5M = FNMS(KP951056516, T5L, T5K);
|
Chris@10
|
361 T72 = FMA(KP951056516, T5O, T5N);
|
Chris@10
|
362 T5P = FNMS(KP951056516, T5O, T5N);
|
Chris@10
|
363 T4u = FNMS(KP634619297, T4t, T4s);
|
Chris@10
|
364 T4D = FMA(KP634619297, T4s, T4t);
|
Chris@10
|
365 T3p = FMA(KP256756360, T2r, T2y);
|
Chris@10
|
366 T2z = FNMS(KP256756360, T2y, T2r);
|
Chris@10
|
367 T74 = FMA(KP126329378, T73, T72);
|
Chris@10
|
368 T7h = FNMS(KP126329378, T72, T73);
|
Chris@10
|
369 T63 = FNMS(KP470564281, T5M, T5P);
|
Chris@10
|
370 T5Q = FMA(KP470564281, T5P, T5M);
|
Chris@10
|
371 T2G = FMA(KP951056516, T2F, T2C);
|
Chris@10
|
372 T4v = FNMS(KP951056516, T2F, T2C);
|
Chris@10
|
373 }
|
Chris@10
|
374 T5y = FMA(KP559016994, T36, T35);
|
Chris@10
|
375 T37 = FNMS(KP559016994, T36, T35);
|
Chris@10
|
376 T3e = FNMS(KP559016994, T3d, T3c);
|
Chris@10
|
377 T5v = FMA(KP559016994, T3d, T3c);
|
Chris@10
|
378 }
|
Chris@10
|
379 {
|
Chris@10
|
380 E T5x, T6Y, T4w, T2N;
|
Chris@10
|
381 T4w = FNMS(KP951056516, T2M, T2J);
|
Chris@10
|
382 T2N = FMA(KP951056516, T2M, T2J);
|
Chris@10
|
383 {
|
Chris@10
|
384 E T4l, T3b, T4m, T3i;
|
Chris@10
|
385 T4l = FMA(KP951056516, T3a, T37);
|
Chris@10
|
386 T3b = FNMS(KP951056516, T3a, T37);
|
Chris@10
|
387 T4m = FMA(KP951056516, T3h, T3e);
|
Chris@10
|
388 T3i = FNMS(KP951056516, T3h, T3e);
|
Chris@10
|
389 T4x = FNMS(KP827271945, T4w, T4v);
|
Chris@10
|
390 T4E = FMA(KP827271945, T4v, T4w);
|
Chris@10
|
391 T3q = FMA(KP634619297, T2G, T2N);
|
Chris@10
|
392 T2O = FNMS(KP634619297, T2N, T2G);
|
Chris@10
|
393 T4n = FNMS(KP126329378, T4m, T4l);
|
Chris@10
|
394 T4G = FMA(KP126329378, T4l, T4m);
|
Chris@10
|
395 T3t = FNMS(KP939062505, T3b, T3i);
|
Chris@10
|
396 T3j = FMA(KP939062505, T3i, T3b);
|
Chris@10
|
397 T5x = FMA(KP951056516, T5w, T5v);
|
Chris@10
|
398 T6Y = FNMS(KP951056516, T5w, T5v);
|
Chris@10
|
399 }
|
Chris@10
|
400 {
|
Chris@10
|
401 E T2S, T2Z, T5C, T6Z, T5A;
|
Chris@10
|
402 T5F = FMA(KP559016994, T2R, T2Q);
|
Chris@10
|
403 T2S = FNMS(KP559016994, T2R, T2Q);
|
Chris@10
|
404 T2Z = FNMS(KP559016994, T2Y, T2X);
|
Chris@10
|
405 T5C = FMA(KP559016994, T2Y, T2X);
|
Chris@10
|
406 T6Z = FNMS(KP951056516, T5z, T5y);
|
Chris@10
|
407 T5A = FMA(KP951056516, T5z, T5y);
|
Chris@10
|
408 {
|
Chris@10
|
409 E T4p, T2W, T4o, T33;
|
Chris@10
|
410 T4p = FMA(KP951056516, T2V, T2S);
|
Chris@10
|
411 T2W = FNMS(KP951056516, T2V, T2S);
|
Chris@10
|
412 T4o = FMA(KP951056516, T32, T2Z);
|
Chris@10
|
413 T33 = FNMS(KP951056516, T32, T2Z);
|
Chris@10
|
414 T70 = FNMS(KP827271945, T6Z, T6Y);
|
Chris@10
|
415 T7f = FMA(KP827271945, T6Y, T6Z);
|
Chris@10
|
416 T66 = FNMS(KP062914667, T5x, T5A);
|
Chris@10
|
417 T5B = FMA(KP062914667, T5A, T5x);
|
Chris@10
|
418 T4q = FNMS(KP470564281, T4p, T4o);
|
Chris@10
|
419 T4H = FMA(KP470564281, T4o, T4p);
|
Chris@10
|
420 T3s = FNMS(KP549754652, T2W, T33);
|
Chris@10
|
421 T34 = FMA(KP549754652, T33, T2W);
|
Chris@10
|
422 T5E = FNMS(KP951056516, T5D, T5C);
|
Chris@10
|
423 T6V = FMA(KP951056516, T5D, T5C);
|
Chris@10
|
424 }
|
Chris@10
|
425 }
|
Chris@10
|
426 }
|
Chris@10
|
427 }
|
Chris@10
|
428 }
|
Chris@10
|
429 }
|
Chris@10
|
430 }
|
Chris@10
|
431 {
|
Chris@10
|
432 E T6X, T7e, T6A, T6F, T6C, T6G, T6B;
|
Chris@10
|
433 cr[0] = T9 + TM;
|
Chris@10
|
434 {
|
Chris@10
|
435 E T67, T5I, T25, T22, T1X, T26, T21;
|
Chris@10
|
436 {
|
Chris@10
|
437 E T1I, T23, T1L, T1Z, T1t, TO, T24, T1O;
|
Chris@10
|
438 {
|
Chris@10
|
439 E T1K, T6W, T5H, T1J;
|
Chris@10
|
440 T1K = T1G - T1H;
|
Chris@10
|
441 T1I = T1G + T1H;
|
Chris@10
|
442 T6W = FNMS(KP951056516, T5G, T5F);
|
Chris@10
|
443 T5H = FMA(KP951056516, T5G, T5F);
|
Chris@10
|
444 T1J = FNMS(KP250000000, T1I, T1F);
|
Chris@10
|
445 T6X = FMA(KP062914667, T6W, T6V);
|
Chris@10
|
446 T7e = FNMS(KP062914667, T6V, T6W);
|
Chris@10
|
447 T67 = FNMS(KP634619297, T5E, T5H);
|
Chris@10
|
448 T5I = FMA(KP634619297, T5H, T5E);
|
Chris@10
|
449 T23 = FNMS(KP559016994, T1K, T1J);
|
Chris@10
|
450 T1L = FMA(KP559016994, T1K, T1J);
|
Chris@10
|
451 T1Z = FNMS(KP618033988, T19, T1s);
|
Chris@10
|
452 T1t = FMA(KP618033988, T1s, T19);
|
Chris@10
|
453 TO = FNMS(KP250000000, TM, T9);
|
Chris@10
|
454 T24 = FNMS(KP618033988, T1M, T1N);
|
Chris@10
|
455 T1O = FMA(KP618033988, T1N, T1M);
|
Chris@10
|
456 }
|
Chris@10
|
457 {
|
Chris@10
|
458 E T2b, T2a, T1Y, TQ, T27;
|
Chris@10
|
459 ci[0] = T1F + T1I;
|
Chris@10
|
460 T2b = FMA(KP951056516, T24, T23);
|
Chris@10
|
461 T25 = FNMS(KP951056516, T24, T23);
|
Chris@10
|
462 T2a = W[29];
|
Chris@10
|
463 T1Y = FNMS(KP559016994, TP, TO);
|
Chris@10
|
464 TQ = FMA(KP559016994, TP, TO);
|
Chris@10
|
465 T27 = W[28];
|
Chris@10
|
466 {
|
Chris@10
|
467 E T1V, T1P, T20, T1S, T1w, T1v, TN, T1Q;
|
Chris@10
|
468 T1V = FNMS(KP951056516, T1O, T1L);
|
Chris@10
|
469 T1P = FMA(KP951056516, T1O, T1L);
|
Chris@10
|
470 {
|
Chris@10
|
471 E T28, T1u, T29, T2c;
|
Chris@10
|
472 T20 = FMA(KP951056516, T1Z, T1Y);
|
Chris@10
|
473 T28 = FNMS(KP951056516, T1Z, T1Y);
|
Chris@10
|
474 T1S = FMA(KP951056516, T1t, TQ);
|
Chris@10
|
475 T1u = FNMS(KP951056516, T1t, TQ);
|
Chris@10
|
476 T1w = W[9];
|
Chris@10
|
477 T29 = T27 * T28;
|
Chris@10
|
478 T2c = T2a * T28;
|
Chris@10
|
479 TN = W[8];
|
Chris@10
|
480 T1Q = T1w * T1u;
|
Chris@10
|
481 cr[WS(rs, 15)] = FNMS(T2a, T2b, T29);
|
Chris@10
|
482 ci[WS(rs, 15)] = FMA(T27, T2b, T2c);
|
Chris@10
|
483 T1v = TN * T1u;
|
Chris@10
|
484 }
|
Chris@10
|
485 ci[WS(rs, 5)] = FMA(TN, T1P, T1Q);
|
Chris@10
|
486 {
|
Chris@10
|
487 E T1U, T1R, T1W, T1T;
|
Chris@10
|
488 T1U = W[39];
|
Chris@10
|
489 cr[WS(rs, 5)] = FNMS(T1w, T1P, T1v);
|
Chris@10
|
490 T1R = W[38];
|
Chris@10
|
491 T1W = T1U * T1S;
|
Chris@10
|
492 T22 = W[19];
|
Chris@10
|
493 T1T = T1R * T1S;
|
Chris@10
|
494 T1X = W[18];
|
Chris@10
|
495 ci[WS(rs, 20)] = FMA(T1R, T1V, T1W);
|
Chris@10
|
496 T26 = T22 * T20;
|
Chris@10
|
497 cr[WS(rs, 20)] = FNMS(T1U, T1V, T1T);
|
Chris@10
|
498 T21 = T1X * T20;
|
Chris@10
|
499 }
|
Chris@10
|
500 }
|
Chris@10
|
501 }
|
Chris@10
|
502 }
|
Chris@10
|
503 {
|
Chris@10
|
504 E T6h, T6g, T5Y, T5J, T6z, T69, T6o, T6E;
|
Chris@10
|
505 {
|
Chris@10
|
506 E T6m, T6n, T65, T68;
|
Chris@10
|
507 T65 = FMA(KP968479752, T64, T63);
|
Chris@10
|
508 T6h = FNMS(KP968479752, T64, T63);
|
Chris@10
|
509 ci[WS(rs, 10)] = FMA(T1X, T25, T26);
|
Chris@10
|
510 T68 = FNMS(KP845997307, T67, T66);
|
Chris@10
|
511 T6g = FMA(KP845997307, T67, T66);
|
Chris@10
|
512 cr[WS(rs, 10)] = FNMS(T22, T25, T21);
|
Chris@10
|
513 T6m = FNMS(KP968479752, T5X, T5Q);
|
Chris@10
|
514 T5Y = FMA(KP968479752, T5X, T5Q);
|
Chris@10
|
515 T5J = FMA(KP845997307, T5I, T5B);
|
Chris@10
|
516 T6n = FNMS(KP845997307, T5I, T5B);
|
Chris@10
|
517 T6z = FMA(KP560319534, T65, T68);
|
Chris@10
|
518 T69 = FNMS(KP681693190, T68, T65);
|
Chris@10
|
519 T6o = FMA(KP681693190, T6n, T6m);
|
Chris@10
|
520 T6E = FNMS(KP560319534, T6m, T6n);
|
Chris@10
|
521 }
|
Chris@10
|
522 {
|
Chris@10
|
523 E T62, T6l, T6I, T6L, T6H, T6K;
|
Chris@10
|
524 {
|
Chris@10
|
525 E T6Q, T6O, T6y, T6D, T6S;
|
Chris@10
|
526 {
|
Chris@10
|
527 E T6N, T5Z, T61, T6i, T6k;
|
Chris@10
|
528 T6N = W[2];
|
Chris@10
|
529 T5Z = FMA(KP906616052, T5Y, T5J);
|
Chris@10
|
530 T61 = FNMS(KP906616052, T5Y, T5J);
|
Chris@10
|
531 T6i = FNMS(KP906616052, T6h, T6g);
|
Chris@10
|
532 T6k = FMA(KP906616052, T6h, T6g);
|
Chris@10
|
533 T6Q = W[3];
|
Chris@10
|
534 {
|
Chris@10
|
535 E T60, T6j, T6R, T6P;
|
Chris@10
|
536 T60 = FNMS(KP249506682, T5Z, T5u);
|
Chris@10
|
537 T6O = FMA(KP998026728, T5Z, T5u);
|
Chris@10
|
538 T6j = FNMS(KP249506682, T6i, T6f);
|
Chris@10
|
539 T6R = FMA(KP998026728, T6i, T6f);
|
Chris@10
|
540 T6y = FMA(KP557913902, T61, T60);
|
Chris@10
|
541 T62 = FNMS(KP557913902, T61, T60);
|
Chris@10
|
542 T6P = T6N * T6O;
|
Chris@10
|
543 T6l = FNMS(KP557913902, T6k, T6j);
|
Chris@10
|
544 T6D = FMA(KP557913902, T6k, T6j);
|
Chris@10
|
545 T6S = T6N * T6R;
|
Chris@10
|
546 cr[WS(rs, 2)] = FNMS(T6Q, T6R, T6P);
|
Chris@10
|
547 }
|
Chris@10
|
548 }
|
Chris@10
|
549 T6A = FNMS(KP949179823, T6z, T6y);
|
Chris@10
|
550 T6I = FMA(KP949179823, T6z, T6y);
|
Chris@10
|
551 T6L = FNMS(KP949179823, T6E, T6D);
|
Chris@10
|
552 T6F = FMA(KP949179823, T6E, T6D);
|
Chris@10
|
553 ci[WS(rs, 2)] = FMA(T6Q, T6O, T6S);
|
Chris@10
|
554 T6H = W[32];
|
Chris@10
|
555 T6K = W[33];
|
Chris@10
|
556 }
|
Chris@10
|
557 {
|
Chris@10
|
558 E T6a, T6s, T6v, T6p, T6c, T6q, T6b, T6M, T6J, T5r;
|
Chris@10
|
559 T6a = FNMS(KP860541664, T69, T62);
|
Chris@10
|
560 T6s = FMA(KP860541664, T69, T62);
|
Chris@10
|
561 T6v = FMA(KP860541664, T6o, T6l);
|
Chris@10
|
562 T6p = FNMS(KP860541664, T6o, T6l);
|
Chris@10
|
563 T6M = T6H * T6L;
|
Chris@10
|
564 T6J = T6H * T6I;
|
Chris@10
|
565 T5r = W[12];
|
Chris@10
|
566 T6c = W[13];
|
Chris@10
|
567 ci[WS(rs, 17)] = FMA(T6K, T6I, T6M);
|
Chris@10
|
568 cr[WS(rs, 17)] = FNMS(T6K, T6L, T6J);
|
Chris@10
|
569 T6q = T5r * T6p;
|
Chris@10
|
570 T6b = T5r * T6a;
|
Chris@10
|
571 {
|
Chris@10
|
572 E T6r, T6u, T6w, T6t, T6x;
|
Chris@10
|
573 ci[WS(rs, 7)] = FMA(T6c, T6a, T6q);
|
Chris@10
|
574 cr[WS(rs, 7)] = FNMS(T6c, T6p, T6b);
|
Chris@10
|
575 T6r = W[42];
|
Chris@10
|
576 T6u = W[43];
|
Chris@10
|
577 T6w = T6r * T6v;
|
Chris@10
|
578 T6t = T6r * T6s;
|
Chris@10
|
579 T6x = W[22];
|
Chris@10
|
580 T6C = W[23];
|
Chris@10
|
581 ci[WS(rs, 22)] = FMA(T6u, T6s, T6w);
|
Chris@10
|
582 cr[WS(rs, 22)] = FNMS(T6u, T6v, T6t);
|
Chris@10
|
583 T6G = T6x * T6F;
|
Chris@10
|
584 T6B = T6x * T6A;
|
Chris@10
|
585 }
|
Chris@10
|
586 }
|
Chris@10
|
587 }
|
Chris@10
|
588 }
|
Chris@10
|
589 }
|
Chris@10
|
590 {
|
Chris@10
|
591 E T7u, T7D, T7n, T7w, T7v, T7E;
|
Chris@10
|
592 {
|
Chris@10
|
593 E T78, T7t, T7N, T71, T7C, T7S, T7y, T7k;
|
Chris@10
|
594 {
|
Chris@10
|
595 E T7j, T7g, T7A, T7B, T7r, T7s;
|
Chris@10
|
596 T7r = FNMS(KP734762448, T7i, T7h);
|
Chris@10
|
597 T7j = FMA(KP734762448, T7i, T7h);
|
Chris@10
|
598 T7g = FMA(KP772036680, T7f, T7e);
|
Chris@10
|
599 T7s = FNMS(KP772036680, T7f, T7e);
|
Chris@10
|
600 ci[WS(rs, 12)] = FMA(T6C, T6A, T6G);
|
Chris@10
|
601 cr[WS(rs, 12)] = FNMS(T6C, T6F, T6B);
|
Chris@10
|
602 T7A = FNMS(KP734762448, T77, T74);
|
Chris@10
|
603 T78 = FMA(KP734762448, T77, T74);
|
Chris@10
|
604 T7t = FNMS(KP621716863, T7s, T7r);
|
Chris@10
|
605 T7N = FMA(KP614372930, T7r, T7s);
|
Chris@10
|
606 T71 = FMA(KP772036680, T70, T6X);
|
Chris@10
|
607 T7B = FNMS(KP772036680, T70, T6X);
|
Chris@10
|
608 T7C = FNMS(KP621716863, T7B, T7A);
|
Chris@10
|
609 T7S = FMA(KP614372930, T7A, T7B);
|
Chris@10
|
610 T7y = FNMS(KP994076283, T7j, T7g);
|
Chris@10
|
611 T7k = FMA(KP994076283, T7j, T7g);
|
Chris@10
|
612 }
|
Chris@10
|
613 {
|
Chris@10
|
614 E T7c, T6T, T7x, T7l, T79, T7p;
|
Chris@10
|
615 T7c = W[5];
|
Chris@10
|
616 T6T = W[4];
|
Chris@10
|
617 T7x = FNMS(KP249506682, T7k, T7d);
|
Chris@10
|
618 T7l = FMA(KP998026728, T7k, T7d);
|
Chris@10
|
619 T79 = FMA(KP994076283, T78, T71);
|
Chris@10
|
620 T7p = FNMS(KP994076283, T78, T71);
|
Chris@10
|
621 {
|
Chris@10
|
622 E T7z, T7Y, T7Z, T7T, T7q, T7O, T7X, T7L, T7Q, T7P, T7U;
|
Chris@10
|
623 {
|
Chris@10
|
624 E T7V, T80, T7b, T7m, T7W;
|
Chris@10
|
625 {
|
Chris@10
|
626 E T7R, T7o, T7a, T7M;
|
Chris@10
|
627 T7V = W[34];
|
Chris@10
|
628 T7R = FMA(KP557913902, T7y, T7x);
|
Chris@10
|
629 T7z = FNMS(KP557913902, T7y, T7x);
|
Chris@10
|
630 T7Y = W[35];
|
Chris@10
|
631 T7o = FNMS(KP249506682, T79, T6U);
|
Chris@10
|
632 T7a = FMA(KP998026728, T79, T6U);
|
Chris@10
|
633 T7Z = FMA(KP949179823, T7S, T7R);
|
Chris@10
|
634 T7T = FNMS(KP949179823, T7S, T7R);
|
Chris@10
|
635 T7M = FMA(KP557913902, T7p, T7o);
|
Chris@10
|
636 T7q = FNMS(KP557913902, T7p, T7o);
|
Chris@10
|
637 T7b = T6T * T7a;
|
Chris@10
|
638 T7m = T7c * T7a;
|
Chris@10
|
639 T7W = FNMS(KP949179823, T7N, T7M);
|
Chris@10
|
640 T7O = FMA(KP949179823, T7N, T7M);
|
Chris@10
|
641 }
|
Chris@10
|
642 cr[WS(rs, 3)] = FNMS(T7c, T7l, T7b);
|
Chris@10
|
643 ci[WS(rs, 3)] = FMA(T6T, T7l, T7m);
|
Chris@10
|
644 T80 = T7Y * T7W;
|
Chris@10
|
645 T7X = T7V * T7W;
|
Chris@10
|
646 T7L = W[24];
|
Chris@10
|
647 T7Q = W[25];
|
Chris@10
|
648 ci[WS(rs, 18)] = FMA(T7V, T7Z, T80);
|
Chris@10
|
649 }
|
Chris@10
|
650 cr[WS(rs, 18)] = FNMS(T7Y, T7Z, T7X);
|
Chris@10
|
651 T7P = T7L * T7O;
|
Chris@10
|
652 T7U = T7Q * T7O;
|
Chris@10
|
653 {
|
Chris@10
|
654 E T7J, T7F, T7I, T7H, T7K, T7G;
|
Chris@10
|
655 T7u = FMA(KP943557151, T7t, T7q);
|
Chris@10
|
656 T7G = FNMS(KP943557151, T7t, T7q);
|
Chris@10
|
657 cr[WS(rs, 13)] = FNMS(T7Q, T7T, T7P);
|
Chris@10
|
658 ci[WS(rs, 13)] = FMA(T7L, T7T, T7U);
|
Chris@10
|
659 T7J = FMA(KP943557151, T7C, T7z);
|
Chris@10
|
660 T7D = FNMS(KP943557151, T7C, T7z);
|
Chris@10
|
661 T7F = W[44];
|
Chris@10
|
662 T7I = W[45];
|
Chris@10
|
663 T7n = W[14];
|
Chris@10
|
664 T7H = T7F * T7G;
|
Chris@10
|
665 T7K = T7I * T7G;
|
Chris@10
|
666 T7w = W[15];
|
Chris@10
|
667 T7v = T7n * T7u;
|
Chris@10
|
668 cr[WS(rs, 23)] = FNMS(T7I, T7J, T7H);
|
Chris@10
|
669 ci[WS(rs, 23)] = FMA(T7F, T7J, T7K);
|
Chris@10
|
670 }
|
Chris@10
|
671 }
|
Chris@10
|
672 }
|
Chris@10
|
673 }
|
Chris@10
|
674 T7E = T7w * T7u;
|
Chris@10
|
675 cr[WS(rs, 8)] = FNMS(T7w, T7D, T7v);
|
Chris@10
|
676 {
|
Chris@10
|
677 E T3F, T4K, T4X, T4j, T4M, T4L, T4Y;
|
Chris@10
|
678 {
|
Chris@10
|
679 E T4P, T4O, T4y, T4r, T4J, T57, T4N, T5c, T4W;
|
Chris@10
|
680 {
|
Chris@10
|
681 E T4U, T4V, T4F, T4I;
|
Chris@10
|
682 T4F = FNMS(KP912575812, T4E, T4D);
|
Chris@10
|
683 T4P = FMA(KP912575812, T4E, T4D);
|
Chris@10
|
684 T4O = FMA(KP912018591, T4H, T4G);
|
Chris@10
|
685 T4I = FNMS(KP912018591, T4H, T4G);
|
Chris@10
|
686 ci[WS(rs, 8)] = FMA(T7n, T7D, T7E);
|
Chris@10
|
687 T4y = FMA(KP912575812, T4x, T4u);
|
Chris@10
|
688 T4U = FNMS(KP912575812, T4x, T4u);
|
Chris@10
|
689 T4V = FMA(KP912018591, T4q, T4n);
|
Chris@10
|
690 T4r = FNMS(KP912018591, T4q, T4n);
|
Chris@10
|
691 T4J = FNMS(KP726211448, T4I, T4F);
|
Chris@10
|
692 T57 = FMA(KP525970792, T4F, T4I);
|
Chris@10
|
693 T3F = FMA(KP951056516, T3E, T3B);
|
Chris@10
|
694 T4N = FNMS(KP951056516, T3E, T3B);
|
Chris@10
|
695 T5c = FMA(KP525970792, T4U, T4V);
|
Chris@10
|
696 T4W = FNMS(KP726211448, T4V, T4U);
|
Chris@10
|
697 }
|
Chris@10
|
698 {
|
Chris@10
|
699 E T5o, T4S, T4B, T5l, T5p, T4R, T4A, T5m, T4Q, T4z;
|
Chris@10
|
700 T5o = W[7];
|
Chris@10
|
701 T4Q = FMA(KP851038619, T4P, T4O);
|
Chris@10
|
702 T4S = FNMS(KP851038619, T4P, T4O);
|
Chris@10
|
703 T4z = FMA(KP851038619, T4y, T4r);
|
Chris@10
|
704 T4B = FNMS(KP851038619, T4y, T4r);
|
Chris@10
|
705 T5l = W[6];
|
Chris@10
|
706 T5p = FMA(KP992114701, T4Q, T4N);
|
Chris@10
|
707 T4R = FNMS(KP248028675, T4Q, T4N);
|
Chris@10
|
708 T4A = FMA(KP248028675, T4z, T4k);
|
Chris@10
|
709 T5m = FNMS(KP992114701, T4z, T4k);
|
Chris@10
|
710 {
|
Chris@10
|
711 E T4T, T4C, T5d, T58, T55, T5a, T59, T5e;
|
Chris@10
|
712 {
|
Chris@10
|
713 E T5f, T5j, T5i, T5h, T5k, T5g;
|
Chris@10
|
714 T5f = W[36];
|
Chris@10
|
715 {
|
Chris@10
|
716 E T5b, T56, T5n, T5q;
|
Chris@10
|
717 T4T = FNMS(KP554608978, T4S, T4R);
|
Chris@10
|
718 T5b = FMA(KP554608978, T4S, T4R);
|
Chris@10
|
719 T56 = FNMS(KP554608978, T4B, T4A);
|
Chris@10
|
720 T4C = FMA(KP554608978, T4B, T4A);
|
Chris@10
|
721 T5n = T5l * T5m;
|
Chris@10
|
722 T5q = T5o * T5m;
|
Chris@10
|
723 T5j = FMA(KP943557151, T5c, T5b);
|
Chris@10
|
724 T5d = FNMS(KP943557151, T5c, T5b);
|
Chris@10
|
725 T5g = FMA(KP943557151, T57, T56);
|
Chris@10
|
726 T58 = FNMS(KP943557151, T57, T56);
|
Chris@10
|
727 cr[WS(rs, 4)] = FNMS(T5o, T5p, T5n);
|
Chris@10
|
728 ci[WS(rs, 4)] = FMA(T5l, T5p, T5q);
|
Chris@10
|
729 }
|
Chris@10
|
730 T5i = W[37];
|
Chris@10
|
731 T5h = T5f * T5g;
|
Chris@10
|
732 T55 = W[26];
|
Chris@10
|
733 T5k = T5i * T5g;
|
Chris@10
|
734 T5a = W[27];
|
Chris@10
|
735 cr[WS(rs, 19)] = FNMS(T5i, T5j, T5h);
|
Chris@10
|
736 T59 = T55 * T58;
|
Chris@10
|
737 ci[WS(rs, 19)] = FMA(T5f, T5j, T5k);
|
Chris@10
|
738 }
|
Chris@10
|
739 T5e = T5a * T58;
|
Chris@10
|
740 {
|
Chris@10
|
741 E T53, T4Z, T52, T51, T54, T50;
|
Chris@10
|
742 cr[WS(rs, 14)] = FNMS(T5a, T5d, T59);
|
Chris@10
|
743 T4K = FNMS(KP803003575, T4J, T4C);
|
Chris@10
|
744 T50 = FMA(KP803003575, T4J, T4C);
|
Chris@10
|
745 ci[WS(rs, 14)] = FMA(T55, T5d, T5e);
|
Chris@10
|
746 T4X = FNMS(KP803003575, T4W, T4T);
|
Chris@10
|
747 T53 = FMA(KP803003575, T4W, T4T);
|
Chris@10
|
748 T4Z = W[46];
|
Chris@10
|
749 T52 = W[47];
|
Chris@10
|
750 T4j = W[16];
|
Chris@10
|
751 T51 = T4Z * T50;
|
Chris@10
|
752 T54 = T52 * T50;
|
Chris@10
|
753 T4M = W[17];
|
Chris@10
|
754 T4L = T4j * T4K;
|
Chris@10
|
755 cr[WS(rs, 24)] = FNMS(T52, T53, T51);
|
Chris@10
|
756 ci[WS(rs, 24)] = FMA(T4Z, T53, T54);
|
Chris@10
|
757 }
|
Chris@10
|
758 }
|
Chris@10
|
759 }
|
Chris@10
|
760 }
|
Chris@10
|
761 T4Y = T4M * T4K;
|
Chris@10
|
762 cr[WS(rs, 9)] = FNMS(T4M, T4X, T4L);
|
Chris@10
|
763 {
|
Chris@10
|
764 E T3G, T3H, T2P, T3k, T3Z, T3v, T3O, T44;
|
Chris@10
|
765 {
|
Chris@10
|
766 E T3M, T3N, T3r, T3u;
|
Chris@10
|
767 T3G = FNMS(KP871714437, T3q, T3p);
|
Chris@10
|
768 T3r = FMA(KP871714437, T3q, T3p);
|
Chris@10
|
769 T3u = FNMS(KP831864738, T3t, T3s);
|
Chris@10
|
770 T3H = FMA(KP831864738, T3t, T3s);
|
Chris@10
|
771 ci[WS(rs, 9)] = FMA(T4j, T4X, T4Y);
|
Chris@10
|
772 T3M = FNMS(KP871714437, T2O, T2z);
|
Chris@10
|
773 T2P = FMA(KP871714437, T2O, T2z);
|
Chris@10
|
774 T3k = FMA(KP831864738, T3j, T34);
|
Chris@10
|
775 T3N = FNMS(KP831864738, T3j, T34);
|
Chris@10
|
776 T3Z = FMA(KP683113946, T3r, T3u);
|
Chris@10
|
777 T3v = FNMS(KP559154169, T3u, T3r);
|
Chris@10
|
778 T3O = FMA(KP559154169, T3N, T3M);
|
Chris@10
|
779 T44 = FNMS(KP683113946, T3M, T3N);
|
Chris@10
|
780 }
|
Chris@10
|
781 {
|
Chris@10
|
782 E T4g, T3K, T3n, T4d, T3J, T4h, T4e, T3m, T3I, T3l;
|
Chris@10
|
783 T4g = W[1];
|
Chris@10
|
784 T3K = FMA(KP904730450, T3H, T3G);
|
Chris@10
|
785 T3I = FNMS(KP904730450, T3H, T3G);
|
Chris@10
|
786 T3n = FNMS(KP904730450, T3k, T2P);
|
Chris@10
|
787 T3l = FMA(KP904730450, T3k, T2P);
|
Chris@10
|
788 T4d = W[0];
|
Chris@10
|
789 T3J = FNMS(KP242145790, T3I, T3F);
|
Chris@10
|
790 T4h = FMA(KP968583161, T3I, T3F);
|
Chris@10
|
791 T4e = FMA(KP968583161, T3l, T2k);
|
Chris@10
|
792 T3m = FNMS(KP242145790, T3l, T2k);
|
Chris@10
|
793 {
|
Chris@10
|
794 E T3L, T3o, T45, T40, T3X, T42, T41, T46;
|
Chris@10
|
795 {
|
Chris@10
|
796 E T47, T4b, T4a, T49, T4c, T48;
|
Chris@10
|
797 T47 = W[30];
|
Chris@10
|
798 {
|
Chris@10
|
799 E T43, T3Y, T4f, T4i;
|
Chris@10
|
800 T43 = FNMS(KP541454447, T3K, T3J);
|
Chris@10
|
801 T3L = FMA(KP541454447, T3K, T3J);
|
Chris@10
|
802 T3o = FMA(KP541454447, T3n, T3m);
|
Chris@10
|
803 T3Y = FNMS(KP541454447, T3n, T3m);
|
Chris@10
|
804 T4f = T4d * T4e;
|
Chris@10
|
805 T4i = T4g * T4e;
|
Chris@10
|
806 T45 = FNMS(KP833417178, T44, T43);
|
Chris@10
|
807 T4b = FMA(KP833417178, T44, T43);
|
Chris@10
|
808 T40 = FNMS(KP833417178, T3Z, T3Y);
|
Chris@10
|
809 T48 = FMA(KP833417178, T3Z, T3Y);
|
Chris@10
|
810 cr[WS(rs, 1)] = FNMS(T4g, T4h, T4f);
|
Chris@10
|
811 ci[WS(rs, 1)] = FMA(T4d, T4h, T4i);
|
Chris@10
|
812 }
|
Chris@10
|
813 T4a = W[31];
|
Chris@10
|
814 T49 = T47 * T48;
|
Chris@10
|
815 T3X = W[20];
|
Chris@10
|
816 T4c = T4a * T48;
|
Chris@10
|
817 T42 = W[21];
|
Chris@10
|
818 cr[WS(rs, 16)] = FNMS(T4a, T4b, T49);
|
Chris@10
|
819 T41 = T3X * T40;
|
Chris@10
|
820 ci[WS(rs, 16)] = FMA(T47, T4b, T4c);
|
Chris@10
|
821 }
|
Chris@10
|
822 T46 = T42 * T40;
|
Chris@10
|
823 {
|
Chris@10
|
824 E T3V, T3R, T3U, T3T, T3W, T3S;
|
Chris@10
|
825 cr[WS(rs, 11)] = FNMS(T42, T45, T41);
|
Chris@10
|
826 T3S = FMA(KP921177326, T3v, T3o);
|
Chris@10
|
827 T3w = FNMS(KP921177326, T3v, T3o);
|
Chris@10
|
828 ci[WS(rs, 11)] = FMA(T3X, T45, T46);
|
Chris@10
|
829 T3V = FNMS(KP921177326, T3O, T3L);
|
Chris@10
|
830 T3P = FMA(KP921177326, T3O, T3L);
|
Chris@10
|
831 T3R = W[40];
|
Chris@10
|
832 T3U = W[41];
|
Chris@10
|
833 T2d = W[10];
|
Chris@10
|
834 T3T = T3R * T3S;
|
Chris@10
|
835 T3W = T3U * T3S;
|
Chris@10
|
836 T3y = W[11];
|
Chris@10
|
837 T3x = T2d * T3w;
|
Chris@10
|
838 cr[WS(rs, 21)] = FNMS(T3U, T3V, T3T);
|
Chris@10
|
839 ci[WS(rs, 21)] = FMA(T3R, T3V, T3W);
|
Chris@10
|
840 }
|
Chris@10
|
841 }
|
Chris@10
|
842 }
|
Chris@10
|
843 }
|
Chris@10
|
844 }
|
Chris@10
|
845 }
|
Chris@10
|
846 }
|
Chris@10
|
847 }
|
Chris@10
|
848 cr[WS(rs, 6)] = FNMS(T3y, T3P, T3x);
|
Chris@10
|
849 T3Q = T3y * T3w;
|
Chris@10
|
850 ci[WS(rs, 6)] = FMA(T2d, T3P, T3Q);
|
Chris@10
|
851 }
|
Chris@10
|
852 }
|
Chris@10
|
853 }
|
Chris@10
|
854
|
Chris@10
|
855 static const tw_instr twinstr[] = {
|
Chris@10
|
856 {TW_FULL, 1, 25},
|
Chris@10
|
857 {TW_NEXT, 1, 0}
|
Chris@10
|
858 };
|
Chris@10
|
859
|
Chris@10
|
860 static const hc2hc_desc desc = { 25, "hb_25", twinstr, &GENUS, {84, 48, 316, 0} };
|
Chris@10
|
861
|
Chris@10
|
862 void X(codelet_hb_25) (planner *p) {
|
Chris@10
|
863 X(khc2hc_register) (p, hb_25, &desc);
|
Chris@10
|
864 }
|
Chris@10
|
865 #else /* HAVE_FMA */
|
Chris@10
|
866
|
Chris@10
|
867 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -dif -name hb_25 -include hb.h */
|
Chris@10
|
868
|
Chris@10
|
869 /*
|
Chris@10
|
870 * This function contains 400 FP additions, 280 FP multiplications,
|
Chris@10
|
871 * (or, 260 additions, 140 multiplications, 140 fused multiply/add),
|
Chris@10
|
872 * 107 stack variables, 20 constants, and 100 memory accesses
|
Chris@10
|
873 */
|
Chris@10
|
874 #include "hb.h"
|
Chris@10
|
875
|
Chris@10
|
876 static void hb_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
877 {
|
Chris@10
|
878 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
879 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
880 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
881 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
882 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
883 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
884 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
885 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
886 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
887 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
888 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
889 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
890 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
891 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
892 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
893 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
894 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
895 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
896 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
897 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
898 {
|
Chris@10
|
899 INT m;
|
Chris@10
|
900 for (m = mb, W = W + ((mb - 1) * 48); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 48, MAKE_VOLATILE_STRIDE(50, rs)) {
|
Chris@10
|
901 E T9, T5Q, T3y, T39, T5v, Ti, Tr, Ts, TZ, T18, T1z, T2k, T4l, T3h, T44;
|
Chris@10
|
902 E T5d, T6C, T5C, T6o, T56, T6B, T5B, T6l, T2z, T4m, T3i, T47, T1K, T5w, T3c;
|
Chris@10
|
903 E T3B, T5R, TB, TK, TL, T1i, T1r, T1A, T2P, T4o, T3k, T4b, T5s, T6F, T5F;
|
Chris@10
|
904 E T6v, T5l, T6E, T5E, T6s, T34, T4p, T3l, T4e;
|
Chris@10
|
905 {
|
Chris@10
|
906 E T1, T4, T7, T8, T3x, T3w, T37, T38;
|
Chris@10
|
907 T1 = cr[0];
|
Chris@10
|
908 {
|
Chris@10
|
909 E T2, T3, T5, T6;
|
Chris@10
|
910 T2 = cr[WS(rs, 5)];
|
Chris@10
|
911 T3 = ci[WS(rs, 4)];
|
Chris@10
|
912 T4 = T2 + T3;
|
Chris@10
|
913 T5 = cr[WS(rs, 10)];
|
Chris@10
|
914 T6 = ci[WS(rs, 9)];
|
Chris@10
|
915 T7 = T5 + T6;
|
Chris@10
|
916 T8 = T4 + T7;
|
Chris@10
|
917 T3x = T5 - T6;
|
Chris@10
|
918 T3w = T2 - T3;
|
Chris@10
|
919 }
|
Chris@10
|
920 T9 = T1 + T8;
|
Chris@10
|
921 T5Q = FMA(KP951056516, T3w, KP587785252 * T3x);
|
Chris@10
|
922 T3y = FNMS(KP951056516, T3x, KP587785252 * T3w);
|
Chris@10
|
923 T37 = FNMS(KP250000000, T8, T1);
|
Chris@10
|
924 T38 = KP559016994 * (T4 - T7);
|
Chris@10
|
925 T39 = T37 - T38;
|
Chris@10
|
926 T5v = T38 + T37;
|
Chris@10
|
927 }
|
Chris@10
|
928 {
|
Chris@10
|
929 E Ta, T27, T53, T2f, Th, T26, T10, T2p, T58, T2x, T17, T2o, Tj, T2n, T5a;
|
Chris@10
|
930 E T2t, Tq, T2s, TR, T2b, T51, T2h, TY, T2g;
|
Chris@10
|
931 {
|
Chris@10
|
932 E Tg, T2e, Td, T2d;
|
Chris@10
|
933 Ta = cr[WS(rs, 1)];
|
Chris@10
|
934 {
|
Chris@10
|
935 E Te, Tf, Tb, Tc;
|
Chris@10
|
936 Te = cr[WS(rs, 11)];
|
Chris@10
|
937 Tf = ci[WS(rs, 8)];
|
Chris@10
|
938 Tg = Te + Tf;
|
Chris@10
|
939 T2e = Te - Tf;
|
Chris@10
|
940 Tb = cr[WS(rs, 6)];
|
Chris@10
|
941 Tc = ci[WS(rs, 3)];
|
Chris@10
|
942 Td = Tb + Tc;
|
Chris@10
|
943 T2d = Tb - Tc;
|
Chris@10
|
944 }
|
Chris@10
|
945 T27 = KP559016994 * (Td - Tg);
|
Chris@10
|
946 T53 = FMA(KP951056516, T2d, KP587785252 * T2e);
|
Chris@10
|
947 T2f = FNMS(KP951056516, T2e, KP587785252 * T2d);
|
Chris@10
|
948 Th = Td + Tg;
|
Chris@10
|
949 T26 = FNMS(KP250000000, Th, Ta);
|
Chris@10
|
950 }
|
Chris@10
|
951 {
|
Chris@10
|
952 E T16, T2w, T13, T2v;
|
Chris@10
|
953 T10 = ci[WS(rs, 20)];
|
Chris@10
|
954 {
|
Chris@10
|
955 E T14, T15, T11, T12;
|
Chris@10
|
956 T14 = cr[WS(rs, 14)];
|
Chris@10
|
957 T15 = cr[WS(rs, 19)];
|
Chris@10
|
958 T16 = T14 + T15;
|
Chris@10
|
959 T2w = T15 - T14;
|
Chris@10
|
960 T11 = ci[WS(rs, 15)];
|
Chris@10
|
961 T12 = cr[WS(rs, 24)];
|
Chris@10
|
962 T13 = T11 - T12;
|
Chris@10
|
963 T2v = T11 + T12;
|
Chris@10
|
964 }
|
Chris@10
|
965 T2p = KP559016994 * (T13 + T16);
|
Chris@10
|
966 T58 = FMA(KP951056516, T2v, KP587785252 * T2w);
|
Chris@10
|
967 T2x = FNMS(KP951056516, T2w, KP587785252 * T2v);
|
Chris@10
|
968 T17 = T13 - T16;
|
Chris@10
|
969 T2o = FNMS(KP250000000, T17, T10);
|
Chris@10
|
970 }
|
Chris@10
|
971 {
|
Chris@10
|
972 E Tp, T2m, Tm, T2l;
|
Chris@10
|
973 Tj = cr[WS(rs, 4)];
|
Chris@10
|
974 {
|
Chris@10
|
975 E Tn, To, Tk, Tl;
|
Chris@10
|
976 Tn = ci[WS(rs, 10)];
|
Chris@10
|
977 To = ci[WS(rs, 5)];
|
Chris@10
|
978 Tp = Tn + To;
|
Chris@10
|
979 T2m = Tn - To;
|
Chris@10
|
980 Tk = cr[WS(rs, 9)];
|
Chris@10
|
981 Tl = ci[0];
|
Chris@10
|
982 Tm = Tk + Tl;
|
Chris@10
|
983 T2l = Tk - Tl;
|
Chris@10
|
984 }
|
Chris@10
|
985 T2n = FNMS(KP951056516, T2m, KP587785252 * T2l);
|
Chris@10
|
986 T5a = FMA(KP951056516, T2l, KP587785252 * T2m);
|
Chris@10
|
987 T2t = KP559016994 * (Tm - Tp);
|
Chris@10
|
988 Tq = Tm + Tp;
|
Chris@10
|
989 T2s = FNMS(KP250000000, Tq, Tj);
|
Chris@10
|
990 }
|
Chris@10
|
991 {
|
Chris@10
|
992 E TX, T2a, TU, T29;
|
Chris@10
|
993 TR = ci[WS(rs, 23)];
|
Chris@10
|
994 {
|
Chris@10
|
995 E TV, TW, TS, TT;
|
Chris@10
|
996 TV = ci[WS(rs, 13)];
|
Chris@10
|
997 TW = cr[WS(rs, 16)];
|
Chris@10
|
998 TX = TV - TW;
|
Chris@10
|
999 T2a = TV + TW;
|
Chris@10
|
1000 TS = ci[WS(rs, 18)];
|
Chris@10
|
1001 TT = cr[WS(rs, 21)];
|
Chris@10
|
1002 TU = TS - TT;
|
Chris@10
|
1003 T29 = TS + TT;
|
Chris@10
|
1004 }
|
Chris@10
|
1005 T2b = FNMS(KP951056516, T2a, KP587785252 * T29);
|
Chris@10
|
1006 T51 = FMA(KP951056516, T29, KP587785252 * T2a);
|
Chris@10
|
1007 T2h = KP559016994 * (TU - TX);
|
Chris@10
|
1008 TY = TU + TX;
|
Chris@10
|
1009 T2g = FNMS(KP250000000, TY, TR);
|
Chris@10
|
1010 }
|
Chris@10
|
1011 Ti = Ta + Th;
|
Chris@10
|
1012 Tr = Tj + Tq;
|
Chris@10
|
1013 Ts = Ti + Tr;
|
Chris@10
|
1014 TZ = TR + TY;
|
Chris@10
|
1015 T18 = T10 + T17;
|
Chris@10
|
1016 T1z = TZ + T18;
|
Chris@10
|
1017 {
|
Chris@10
|
1018 E T2c, T42, T2j, T43, T28, T2i;
|
Chris@10
|
1019 T28 = T26 - T27;
|
Chris@10
|
1020 T2c = T28 - T2b;
|
Chris@10
|
1021 T42 = T28 + T2b;
|
Chris@10
|
1022 T2i = T2g - T2h;
|
Chris@10
|
1023 T2j = T2f + T2i;
|
Chris@10
|
1024 T43 = T2i - T2f;
|
Chris@10
|
1025 T2k = FNMS(KP481753674, T2j, KP876306680 * T2c);
|
Chris@10
|
1026 T4l = FMA(KP728968627, T43, KP684547105 * T42);
|
Chris@10
|
1027 T3h = FMA(KP876306680, T2j, KP481753674 * T2c);
|
Chris@10
|
1028 T44 = FNMS(KP684547105, T43, KP728968627 * T42);
|
Chris@10
|
1029 }
|
Chris@10
|
1030 {
|
Chris@10
|
1031 E T59, T6n, T5c, T6m, T57, T5b;
|
Chris@10
|
1032 T57 = T2t + T2s;
|
Chris@10
|
1033 T59 = T57 - T58;
|
Chris@10
|
1034 T6n = T57 + T58;
|
Chris@10
|
1035 T5b = T2o + T2p;
|
Chris@10
|
1036 T5c = T5a + T5b;
|
Chris@10
|
1037 T6m = T5b - T5a;
|
Chris@10
|
1038 T5d = FNMS(KP844327925, T5c, KP535826794 * T59);
|
Chris@10
|
1039 T6C = FMA(KP637423989, T6m, KP770513242 * T6n);
|
Chris@10
|
1040 T5C = FMA(KP535826794, T5c, KP844327925 * T59);
|
Chris@10
|
1041 T6o = FNMS(KP637423989, T6n, KP770513242 * T6m);
|
Chris@10
|
1042 }
|
Chris@10
|
1043 {
|
Chris@10
|
1044 E T52, T6j, T55, T6k, T50, T54;
|
Chris@10
|
1045 T50 = T27 + T26;
|
Chris@10
|
1046 T52 = T50 - T51;
|
Chris@10
|
1047 T6j = T50 + T51;
|
Chris@10
|
1048 T54 = T2h + T2g;
|
Chris@10
|
1049 T55 = T53 + T54;
|
Chris@10
|
1050 T6k = T54 - T53;
|
Chris@10
|
1051 T56 = FNMS(KP248689887, T55, KP968583161 * T52);
|
Chris@10
|
1052 T6B = FMA(KP535826794, T6k, KP844327925 * T6j);
|
Chris@10
|
1053 T5B = FMA(KP968583161, T55, KP248689887 * T52);
|
Chris@10
|
1054 T6l = FNMS(KP844327925, T6k, KP535826794 * T6j);
|
Chris@10
|
1055 }
|
Chris@10
|
1056 {
|
Chris@10
|
1057 E T2r, T45, T2y, T46, T2q, T2u;
|
Chris@10
|
1058 T2q = T2o - T2p;
|
Chris@10
|
1059 T2r = T2n + T2q;
|
Chris@10
|
1060 T45 = T2q - T2n;
|
Chris@10
|
1061 T2u = T2s - T2t;
|
Chris@10
|
1062 T2y = T2u - T2x;
|
Chris@10
|
1063 T46 = T2u + T2x;
|
Chris@10
|
1064 T2z = FMA(KP904827052, T2r, KP425779291 * T2y);
|
Chris@10
|
1065 T4m = FNMS(KP992114701, T45, KP125333233 * T46);
|
Chris@10
|
1066 T3i = FNMS(KP425779291, T2r, KP904827052 * T2y);
|
Chris@10
|
1067 T47 = FMA(KP125333233, T45, KP992114701 * T46);
|
Chris@10
|
1068 }
|
Chris@10
|
1069 }
|
Chris@10
|
1070 {
|
Chris@10
|
1071 E T1C, T1F, T1I, T1J, T3b, T3a, T3z, T3A;
|
Chris@10
|
1072 T1C = ci[WS(rs, 24)];
|
Chris@10
|
1073 {
|
Chris@10
|
1074 E T1D, T1E, T1G, T1H;
|
Chris@10
|
1075 T1D = ci[WS(rs, 19)];
|
Chris@10
|
1076 T1E = cr[WS(rs, 20)];
|
Chris@10
|
1077 T1F = T1D - T1E;
|
Chris@10
|
1078 T1G = ci[WS(rs, 14)];
|
Chris@10
|
1079 T1H = cr[WS(rs, 15)];
|
Chris@10
|
1080 T1I = T1G - T1H;
|
Chris@10
|
1081 T1J = T1F + T1I;
|
Chris@10
|
1082 T3b = T1G + T1H;
|
Chris@10
|
1083 T3a = T1D + T1E;
|
Chris@10
|
1084 }
|
Chris@10
|
1085 T1K = T1C + T1J;
|
Chris@10
|
1086 T5w = FMA(KP951056516, T3a, KP587785252 * T3b);
|
Chris@10
|
1087 T3c = FNMS(KP951056516, T3b, KP587785252 * T3a);
|
Chris@10
|
1088 T3z = FNMS(KP250000000, T1J, T1C);
|
Chris@10
|
1089 T3A = KP559016994 * (T1F - T1I);
|
Chris@10
|
1090 T3B = T3z - T3A;
|
Chris@10
|
1091 T5R = T3A + T3z;
|
Chris@10
|
1092 }
|
Chris@10
|
1093 {
|
Chris@10
|
1094 E Tt, T2C, T5i, T2K, TA, T2B, T1a, T2G, T5g, T2M, T1h, T2L, TC, T2R, T5p;
|
Chris@10
|
1095 E T2Z, TJ, T2Q, T1j, T2V, T5n, T31, T1q, T30;
|
Chris@10
|
1096 {
|
Chris@10
|
1097 E Tw, T2I, Tz, T2J;
|
Chris@10
|
1098 Tt = cr[WS(rs, 2)];
|
Chris@10
|
1099 {
|
Chris@10
|
1100 E Tu, Tv, Tx, Ty;
|
Chris@10
|
1101 Tu = cr[WS(rs, 7)];
|
Chris@10
|
1102 Tv = ci[WS(rs, 2)];
|
Chris@10
|
1103 Tw = Tu + Tv;
|
Chris@10
|
1104 T2I = Tu - Tv;
|
Chris@10
|
1105 Tx = cr[WS(rs, 12)];
|
Chris@10
|
1106 Ty = ci[WS(rs, 7)];
|
Chris@10
|
1107 Tz = Tx + Ty;
|
Chris@10
|
1108 T2J = Tx - Ty;
|
Chris@10
|
1109 }
|
Chris@10
|
1110 T2C = KP559016994 * (Tw - Tz);
|
Chris@10
|
1111 T5i = FMA(KP951056516, T2I, KP587785252 * T2J);
|
Chris@10
|
1112 T2K = FNMS(KP951056516, T2J, KP587785252 * T2I);
|
Chris@10
|
1113 TA = Tw + Tz;
|
Chris@10
|
1114 T2B = FNMS(KP250000000, TA, Tt);
|
Chris@10
|
1115 }
|
Chris@10
|
1116 {
|
Chris@10
|
1117 E T1d, T2E, T1g, T2F;
|
Chris@10
|
1118 T1a = ci[WS(rs, 22)];
|
Chris@10
|
1119 {
|
Chris@10
|
1120 E T1b, T1c, T1e, T1f;
|
Chris@10
|
1121 T1b = ci[WS(rs, 17)];
|
Chris@10
|
1122 T1c = cr[WS(rs, 22)];
|
Chris@10
|
1123 T1d = T1b - T1c;
|
Chris@10
|
1124 T2E = T1b + T1c;
|
Chris@10
|
1125 T1e = ci[WS(rs, 12)];
|
Chris@10
|
1126 T1f = cr[WS(rs, 17)];
|
Chris@10
|
1127 T1g = T1e - T1f;
|
Chris@10
|
1128 T2F = T1e + T1f;
|
Chris@10
|
1129 }
|
Chris@10
|
1130 T2G = FNMS(KP951056516, T2F, KP587785252 * T2E);
|
Chris@10
|
1131 T5g = FMA(KP951056516, T2E, KP587785252 * T2F);
|
Chris@10
|
1132 T2M = KP559016994 * (T1d - T1g);
|
Chris@10
|
1133 T1h = T1d + T1g;
|
Chris@10
|
1134 T2L = FNMS(KP250000000, T1h, T1a);
|
Chris@10
|
1135 }
|
Chris@10
|
1136 {
|
Chris@10
|
1137 E TI, T2Y, TF, T2X;
|
Chris@10
|
1138 TC = cr[WS(rs, 3)];
|
Chris@10
|
1139 {
|
Chris@10
|
1140 E TG, TH, TD, TE;
|
Chris@10
|
1141 TG = ci[WS(rs, 11)];
|
Chris@10
|
1142 TH = ci[WS(rs, 6)];
|
Chris@10
|
1143 TI = TG + TH;
|
Chris@10
|
1144 T2Y = TG - TH;
|
Chris@10
|
1145 TD = cr[WS(rs, 8)];
|
Chris@10
|
1146 TE = ci[WS(rs, 1)];
|
Chris@10
|
1147 TF = TD + TE;
|
Chris@10
|
1148 T2X = TD - TE;
|
Chris@10
|
1149 }
|
Chris@10
|
1150 T2R = KP559016994 * (TF - TI);
|
Chris@10
|
1151 T5p = FMA(KP951056516, T2X, KP587785252 * T2Y);
|
Chris@10
|
1152 T2Z = FNMS(KP951056516, T2Y, KP587785252 * T2X);
|
Chris@10
|
1153 TJ = TF + TI;
|
Chris@10
|
1154 T2Q = FNMS(KP250000000, TJ, TC);
|
Chris@10
|
1155 }
|
Chris@10
|
1156 {
|
Chris@10
|
1157 E T1p, T2U, T1m, T2T;
|
Chris@10
|
1158 T1j = ci[WS(rs, 21)];
|
Chris@10
|
1159 {
|
Chris@10
|
1160 E T1n, T1o, T1k, T1l;
|
Chris@10
|
1161 T1n = cr[WS(rs, 13)];
|
Chris@10
|
1162 T1o = cr[WS(rs, 18)];
|
Chris@10
|
1163 T1p = T1n + T1o;
|
Chris@10
|
1164 T2U = T1o - T1n;
|
Chris@10
|
1165 T1k = ci[WS(rs, 16)];
|
Chris@10
|
1166 T1l = cr[WS(rs, 23)];
|
Chris@10
|
1167 T1m = T1k - T1l;
|
Chris@10
|
1168 T2T = T1k + T1l;
|
Chris@10
|
1169 }
|
Chris@10
|
1170 T2V = FNMS(KP951056516, T2U, KP587785252 * T2T);
|
Chris@10
|
1171 T5n = FMA(KP951056516, T2T, KP587785252 * T2U);
|
Chris@10
|
1172 T31 = KP559016994 * (T1m + T1p);
|
Chris@10
|
1173 T1q = T1m - T1p;
|
Chris@10
|
1174 T30 = FNMS(KP250000000, T1q, T1j);
|
Chris@10
|
1175 }
|
Chris@10
|
1176 TB = Tt + TA;
|
Chris@10
|
1177 TK = TC + TJ;
|
Chris@10
|
1178 TL = TB + TK;
|
Chris@10
|
1179 T1i = T1a + T1h;
|
Chris@10
|
1180 T1r = T1j + T1q;
|
Chris@10
|
1181 T1A = T1i + T1r;
|
Chris@10
|
1182 {
|
Chris@10
|
1183 E T2H, T49, T2O, T4a, T2D, T2N;
|
Chris@10
|
1184 T2D = T2B - T2C;
|
Chris@10
|
1185 T2H = T2D - T2G;
|
Chris@10
|
1186 T49 = T2D + T2G;
|
Chris@10
|
1187 T2N = T2L - T2M;
|
Chris@10
|
1188 T2O = T2K + T2N;
|
Chris@10
|
1189 T4a = T2N - T2K;
|
Chris@10
|
1190 T2P = FNMS(KP844327925, T2O, KP535826794 * T2H);
|
Chris@10
|
1191 T4o = FMA(KP062790519, T4a, KP998026728 * T49);
|
Chris@10
|
1192 T3k = FMA(KP535826794, T2O, KP844327925 * T2H);
|
Chris@10
|
1193 T4b = FNMS(KP998026728, T4a, KP062790519 * T49);
|
Chris@10
|
1194 }
|
Chris@10
|
1195 {
|
Chris@10
|
1196 E T5o, T6u, T5r, T6t, T5m, T5q;
|
Chris@10
|
1197 T5m = T2R + T2Q;
|
Chris@10
|
1198 T5o = T5m - T5n;
|
Chris@10
|
1199 T6u = T5m + T5n;
|
Chris@10
|
1200 T5q = T30 + T31;
|
Chris@10
|
1201 T5r = T5p + T5q;
|
Chris@10
|
1202 T6t = T5q - T5p;
|
Chris@10
|
1203 T5s = FNMS(KP684547105, T5r, KP728968627 * T5o);
|
Chris@10
|
1204 T6F = FNMS(KP992114701, T6t, KP125333233 * T6u);
|
Chris@10
|
1205 T5F = FMA(KP728968627, T5r, KP684547105 * T5o);
|
Chris@10
|
1206 T6v = FMA(KP125333233, T6t, KP992114701 * T6u);
|
Chris@10
|
1207 }
|
Chris@10
|
1208 {
|
Chris@10
|
1209 E T5h, T6r, T5k, T6q, T5f, T5j;
|
Chris@10
|
1210 T5f = T2C + T2B;
|
Chris@10
|
1211 T5h = T5f - T5g;
|
Chris@10
|
1212 T6r = T5f + T5g;
|
Chris@10
|
1213 T5j = T2M + T2L;
|
Chris@10
|
1214 T5k = T5i + T5j;
|
Chris@10
|
1215 T6q = T5j - T5i;
|
Chris@10
|
1216 T5l = FNMS(KP481753674, T5k, KP876306680 * T5h);
|
Chris@10
|
1217 T6E = FNMS(KP425779291, T6q, KP904827052 * T6r);
|
Chris@10
|
1218 T5E = FMA(KP876306680, T5k, KP481753674 * T5h);
|
Chris@10
|
1219 T6s = FMA(KP904827052, T6q, KP425779291 * T6r);
|
Chris@10
|
1220 }
|
Chris@10
|
1221 {
|
Chris@10
|
1222 E T2W, T4d, T33, T4c, T2S, T32;
|
Chris@10
|
1223 T2S = T2Q - T2R;
|
Chris@10
|
1224 T2W = T2S - T2V;
|
Chris@10
|
1225 T4d = T2S + T2V;
|
Chris@10
|
1226 T32 = T30 - T31;
|
Chris@10
|
1227 T33 = T2Z + T32;
|
Chris@10
|
1228 T4c = T32 - T2Z;
|
Chris@10
|
1229 T34 = FNMS(KP998026728, T33, KP062790519 * T2W);
|
Chris@10
|
1230 T4p = FNMS(KP637423989, T4c, KP770513242 * T4d);
|
Chris@10
|
1231 T3l = FMA(KP062790519, T33, KP998026728 * T2W);
|
Chris@10
|
1232 T4e = FMA(KP770513242, T4c, KP637423989 * T4d);
|
Chris@10
|
1233 }
|
Chris@10
|
1234 }
|
Chris@10
|
1235 {
|
Chris@10
|
1236 E TM, TQ, T1U, T1L, T1N, T1Z, T1t, T1V, T1y, T1Y;
|
Chris@10
|
1237 {
|
Chris@10
|
1238 E TO, TP, T1B, T1M;
|
Chris@10
|
1239 TO = KP559016994 * (Ts - TL);
|
Chris@10
|
1240 TM = Ts + TL;
|
Chris@10
|
1241 TP = FNMS(KP250000000, TM, T9);
|
Chris@10
|
1242 TQ = TO + TP;
|
Chris@10
|
1243 T1U = TP - TO;
|
Chris@10
|
1244 T1B = KP559016994 * (T1z - T1A);
|
Chris@10
|
1245 T1L = T1z + T1A;
|
Chris@10
|
1246 T1M = FNMS(KP250000000, T1L, T1K);
|
Chris@10
|
1247 T1N = T1B + T1M;
|
Chris@10
|
1248 T1Z = T1M - T1B;
|
Chris@10
|
1249 }
|
Chris@10
|
1250 {
|
Chris@10
|
1251 E T19, T1s, T1w, T1x;
|
Chris@10
|
1252 T19 = TZ - T18;
|
Chris@10
|
1253 T1s = T1i - T1r;
|
Chris@10
|
1254 T1t = FMA(KP951056516, T19, KP587785252 * T1s);
|
Chris@10
|
1255 T1V = FNMS(KP951056516, T1s, KP587785252 * T19);
|
Chris@10
|
1256 T1w = Ti - Tr;
|
Chris@10
|
1257 T1x = TB - TK;
|
Chris@10
|
1258 T1y = FMA(KP951056516, T1w, KP587785252 * T1x);
|
Chris@10
|
1259 T1Y = FNMS(KP951056516, T1x, KP587785252 * T1w);
|
Chris@10
|
1260 }
|
Chris@10
|
1261 cr[0] = T9 + TM;
|
Chris@10
|
1262 ci[0] = T1K + T1L;
|
Chris@10
|
1263 {
|
Chris@10
|
1264 E T1u, T1O, TN, T1v;
|
Chris@10
|
1265 T1u = TQ - T1t;
|
Chris@10
|
1266 T1O = T1y + T1N;
|
Chris@10
|
1267 TN = W[8];
|
Chris@10
|
1268 T1v = W[9];
|
Chris@10
|
1269 cr[WS(rs, 5)] = FNMS(T1v, T1O, TN * T1u);
|
Chris@10
|
1270 ci[WS(rs, 5)] = FMA(T1v, T1u, TN * T1O);
|
Chris@10
|
1271 }
|
Chris@10
|
1272 {
|
Chris@10
|
1273 E T22, T24, T21, T23;
|
Chris@10
|
1274 T22 = T1U + T1V;
|
Chris@10
|
1275 T24 = T1Z - T1Y;
|
Chris@10
|
1276 T21 = W[28];
|
Chris@10
|
1277 T23 = W[29];
|
Chris@10
|
1278 cr[WS(rs, 15)] = FNMS(T23, T24, T21 * T22);
|
Chris@10
|
1279 ci[WS(rs, 15)] = FMA(T23, T22, T21 * T24);
|
Chris@10
|
1280 }
|
Chris@10
|
1281 {
|
Chris@10
|
1282 E T1W, T20, T1T, T1X;
|
Chris@10
|
1283 T1W = T1U - T1V;
|
Chris@10
|
1284 T20 = T1Y + T1Z;
|
Chris@10
|
1285 T1T = W[18];
|
Chris@10
|
1286 T1X = W[19];
|
Chris@10
|
1287 cr[WS(rs, 10)] = FNMS(T1X, T20, T1T * T1W);
|
Chris@10
|
1288 ci[WS(rs, 10)] = FMA(T1X, T1W, T1T * T20);
|
Chris@10
|
1289 }
|
Chris@10
|
1290 {
|
Chris@10
|
1291 E T1Q, T1S, T1P, T1R;
|
Chris@10
|
1292 T1Q = TQ + T1t;
|
Chris@10
|
1293 T1S = T1N - T1y;
|
Chris@10
|
1294 T1P = W[38];
|
Chris@10
|
1295 T1R = W[39];
|
Chris@10
|
1296 cr[WS(rs, 20)] = FNMS(T1R, T1S, T1P * T1Q);
|
Chris@10
|
1297 ci[WS(rs, 20)] = FMA(T1R, T1Q, T1P * T1S);
|
Chris@10
|
1298 }
|
Chris@10
|
1299 }
|
Chris@10
|
1300 {
|
Chris@10
|
1301 E T6H, T71, T6M, T74, T6i, T6x, T6y, T6z, T6Q, T6R, T6P, T6S;
|
Chris@10
|
1302 {
|
Chris@10
|
1303 E T6D, T6G, T6K, T6L;
|
Chris@10
|
1304 T6D = T6B + T6C;
|
Chris@10
|
1305 T6G = T6E - T6F;
|
Chris@10
|
1306 T6H = FMA(KP951056516, T6D, KP587785252 * T6G);
|
Chris@10
|
1307 T71 = FNMS(KP951056516, T6G, KP587785252 * T6D);
|
Chris@10
|
1308 T6K = T6l - T6o;
|
Chris@10
|
1309 T6L = T6v - T6s;
|
Chris@10
|
1310 T6M = FMA(KP951056516, T6K, KP587785252 * T6L);
|
Chris@10
|
1311 T74 = FNMS(KP951056516, T6L, KP587785252 * T6K);
|
Chris@10
|
1312 }
|
Chris@10
|
1313 {
|
Chris@10
|
1314 E T6p, T6w, T6N, T6O;
|
Chris@10
|
1315 T6i = T5v + T5w;
|
Chris@10
|
1316 T6p = T6l + T6o;
|
Chris@10
|
1317 T6w = T6s + T6v;
|
Chris@10
|
1318 T6x = T6p - T6w;
|
Chris@10
|
1319 T6y = FNMS(KP250000000, T6x, T6i);
|
Chris@10
|
1320 T6z = KP559016994 * (T6p + T6w);
|
Chris@10
|
1321 T6Q = T5R - T5Q;
|
Chris@10
|
1322 T6N = T6B - T6C;
|
Chris@10
|
1323 T6O = T6E + T6F;
|
Chris@10
|
1324 T6R = T6N + T6O;
|
Chris@10
|
1325 T6P = KP559016994 * (T6N - T6O);
|
Chris@10
|
1326 T6S = FNMS(KP250000000, T6R, T6Q);
|
Chris@10
|
1327 }
|
Chris@10
|
1328 {
|
Chris@10
|
1329 E T7c, T7e, T7b, T7d;
|
Chris@10
|
1330 T7c = T6i + T6x;
|
Chris@10
|
1331 T7e = T6Q + T6R;
|
Chris@10
|
1332 T7b = W[6];
|
Chris@10
|
1333 T7d = W[7];
|
Chris@10
|
1334 cr[WS(rs, 4)] = FNMS(T7d, T7e, T7b * T7c);
|
Chris@10
|
1335 ci[WS(rs, 4)] = FMA(T7d, T7c, T7b * T7e);
|
Chris@10
|
1336 }
|
Chris@10
|
1337 {
|
Chris@10
|
1338 E T72, T78, T76, T7a, T70, T75;
|
Chris@10
|
1339 T70 = T6y - T6z;
|
Chris@10
|
1340 T72 = T70 - T71;
|
Chris@10
|
1341 T78 = T70 + T71;
|
Chris@10
|
1342 T75 = T6S - T6P;
|
Chris@10
|
1343 T76 = T74 + T75;
|
Chris@10
|
1344 T7a = T75 - T74;
|
Chris@10
|
1345 {
|
Chris@10
|
1346 E T6Z, T73, T77, T79;
|
Chris@10
|
1347 T6Z = W[26];
|
Chris@10
|
1348 T73 = W[27];
|
Chris@10
|
1349 cr[WS(rs, 14)] = FNMS(T73, T76, T6Z * T72);
|
Chris@10
|
1350 ci[WS(rs, 14)] = FMA(T73, T72, T6Z * T76);
|
Chris@10
|
1351 T77 = W[36];
|
Chris@10
|
1352 T79 = W[37];
|
Chris@10
|
1353 cr[WS(rs, 19)] = FNMS(T79, T7a, T77 * T78);
|
Chris@10
|
1354 ci[WS(rs, 19)] = FMA(T79, T78, T77 * T7a);
|
Chris@10
|
1355 }
|
Chris@10
|
1356 }
|
Chris@10
|
1357 {
|
Chris@10
|
1358 E T6I, T6W, T6U, T6Y, T6A, T6T;
|
Chris@10
|
1359 T6A = T6y + T6z;
|
Chris@10
|
1360 T6I = T6A - T6H;
|
Chris@10
|
1361 T6W = T6A + T6H;
|
Chris@10
|
1362 T6T = T6P + T6S;
|
Chris@10
|
1363 T6U = T6M + T6T;
|
Chris@10
|
1364 T6Y = T6T - T6M;
|
Chris@10
|
1365 {
|
Chris@10
|
1366 E T6h, T6J, T6V, T6X;
|
Chris@10
|
1367 T6h = W[16];
|
Chris@10
|
1368 T6J = W[17];
|
Chris@10
|
1369 cr[WS(rs, 9)] = FNMS(T6J, T6U, T6h * T6I);
|
Chris@10
|
1370 ci[WS(rs, 9)] = FMA(T6J, T6I, T6h * T6U);
|
Chris@10
|
1371 T6V = W[46];
|
Chris@10
|
1372 T6X = W[47];
|
Chris@10
|
1373 cr[WS(rs, 24)] = FNMS(T6X, T6Y, T6V * T6W);
|
Chris@10
|
1374 ci[WS(rs, 24)] = FMA(T6X, T6W, T6V * T6Y);
|
Chris@10
|
1375 }
|
Chris@10
|
1376 }
|
Chris@10
|
1377 }
|
Chris@10
|
1378 {
|
Chris@10
|
1379 E T3n, T3N, T3s, T3Q, T3d, T3e, T36, T3f, T3C, T3D, T3v, T3E;
|
Chris@10
|
1380 {
|
Chris@10
|
1381 E T3j, T3m, T3q, T3r;
|
Chris@10
|
1382 T3j = T3h - T3i;
|
Chris@10
|
1383 T3m = T3k - T3l;
|
Chris@10
|
1384 T3n = FMA(KP951056516, T3j, KP587785252 * T3m);
|
Chris@10
|
1385 T3N = FNMS(KP951056516, T3m, KP587785252 * T3j);
|
Chris@10
|
1386 T3q = T2k + T2z;
|
Chris@10
|
1387 T3r = T2P - T34;
|
Chris@10
|
1388 T3s = FMA(KP951056516, T3q, KP587785252 * T3r);
|
Chris@10
|
1389 T3Q = FNMS(KP951056516, T3r, KP587785252 * T3q);
|
Chris@10
|
1390 }
|
Chris@10
|
1391 {
|
Chris@10
|
1392 E T2A, T35, T3t, T3u;
|
Chris@10
|
1393 T3d = T39 - T3c;
|
Chris@10
|
1394 T2A = T2k - T2z;
|
Chris@10
|
1395 T35 = T2P + T34;
|
Chris@10
|
1396 T3e = T2A + T35;
|
Chris@10
|
1397 T36 = KP559016994 * (T2A - T35);
|
Chris@10
|
1398 T3f = FNMS(KP250000000, T3e, T3d);
|
Chris@10
|
1399 T3C = T3y + T3B;
|
Chris@10
|
1400 T3t = T3h + T3i;
|
Chris@10
|
1401 T3u = T3k + T3l;
|
Chris@10
|
1402 T3D = T3t + T3u;
|
Chris@10
|
1403 T3v = KP559016994 * (T3t - T3u);
|
Chris@10
|
1404 T3E = FNMS(KP250000000, T3D, T3C);
|
Chris@10
|
1405 }
|
Chris@10
|
1406 {
|
Chris@10
|
1407 E T3Y, T40, T3X, T3Z;
|
Chris@10
|
1408 T3Y = T3d + T3e;
|
Chris@10
|
1409 T40 = T3C + T3D;
|
Chris@10
|
1410 T3X = W[2];
|
Chris@10
|
1411 T3Z = W[3];
|
Chris@10
|
1412 cr[WS(rs, 2)] = FNMS(T3Z, T40, T3X * T3Y);
|
Chris@10
|
1413 ci[WS(rs, 2)] = FMA(T3Z, T3Y, T3X * T40);
|
Chris@10
|
1414 }
|
Chris@10
|
1415 {
|
Chris@10
|
1416 E T3O, T3U, T3S, T3W, T3M, T3R;
|
Chris@10
|
1417 T3M = T3f - T36;
|
Chris@10
|
1418 T3O = T3M - T3N;
|
Chris@10
|
1419 T3U = T3M + T3N;
|
Chris@10
|
1420 T3R = T3E - T3v;
|
Chris@10
|
1421 T3S = T3Q + T3R;
|
Chris@10
|
1422 T3W = T3R - T3Q;
|
Chris@10
|
1423 {
|
Chris@10
|
1424 E T3L, T3P, T3T, T3V;
|
Chris@10
|
1425 T3L = W[22];
|
Chris@10
|
1426 T3P = W[23];
|
Chris@10
|
1427 cr[WS(rs, 12)] = FNMS(T3P, T3S, T3L * T3O);
|
Chris@10
|
1428 ci[WS(rs, 12)] = FMA(T3P, T3O, T3L * T3S);
|
Chris@10
|
1429 T3T = W[32];
|
Chris@10
|
1430 T3V = W[33];
|
Chris@10
|
1431 cr[WS(rs, 17)] = FNMS(T3V, T3W, T3T * T3U);
|
Chris@10
|
1432 ci[WS(rs, 17)] = FMA(T3V, T3U, T3T * T3W);
|
Chris@10
|
1433 }
|
Chris@10
|
1434 }
|
Chris@10
|
1435 {
|
Chris@10
|
1436 E T3o, T3I, T3G, T3K, T3g, T3F;
|
Chris@10
|
1437 T3g = T36 + T3f;
|
Chris@10
|
1438 T3o = T3g - T3n;
|
Chris@10
|
1439 T3I = T3g + T3n;
|
Chris@10
|
1440 T3F = T3v + T3E;
|
Chris@10
|
1441 T3G = T3s + T3F;
|
Chris@10
|
1442 T3K = T3F - T3s;
|
Chris@10
|
1443 {
|
Chris@10
|
1444 E T25, T3p, T3H, T3J;
|
Chris@10
|
1445 T25 = W[12];
|
Chris@10
|
1446 T3p = W[13];
|
Chris@10
|
1447 cr[WS(rs, 7)] = FNMS(T3p, T3G, T25 * T3o);
|
Chris@10
|
1448 ci[WS(rs, 7)] = FMA(T3p, T3o, T25 * T3G);
|
Chris@10
|
1449 T3H = W[42];
|
Chris@10
|
1450 T3J = W[43];
|
Chris@10
|
1451 cr[WS(rs, 22)] = FNMS(T3J, T3K, T3H * T3I);
|
Chris@10
|
1452 ci[WS(rs, 22)] = FMA(T3J, T3I, T3H * T3K);
|
Chris@10
|
1453 }
|
Chris@10
|
1454 }
|
Chris@10
|
1455 }
|
Chris@10
|
1456 {
|
Chris@10
|
1457 E T4r, T4L, T4w, T4O, T4h, T4i, T4g, T4j, T4A, T4B, T4z, T4C;
|
Chris@10
|
1458 {
|
Chris@10
|
1459 E T4n, T4q, T4u, T4v;
|
Chris@10
|
1460 T4n = T4l - T4m;
|
Chris@10
|
1461 T4q = T4o - T4p;
|
Chris@10
|
1462 T4r = FMA(KP951056516, T4n, KP587785252 * T4q);
|
Chris@10
|
1463 T4L = FNMS(KP951056516, T4q, KP587785252 * T4n);
|
Chris@10
|
1464 T4u = T44 + T47;
|
Chris@10
|
1465 T4v = T4b + T4e;
|
Chris@10
|
1466 T4w = FMA(KP951056516, T4u, KP587785252 * T4v);
|
Chris@10
|
1467 T4O = FNMS(KP951056516, T4v, KP587785252 * T4u);
|
Chris@10
|
1468 }
|
Chris@10
|
1469 {
|
Chris@10
|
1470 E T48, T4f, T4x, T4y;
|
Chris@10
|
1471 T4h = T39 + T3c;
|
Chris@10
|
1472 T48 = T44 - T47;
|
Chris@10
|
1473 T4f = T4b - T4e;
|
Chris@10
|
1474 T4i = T48 + T4f;
|
Chris@10
|
1475 T4g = KP559016994 * (T48 - T4f);
|
Chris@10
|
1476 T4j = FNMS(KP250000000, T4i, T4h);
|
Chris@10
|
1477 T4A = T3B - T3y;
|
Chris@10
|
1478 T4x = T4l + T4m;
|
Chris@10
|
1479 T4y = T4o + T4p;
|
Chris@10
|
1480 T4B = T4x + T4y;
|
Chris@10
|
1481 T4z = KP559016994 * (T4x - T4y);
|
Chris@10
|
1482 T4C = FNMS(KP250000000, T4B, T4A);
|
Chris@10
|
1483 }
|
Chris@10
|
1484 {
|
Chris@10
|
1485 E T4W, T4Y, T4V, T4X;
|
Chris@10
|
1486 T4W = T4h + T4i;
|
Chris@10
|
1487 T4Y = T4A + T4B;
|
Chris@10
|
1488 T4V = W[4];
|
Chris@10
|
1489 T4X = W[5];
|
Chris@10
|
1490 cr[WS(rs, 3)] = FNMS(T4X, T4Y, T4V * T4W);
|
Chris@10
|
1491 ci[WS(rs, 3)] = FMA(T4X, T4W, T4V * T4Y);
|
Chris@10
|
1492 }
|
Chris@10
|
1493 {
|
Chris@10
|
1494 E T4M, T4S, T4Q, T4U, T4K, T4P;
|
Chris@10
|
1495 T4K = T4j - T4g;
|
Chris@10
|
1496 T4M = T4K - T4L;
|
Chris@10
|
1497 T4S = T4K + T4L;
|
Chris@10
|
1498 T4P = T4C - T4z;
|
Chris@10
|
1499 T4Q = T4O + T4P;
|
Chris@10
|
1500 T4U = T4P - T4O;
|
Chris@10
|
1501 {
|
Chris@10
|
1502 E T4J, T4N, T4R, T4T;
|
Chris@10
|
1503 T4J = W[24];
|
Chris@10
|
1504 T4N = W[25];
|
Chris@10
|
1505 cr[WS(rs, 13)] = FNMS(T4N, T4Q, T4J * T4M);
|
Chris@10
|
1506 ci[WS(rs, 13)] = FMA(T4N, T4M, T4J * T4Q);
|
Chris@10
|
1507 T4R = W[34];
|
Chris@10
|
1508 T4T = W[35];
|
Chris@10
|
1509 cr[WS(rs, 18)] = FNMS(T4T, T4U, T4R * T4S);
|
Chris@10
|
1510 ci[WS(rs, 18)] = FMA(T4T, T4S, T4R * T4U);
|
Chris@10
|
1511 }
|
Chris@10
|
1512 }
|
Chris@10
|
1513 {
|
Chris@10
|
1514 E T4s, T4G, T4E, T4I, T4k, T4D;
|
Chris@10
|
1515 T4k = T4g + T4j;
|
Chris@10
|
1516 T4s = T4k - T4r;
|
Chris@10
|
1517 T4G = T4k + T4r;
|
Chris@10
|
1518 T4D = T4z + T4C;
|
Chris@10
|
1519 T4E = T4w + T4D;
|
Chris@10
|
1520 T4I = T4D - T4w;
|
Chris@10
|
1521 {
|
Chris@10
|
1522 E T41, T4t, T4F, T4H;
|
Chris@10
|
1523 T41 = W[14];
|
Chris@10
|
1524 T4t = W[15];
|
Chris@10
|
1525 cr[WS(rs, 8)] = FNMS(T4t, T4E, T41 * T4s);
|
Chris@10
|
1526 ci[WS(rs, 8)] = FMA(T4t, T4s, T41 * T4E);
|
Chris@10
|
1527 T4F = W[44];
|
Chris@10
|
1528 T4H = W[45];
|
Chris@10
|
1529 cr[WS(rs, 23)] = FNMS(T4H, T4I, T4F * T4G);
|
Chris@10
|
1530 ci[WS(rs, 23)] = FMA(T4H, T4G, T4F * T4I);
|
Chris@10
|
1531 }
|
Chris@10
|
1532 }
|
Chris@10
|
1533 }
|
Chris@10
|
1534 {
|
Chris@10
|
1535 E T5H, T63, T5M, T66, T5x, T5y, T5u, T5z, T5S, T5T, T5P, T5U;
|
Chris@10
|
1536 {
|
Chris@10
|
1537 E T5D, T5G, T5K, T5L;
|
Chris@10
|
1538 T5D = T5B - T5C;
|
Chris@10
|
1539 T5G = T5E - T5F;
|
Chris@10
|
1540 T5H = FMA(KP951056516, T5D, KP587785252 * T5G);
|
Chris@10
|
1541 T63 = FNMS(KP951056516, T5G, KP587785252 * T5D);
|
Chris@10
|
1542 T5K = T56 - T5d;
|
Chris@10
|
1543 T5L = T5l - T5s;
|
Chris@10
|
1544 T5M = FMA(KP951056516, T5K, KP587785252 * T5L);
|
Chris@10
|
1545 T66 = FNMS(KP951056516, T5L, KP587785252 * T5K);
|
Chris@10
|
1546 }
|
Chris@10
|
1547 {
|
Chris@10
|
1548 E T5e, T5t, T5N, T5O;
|
Chris@10
|
1549 T5x = T5v - T5w;
|
Chris@10
|
1550 T5e = T56 + T5d;
|
Chris@10
|
1551 T5t = T5l + T5s;
|
Chris@10
|
1552 T5y = T5e + T5t;
|
Chris@10
|
1553 T5u = KP559016994 * (T5e - T5t);
|
Chris@10
|
1554 T5z = FNMS(KP250000000, T5y, T5x);
|
Chris@10
|
1555 T5S = T5Q + T5R;
|
Chris@10
|
1556 T5N = T5B + T5C;
|
Chris@10
|
1557 T5O = T5E + T5F;
|
Chris@10
|
1558 T5T = T5N + T5O;
|
Chris@10
|
1559 T5P = KP559016994 * (T5N - T5O);
|
Chris@10
|
1560 T5U = FNMS(KP250000000, T5T, T5S);
|
Chris@10
|
1561 }
|
Chris@10
|
1562 {
|
Chris@10
|
1563 E T6e, T6g, T6d, T6f;
|
Chris@10
|
1564 T6e = T5x + T5y;
|
Chris@10
|
1565 T6g = T5S + T5T;
|
Chris@10
|
1566 T6d = W[0];
|
Chris@10
|
1567 T6f = W[1];
|
Chris@10
|
1568 cr[WS(rs, 1)] = FNMS(T6f, T6g, T6d * T6e);
|
Chris@10
|
1569 ci[WS(rs, 1)] = FMA(T6f, T6e, T6d * T6g);
|
Chris@10
|
1570 }
|
Chris@10
|
1571 {
|
Chris@10
|
1572 E T64, T6a, T68, T6c, T62, T67;
|
Chris@10
|
1573 T62 = T5z - T5u;
|
Chris@10
|
1574 T64 = T62 - T63;
|
Chris@10
|
1575 T6a = T62 + T63;
|
Chris@10
|
1576 T67 = T5U - T5P;
|
Chris@10
|
1577 T68 = T66 + T67;
|
Chris@10
|
1578 T6c = T67 - T66;
|
Chris@10
|
1579 {
|
Chris@10
|
1580 E T61, T65, T69, T6b;
|
Chris@10
|
1581 T61 = W[20];
|
Chris@10
|
1582 T65 = W[21];
|
Chris@10
|
1583 cr[WS(rs, 11)] = FNMS(T65, T68, T61 * T64);
|
Chris@10
|
1584 ci[WS(rs, 11)] = FMA(T65, T64, T61 * T68);
|
Chris@10
|
1585 T69 = W[30];
|
Chris@10
|
1586 T6b = W[31];
|
Chris@10
|
1587 cr[WS(rs, 16)] = FNMS(T6b, T6c, T69 * T6a);
|
Chris@10
|
1588 ci[WS(rs, 16)] = FMA(T6b, T6a, T69 * T6c);
|
Chris@10
|
1589 }
|
Chris@10
|
1590 }
|
Chris@10
|
1591 {
|
Chris@10
|
1592 E T5I, T5Y, T5W, T60, T5A, T5V;
|
Chris@10
|
1593 T5A = T5u + T5z;
|
Chris@10
|
1594 T5I = T5A - T5H;
|
Chris@10
|
1595 T5Y = T5A + T5H;
|
Chris@10
|
1596 T5V = T5P + T5U;
|
Chris@10
|
1597 T5W = T5M + T5V;
|
Chris@10
|
1598 T60 = T5V - T5M;
|
Chris@10
|
1599 {
|
Chris@10
|
1600 E T4Z, T5J, T5X, T5Z;
|
Chris@10
|
1601 T4Z = W[10];
|
Chris@10
|
1602 T5J = W[11];
|
Chris@10
|
1603 cr[WS(rs, 6)] = FNMS(T5J, T5W, T4Z * T5I);
|
Chris@10
|
1604 ci[WS(rs, 6)] = FMA(T5J, T5I, T4Z * T5W);
|
Chris@10
|
1605 T5X = W[40];
|
Chris@10
|
1606 T5Z = W[41];
|
Chris@10
|
1607 cr[WS(rs, 21)] = FNMS(T5Z, T60, T5X * T5Y);
|
Chris@10
|
1608 ci[WS(rs, 21)] = FMA(T5Z, T5Y, T5X * T60);
|
Chris@10
|
1609 }
|
Chris@10
|
1610 }
|
Chris@10
|
1611 }
|
Chris@10
|
1612 }
|
Chris@10
|
1613 }
|
Chris@10
|
1614 }
|
Chris@10
|
1615
|
Chris@10
|
1616 static const tw_instr twinstr[] = {
|
Chris@10
|
1617 {TW_FULL, 1, 25},
|
Chris@10
|
1618 {TW_NEXT, 1, 0}
|
Chris@10
|
1619 };
|
Chris@10
|
1620
|
Chris@10
|
1621 static const hc2hc_desc desc = { 25, "hb_25", twinstr, &GENUS, {260, 140, 140, 0} };
|
Chris@10
|
1622
|
Chris@10
|
1623 void X(codelet_hb_25) (planner *p) {
|
Chris@10
|
1624 X(khc2hc_register) (p, hb_25, &desc);
|
Chris@10
|
1625 }
|
Chris@10
|
1626 #endif /* HAVE_FMA */
|