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 07:40:09 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 -twiddle-log3 -precompute-twiddles -n 25 -dit -name hf2_25 -include hf.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 * 215 stack variables, 47 constants, and 100 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hf2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
Chris@10
|
40 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
Chris@10
|
41 DK(KP621716863, +0.621716863012209892444754556304102309693593202);
|
Chris@10
|
42 DK(KP614372930, +0.614372930789563808870829930444362096004872855);
|
Chris@10
|
43 DK(KP557913902, +0.557913902031834264187699648465567037992437152);
|
Chris@10
|
44 DK(KP249506682, +0.249506682107067890488084201715862638334226305);
|
Chris@10
|
45 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
Chris@10
|
46 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
Chris@10
|
47 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
Chris@10
|
48 DK(KP968479752, +0.968479752739016373193524836781420152702090879);
|
Chris@10
|
49 DK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
Chris@10
|
50 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
51 DK(KP994076283, +0.994076283785401014123185814696322018529298887);
|
Chris@10
|
52 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
Chris@10
|
53 DK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
Chris@10
|
54 DK(KP062914667, +0.062914667253649757225485955897349402364686947);
|
Chris@10
|
55 DK(KP921177326, +0.921177326965143320250447435415066029359282231);
|
Chris@10
|
56 DK(KP833417178, +0.833417178328688677408962550243238843138996060);
|
Chris@10
|
57 DK(KP541454447, +0.541454447536312777046285590082819509052033189);
|
Chris@10
|
58 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
Chris@10
|
59 DK(KP943557151, +0.943557151597354104399655195398983005179443399);
|
Chris@10
|
60 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
Chris@10
|
61 DK(KP554608978, +0.554608978404018097464974850792216217022558774);
|
Chris@10
|
62 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
63 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
64 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
65 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
Chris@10
|
66 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
67 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
68 DK(KP525970792, +0.525970792408939708442463226536226366643874659);
|
Chris@10
|
69 DK(KP726211448, +0.726211448929902658173535992263577167607493062);
|
Chris@10
|
70 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
Chris@10
|
71 DK(KP871714437, +0.871714437527667770979999223229522602943903653);
|
Chris@10
|
72 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
73 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
Chris@10
|
74 DK(KP851038619, +0.851038619207379630836264138867114231259902550);
|
Chris@10
|
75 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
Chris@10
|
76 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
77 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
Chris@10
|
78 DK(KP912018591, +0.912018591466481957908415381764119056233607330);
|
Chris@10
|
79 DK(KP470564281, +0.470564281212251493087595091036643380879947982);
|
Chris@10
|
80 DK(KP827271945, +0.827271945972475634034355757144307982555673741);
|
Chris@10
|
81 DK(KP126329378, +0.126329378446108174786050455341811215027378105);
|
Chris@10
|
82 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
83 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
84 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
85 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
86 {
|
Chris@10
|
87 INT 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 T7M, T6S, T6Q, T7S, T7Q, T7L, T6R, T6J, T7N, T7R;
|
Chris@10
|
90 {
|
Chris@10
|
91 E T2, T8, T3, T6, Tk, Tv, TS, T4, Ta, TD, T2L, T10, Tm, T5, Tc;
|
Chris@10
|
92 T2 = W[0];
|
Chris@10
|
93 T8 = W[4];
|
Chris@10
|
94 T3 = W[2];
|
Chris@10
|
95 T6 = W[3];
|
Chris@10
|
96 Tk = W[6];
|
Chris@10
|
97 Tv = T2 * T8;
|
Chris@10
|
98 TS = T3 * T8;
|
Chris@10
|
99 T4 = T2 * T3;
|
Chris@10
|
100 Ta = T2 * T6;
|
Chris@10
|
101 TD = T8 * Tk;
|
Chris@10
|
102 T2L = T2 * Tk;
|
Chris@10
|
103 T10 = T3 * Tk;
|
Chris@10
|
104 Tm = W[7];
|
Chris@10
|
105 T5 = W[1];
|
Chris@10
|
106 Tc = W[5];
|
Chris@10
|
107 {
|
Chris@10
|
108 E T7u, T7U, T4s, T6a, T4g, TN, T4f, T7q, T8j, T7p, T4G, T6k, T3a, T4z, T6n;
|
Chris@10
|
109 E T6m, T4w, T4a, T4D, T6j, T6C, T54, T6z, T5b, T1v, T3t, T6y, T58, T6B, T51;
|
Chris@10
|
110 E T6v, T5j, T6s, T5q, T21, T3H, T6r, T5n, T6u, T5g, T26, T3K, T4N, T2A, T3U;
|
Chris@10
|
111 E T4U, T2c, T3M, T2k, T3O;
|
Chris@10
|
112 {
|
Chris@10
|
113 E T11, T1b, Tb, T19, T7, T2m, TT, T15, T2Q, TX, T2p, T1g, T2a, T2e, T2i;
|
Chris@10
|
114 E T27, T1c, T1O, T1K, T1q, T1m, T2x, T2t, T1W, T1S, T2G, T3Y, T2N, T4F, T38;
|
Chris@10
|
115 E T48, T4y, T2K, T40, T2S, T41;
|
Chris@10
|
116 {
|
Chris@10
|
117 E T2M, T1j, T1l, T2X, T2U, T35, T31, T7l, T7n, T7m, T2O, T2R;
|
Chris@10
|
118 {
|
Chris@10
|
119 E T1, Tj, T4j, TK, T4q, TC, T4o, Tt, T4l;
|
Chris@10
|
120 {
|
Chris@10
|
121 E TE, Tw, TI, TA, Th, Tr, Tn, Td, Te, Ti, T14, T2P, TH, Tx, TB;
|
Chris@10
|
122 T1 = cr[0];
|
Chris@10
|
123 T11 = FMA(T6, Tm, T10);
|
Chris@10
|
124 T14 = T3 * Tm;
|
Chris@10
|
125 T2P = T2 * Tm;
|
Chris@10
|
126 TH = T8 * Tm;
|
Chris@10
|
127 T2M = FMA(T5, Tm, T2L);
|
Chris@10
|
128 T1b = FNMS(T5, T3, Ta);
|
Chris@10
|
129 Tb = FMA(T5, T3, Ta);
|
Chris@10
|
130 T19 = FMA(T5, T6, T4);
|
Chris@10
|
131 T7 = FNMS(T5, T6, T4);
|
Chris@10
|
132 T2m = FNMS(T6, Tc, TS);
|
Chris@10
|
133 TT = FMA(T6, Tc, TS);
|
Chris@10
|
134 TE = FMA(Tc, Tm, TD);
|
Chris@10
|
135 T1j = FMA(T5, Tc, Tv);
|
Chris@10
|
136 Tw = FNMS(T5, Tc, Tv);
|
Chris@10
|
137 {
|
Chris@10
|
138 E TW, Tz, T1f, T2d;
|
Chris@10
|
139 TW = T3 * Tc;
|
Chris@10
|
140 Tz = T2 * Tc;
|
Chris@10
|
141 T15 = FNMS(T6, Tk, T14);
|
Chris@10
|
142 T2Q = FNMS(T5, Tk, T2P);
|
Chris@10
|
143 TI = FNMS(Tc, Tk, TH);
|
Chris@10
|
144 T1f = T19 * Tc;
|
Chris@10
|
145 T2d = T19 * Tk;
|
Chris@10
|
146 {
|
Chris@10
|
147 E T2h, T1a, Tg, Tq;
|
Chris@10
|
148 T2h = T19 * Tm;
|
Chris@10
|
149 T1a = T19 * T8;
|
Chris@10
|
150 Tg = T7 * Tc;
|
Chris@10
|
151 Tq = T7 * Tm;
|
Chris@10
|
152 {
|
Chris@10
|
153 E Tl, T9, T1p, T1k;
|
Chris@10
|
154 Tl = T7 * Tk;
|
Chris@10
|
155 T9 = T7 * T8;
|
Chris@10
|
156 T1p = T1j * Tm;
|
Chris@10
|
157 T1k = T1j * Tk;
|
Chris@10
|
158 {
|
Chris@10
|
159 E T34, T30, T1N, T1J;
|
Chris@10
|
160 T34 = TT * Tm;
|
Chris@10
|
161 T30 = TT * Tk;
|
Chris@10
|
162 T1N = Tw * Tm;
|
Chris@10
|
163 T1J = Tw * Tk;
|
Chris@10
|
164 TX = FNMS(T6, T8, TW);
|
Chris@10
|
165 T2p = FMA(T6, T8, TW);
|
Chris@10
|
166 TA = FMA(T5, T8, Tz);
|
Chris@10
|
167 T1l = FNMS(T5, T8, Tz);
|
Chris@10
|
168 T1g = FMA(T1b, T8, T1f);
|
Chris@10
|
169 T2a = FNMS(T1b, T8, T1f);
|
Chris@10
|
170 T2e = FMA(T1b, Tm, T2d);
|
Chris@10
|
171 T2i = FNMS(T1b, Tk, T2h);
|
Chris@10
|
172 T27 = FMA(T1b, Tc, T1a);
|
Chris@10
|
173 T1c = FNMS(T1b, Tc, T1a);
|
Chris@10
|
174 T2X = FMA(Tb, T8, Tg);
|
Chris@10
|
175 Th = FNMS(Tb, T8, Tg);
|
Chris@10
|
176 Tr = FNMS(Tb, Tk, Tq);
|
Chris@10
|
177 Tn = FMA(Tb, Tm, Tl);
|
Chris@10
|
178 Td = FMA(Tb, Tc, T9);
|
Chris@10
|
179 T2U = FNMS(Tb, Tc, T9);
|
Chris@10
|
180 T35 = FNMS(TX, Tk, T34);
|
Chris@10
|
181 T31 = FMA(TX, Tm, T30);
|
Chris@10
|
182 T1O = FNMS(TA, Tk, T1N);
|
Chris@10
|
183 T1K = FMA(TA, Tm, T1J);
|
Chris@10
|
184 T1q = FNMS(T1l, Tk, T1p);
|
Chris@10
|
185 T1m = FMA(T1l, Tm, T1k);
|
Chris@10
|
186 {
|
Chris@10
|
187 E T2w, T2s, T1V, T1R;
|
Chris@10
|
188 T2w = T27 * Tm;
|
Chris@10
|
189 T2s = T27 * Tk;
|
Chris@10
|
190 T1V = Td * Tm;
|
Chris@10
|
191 T1R = Td * Tk;
|
Chris@10
|
192 T2x = FNMS(T2a, Tk, T2w);
|
Chris@10
|
193 T2t = FMA(T2a, Tm, T2s);
|
Chris@10
|
194 T1W = FNMS(Th, Tk, T1V);
|
Chris@10
|
195 T1S = FMA(Th, Tm, T1R);
|
Chris@10
|
196 T7l = ci[0];
|
Chris@10
|
197 Te = cr[WS(rs, 5)];
|
Chris@10
|
198 Ti = ci[WS(rs, 5)];
|
Chris@10
|
199 }
|
Chris@10
|
200 }
|
Chris@10
|
201 }
|
Chris@10
|
202 }
|
Chris@10
|
203 }
|
Chris@10
|
204 {
|
Chris@10
|
205 E TF, TJ, Tf, T4i, TG, T4p;
|
Chris@10
|
206 TF = cr[WS(rs, 15)];
|
Chris@10
|
207 TJ = ci[WS(rs, 15)];
|
Chris@10
|
208 Tf = Td * Te;
|
Chris@10
|
209 T4i = Td * Ti;
|
Chris@10
|
210 TG = TE * TF;
|
Chris@10
|
211 T4p = TE * TJ;
|
Chris@10
|
212 Tj = FMA(Th, Ti, Tf);
|
Chris@10
|
213 T4j = FNMS(Th, Te, T4i);
|
Chris@10
|
214 TK = FMA(TI, TJ, TG);
|
Chris@10
|
215 T4q = FNMS(TI, TF, T4p);
|
Chris@10
|
216 }
|
Chris@10
|
217 Tx = cr[WS(rs, 10)];
|
Chris@10
|
218 TB = ci[WS(rs, 10)];
|
Chris@10
|
219 {
|
Chris@10
|
220 E To, Ts, Ty, T4n, Tp, T4k;
|
Chris@10
|
221 To = cr[WS(rs, 20)];
|
Chris@10
|
222 Ts = ci[WS(rs, 20)];
|
Chris@10
|
223 Ty = Tw * Tx;
|
Chris@10
|
224 T4n = Tw * TB;
|
Chris@10
|
225 Tp = Tn * To;
|
Chris@10
|
226 T4k = Tn * Ts;
|
Chris@10
|
227 TC = FMA(TA, TB, Ty);
|
Chris@10
|
228 T4o = FNMS(TA, Tx, T4n);
|
Chris@10
|
229 Tt = FMA(Tr, Ts, Tp);
|
Chris@10
|
230 T4l = FNMS(Tr, To, T4k);
|
Chris@10
|
231 }
|
Chris@10
|
232 }
|
Chris@10
|
233 {
|
Chris@10
|
234 E TL, T7s, T4r, Tu, T7t, T4m, TM;
|
Chris@10
|
235 TL = TC + TK;
|
Chris@10
|
236 T7s = TC - TK;
|
Chris@10
|
237 T4r = T4o - T4q;
|
Chris@10
|
238 T7n = T4o + T4q;
|
Chris@10
|
239 Tu = Tj + Tt;
|
Chris@10
|
240 T7t = Tj - Tt;
|
Chris@10
|
241 T4m = T4j - T4l;
|
Chris@10
|
242 T7m = T4j + T4l;
|
Chris@10
|
243 T7u = FNMS(KP618033988, T7t, T7s);
|
Chris@10
|
244 T7U = FMA(KP618033988, T7s, T7t);
|
Chris@10
|
245 T4s = FMA(KP618033988, T4r, T4m);
|
Chris@10
|
246 T6a = FNMS(KP618033988, T4m, T4r);
|
Chris@10
|
247 T4g = Tu - TL;
|
Chris@10
|
248 TM = Tu + TL;
|
Chris@10
|
249 TN = T1 + TM;
|
Chris@10
|
250 T4f = FNMS(KP250000000, TM, T1);
|
Chris@10
|
251 }
|
Chris@10
|
252 }
|
Chris@10
|
253 {
|
Chris@10
|
254 E T2D, T2F, T7o, T2E, T3X;
|
Chris@10
|
255 T2D = cr[WS(rs, 3)];
|
Chris@10
|
256 T2F = ci[WS(rs, 3)];
|
Chris@10
|
257 T7q = T7m - T7n;
|
Chris@10
|
258 T7o = T7m + T7n;
|
Chris@10
|
259 T2E = T3 * T2D;
|
Chris@10
|
260 T3X = T3 * T2F;
|
Chris@10
|
261 {
|
Chris@10
|
262 E T2V, T2W, T2Y, T32, T36;
|
Chris@10
|
263 T2V = cr[WS(rs, 13)];
|
Chris@10
|
264 T8j = T7o + T7l;
|
Chris@10
|
265 T7p = FNMS(KP250000000, T7o, T7l);
|
Chris@10
|
266 T2G = FMA(T6, T2F, T2E);
|
Chris@10
|
267 T3Y = FNMS(T6, T2D, T3X);
|
Chris@10
|
268 T2W = T2U * T2V;
|
Chris@10
|
269 T2Y = ci[WS(rs, 13)];
|
Chris@10
|
270 T32 = cr[WS(rs, 18)];
|
Chris@10
|
271 T36 = ci[WS(rs, 18)];
|
Chris@10
|
272 {
|
Chris@10
|
273 E T2H, T2I, T2J, T3Z;
|
Chris@10
|
274 {
|
Chris@10
|
275 E T2Z, T45, T37, T47, T44, T33, T46;
|
Chris@10
|
276 T2H = cr[WS(rs, 8)];
|
Chris@10
|
277 T2Z = FMA(T2X, T2Y, T2W);
|
Chris@10
|
278 T44 = T2U * T2Y;
|
Chris@10
|
279 T33 = T31 * T32;
|
Chris@10
|
280 T46 = T31 * T36;
|
Chris@10
|
281 T2I = T1j * T2H;
|
Chris@10
|
282 T45 = FNMS(T2X, T2V, T44);
|
Chris@10
|
283 T37 = FMA(T35, T36, T33);
|
Chris@10
|
284 T47 = FNMS(T35, T32, T46);
|
Chris@10
|
285 T2J = ci[WS(rs, 8)];
|
Chris@10
|
286 T2N = cr[WS(rs, 23)];
|
Chris@10
|
287 T4F = T2Z - T37;
|
Chris@10
|
288 T38 = T2Z + T37;
|
Chris@10
|
289 T48 = T45 + T47;
|
Chris@10
|
290 T4y = T47 - T45;
|
Chris@10
|
291 T3Z = T1j * T2J;
|
Chris@10
|
292 T2O = T2M * T2N;
|
Chris@10
|
293 T2R = ci[WS(rs, 23)];
|
Chris@10
|
294 }
|
Chris@10
|
295 T2K = FMA(T1l, T2J, T2I);
|
Chris@10
|
296 T40 = FNMS(T1l, T2H, T3Z);
|
Chris@10
|
297 }
|
Chris@10
|
298 }
|
Chris@10
|
299 }
|
Chris@10
|
300 T2S = FMA(T2Q, T2R, T2O);
|
Chris@10
|
301 T41 = T2M * T2R;
|
Chris@10
|
302 }
|
Chris@10
|
303 {
|
Chris@10
|
304 E TR, T3h, T1t, T53, T3r, T5a, TZ, T3j, T17, T3l;
|
Chris@10
|
305 {
|
Chris@10
|
306 E T12, T16, T13, T3k;
|
Chris@10
|
307 {
|
Chris@10
|
308 E TO, TP, T4C, T4B, TQ;
|
Chris@10
|
309 {
|
Chris@10
|
310 E T2T, T4E, T42, T4v, T39;
|
Chris@10
|
311 TO = cr[WS(rs, 1)];
|
Chris@10
|
312 T2T = T2K + T2S;
|
Chris@10
|
313 T4E = T2K - T2S;
|
Chris@10
|
314 T42 = FNMS(T2Q, T2N, T41);
|
Chris@10
|
315 TP = T2 * TO;
|
Chris@10
|
316 T4G = FMA(KP618033988, T4F, T4E);
|
Chris@10
|
317 T6k = FNMS(KP618033988, T4E, T4F);
|
Chris@10
|
318 T4v = T38 - T2T;
|
Chris@10
|
319 T39 = T2T + T38;
|
Chris@10
|
320 {
|
Chris@10
|
321 E T43, T4x, T4u, T49;
|
Chris@10
|
322 T43 = T40 + T42;
|
Chris@10
|
323 T4x = T42 - T40;
|
Chris@10
|
324 T4u = FNMS(KP250000000, T39, T2G);
|
Chris@10
|
325 T3a = T2G + T39;
|
Chris@10
|
326 T4z = FMA(KP618033988, T4y, T4x);
|
Chris@10
|
327 T6n = FNMS(KP618033988, T4x, T4y);
|
Chris@10
|
328 T4C = T48 - T43;
|
Chris@10
|
329 T49 = T43 + T48;
|
Chris@10
|
330 T6m = FMA(KP559016994, T4v, T4u);
|
Chris@10
|
331 T4w = FNMS(KP559016994, T4v, T4u);
|
Chris@10
|
332 T4B = FNMS(KP250000000, T49, T3Y);
|
Chris@10
|
333 T4a = T3Y + T49;
|
Chris@10
|
334 TQ = ci[WS(rs, 1)];
|
Chris@10
|
335 }
|
Chris@10
|
336 }
|
Chris@10
|
337 {
|
Chris@10
|
338 E T1n, T1r, T1i, T1o, T3o, T3p;
|
Chris@10
|
339 {
|
Chris@10
|
340 E T1d, T1h, T1e, T3n, T3g;
|
Chris@10
|
341 T1d = cr[WS(rs, 11)];
|
Chris@10
|
342 T1h = ci[WS(rs, 11)];
|
Chris@10
|
343 T4D = FNMS(KP559016994, T4C, T4B);
|
Chris@10
|
344 T6j = FMA(KP559016994, T4C, T4B);
|
Chris@10
|
345 TR = FMA(T5, TQ, TP);
|
Chris@10
|
346 T3g = T2 * TQ;
|
Chris@10
|
347 T1e = T1c * T1d;
|
Chris@10
|
348 T3n = T1c * T1h;
|
Chris@10
|
349 T1n = cr[WS(rs, 16)];
|
Chris@10
|
350 T3h = FNMS(T5, TO, T3g);
|
Chris@10
|
351 T1r = ci[WS(rs, 16)];
|
Chris@10
|
352 T1i = FMA(T1g, T1h, T1e);
|
Chris@10
|
353 T1o = T1m * T1n;
|
Chris@10
|
354 T3o = FNMS(T1g, T1d, T3n);
|
Chris@10
|
355 T3p = T1m * T1r;
|
Chris@10
|
356 }
|
Chris@10
|
357 {
|
Chris@10
|
358 E TU, TY, TV, T3i, T3q, T1s;
|
Chris@10
|
359 TU = cr[WS(rs, 6)];
|
Chris@10
|
360 T1s = FMA(T1q, T1r, T1o);
|
Chris@10
|
361 TY = ci[WS(rs, 6)];
|
Chris@10
|
362 T3q = FNMS(T1q, T1n, T3p);
|
Chris@10
|
363 TV = TT * TU;
|
Chris@10
|
364 T1t = T1i + T1s;
|
Chris@10
|
365 T53 = T1s - T1i;
|
Chris@10
|
366 T3i = TT * TY;
|
Chris@10
|
367 T3r = T3o + T3q;
|
Chris@10
|
368 T5a = T3q - T3o;
|
Chris@10
|
369 T12 = cr[WS(rs, 21)];
|
Chris@10
|
370 T16 = ci[WS(rs, 21)];
|
Chris@10
|
371 TZ = FMA(TX, TY, TV);
|
Chris@10
|
372 T3j = FNMS(TX, TU, T3i);
|
Chris@10
|
373 T13 = T11 * T12;
|
Chris@10
|
374 T3k = T11 * T16;
|
Chris@10
|
375 }
|
Chris@10
|
376 }
|
Chris@10
|
377 }
|
Chris@10
|
378 T17 = FMA(T15, T16, T13);
|
Chris@10
|
379 T3l = FNMS(T15, T12, T3k);
|
Chris@10
|
380 }
|
Chris@10
|
381 {
|
Chris@10
|
382 E T1z, T3v, T5i, T1Z, T3F, T5p, T1D, T3x, T1H, T3z;
|
Chris@10
|
383 {
|
Chris@10
|
384 E T1E, T1G, T1F, T3y;
|
Chris@10
|
385 {
|
Chris@10
|
386 E T1w, T1y, T1x, T57, T50, T56, T4Z, T3u, T18, T52;
|
Chris@10
|
387 T1w = cr[WS(rs, 4)];
|
Chris@10
|
388 T1y = ci[WS(rs, 4)];
|
Chris@10
|
389 T18 = TZ + T17;
|
Chris@10
|
390 T52 = T17 - TZ;
|
Chris@10
|
391 {
|
Chris@10
|
392 E T3m, T59, T1u, T3s;
|
Chris@10
|
393 T3m = T3j + T3l;
|
Chris@10
|
394 T59 = T3j - T3l;
|
Chris@10
|
395 T1x = T7 * T1w;
|
Chris@10
|
396 T6C = FNMS(KP618033988, T52, T53);
|
Chris@10
|
397 T54 = FMA(KP618033988, T53, T52);
|
Chris@10
|
398 T1u = T18 + T1t;
|
Chris@10
|
399 T57 = T18 - T1t;
|
Chris@10
|
400 T6z = FMA(KP618033988, T59, T5a);
|
Chris@10
|
401 T5b = FNMS(KP618033988, T5a, T59);
|
Chris@10
|
402 T3s = T3m + T3r;
|
Chris@10
|
403 T50 = T3m - T3r;
|
Chris@10
|
404 T1v = TR + T1u;
|
Chris@10
|
405 T56 = FNMS(KP250000000, T1u, TR);
|
Chris@10
|
406 T3t = T3h + T3s;
|
Chris@10
|
407 T4Z = FNMS(KP250000000, T3s, T3h);
|
Chris@10
|
408 T3u = T7 * T1y;
|
Chris@10
|
409 }
|
Chris@10
|
410 T6y = FNMS(KP559016994, T57, T56);
|
Chris@10
|
411 T58 = FMA(KP559016994, T57, T56);
|
Chris@10
|
412 T6B = FNMS(KP559016994, T50, T4Z);
|
Chris@10
|
413 T51 = FMA(KP559016994, T50, T4Z);
|
Chris@10
|
414 T1z = FMA(Tb, T1y, T1x);
|
Chris@10
|
415 T3v = FNMS(Tb, T1w, T3u);
|
Chris@10
|
416 }
|
Chris@10
|
417 {
|
Chris@10
|
418 E T1Q, T3C, T1Y, T3E;
|
Chris@10
|
419 {
|
Chris@10
|
420 E T1L, T1P, T1T, T1X, T1M, T3B, T1U, T3D;
|
Chris@10
|
421 T1L = cr[WS(rs, 14)];
|
Chris@10
|
422 T1P = ci[WS(rs, 14)];
|
Chris@10
|
423 T1T = cr[WS(rs, 19)];
|
Chris@10
|
424 T1X = ci[WS(rs, 19)];
|
Chris@10
|
425 T1M = T1K * T1L;
|
Chris@10
|
426 T3B = T1K * T1P;
|
Chris@10
|
427 T1U = T1S * T1T;
|
Chris@10
|
428 T3D = T1S * T1X;
|
Chris@10
|
429 T1Q = FMA(T1O, T1P, T1M);
|
Chris@10
|
430 T3C = FNMS(T1O, T1L, T3B);
|
Chris@10
|
431 T1Y = FMA(T1W, T1X, T1U);
|
Chris@10
|
432 T3E = FNMS(T1W, T1T, T3D);
|
Chris@10
|
433 }
|
Chris@10
|
434 {
|
Chris@10
|
435 E T1A, T1C, T1B, T3w;
|
Chris@10
|
436 T1A = cr[WS(rs, 9)];
|
Chris@10
|
437 T1C = ci[WS(rs, 9)];
|
Chris@10
|
438 T5i = T1Y - T1Q;
|
Chris@10
|
439 T1Z = T1Q + T1Y;
|
Chris@10
|
440 T3F = T3C + T3E;
|
Chris@10
|
441 T5p = T3E - T3C;
|
Chris@10
|
442 T1B = T8 * T1A;
|
Chris@10
|
443 T3w = T8 * T1C;
|
Chris@10
|
444 T1E = cr[WS(rs, 24)];
|
Chris@10
|
445 T1G = ci[WS(rs, 24)];
|
Chris@10
|
446 T1D = FMA(Tc, T1C, T1B);
|
Chris@10
|
447 T3x = FNMS(Tc, T1A, T3w);
|
Chris@10
|
448 T1F = Tk * T1E;
|
Chris@10
|
449 T3y = Tk * T1G;
|
Chris@10
|
450 }
|
Chris@10
|
451 }
|
Chris@10
|
452 T1H = FMA(Tm, T1G, T1F);
|
Chris@10
|
453 T3z = FNMS(Tm, T1E, T3y);
|
Chris@10
|
454 }
|
Chris@10
|
455 {
|
Chris@10
|
456 E T2f, T2j, T2g, T3N;
|
Chris@10
|
457 {
|
Chris@10
|
458 E T23, T25, T24, T5m, T5f, T5l, T5e, T3J, T1I, T5h;
|
Chris@10
|
459 T23 = cr[WS(rs, 2)];
|
Chris@10
|
460 T25 = ci[WS(rs, 2)];
|
Chris@10
|
461 T1I = T1D + T1H;
|
Chris@10
|
462 T5h = T1H - T1D;
|
Chris@10
|
463 {
|
Chris@10
|
464 E T3A, T5o, T20, T3G;
|
Chris@10
|
465 T3A = T3x + T3z;
|
Chris@10
|
466 T5o = T3z - T3x;
|
Chris@10
|
467 T24 = T19 * T23;
|
Chris@10
|
468 T6v = FNMS(KP618033988, T5h, T5i);
|
Chris@10
|
469 T5j = FMA(KP618033988, T5i, T5h);
|
Chris@10
|
470 T20 = T1I + T1Z;
|
Chris@10
|
471 T5m = T1I - T1Z;
|
Chris@10
|
472 T6s = FNMS(KP618033988, T5o, T5p);
|
Chris@10
|
473 T5q = FMA(KP618033988, T5p, T5o);
|
Chris@10
|
474 T3G = T3A + T3F;
|
Chris@10
|
475 T5f = T3F - T3A;
|
Chris@10
|
476 T21 = T1z + T20;
|
Chris@10
|
477 T5l = FNMS(KP250000000, T20, T1z);
|
Chris@10
|
478 T3H = T3v + T3G;
|
Chris@10
|
479 T5e = FNMS(KP250000000, T3G, T3v);
|
Chris@10
|
480 T3J = T19 * T25;
|
Chris@10
|
481 }
|
Chris@10
|
482 T6r = FNMS(KP559016994, T5m, T5l);
|
Chris@10
|
483 T5n = FMA(KP559016994, T5m, T5l);
|
Chris@10
|
484 T6u = FMA(KP559016994, T5f, T5e);
|
Chris@10
|
485 T5g = FNMS(KP559016994, T5f, T5e);
|
Chris@10
|
486 T26 = FMA(T1b, T25, T24);
|
Chris@10
|
487 T3K = FNMS(T1b, T23, T3J);
|
Chris@10
|
488 }
|
Chris@10
|
489 {
|
Chris@10
|
490 E T2r, T3R, T2z, T3T;
|
Chris@10
|
491 {
|
Chris@10
|
492 E T2n, T2q, T2u, T2y, T2o, T3Q, T2v, T3S;
|
Chris@10
|
493 T2n = cr[WS(rs, 12)];
|
Chris@10
|
494 T2q = ci[WS(rs, 12)];
|
Chris@10
|
495 T2u = cr[WS(rs, 17)];
|
Chris@10
|
496 T2y = ci[WS(rs, 17)];
|
Chris@10
|
497 T2o = T2m * T2n;
|
Chris@10
|
498 T3Q = T2m * T2q;
|
Chris@10
|
499 T2v = T2t * T2u;
|
Chris@10
|
500 T3S = T2t * T2y;
|
Chris@10
|
501 T2r = FMA(T2p, T2q, T2o);
|
Chris@10
|
502 T3R = FNMS(T2p, T2n, T3Q);
|
Chris@10
|
503 T2z = FMA(T2x, T2y, T2v);
|
Chris@10
|
504 T3T = FNMS(T2x, T2u, T3S);
|
Chris@10
|
505 }
|
Chris@10
|
506 {
|
Chris@10
|
507 E T28, T2b, T29, T3L;
|
Chris@10
|
508 T28 = cr[WS(rs, 7)];
|
Chris@10
|
509 T2b = ci[WS(rs, 7)];
|
Chris@10
|
510 T4N = T2z - T2r;
|
Chris@10
|
511 T2A = T2r + T2z;
|
Chris@10
|
512 T3U = T3R + T3T;
|
Chris@10
|
513 T4U = T3R - T3T;
|
Chris@10
|
514 T29 = T27 * T28;
|
Chris@10
|
515 T3L = T27 * T2b;
|
Chris@10
|
516 T2f = cr[WS(rs, 22)];
|
Chris@10
|
517 T2j = ci[WS(rs, 22)];
|
Chris@10
|
518 T2c = FMA(T2a, T2b, T29);
|
Chris@10
|
519 T3M = FNMS(T2a, T28, T3L);
|
Chris@10
|
520 T2g = T2e * T2f;
|
Chris@10
|
521 T3N = T2e * T2j;
|
Chris@10
|
522 }
|
Chris@10
|
523 }
|
Chris@10
|
524 T2k = FMA(T2i, T2j, T2g);
|
Chris@10
|
525 T3O = FNMS(T2i, T2f, T3N);
|
Chris@10
|
526 }
|
Chris@10
|
527 }
|
Chris@10
|
528 }
|
Chris@10
|
529 }
|
Chris@10
|
530 {
|
Chris@10
|
531 E T8k, T6d, T6g, T8r, T6f, T8l, T6c, T8q, T69, T7r, T5Y, T8g, T8i, T66, T68;
|
Chris@10
|
532 E T5X, T8d, T8h;
|
Chris@10
|
533 {
|
Chris@10
|
534 E T4O, T4V, T22, T4S, T4L, T3b, T4e, T4c, T3I;
|
Chris@10
|
535 T8k = T3t + T3H;
|
Chris@10
|
536 T3I = T3t - T3H;
|
Chris@10
|
537 {
|
Chris@10
|
538 E T2l, T4M, T3P, T4T;
|
Chris@10
|
539 T2l = T2c + T2k;
|
Chris@10
|
540 T4M = T2k - T2c;
|
Chris@10
|
541 T3P = T3M + T3O;
|
Chris@10
|
542 T4T = T3O - T3M;
|
Chris@10
|
543 T4O = FMA(KP618033988, T4N, T4M);
|
Chris@10
|
544 T6d = FNMS(KP618033988, T4M, T4N);
|
Chris@10
|
545 {
|
Chris@10
|
546 E T4R, T2B, T4K, T3V;
|
Chris@10
|
547 T4R = T2A - T2l;
|
Chris@10
|
548 T2B = T2l + T2A;
|
Chris@10
|
549 T4V = FNMS(KP618033988, T4U, T4T);
|
Chris@10
|
550 T6g = FMA(KP618033988, T4T, T4U);
|
Chris@10
|
551 T4K = T3U - T3P;
|
Chris@10
|
552 T3V = T3P + T3U;
|
Chris@10
|
553 {
|
Chris@10
|
554 E T4Q, T2C, T4J, T3W, T4b;
|
Chris@10
|
555 T4Q = FNMS(KP250000000, T2B, T26);
|
Chris@10
|
556 T2C = T26 + T2B;
|
Chris@10
|
557 T4J = FNMS(KP250000000, T3V, T3K);
|
Chris@10
|
558 T3W = T3K + T3V;
|
Chris@10
|
559 T8r = T21 - T1v;
|
Chris@10
|
560 T22 = T1v + T21;
|
Chris@10
|
561 T4S = FNMS(KP559016994, T4R, T4Q);
|
Chris@10
|
562 T6f = FMA(KP559016994, T4R, T4Q);
|
Chris@10
|
563 T4b = T3W - T4a;
|
Chris@10
|
564 T8l = T3W + T4a;
|
Chris@10
|
565 T6c = FMA(KP559016994, T4K, T4J);
|
Chris@10
|
566 T4L = FNMS(KP559016994, T4K, T4J);
|
Chris@10
|
567 T8q = T2C - T3a;
|
Chris@10
|
568 T3b = T2C + T3a;
|
Chris@10
|
569 T4e = FNMS(KP618033988, T3I, T4b);
|
Chris@10
|
570 T4c = FMA(KP618033988, T4b, T3I);
|
Chris@10
|
571 }
|
Chris@10
|
572 }
|
Chris@10
|
573 }
|
Chris@10
|
574 {
|
Chris@10
|
575 E T5H, T4t, T7V, T87, T5Q, T5P, T5D, T8e, T5A, T8f, T5K, T60, T8c, T8a, T5u;
|
Chris@10
|
576 E T5w, T5U, T64, T5N, T61;
|
Chris@10
|
577 {
|
Chris@10
|
578 E T3e, T3d, T4h, T3c, T7T;
|
Chris@10
|
579 T4h = FMA(KP559016994, T4g, T4f);
|
Chris@10
|
580 T69 = FNMS(KP559016994, T4g, T4f);
|
Chris@10
|
581 T3c = T22 + T3b;
|
Chris@10
|
582 T3e = T22 - T3b;
|
Chris@10
|
583 T7r = FNMS(KP559016994, T7q, T7p);
|
Chris@10
|
584 T7T = FMA(KP559016994, T7q, T7p);
|
Chris@10
|
585 T5H = FMA(KP951056516, T4s, T4h);
|
Chris@10
|
586 T4t = FNMS(KP951056516, T4s, T4h);
|
Chris@10
|
587 cr[0] = TN + T3c;
|
Chris@10
|
588 T3d = FNMS(KP250000000, T3c, TN);
|
Chris@10
|
589 T7V = FNMS(KP951056516, T7U, T7T);
|
Chris@10
|
590 T87 = FMA(KP951056516, T7U, T7T);
|
Chris@10
|
591 {
|
Chris@10
|
592 E T5S, T5T, T5L, T4I, T5B, T5M, T55, T5J, T5s, T5z, T4X, T5C, T5I, T5c;
|
Chris@10
|
593 {
|
Chris@10
|
594 E T5k, T5r, T4P, T4W;
|
Chris@10
|
595 {
|
Chris@10
|
596 E T4A, T4d, T3f, T4H;
|
Chris@10
|
597 T4A = FMA(KP951056516, T4z, T4w);
|
Chris@10
|
598 T5S = FNMS(KP951056516, T4z, T4w);
|
Chris@10
|
599 T4d = FNMS(KP559016994, T3e, T3d);
|
Chris@10
|
600 T3f = FMA(KP559016994, T3e, T3d);
|
Chris@10
|
601 T5T = FNMS(KP951056516, T4G, T4D);
|
Chris@10
|
602 T4H = FMA(KP951056516, T4G, T4D);
|
Chris@10
|
603 T5k = FNMS(KP951056516, T5j, T5g);
|
Chris@10
|
604 T5L = FMA(KP951056516, T5j, T5g);
|
Chris@10
|
605 cr[WS(rs, 5)] = FMA(KP951056516, T4c, T3f);
|
Chris@10
|
606 ci[WS(rs, 4)] = FNMS(KP951056516, T4c, T3f);
|
Chris@10
|
607 ci[WS(rs, 9)] = FMA(KP951056516, T4e, T4d);
|
Chris@10
|
608 cr[WS(rs, 10)] = FNMS(KP951056516, T4e, T4d);
|
Chris@10
|
609 T4I = FNMS(KP126329378, T4H, T4A);
|
Chris@10
|
610 T5B = FMA(KP126329378, T4A, T4H);
|
Chris@10
|
611 T5M = FNMS(KP951056516, T5q, T5n);
|
Chris@10
|
612 T5r = FMA(KP951056516, T5q, T5n);
|
Chris@10
|
613 }
|
Chris@10
|
614 T4P = FNMS(KP951056516, T4O, T4L);
|
Chris@10
|
615 T5Q = FMA(KP951056516, T4O, T4L);
|
Chris@10
|
616 T5P = FNMS(KP951056516, T4V, T4S);
|
Chris@10
|
617 T4W = FMA(KP951056516, T4V, T4S);
|
Chris@10
|
618 T55 = FNMS(KP951056516, T54, T51);
|
Chris@10
|
619 T5J = FMA(KP951056516, T54, T51);
|
Chris@10
|
620 T5s = FMA(KP827271945, T5r, T5k);
|
Chris@10
|
621 T5z = FNMS(KP827271945, T5k, T5r);
|
Chris@10
|
622 T4X = FNMS(KP470564281, T4W, T4P);
|
Chris@10
|
623 T5C = FMA(KP470564281, T4P, T4W);
|
Chris@10
|
624 T5I = FMA(KP951056516, T5b, T58);
|
Chris@10
|
625 T5c = FNMS(KP951056516, T5b, T58);
|
Chris@10
|
626 }
|
Chris@10
|
627 {
|
Chris@10
|
628 E T88, T4Y, T5d, T5y, T89, T5t;
|
Chris@10
|
629 T5D = FNMS(KP912018591, T5C, T5B);
|
Chris@10
|
630 T88 = FMA(KP912018591, T5C, T5B);
|
Chris@10
|
631 T8e = FMA(KP912018591, T4X, T4I);
|
Chris@10
|
632 T4Y = FNMS(KP912018591, T4X, T4I);
|
Chris@10
|
633 T5d = FMA(KP634619297, T5c, T55);
|
Chris@10
|
634 T5y = FNMS(KP634619297, T55, T5c);
|
Chris@10
|
635 T5A = FMA(KP912575812, T5z, T5y);
|
Chris@10
|
636 T89 = FNMS(KP912575812, T5z, T5y);
|
Chris@10
|
637 T8f = FMA(KP912575812, T5s, T5d);
|
Chris@10
|
638 T5t = FNMS(KP912575812, T5s, T5d);
|
Chris@10
|
639 T5K = FMA(KP256756360, T5J, T5I);
|
Chris@10
|
640 T60 = FNMS(KP256756360, T5I, T5J);
|
Chris@10
|
641 T8c = FNMS(KP851038619, T89, T88);
|
Chris@10
|
642 T8a = FMA(KP851038619, T89, T88);
|
Chris@10
|
643 T5u = FNMS(KP851038619, T5t, T4Y);
|
Chris@10
|
644 T5w = FMA(KP851038619, T5t, T4Y);
|
Chris@10
|
645 }
|
Chris@10
|
646 T5U = FMA(KP939062505, T5T, T5S);
|
Chris@10
|
647 T64 = FNMS(KP939062505, T5S, T5T);
|
Chris@10
|
648 T5N = FMA(KP634619297, T5M, T5L);
|
Chris@10
|
649 T61 = FNMS(KP634619297, T5L, T5M);
|
Chris@10
|
650 }
|
Chris@10
|
651 }
|
Chris@10
|
652 {
|
Chris@10
|
653 E T62, T7W, T83, T5O, T5R, T63;
|
Chris@10
|
654 cr[WS(rs, 4)] = FNMS(KP992114701, T5u, T4t);
|
Chris@10
|
655 T62 = FMA(KP871714437, T61, T60);
|
Chris@10
|
656 T7W = FNMS(KP871714437, T61, T60);
|
Chris@10
|
657 T83 = FNMS(KP871714437, T5N, T5K);
|
Chris@10
|
658 T5O = FMA(KP871714437, T5N, T5K);
|
Chris@10
|
659 T5R = FMA(KP549754652, T5Q, T5P);
|
Chris@10
|
660 T63 = FNMS(KP549754652, T5P, T5Q);
|
Chris@10
|
661 ci[WS(rs, 20)] = FNMS(KP992114701, T8a, T87);
|
Chris@10
|
662 {
|
Chris@10
|
663 E T65, T5W, T84, T86, T81, T85, T8b;
|
Chris@10
|
664 {
|
Chris@10
|
665 E T5E, T5G, T82, T80, T7Y, T5v, T7X, T5V, T5F, T5x, T7Z;
|
Chris@10
|
666 T5E = FNMS(KP726211448, T5D, T5A);
|
Chris@10
|
667 T5G = FMA(KP525970792, T5A, T5D);
|
Chris@10
|
668 T65 = FNMS(KP831864738, T64, T63);
|
Chris@10
|
669 T7X = FMA(KP831864738, T64, T63);
|
Chris@10
|
670 T82 = FNMS(KP831864738, T5U, T5R);
|
Chris@10
|
671 T5V = FMA(KP831864738, T5U, T5R);
|
Chris@10
|
672 T80 = FNMS(KP904730450, T7X, T7W);
|
Chris@10
|
673 T7Y = FMA(KP904730450, T7X, T7W);
|
Chris@10
|
674 T5Y = FNMS(KP904730450, T5V, T5O);
|
Chris@10
|
675 T5W = FMA(KP904730450, T5V, T5O);
|
Chris@10
|
676 T5v = FMA(KP248028675, T5u, T4t);
|
Chris@10
|
677 ci[WS(rs, 23)] = FMA(KP968583161, T7Y, T7V);
|
Chris@10
|
678 cr[WS(rs, 1)] = FMA(KP968583161, T5W, T5H);
|
Chris@10
|
679 T84 = FNMS(KP683113946, T83, T82);
|
Chris@10
|
680 T86 = FMA(KP559154169, T82, T83);
|
Chris@10
|
681 T5F = FNMS(KP554608978, T5w, T5v);
|
Chris@10
|
682 T5x = FMA(KP554608978, T5w, T5v);
|
Chris@10
|
683 T7Z = FNMS(KP242145790, T7Y, T7V);
|
Chris@10
|
684 ci[WS(rs, 10)] = FNMS(KP943557151, T5G, T5F);
|
Chris@10
|
685 ci[WS(rs, 5)] = FMA(KP943557151, T5G, T5F);
|
Chris@10
|
686 ci[0] = FMA(KP803003575, T5E, T5x);
|
Chris@10
|
687 cr[WS(rs, 9)] = FNMS(KP803003575, T5E, T5x);
|
Chris@10
|
688 T81 = FNMS(KP541454447, T80, T7Z);
|
Chris@10
|
689 T85 = FMA(KP541454447, T80, T7Z);
|
Chris@10
|
690 }
|
Chris@10
|
691 T8g = FNMS(KP525970792, T8f, T8e);
|
Chris@10
|
692 T8i = FMA(KP726211448, T8e, T8f);
|
Chris@10
|
693 ci[WS(rs, 13)] = FMA(KP833417178, T84, T81);
|
Chris@10
|
694 cr[WS(rs, 16)] = FMS(KP833417178, T84, T81);
|
Chris@10
|
695 cr[WS(rs, 21)] = -(FMA(KP921177326, T86, T85));
|
Chris@10
|
696 ci[WS(rs, 18)] = FNMS(KP921177326, T86, T85);
|
Chris@10
|
697 T8b = FMA(KP248028675, T8a, T87);
|
Chris@10
|
698 T66 = FMA(KP559154169, T65, T62);
|
Chris@10
|
699 T68 = FNMS(KP683113946, T62, T65);
|
Chris@10
|
700 T5X = FNMS(KP242145790, T5W, T5H);
|
Chris@10
|
701 T8d = FNMS(KP554608978, T8c, T8b);
|
Chris@10
|
702 T8h = FMA(KP554608978, T8c, T8b);
|
Chris@10
|
703 }
|
Chris@10
|
704 }
|
Chris@10
|
705 }
|
Chris@10
|
706 }
|
Chris@10
|
707 {
|
Chris@10
|
708 E T8s, T8u, T5Z, T67;
|
Chris@10
|
709 cr[WS(rs, 24)] = -(FMA(KP803003575, T8i, T8h));
|
Chris@10
|
710 ci[WS(rs, 15)] = FNMS(KP803003575, T8i, T8h);
|
Chris@10
|
711 cr[WS(rs, 19)] = FMS(KP943557151, T8g, T8d);
|
Chris@10
|
712 cr[WS(rs, 14)] = -(FMA(KP943557151, T8g, T8d));
|
Chris@10
|
713 T5Z = FMA(KP541454447, T5Y, T5X);
|
Chris@10
|
714 T67 = FNMS(KP541454447, T5Y, T5X);
|
Chris@10
|
715 cr[WS(rs, 11)] = FNMS(KP833417178, T68, T67);
|
Chris@10
|
716 ci[WS(rs, 8)] = FMA(KP833417178, T68, T67);
|
Chris@10
|
717 cr[WS(rs, 6)] = FMA(KP921177326, T66, T5Z);
|
Chris@10
|
718 ci[WS(rs, 3)] = FNMS(KP921177326, T66, T5Z);
|
Chris@10
|
719 T8s = FMA(KP618033988, T8r, T8q);
|
Chris@10
|
720 T8u = FNMS(KP618033988, T8q, T8r);
|
Chris@10
|
721 {
|
Chris@10
|
722 E T6X, T6T, T6b, T7H, T7v, T6Y, T72, T71, T6P, T7O, T6M, T7P, T7K, T6G, T6I;
|
Chris@10
|
723 E T6W, T7f, T7d, T76;
|
Chris@10
|
724 {
|
Chris@10
|
725 E T74, T75, T6i, T6N, T6L, T6E, T6U, T6l, T6o, T6V, T6t, T6w;
|
Chris@10
|
726 {
|
Chris@10
|
727 E T6e, T8o, T8n, T6h, T8m;
|
Chris@10
|
728 T6X = FNMS(KP951056516, T6d, T6c);
|
Chris@10
|
729 T6e = FMA(KP951056516, T6d, T6c);
|
Chris@10
|
730 T8o = T8k - T8l;
|
Chris@10
|
731 T8m = T8k + T8l;
|
Chris@10
|
732 T6T = FNMS(KP951056516, T6a, T69);
|
Chris@10
|
733 T6b = FMA(KP951056516, T6a, T69);
|
Chris@10
|
734 T7H = FNMS(KP951056516, T7u, T7r);
|
Chris@10
|
735 T7v = FMA(KP951056516, T7u, T7r);
|
Chris@10
|
736 ci[WS(rs, 24)] = T8m + T8j;
|
Chris@10
|
737 T8n = FNMS(KP250000000, T8m, T8j);
|
Chris@10
|
738 T6h = FMA(KP951056516, T6g, T6f);
|
Chris@10
|
739 T6Y = FNMS(KP951056516, T6g, T6f);
|
Chris@10
|
740 {
|
Chris@10
|
741 E T6A, T6D, T8t, T8p;
|
Chris@10
|
742 T74 = FMA(KP951056516, T6z, T6y);
|
Chris@10
|
743 T6A = FNMS(KP951056516, T6z, T6y);
|
Chris@10
|
744 T6D = FMA(KP951056516, T6C, T6B);
|
Chris@10
|
745 T75 = FNMS(KP951056516, T6C, T6B);
|
Chris@10
|
746 T8t = FMA(KP559016994, T8o, T8n);
|
Chris@10
|
747 T8p = FNMS(KP559016994, T8o, T8n);
|
Chris@10
|
748 T6i = FMA(KP062914667, T6h, T6e);
|
Chris@10
|
749 T6N = FNMS(KP062914667, T6e, T6h);
|
Chris@10
|
750 ci[WS(rs, 14)] = FMA(KP951056516, T8s, T8p);
|
Chris@10
|
751 cr[WS(rs, 15)] = FMS(KP951056516, T8s, T8p);
|
Chris@10
|
752 ci[WS(rs, 19)] = FMA(KP951056516, T8u, T8t);
|
Chris@10
|
753 cr[WS(rs, 20)] = FMS(KP951056516, T8u, T8t);
|
Chris@10
|
754 T6L = FNMS(KP939062505, T6A, T6D);
|
Chris@10
|
755 T6E = FMA(KP939062505, T6D, T6A);
|
Chris@10
|
756 }
|
Chris@10
|
757 }
|
Chris@10
|
758 T6U = FMA(KP951056516, T6k, T6j);
|
Chris@10
|
759 T6l = FNMS(KP951056516, T6k, T6j);
|
Chris@10
|
760 T6o = FNMS(KP951056516, T6n, T6m);
|
Chris@10
|
761 T6V = FMA(KP951056516, T6n, T6m);
|
Chris@10
|
762 T72 = FMA(KP951056516, T6s, T6r);
|
Chris@10
|
763 T6t = FNMS(KP951056516, T6s, T6r);
|
Chris@10
|
764 T6w = FMA(KP951056516, T6v, T6u);
|
Chris@10
|
765 T71 = FNMS(KP951056516, T6v, T6u);
|
Chris@10
|
766 {
|
Chris@10
|
767 E T6q, T6F, T6O, T6p;
|
Chris@10
|
768 T6O = FMA(KP827271945, T6l, T6o);
|
Chris@10
|
769 T6p = FNMS(KP827271945, T6o, T6l);
|
Chris@10
|
770 {
|
Chris@10
|
771 E T6K, T6x, T7I, T7J;
|
Chris@10
|
772 T6K = FMA(KP126329378, T6t, T6w);
|
Chris@10
|
773 T6x = FNMS(KP126329378, T6w, T6t);
|
Chris@10
|
774 T7I = FMA(KP772036680, T6O, T6N);
|
Chris@10
|
775 T6P = FNMS(KP772036680, T6O, T6N);
|
Chris@10
|
776 T6q = FMA(KP772036680, T6p, T6i);
|
Chris@10
|
777 T7O = FNMS(KP772036680, T6p, T6i);
|
Chris@10
|
778 T7J = FNMS(KP734762448, T6L, T6K);
|
Chris@10
|
779 T6M = FMA(KP734762448, T6L, T6K);
|
Chris@10
|
780 T6F = FNMS(KP734762448, T6E, T6x);
|
Chris@10
|
781 T7P = FMA(KP734762448, T6E, T6x);
|
Chris@10
|
782 T7K = FMA(KP994076283, T7J, T7I);
|
Chris@10
|
783 T7M = FNMS(KP994076283, T7J, T7I);
|
Chris@10
|
784 }
|
Chris@10
|
785 T6G = FNMS(KP994076283, T6F, T6q);
|
Chris@10
|
786 T6I = FMA(KP994076283, T6F, T6q);
|
Chris@10
|
787 }
|
Chris@10
|
788 T6W = FMA(KP062914667, T6V, T6U);
|
Chris@10
|
789 T7f = FNMS(KP062914667, T6U, T6V);
|
Chris@10
|
790 T7d = FNMS(KP549754652, T74, T75);
|
Chris@10
|
791 T76 = FMA(KP549754652, T75, T74);
|
Chris@10
|
792 }
|
Chris@10
|
793 {
|
Chris@10
|
794 E T7h, T7C, T7e, T7D, T7y, T7A, T78, T7a;
|
Chris@10
|
795 {
|
Chris@10
|
796 E T70, T77, T7g, T6Z;
|
Chris@10
|
797 cr[WS(rs, 3)] = FMA(KP998026728, T6G, T6b);
|
Chris@10
|
798 T7g = FNMS(KP634619297, T6X, T6Y);
|
Chris@10
|
799 T6Z = FMA(KP634619297, T6Y, T6X);
|
Chris@10
|
800 {
|
Chris@10
|
801 E T7c, T73, T7w, T7x;
|
Chris@10
|
802 T7c = FMA(KP470564281, T71, T72);
|
Chris@10
|
803 T73 = FNMS(KP470564281, T72, T71);
|
Chris@10
|
804 T7w = FMA(KP845997307, T7g, T7f);
|
Chris@10
|
805 T7h = FNMS(KP845997307, T7g, T7f);
|
Chris@10
|
806 T70 = FMA(KP845997307, T6Z, T6W);
|
Chris@10
|
807 T7C = FNMS(KP845997307, T6Z, T6W);
|
Chris@10
|
808 T7x = FNMS(KP968479752, T7d, T7c);
|
Chris@10
|
809 T7e = FMA(KP968479752, T7d, T7c);
|
Chris@10
|
810 T77 = FMA(KP968479752, T76, T73);
|
Chris@10
|
811 T7D = FNMS(KP968479752, T76, T73);
|
Chris@10
|
812 T7y = FMA(KP906616052, T7x, T7w);
|
Chris@10
|
813 T7A = FNMS(KP906616052, T7x, T7w);
|
Chris@10
|
814 }
|
Chris@10
|
815 ci[WS(rs, 21)] = FNMS(KP998026728, T7K, T7H);
|
Chris@10
|
816 T78 = FMA(KP906616052, T77, T70);
|
Chris@10
|
817 T7a = FNMS(KP906616052, T77, T70);
|
Chris@10
|
818 }
|
Chris@10
|
819 {
|
Chris@10
|
820 E T7G, T7E, T7k, T7i, T79, T7F, T7B, T7z, T6H, T7j, T7b;
|
Chris@10
|
821 T7G = FMA(KP681693190, T7C, T7D);
|
Chris@10
|
822 T7E = FNMS(KP560319534, T7D, T7C);
|
Chris@10
|
823 ci[WS(rs, 22)] = FNMS(KP998026728, T7y, T7v);
|
Chris@10
|
824 cr[WS(rs, 2)] = FMA(KP998026728, T78, T6T);
|
Chris@10
|
825 T7z = FMA(KP249506682, T7y, T7v);
|
Chris@10
|
826 T7k = FNMS(KP560319534, T7e, T7h);
|
Chris@10
|
827 T7i = FMA(KP681693190, T7h, T7e);
|
Chris@10
|
828 T79 = FNMS(KP249506682, T78, T6T);
|
Chris@10
|
829 T7F = FMA(KP557913902, T7A, T7z);
|
Chris@10
|
830 T7B = FNMS(KP557913902, T7A, T7z);
|
Chris@10
|
831 T6S = FMA(KP614372930, T6M, T6P);
|
Chris@10
|
832 T6Q = FNMS(KP621716863, T6P, T6M);
|
Chris@10
|
833 cr[WS(rs, 22)] = FMS(KP860541664, T7G, T7F);
|
Chris@10
|
834 ci[WS(rs, 17)] = FMA(KP860541664, T7G, T7F);
|
Chris@10
|
835 ci[WS(rs, 12)] = FNMS(KP949179823, T7E, T7B);
|
Chris@10
|
836 cr[WS(rs, 17)] = -(FMA(KP949179823, T7E, T7B));
|
Chris@10
|
837 T7j = FMA(KP557913902, T7a, T79);
|
Chris@10
|
838 T7b = FNMS(KP557913902, T7a, T79);
|
Chris@10
|
839 T6H = FNMS(KP249506682, T6G, T6b);
|
Chris@10
|
840 ci[WS(rs, 7)] = FMA(KP949179823, T7k, T7j);
|
Chris@10
|
841 cr[WS(rs, 12)] = FNMS(KP949179823, T7k, T7j);
|
Chris@10
|
842 cr[WS(rs, 7)] = FMA(KP860541664, T7i, T7b);
|
Chris@10
|
843 ci[WS(rs, 2)] = FNMS(KP860541664, T7i, T7b);
|
Chris@10
|
844 T7S = FMA(KP621716863, T7O, T7P);
|
Chris@10
|
845 T7Q = FNMS(KP614372930, T7P, T7O);
|
Chris@10
|
846 T7L = FMA(KP249506682, T7K, T7H);
|
Chris@10
|
847 T6R = FMA(KP557913902, T6I, T6H);
|
Chris@10
|
848 T6J = FNMS(KP557913902, T6I, T6H);
|
Chris@10
|
849 }
|
Chris@10
|
850 }
|
Chris@10
|
851 }
|
Chris@10
|
852 }
|
Chris@10
|
853 }
|
Chris@10
|
854 }
|
Chris@10
|
855 }
|
Chris@10
|
856 ci[WS(rs, 6)] = FNMS(KP949179823, T6S, T6R);
|
Chris@10
|
857 ci[WS(rs, 11)] = FMA(KP949179823, T6S, T6R);
|
Chris@10
|
858 cr[WS(rs, 8)] = FMA(KP943557151, T6Q, T6J);
|
Chris@10
|
859 ci[WS(rs, 1)] = FNMS(KP943557151, T6Q, T6J);
|
Chris@10
|
860 T7N = FNMS(KP557913902, T7M, T7L);
|
Chris@10
|
861 T7R = FMA(KP557913902, T7M, T7L);
|
Chris@10
|
862 cr[WS(rs, 23)] = -(FMA(KP943557151, T7S, T7R));
|
Chris@10
|
863 ci[WS(rs, 16)] = FNMS(KP943557151, T7S, T7R);
|
Chris@10
|
864 cr[WS(rs, 18)] = FMS(KP949179823, T7Q, T7N);
|
Chris@10
|
865 cr[WS(rs, 13)] = -(FMA(KP949179823, T7Q, T7N));
|
Chris@10
|
866 }
|
Chris@10
|
867 }
|
Chris@10
|
868 }
|
Chris@10
|
869
|
Chris@10
|
870 static const tw_instr twinstr[] = {
|
Chris@10
|
871 {TW_CEXP, 1, 1},
|
Chris@10
|
872 {TW_CEXP, 1, 3},
|
Chris@10
|
873 {TW_CEXP, 1, 9},
|
Chris@10
|
874 {TW_CEXP, 1, 24},
|
Chris@10
|
875 {TW_NEXT, 1, 0}
|
Chris@10
|
876 };
|
Chris@10
|
877
|
Chris@10
|
878 static const hc2hc_desc desc = { 25, "hf2_25", twinstr, &GENUS, {84, 78, 356, 0} };
|
Chris@10
|
879
|
Chris@10
|
880 void X(codelet_hf2_25) (planner *p) {
|
Chris@10
|
881 X(khc2hc_register) (p, hf2_25, &desc);
|
Chris@10
|
882 }
|
Chris@10
|
883 #else /* HAVE_FMA */
|
Chris@10
|
884
|
Chris@10
|
885 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 25 -dit -name hf2_25 -include hf.h */
|
Chris@10
|
886
|
Chris@10
|
887 /*
|
Chris@10
|
888 * This function contains 440 FP additions, 340 FP multiplications,
|
Chris@10
|
889 * (or, 280 additions, 180 multiplications, 160 fused multiply/add),
|
Chris@10
|
890 * 149 stack variables, 20 constants, and 100 memory accesses
|
Chris@10
|
891 */
|
Chris@10
|
892 #include "hf.h"
|
Chris@10
|
893
|
Chris@10
|
894 static void hf2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
895 {
|
Chris@10
|
896 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
897 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
898 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
899 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
900 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
901 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
902 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
903 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
904 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
905 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
906 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
907 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
908 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
909 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
910 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
911 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
912 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
913 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
914 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
915 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
916 {
|
Chris@10
|
917 INT m;
|
Chris@10
|
918 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
|
919 E T2, T5, T3, T6, T8, Td, T16, T14, Te, T9, T21, T23, Tx, TR, T1g;
|
Chris@10
|
920 E TB, T1f, TV, T1Q, Tg, T1S, Tk, T18, T2s, T1c, T2q, Tn, To, Tp, Tr;
|
Chris@10
|
921 E T28, T2x, TY, T2k, T2m, T2v, TG, TE, T10, T1h, T1E, T26, T1B, T1G, T1V;
|
Chris@10
|
922 E T1X, T1z, T1j;
|
Chris@10
|
923 {
|
Chris@10
|
924 E Tw, TT, Tz, TQ, Tv, TU, TA, TP;
|
Chris@10
|
925 {
|
Chris@10
|
926 E T4, Tc, T7, Tb;
|
Chris@10
|
927 T2 = W[0];
|
Chris@10
|
928 T5 = W[1];
|
Chris@10
|
929 T3 = W[2];
|
Chris@10
|
930 T6 = W[3];
|
Chris@10
|
931 T4 = T2 * T3;
|
Chris@10
|
932 Tc = T5 * T3;
|
Chris@10
|
933 T7 = T5 * T6;
|
Chris@10
|
934 Tb = T2 * T6;
|
Chris@10
|
935 T8 = T4 - T7;
|
Chris@10
|
936 Td = Tb + Tc;
|
Chris@10
|
937 T16 = Tb - Tc;
|
Chris@10
|
938 T14 = T4 + T7;
|
Chris@10
|
939 Te = W[5];
|
Chris@10
|
940 Tw = T5 * Te;
|
Chris@10
|
941 TT = T3 * Te;
|
Chris@10
|
942 Tz = T2 * Te;
|
Chris@10
|
943 TQ = T6 * Te;
|
Chris@10
|
944 T9 = W[4];
|
Chris@10
|
945 Tv = T2 * T9;
|
Chris@10
|
946 TU = T6 * T9;
|
Chris@10
|
947 TA = T5 * T9;
|
Chris@10
|
948 TP = T3 * T9;
|
Chris@10
|
949 }
|
Chris@10
|
950 T21 = TP - TQ;
|
Chris@10
|
951 T23 = TT + TU;
|
Chris@10
|
952 {
|
Chris@10
|
953 E T15, T17, Ta, Tf, T1a, T1b, Ti, Tj;
|
Chris@10
|
954 Tx = Tv - Tw;
|
Chris@10
|
955 TR = TP + TQ;
|
Chris@10
|
956 T1g = Tz - TA;
|
Chris@10
|
957 TB = Tz + TA;
|
Chris@10
|
958 T1f = Tv + Tw;
|
Chris@10
|
959 TV = TT - TU;
|
Chris@10
|
960 T15 = T14 * T9;
|
Chris@10
|
961 T17 = T16 * Te;
|
Chris@10
|
962 T1Q = T15 + T17;
|
Chris@10
|
963 Ta = T8 * T9;
|
Chris@10
|
964 Tf = Td * Te;
|
Chris@10
|
965 Tg = Ta + Tf;
|
Chris@10
|
966 T1a = T14 * Te;
|
Chris@10
|
967 T1b = T16 * T9;
|
Chris@10
|
968 T1S = T1a - T1b;
|
Chris@10
|
969 Ti = T8 * Te;
|
Chris@10
|
970 Tj = Td * T9;
|
Chris@10
|
971 Tk = Ti - Tj;
|
Chris@10
|
972 T18 = T15 - T17;
|
Chris@10
|
973 T2s = Ti + Tj;
|
Chris@10
|
974 T1c = T1a + T1b;
|
Chris@10
|
975 T2q = Ta - Tf;
|
Chris@10
|
976 Tn = W[6];
|
Chris@10
|
977 To = W[7];
|
Chris@10
|
978 Tp = FMA(T8, Tn, Td * To);
|
Chris@10
|
979 Tr = FNMS(Td, Tn, T8 * To);
|
Chris@10
|
980 T28 = FNMS(T1S, Tn, T1Q * To);
|
Chris@10
|
981 T2x = FNMS(TV, Tn, TR * To);
|
Chris@10
|
982 TY = FMA(T3, Tn, T6 * To);
|
Chris@10
|
983 T2k = FMA(T2, Tn, T5 * To);
|
Chris@10
|
984 T2m = FNMS(T5, Tn, T2 * To);
|
Chris@10
|
985 T2v = FMA(TR, Tn, TV * To);
|
Chris@10
|
986 TG = FNMS(Te, Tn, T9 * To);
|
Chris@10
|
987 TE = FMA(T9, Tn, Te * To);
|
Chris@10
|
988 T10 = FNMS(T6, Tn, T3 * To);
|
Chris@10
|
989 T1h = FMA(T1f, Tn, T1g * To);
|
Chris@10
|
990 T1E = FMA(Tg, Tn, Tk * To);
|
Chris@10
|
991 T26 = FMA(T1Q, Tn, T1S * To);
|
Chris@10
|
992 T1B = FNMS(TB, Tn, Tx * To);
|
Chris@10
|
993 T1G = FNMS(Tk, Tn, Tg * To);
|
Chris@10
|
994 T1V = FMA(T14, Tn, T16 * To);
|
Chris@10
|
995 T1X = FNMS(T16, Tn, T14 * To);
|
Chris@10
|
996 T1z = FMA(Tx, Tn, TB * To);
|
Chris@10
|
997 T1j = FNMS(T1g, Tn, T1f * To);
|
Chris@10
|
998 }
|
Chris@10
|
999 }
|
Chris@10
|
1000 {
|
Chris@10
|
1001 E T1, T6v, T2F, T6A, TK, T2G, T6y, T6z, T6u, T71, T2O, T52, T2C, T6k, T4c;
|
Chris@10
|
1002 E T5X, T4L, T5s, T4j, T5W, T4K, T5v, T1o, T6g, T30, T5M, T4A, T56, T3b, T5N;
|
Chris@10
|
1003 E T4B, T59, T1L, T6h, T3r, T5P, T4E, T5d, T3y, T5Q, T4D, T5g, T2d, T6j, T3P;
|
Chris@10
|
1004 E T5U, T4I, T5o, T3W, T5T, T4H, T5l;
|
Chris@10
|
1005 {
|
Chris@10
|
1006 E Tm, T2I, Tt, T2J, Tu, T6w, TD, T2L, TI, T2M, TJ, T6x;
|
Chris@10
|
1007 T1 = cr[0];
|
Chris@10
|
1008 T6v = ci[0];
|
Chris@10
|
1009 {
|
Chris@10
|
1010 E Th, Tl, Tq, Ts;
|
Chris@10
|
1011 Th = cr[WS(rs, 5)];
|
Chris@10
|
1012 Tl = ci[WS(rs, 5)];
|
Chris@10
|
1013 Tm = FMA(Tg, Th, Tk * Tl);
|
Chris@10
|
1014 T2I = FNMS(Tk, Th, Tg * Tl);
|
Chris@10
|
1015 Tq = cr[WS(rs, 20)];
|
Chris@10
|
1016 Ts = ci[WS(rs, 20)];
|
Chris@10
|
1017 Tt = FMA(Tp, Tq, Tr * Ts);
|
Chris@10
|
1018 T2J = FNMS(Tr, Tq, Tp * Ts);
|
Chris@10
|
1019 }
|
Chris@10
|
1020 Tu = Tm + Tt;
|
Chris@10
|
1021 T6w = T2I + T2J;
|
Chris@10
|
1022 {
|
Chris@10
|
1023 E Ty, TC, TF, TH;
|
Chris@10
|
1024 Ty = cr[WS(rs, 10)];
|
Chris@10
|
1025 TC = ci[WS(rs, 10)];
|
Chris@10
|
1026 TD = FMA(Tx, Ty, TB * TC);
|
Chris@10
|
1027 T2L = FNMS(TB, Ty, Tx * TC);
|
Chris@10
|
1028 TF = cr[WS(rs, 15)];
|
Chris@10
|
1029 TH = ci[WS(rs, 15)];
|
Chris@10
|
1030 TI = FMA(TE, TF, TG * TH);
|
Chris@10
|
1031 T2M = FNMS(TG, TF, TE * TH);
|
Chris@10
|
1032 }
|
Chris@10
|
1033 TJ = TD + TI;
|
Chris@10
|
1034 T6x = T2L + T2M;
|
Chris@10
|
1035 T2F = KP559016994 * (Tu - TJ);
|
Chris@10
|
1036 T6A = KP559016994 * (T6w - T6x);
|
Chris@10
|
1037 TK = Tu + TJ;
|
Chris@10
|
1038 T2G = FNMS(KP250000000, TK, T1);
|
Chris@10
|
1039 T6y = T6w + T6x;
|
Chris@10
|
1040 T6z = FNMS(KP250000000, T6y, T6v);
|
Chris@10
|
1041 {
|
Chris@10
|
1042 E T6s, T6t, T2K, T2N;
|
Chris@10
|
1043 T6s = TD - TI;
|
Chris@10
|
1044 T6t = Tm - Tt;
|
Chris@10
|
1045 T6u = FNMS(KP587785252, T6t, KP951056516 * T6s);
|
Chris@10
|
1046 T71 = FMA(KP951056516, T6t, KP587785252 * T6s);
|
Chris@10
|
1047 T2K = T2I - T2J;
|
Chris@10
|
1048 T2N = T2L - T2M;
|
Chris@10
|
1049 T2O = FMA(KP951056516, T2K, KP587785252 * T2N);
|
Chris@10
|
1050 T52 = FNMS(KP587785252, T2K, KP951056516 * T2N);
|
Chris@10
|
1051 }
|
Chris@10
|
1052 }
|
Chris@10
|
1053 {
|
Chris@10
|
1054 E T2g, T48, T3Y, T3Z, T4h, T4g, T43, T46, T49, T2p, T2A, T2B, T2e, T2f;
|
Chris@10
|
1055 T2e = cr[WS(rs, 3)];
|
Chris@10
|
1056 T2f = ci[WS(rs, 3)];
|
Chris@10
|
1057 T2g = FMA(T3, T2e, T6 * T2f);
|
Chris@10
|
1058 T48 = FNMS(T6, T2e, T3 * T2f);
|
Chris@10
|
1059 {
|
Chris@10
|
1060 E T2j, T41, T2z, T45, T2o, T42, T2u, T44;
|
Chris@10
|
1061 {
|
Chris@10
|
1062 E T2h, T2i, T2w, T2y;
|
Chris@10
|
1063 T2h = cr[WS(rs, 8)];
|
Chris@10
|
1064 T2i = ci[WS(rs, 8)];
|
Chris@10
|
1065 T2j = FMA(T1f, T2h, T1g * T2i);
|
Chris@10
|
1066 T41 = FNMS(T1g, T2h, T1f * T2i);
|
Chris@10
|
1067 T2w = cr[WS(rs, 18)];
|
Chris@10
|
1068 T2y = ci[WS(rs, 18)];
|
Chris@10
|
1069 T2z = FMA(T2v, T2w, T2x * T2y);
|
Chris@10
|
1070 T45 = FNMS(T2x, T2w, T2v * T2y);
|
Chris@10
|
1071 }
|
Chris@10
|
1072 {
|
Chris@10
|
1073 E T2l, T2n, T2r, T2t;
|
Chris@10
|
1074 T2l = cr[WS(rs, 23)];
|
Chris@10
|
1075 T2n = ci[WS(rs, 23)];
|
Chris@10
|
1076 T2o = FMA(T2k, T2l, T2m * T2n);
|
Chris@10
|
1077 T42 = FNMS(T2m, T2l, T2k * T2n);
|
Chris@10
|
1078 T2r = cr[WS(rs, 13)];
|
Chris@10
|
1079 T2t = ci[WS(rs, 13)];
|
Chris@10
|
1080 T2u = FMA(T2q, T2r, T2s * T2t);
|
Chris@10
|
1081 T44 = FNMS(T2s, T2r, T2q * T2t);
|
Chris@10
|
1082 }
|
Chris@10
|
1083 T3Y = T2j - T2o;
|
Chris@10
|
1084 T3Z = T2u - T2z;
|
Chris@10
|
1085 T4h = T44 - T45;
|
Chris@10
|
1086 T4g = T41 - T42;
|
Chris@10
|
1087 T43 = T41 + T42;
|
Chris@10
|
1088 T46 = T44 + T45;
|
Chris@10
|
1089 T49 = T43 + T46;
|
Chris@10
|
1090 T2p = T2j + T2o;
|
Chris@10
|
1091 T2A = T2u + T2z;
|
Chris@10
|
1092 T2B = T2p + T2A;
|
Chris@10
|
1093 }
|
Chris@10
|
1094 T2C = T2g + T2B;
|
Chris@10
|
1095 T6k = T48 + T49;
|
Chris@10
|
1096 {
|
Chris@10
|
1097 E T40, T5r, T4b, T5q, T47, T4a;
|
Chris@10
|
1098 T40 = FMA(KP951056516, T3Y, KP587785252 * T3Z);
|
Chris@10
|
1099 T5r = FNMS(KP587785252, T3Y, KP951056516 * T3Z);
|
Chris@10
|
1100 T47 = KP559016994 * (T43 - T46);
|
Chris@10
|
1101 T4a = FNMS(KP250000000, T49, T48);
|
Chris@10
|
1102 T4b = T47 + T4a;
|
Chris@10
|
1103 T5q = T4a - T47;
|
Chris@10
|
1104 T4c = T40 + T4b;
|
Chris@10
|
1105 T5X = T5r + T5q;
|
Chris@10
|
1106 T4L = T4b - T40;
|
Chris@10
|
1107 T5s = T5q - T5r;
|
Chris@10
|
1108 }
|
Chris@10
|
1109 {
|
Chris@10
|
1110 E T4i, T5u, T4f, T5t, T4d, T4e;
|
Chris@10
|
1111 T4i = FMA(KP951056516, T4g, KP587785252 * T4h);
|
Chris@10
|
1112 T5u = FNMS(KP587785252, T4g, KP951056516 * T4h);
|
Chris@10
|
1113 T4d = KP559016994 * (T2p - T2A);
|
Chris@10
|
1114 T4e = FNMS(KP250000000, T2B, T2g);
|
Chris@10
|
1115 T4f = T4d + T4e;
|
Chris@10
|
1116 T5t = T4e - T4d;
|
Chris@10
|
1117 T4j = T4f - T4i;
|
Chris@10
|
1118 T5W = T5t - T5u;
|
Chris@10
|
1119 T4K = T4f + T4i;
|
Chris@10
|
1120 T5v = T5t + T5u;
|
Chris@10
|
1121 }
|
Chris@10
|
1122 }
|
Chris@10
|
1123 {
|
Chris@10
|
1124 E TO, T37, T2V, T2Y, T32, T31, T34, T35, T38, T13, T1m, T1n, TM, TN;
|
Chris@10
|
1125 TM = cr[WS(rs, 1)];
|
Chris@10
|
1126 TN = ci[WS(rs, 1)];
|
Chris@10
|
1127 TO = FMA(T2, TM, T5 * TN);
|
Chris@10
|
1128 T37 = FNMS(T5, TM, T2 * TN);
|
Chris@10
|
1129 {
|
Chris@10
|
1130 E TX, T2T, T1l, T2X, T12, T2U, T1e, T2W;
|
Chris@10
|
1131 {
|
Chris@10
|
1132 E TS, TW, T1i, T1k;
|
Chris@10
|
1133 TS = cr[WS(rs, 6)];
|
Chris@10
|
1134 TW = ci[WS(rs, 6)];
|
Chris@10
|
1135 TX = FMA(TR, TS, TV * TW);
|
Chris@10
|
1136 T2T = FNMS(TV, TS, TR * TW);
|
Chris@10
|
1137 T1i = cr[WS(rs, 16)];
|
Chris@10
|
1138 T1k = ci[WS(rs, 16)];
|
Chris@10
|
1139 T1l = FMA(T1h, T1i, T1j * T1k);
|
Chris@10
|
1140 T2X = FNMS(T1j, T1i, T1h * T1k);
|
Chris@10
|
1141 }
|
Chris@10
|
1142 {
|
Chris@10
|
1143 E TZ, T11, T19, T1d;
|
Chris@10
|
1144 TZ = cr[WS(rs, 21)];
|
Chris@10
|
1145 T11 = ci[WS(rs, 21)];
|
Chris@10
|
1146 T12 = FMA(TY, TZ, T10 * T11);
|
Chris@10
|
1147 T2U = FNMS(T10, TZ, TY * T11);
|
Chris@10
|
1148 T19 = cr[WS(rs, 11)];
|
Chris@10
|
1149 T1d = ci[WS(rs, 11)];
|
Chris@10
|
1150 T1e = FMA(T18, T19, T1c * T1d);
|
Chris@10
|
1151 T2W = FNMS(T1c, T19, T18 * T1d);
|
Chris@10
|
1152 }
|
Chris@10
|
1153 T2V = T2T - T2U;
|
Chris@10
|
1154 T2Y = T2W - T2X;
|
Chris@10
|
1155 T32 = T1e - T1l;
|
Chris@10
|
1156 T31 = TX - T12;
|
Chris@10
|
1157 T34 = T2T + T2U;
|
Chris@10
|
1158 T35 = T2W + T2X;
|
Chris@10
|
1159 T38 = T34 + T35;
|
Chris@10
|
1160 T13 = TX + T12;
|
Chris@10
|
1161 T1m = T1e + T1l;
|
Chris@10
|
1162 T1n = T13 + T1m;
|
Chris@10
|
1163 }
|
Chris@10
|
1164 T1o = TO + T1n;
|
Chris@10
|
1165 T6g = T37 + T38;
|
Chris@10
|
1166 {
|
Chris@10
|
1167 E T2Z, T55, T2S, T54, T2Q, T2R;
|
Chris@10
|
1168 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
|
Chris@10
|
1169 T55 = FNMS(KP587785252, T2V, KP951056516 * T2Y);
|
Chris@10
|
1170 T2Q = KP559016994 * (T13 - T1m);
|
Chris@10
|
1171 T2R = FNMS(KP250000000, T1n, TO);
|
Chris@10
|
1172 T2S = T2Q + T2R;
|
Chris@10
|
1173 T54 = T2R - T2Q;
|
Chris@10
|
1174 T30 = T2S - T2Z;
|
Chris@10
|
1175 T5M = T54 - T55;
|
Chris@10
|
1176 T4A = T2S + T2Z;
|
Chris@10
|
1177 T56 = T54 + T55;
|
Chris@10
|
1178 }
|
Chris@10
|
1179 {
|
Chris@10
|
1180 E T33, T58, T3a, T57, T36, T39;
|
Chris@10
|
1181 T33 = FMA(KP951056516, T31, KP587785252 * T32);
|
Chris@10
|
1182 T58 = FNMS(KP587785252, T31, KP951056516 * T32);
|
Chris@10
|
1183 T36 = KP559016994 * (T34 - T35);
|
Chris@10
|
1184 T39 = FNMS(KP250000000, T38, T37);
|
Chris@10
|
1185 T3a = T36 + T39;
|
Chris@10
|
1186 T57 = T39 - T36;
|
Chris@10
|
1187 T3b = T33 + T3a;
|
Chris@10
|
1188 T5N = T58 + T57;
|
Chris@10
|
1189 T4B = T3a - T33;
|
Chris@10
|
1190 T59 = T57 - T58;
|
Chris@10
|
1191 }
|
Chris@10
|
1192 }
|
Chris@10
|
1193 {
|
Chris@10
|
1194 E T1r, T3n, T3d, T3e, T3w, T3v, T3i, T3l, T3o, T1y, T1J, T1K, T1p, T1q;
|
Chris@10
|
1195 T1p = cr[WS(rs, 4)];
|
Chris@10
|
1196 T1q = ci[WS(rs, 4)];
|
Chris@10
|
1197 T1r = FMA(T8, T1p, Td * T1q);
|
Chris@10
|
1198 T3n = FNMS(Td, T1p, T8 * T1q);
|
Chris@10
|
1199 {
|
Chris@10
|
1200 E T1u, T3g, T1I, T3k, T1x, T3h, T1D, T3j;
|
Chris@10
|
1201 {
|
Chris@10
|
1202 E T1s, T1t, T1F, T1H;
|
Chris@10
|
1203 T1s = cr[WS(rs, 9)];
|
Chris@10
|
1204 T1t = ci[WS(rs, 9)];
|
Chris@10
|
1205 T1u = FMA(T9, T1s, Te * T1t);
|
Chris@10
|
1206 T3g = FNMS(Te, T1s, T9 * T1t);
|
Chris@10
|
1207 T1F = cr[WS(rs, 19)];
|
Chris@10
|
1208 T1H = ci[WS(rs, 19)];
|
Chris@10
|
1209 T1I = FMA(T1E, T1F, T1G * T1H);
|
Chris@10
|
1210 T3k = FNMS(T1G, T1F, T1E * T1H);
|
Chris@10
|
1211 }
|
Chris@10
|
1212 {
|
Chris@10
|
1213 E T1v, T1w, T1A, T1C;
|
Chris@10
|
1214 T1v = cr[WS(rs, 24)];
|
Chris@10
|
1215 T1w = ci[WS(rs, 24)];
|
Chris@10
|
1216 T1x = FMA(Tn, T1v, To * T1w);
|
Chris@10
|
1217 T3h = FNMS(To, T1v, Tn * T1w);
|
Chris@10
|
1218 T1A = cr[WS(rs, 14)];
|
Chris@10
|
1219 T1C = ci[WS(rs, 14)];
|
Chris@10
|
1220 T1D = FMA(T1z, T1A, T1B * T1C);
|
Chris@10
|
1221 T3j = FNMS(T1B, T1A, T1z * T1C);
|
Chris@10
|
1222 }
|
Chris@10
|
1223 T3d = T1x - T1u;
|
Chris@10
|
1224 T3e = T1D - T1I;
|
Chris@10
|
1225 T3w = T3j - T3k;
|
Chris@10
|
1226 T3v = T3g - T3h;
|
Chris@10
|
1227 T3i = T3g + T3h;
|
Chris@10
|
1228 T3l = T3j + T3k;
|
Chris@10
|
1229 T3o = T3i + T3l;
|
Chris@10
|
1230 T1y = T1u + T1x;
|
Chris@10
|
1231 T1J = T1D + T1I;
|
Chris@10
|
1232 T1K = T1y + T1J;
|
Chris@10
|
1233 }
|
Chris@10
|
1234 T1L = T1r + T1K;
|
Chris@10
|
1235 T6h = T3n + T3o;
|
Chris@10
|
1236 {
|
Chris@10
|
1237 E T3f, T5c, T3q, T5b, T3m, T3p;
|
Chris@10
|
1238 T3f = FNMS(KP587785252, T3e, KP951056516 * T3d);
|
Chris@10
|
1239 T5c = FMA(KP587785252, T3d, KP951056516 * T3e);
|
Chris@10
|
1240 T3m = KP559016994 * (T3i - T3l);
|
Chris@10
|
1241 T3p = FNMS(KP250000000, T3o, T3n);
|
Chris@10
|
1242 T3q = T3m + T3p;
|
Chris@10
|
1243 T5b = T3p - T3m;
|
Chris@10
|
1244 T3r = T3f - T3q;
|
Chris@10
|
1245 T5P = T5c + T5b;
|
Chris@10
|
1246 T4E = T3f + T3q;
|
Chris@10
|
1247 T5d = T5b - T5c;
|
Chris@10
|
1248 }
|
Chris@10
|
1249 {
|
Chris@10
|
1250 E T3x, T5f, T3u, T5e, T3s, T3t;
|
Chris@10
|
1251 T3x = FMA(KP951056516, T3v, KP587785252 * T3w);
|
Chris@10
|
1252 T5f = FNMS(KP587785252, T3v, KP951056516 * T3w);
|
Chris@10
|
1253 T3s = KP559016994 * (T1y - T1J);
|
Chris@10
|
1254 T3t = FNMS(KP250000000, T1K, T1r);
|
Chris@10
|
1255 T3u = T3s + T3t;
|
Chris@10
|
1256 T5e = T3t - T3s;
|
Chris@10
|
1257 T3y = T3u - T3x;
|
Chris@10
|
1258 T5Q = T5e - T5f;
|
Chris@10
|
1259 T4D = T3u + T3x;
|
Chris@10
|
1260 T5g = T5e + T5f;
|
Chris@10
|
1261 }
|
Chris@10
|
1262 }
|
Chris@10
|
1263 {
|
Chris@10
|
1264 E T1P, T3L, T3B, T3C, T3U, T3T, T3G, T3J, T3M, T20, T2b, T2c, T1N, T1O;
|
Chris@10
|
1265 T1N = cr[WS(rs, 2)];
|
Chris@10
|
1266 T1O = ci[WS(rs, 2)];
|
Chris@10
|
1267 T1P = FMA(T14, T1N, T16 * T1O);
|
Chris@10
|
1268 T3L = FNMS(T16, T1N, T14 * T1O);
|
Chris@10
|
1269 {
|
Chris@10
|
1270 E T1U, T3E, T2a, T3I, T1Z, T3F, T25, T3H;
|
Chris@10
|
1271 {
|
Chris@10
|
1272 E T1R, T1T, T27, T29;
|
Chris@10
|
1273 T1R = cr[WS(rs, 7)];
|
Chris@10
|
1274 T1T = ci[WS(rs, 7)];
|
Chris@10
|
1275 T1U = FMA(T1Q, T1R, T1S * T1T);
|
Chris@10
|
1276 T3E = FNMS(T1S, T1R, T1Q * T1T);
|
Chris@10
|
1277 T27 = cr[WS(rs, 17)];
|
Chris@10
|
1278 T29 = ci[WS(rs, 17)];
|
Chris@10
|
1279 T2a = FMA(T26, T27, T28 * T29);
|
Chris@10
|
1280 T3I = FNMS(T28, T27, T26 * T29);
|
Chris@10
|
1281 }
|
Chris@10
|
1282 {
|
Chris@10
|
1283 E T1W, T1Y, T22, T24;
|
Chris@10
|
1284 T1W = cr[WS(rs, 22)];
|
Chris@10
|
1285 T1Y = ci[WS(rs, 22)];
|
Chris@10
|
1286 T1Z = FMA(T1V, T1W, T1X * T1Y);
|
Chris@10
|
1287 T3F = FNMS(T1X, T1W, T1V * T1Y);
|
Chris@10
|
1288 T22 = cr[WS(rs, 12)];
|
Chris@10
|
1289 T24 = ci[WS(rs, 12)];
|
Chris@10
|
1290 T25 = FMA(T21, T22, T23 * T24);
|
Chris@10
|
1291 T3H = FNMS(T23, T22, T21 * T24);
|
Chris@10
|
1292 }
|
Chris@10
|
1293 T3B = T1U - T1Z;
|
Chris@10
|
1294 T3C = T25 - T2a;
|
Chris@10
|
1295 T3U = T3H - T3I;
|
Chris@10
|
1296 T3T = T3E - T3F;
|
Chris@10
|
1297 T3G = T3E + T3F;
|
Chris@10
|
1298 T3J = T3H + T3I;
|
Chris@10
|
1299 T3M = T3G + T3J;
|
Chris@10
|
1300 T20 = T1U + T1Z;
|
Chris@10
|
1301 T2b = T25 + T2a;
|
Chris@10
|
1302 T2c = T20 + T2b;
|
Chris@10
|
1303 }
|
Chris@10
|
1304 T2d = T1P + T2c;
|
Chris@10
|
1305 T6j = T3L + T3M;
|
Chris@10
|
1306 {
|
Chris@10
|
1307 E T3D, T5n, T3O, T5m, T3K, T3N;
|
Chris@10
|
1308 T3D = FMA(KP951056516, T3B, KP587785252 * T3C);
|
Chris@10
|
1309 T5n = FNMS(KP587785252, T3B, KP951056516 * T3C);
|
Chris@10
|
1310 T3K = KP559016994 * (T3G - T3J);
|
Chris@10
|
1311 T3N = FNMS(KP250000000, T3M, T3L);
|
Chris@10
|
1312 T3O = T3K + T3N;
|
Chris@10
|
1313 T5m = T3N - T3K;
|
Chris@10
|
1314 T3P = T3D + T3O;
|
Chris@10
|
1315 T5U = T5n + T5m;
|
Chris@10
|
1316 T4I = T3O - T3D;
|
Chris@10
|
1317 T5o = T5m - T5n;
|
Chris@10
|
1318 }
|
Chris@10
|
1319 {
|
Chris@10
|
1320 E T3V, T5k, T3S, T5j, T3Q, T3R;
|
Chris@10
|
1321 T3V = FMA(KP951056516, T3T, KP587785252 * T3U);
|
Chris@10
|
1322 T5k = FNMS(KP587785252, T3T, KP951056516 * T3U);
|
Chris@10
|
1323 T3Q = KP559016994 * (T20 - T2b);
|
Chris@10
|
1324 T3R = FNMS(KP250000000, T2c, T1P);
|
Chris@10
|
1325 T3S = T3Q + T3R;
|
Chris@10
|
1326 T5j = T3R - T3Q;
|
Chris@10
|
1327 T3W = T3S - T3V;
|
Chris@10
|
1328 T5T = T5j - T5k;
|
Chris@10
|
1329 T4H = T3S + T3V;
|
Chris@10
|
1330 T5l = T5j + T5k;
|
Chris@10
|
1331 }
|
Chris@10
|
1332 }
|
Chris@10
|
1333 {
|
Chris@10
|
1334 E T6m, T6o, TL, T2E, T6d, T6e, T6n, T6f;
|
Chris@10
|
1335 {
|
Chris@10
|
1336 E T6i, T6l, T1M, T2D;
|
Chris@10
|
1337 T6i = T6g - T6h;
|
Chris@10
|
1338 T6l = T6j - T6k;
|
Chris@10
|
1339 T6m = FMA(KP951056516, T6i, KP587785252 * T6l);
|
Chris@10
|
1340 T6o = FNMS(KP587785252, T6i, KP951056516 * T6l);
|
Chris@10
|
1341 TL = T1 + TK;
|
Chris@10
|
1342 T1M = T1o + T1L;
|
Chris@10
|
1343 T2D = T2d + T2C;
|
Chris@10
|
1344 T2E = T1M + T2D;
|
Chris@10
|
1345 T6d = KP559016994 * (T1M - T2D);
|
Chris@10
|
1346 T6e = FNMS(KP250000000, T2E, TL);
|
Chris@10
|
1347 }
|
Chris@10
|
1348 cr[0] = TL + T2E;
|
Chris@10
|
1349 T6n = T6e - T6d;
|
Chris@10
|
1350 cr[WS(rs, 10)] = T6n - T6o;
|
Chris@10
|
1351 ci[WS(rs, 9)] = T6n + T6o;
|
Chris@10
|
1352 T6f = T6d + T6e;
|
Chris@10
|
1353 ci[WS(rs, 4)] = T6f - T6m;
|
Chris@10
|
1354 cr[WS(rs, 5)] = T6f + T6m;
|
Chris@10
|
1355 }
|
Chris@10
|
1356 {
|
Chris@10
|
1357 E T2P, T4z, T72, T7e, T4m, T7j, T4n, T7i, T4U, T77, T4X, T75, T4O, T6Y, T4P;
|
Chris@10
|
1358 E T6X, T4s, T7f, T4v, T7d, T2H, T70;
|
Chris@10
|
1359 T2H = T2F + T2G;
|
Chris@10
|
1360 T2P = T2H - T2O;
|
Chris@10
|
1361 T4z = T2H + T2O;
|
Chris@10
|
1362 T70 = T6A + T6z;
|
Chris@10
|
1363 T72 = T70 - T71;
|
Chris@10
|
1364 T7e = T71 + T70;
|
Chris@10
|
1365 {
|
Chris@10
|
1366 E T3c, T3z, T3A, T3X, T4k, T4l;
|
Chris@10
|
1367 T3c = FMA(KP535826794, T30, KP844327925 * T3b);
|
Chris@10
|
1368 T3z = FNMS(KP637423989, T3y, KP770513242 * T3r);
|
Chris@10
|
1369 T3A = T3c + T3z;
|
Chris@10
|
1370 T3X = FNMS(KP425779291, T3W, KP904827052 * T3P);
|
Chris@10
|
1371 T4k = FNMS(KP992114701, T4j, KP125333233 * T4c);
|
Chris@10
|
1372 T4l = T3X + T4k;
|
Chris@10
|
1373 T4m = T3A + T4l;
|
Chris@10
|
1374 T7j = T3X - T4k;
|
Chris@10
|
1375 T4n = KP559016994 * (T3A - T4l);
|
Chris@10
|
1376 T7i = T3z - T3c;
|
Chris@10
|
1377 }
|
Chris@10
|
1378 {
|
Chris@10
|
1379 E T4S, T4T, T73, T4V, T4W, T74;
|
Chris@10
|
1380 T4S = FNMS(KP248689887, T4A, KP968583161 * T4B);
|
Chris@10
|
1381 T4T = FNMS(KP844327925, T4D, KP535826794 * T4E);
|
Chris@10
|
1382 T73 = T4S + T4T;
|
Chris@10
|
1383 T4V = FNMS(KP481753674, T4H, KP876306680 * T4I);
|
Chris@10
|
1384 T4W = FNMS(KP684547105, T4K, KP728968627 * T4L);
|
Chris@10
|
1385 T74 = T4V + T4W;
|
Chris@10
|
1386 T4U = T4S - T4T;
|
Chris@10
|
1387 T77 = KP559016994 * (T73 - T74);
|
Chris@10
|
1388 T4X = T4V - T4W;
|
Chris@10
|
1389 T75 = T73 + T74;
|
Chris@10
|
1390 }
|
Chris@10
|
1391 {
|
Chris@10
|
1392 E T4C, T4F, T4G, T4J, T4M, T4N;
|
Chris@10
|
1393 T4C = FMA(KP968583161, T4A, KP248689887 * T4B);
|
Chris@10
|
1394 T4F = FMA(KP535826794, T4D, KP844327925 * T4E);
|
Chris@10
|
1395 T4G = T4C + T4F;
|
Chris@10
|
1396 T4J = FMA(KP876306680, T4H, KP481753674 * T4I);
|
Chris@10
|
1397 T4M = FMA(KP728968627, T4K, KP684547105 * T4L);
|
Chris@10
|
1398 T4N = T4J + T4M;
|
Chris@10
|
1399 T4O = T4G + T4N;
|
Chris@10
|
1400 T6Y = T4J - T4M;
|
Chris@10
|
1401 T4P = KP559016994 * (T4G - T4N);
|
Chris@10
|
1402 T6X = T4F - T4C;
|
Chris@10
|
1403 }
|
Chris@10
|
1404 {
|
Chris@10
|
1405 E T4q, T4r, T7b, T4t, T4u, T7c;
|
Chris@10
|
1406 T4q = FNMS(KP844327925, T30, KP535826794 * T3b);
|
Chris@10
|
1407 T4r = FMA(KP770513242, T3y, KP637423989 * T3r);
|
Chris@10
|
1408 T7b = T4q + T4r;
|
Chris@10
|
1409 T4t = FMA(KP125333233, T4j, KP992114701 * T4c);
|
Chris@10
|
1410 T4u = FMA(KP904827052, T3W, KP425779291 * T3P);
|
Chris@10
|
1411 T7c = T4u + T4t;
|
Chris@10
|
1412 T4s = T4q - T4r;
|
Chris@10
|
1413 T7f = T7b - T7c;
|
Chris@10
|
1414 T4v = T4t - T4u;
|
Chris@10
|
1415 T7d = KP559016994 * (T7b + T7c);
|
Chris@10
|
1416 }
|
Chris@10
|
1417 cr[WS(rs, 4)] = T2P + T4m;
|
Chris@10
|
1418 ci[WS(rs, 23)] = T75 + T72;
|
Chris@10
|
1419 ci[WS(rs, 20)] = T7f + T7e;
|
Chris@10
|
1420 cr[WS(rs, 1)] = T4z + T4O;
|
Chris@10
|
1421 {
|
Chris@10
|
1422 E T4w, T4y, T4p, T4x, T4o;
|
Chris@10
|
1423 T4w = FMA(KP951056516, T4s, KP587785252 * T4v);
|
Chris@10
|
1424 T4y = FNMS(KP587785252, T4s, KP951056516 * T4v);
|
Chris@10
|
1425 T4o = FNMS(KP250000000, T4m, T2P);
|
Chris@10
|
1426 T4p = T4n + T4o;
|
Chris@10
|
1427 T4x = T4o - T4n;
|
Chris@10
|
1428 ci[0] = T4p - T4w;
|
Chris@10
|
1429 ci[WS(rs, 5)] = T4x + T4y;
|
Chris@10
|
1430 cr[WS(rs, 9)] = T4p + T4w;
|
Chris@10
|
1431 ci[WS(rs, 10)] = T4x - T4y;
|
Chris@10
|
1432 }
|
Chris@10
|
1433 {
|
Chris@10
|
1434 E T6Z, T79, T78, T7a, T76;
|
Chris@10
|
1435 T6Z = FMA(KP587785252, T6X, KP951056516 * T6Y);
|
Chris@10
|
1436 T79 = FNMS(KP587785252, T6Y, KP951056516 * T6X);
|
Chris@10
|
1437 T76 = FNMS(KP250000000, T75, T72);
|
Chris@10
|
1438 T78 = T76 - T77;
|
Chris@10
|
1439 T7a = T77 + T76;
|
Chris@10
|
1440 cr[WS(rs, 16)] = T6Z - T78;
|
Chris@10
|
1441 ci[WS(rs, 18)] = T79 + T7a;
|
Chris@10
|
1442 ci[WS(rs, 13)] = T6Z + T78;
|
Chris@10
|
1443 cr[WS(rs, 21)] = T79 - T7a;
|
Chris@10
|
1444 }
|
Chris@10
|
1445 {
|
Chris@10
|
1446 E T7k, T7l, T7h, T7m, T7g;
|
Chris@10
|
1447 T7k = FMA(KP587785252, T7i, KP951056516 * T7j);
|
Chris@10
|
1448 T7l = FNMS(KP587785252, T7j, KP951056516 * T7i);
|
Chris@10
|
1449 T7g = FNMS(KP250000000, T7f, T7e);
|
Chris@10
|
1450 T7h = T7d - T7g;
|
Chris@10
|
1451 T7m = T7d + T7g;
|
Chris@10
|
1452 cr[WS(rs, 14)] = T7h - T7k;
|
Chris@10
|
1453 ci[WS(rs, 15)] = T7l + T7m;
|
Chris@10
|
1454 cr[WS(rs, 19)] = T7k + T7h;
|
Chris@10
|
1455 cr[WS(rs, 24)] = T7l - T7m;
|
Chris@10
|
1456 }
|
Chris@10
|
1457 {
|
Chris@10
|
1458 E T4Y, T50, T4R, T4Z, T4Q;
|
Chris@10
|
1459 T4Y = FMA(KP951056516, T4U, KP587785252 * T4X);
|
Chris@10
|
1460 T50 = FNMS(KP587785252, T4U, KP951056516 * T4X);
|
Chris@10
|
1461 T4Q = FNMS(KP250000000, T4O, T4z);
|
Chris@10
|
1462 T4R = T4P + T4Q;
|
Chris@10
|
1463 T4Z = T4Q - T4P;
|
Chris@10
|
1464 ci[WS(rs, 3)] = T4R - T4Y;
|
Chris@10
|
1465 ci[WS(rs, 8)] = T4Z + T50;
|
Chris@10
|
1466 cr[WS(rs, 6)] = T4R + T4Y;
|
Chris@10
|
1467 cr[WS(rs, 11)] = T4Z - T50;
|
Chris@10
|
1468 }
|
Chris@10
|
1469 }
|
Chris@10
|
1470 {
|
Chris@10
|
1471 E T7p, T7x, T7q, T7t, T7u, T7v, T7y, T7w;
|
Chris@10
|
1472 {
|
Chris@10
|
1473 E T7n, T7o, T7r, T7s;
|
Chris@10
|
1474 T7n = T1L - T1o;
|
Chris@10
|
1475 T7o = T2d - T2C;
|
Chris@10
|
1476 T7p = FMA(KP587785252, T7n, KP951056516 * T7o);
|
Chris@10
|
1477 T7x = FNMS(KP587785252, T7o, KP951056516 * T7n);
|
Chris@10
|
1478 T7q = T6y + T6v;
|
Chris@10
|
1479 T7r = T6g + T6h;
|
Chris@10
|
1480 T7s = T6j + T6k;
|
Chris@10
|
1481 T7t = T7r + T7s;
|
Chris@10
|
1482 T7u = FNMS(KP250000000, T7t, T7q);
|
Chris@10
|
1483 T7v = KP559016994 * (T7r - T7s);
|
Chris@10
|
1484 }
|
Chris@10
|
1485 ci[WS(rs, 24)] = T7t + T7q;
|
Chris@10
|
1486 T7y = T7v + T7u;
|
Chris@10
|
1487 cr[WS(rs, 20)] = T7x - T7y;
|
Chris@10
|
1488 ci[WS(rs, 19)] = T7x + T7y;
|
Chris@10
|
1489 T7w = T7u - T7v;
|
Chris@10
|
1490 cr[WS(rs, 15)] = T7p - T7w;
|
Chris@10
|
1491 ci[WS(rs, 14)] = T7p + T7w;
|
Chris@10
|
1492 }
|
Chris@10
|
1493 {
|
Chris@10
|
1494 E T53, T5L, T6C, T6O, T5y, T6T, T5z, T6S, T66, T6H, T69, T6F, T60, T6q, T61;
|
Chris@10
|
1495 E T6p, T5E, T6P, T5H, T6N, T51, T6B;
|
Chris@10
|
1496 T51 = T2G - T2F;
|
Chris@10
|
1497 T53 = T51 + T52;
|
Chris@10
|
1498 T5L = T51 - T52;
|
Chris@10
|
1499 T6B = T6z - T6A;
|
Chris@10
|
1500 T6C = T6u + T6B;
|
Chris@10
|
1501 T6O = T6B - T6u;
|
Chris@10
|
1502 {
|
Chris@10
|
1503 E T5a, T5h, T5i, T5p, T5w, T5x;
|
Chris@10
|
1504 T5a = FMA(KP728968627, T56, KP684547105 * T59);
|
Chris@10
|
1505 T5h = FNMS(KP992114701, T5g, KP125333233 * T5d);
|
Chris@10
|
1506 T5i = T5a + T5h;
|
Chris@10
|
1507 T5p = FMA(KP062790519, T5l, KP998026728 * T5o);
|
Chris@10
|
1508 T5w = FNMS(KP637423989, T5v, KP770513242 * T5s);
|
Chris@10
|
1509 T5x = T5p + T5w;
|
Chris@10
|
1510 T5y = T5i + T5x;
|
Chris@10
|
1511 T6T = T5p - T5w;
|
Chris@10
|
1512 T5z = KP559016994 * (T5i - T5x);
|
Chris@10
|
1513 T6S = T5h - T5a;
|
Chris@10
|
1514 }
|
Chris@10
|
1515 {
|
Chris@10
|
1516 E T64, T65, T6D, T67, T68, T6E;
|
Chris@10
|
1517 T64 = FNMS(KP481753674, T5M, KP876306680 * T5N);
|
Chris@10
|
1518 T65 = FMA(KP904827052, T5Q, KP425779291 * T5P);
|
Chris@10
|
1519 T6D = T64 - T65;
|
Chris@10
|
1520 T67 = FNMS(KP844327925, T5T, KP535826794 * T5U);
|
Chris@10
|
1521 T68 = FNMS(KP998026728, T5W, KP062790519 * T5X);
|
Chris@10
|
1522 T6E = T67 + T68;
|
Chris@10
|
1523 T66 = T64 + T65;
|
Chris@10
|
1524 T6H = KP559016994 * (T6D - T6E);
|
Chris@10
|
1525 T69 = T67 - T68;
|
Chris@10
|
1526 T6F = T6D + T6E;
|
Chris@10
|
1527 }
|
Chris@10
|
1528 {
|
Chris@10
|
1529 E T5O, T5R, T5S, T5V, T5Y, T5Z;
|
Chris@10
|
1530 T5O = FMA(KP876306680, T5M, KP481753674 * T5N);
|
Chris@10
|
1531 T5R = FNMS(KP425779291, T5Q, KP904827052 * T5P);
|
Chris@10
|
1532 T5S = T5O + T5R;
|
Chris@10
|
1533 T5V = FMA(KP535826794, T5T, KP844327925 * T5U);
|
Chris@10
|
1534 T5Y = FMA(KP062790519, T5W, KP998026728 * T5X);
|
Chris@10
|
1535 T5Z = T5V + T5Y;
|
Chris@10
|
1536 T60 = T5S + T5Z;
|
Chris@10
|
1537 T6q = T5V - T5Y;
|
Chris@10
|
1538 T61 = KP559016994 * (T5S - T5Z);
|
Chris@10
|
1539 T6p = T5R - T5O;
|
Chris@10
|
1540 }
|
Chris@10
|
1541 {
|
Chris@10
|
1542 E T5C, T5D, T6L, T5F, T5G, T6M;
|
Chris@10
|
1543 T5C = FNMS(KP684547105, T56, KP728968627 * T59);
|
Chris@10
|
1544 T5D = FMA(KP125333233, T5g, KP992114701 * T5d);
|
Chris@10
|
1545 T6L = T5C - T5D;
|
Chris@10
|
1546 T5F = FNMS(KP998026728, T5l, KP062790519 * T5o);
|
Chris@10
|
1547 T5G = FMA(KP770513242, T5v, KP637423989 * T5s);
|
Chris@10
|
1548 T6M = T5F - T5G;
|
Chris@10
|
1549 T5E = T5C + T5D;
|
Chris@10
|
1550 T6P = T6L + T6M;
|
Chris@10
|
1551 T5H = T5F + T5G;
|
Chris@10
|
1552 T6N = KP559016994 * (T6L - T6M);
|
Chris@10
|
1553 }
|
Chris@10
|
1554 cr[WS(rs, 3)] = T53 + T5y;
|
Chris@10
|
1555 ci[WS(rs, 22)] = T6F + T6C;
|
Chris@10
|
1556 ci[WS(rs, 21)] = T6P + T6O;
|
Chris@10
|
1557 cr[WS(rs, 2)] = T5L + T60;
|
Chris@10
|
1558 {
|
Chris@10
|
1559 E T6r, T6J, T6I, T6K, T6G;
|
Chris@10
|
1560 T6r = FMA(KP587785252, T6p, KP951056516 * T6q);
|
Chris@10
|
1561 T6J = FNMS(KP587785252, T6q, KP951056516 * T6p);
|
Chris@10
|
1562 T6G = FNMS(KP250000000, T6F, T6C);
|
Chris@10
|
1563 T6I = T6G - T6H;
|
Chris@10
|
1564 T6K = T6H + T6G;
|
Chris@10
|
1565 cr[WS(rs, 17)] = T6r - T6I;
|
Chris@10
|
1566 ci[WS(rs, 17)] = T6J + T6K;
|
Chris@10
|
1567 ci[WS(rs, 12)] = T6r + T6I;
|
Chris@10
|
1568 cr[WS(rs, 22)] = T6J - T6K;
|
Chris@10
|
1569 }
|
Chris@10
|
1570 {
|
Chris@10
|
1571 E T6a, T6c, T63, T6b, T62;
|
Chris@10
|
1572 T6a = FMA(KP951056516, T66, KP587785252 * T69);
|
Chris@10
|
1573 T6c = FNMS(KP587785252, T66, KP951056516 * T69);
|
Chris@10
|
1574 T62 = FNMS(KP250000000, T60, T5L);
|
Chris@10
|
1575 T63 = T61 + T62;
|
Chris@10
|
1576 T6b = T62 - T61;
|
Chris@10
|
1577 ci[WS(rs, 2)] = T63 - T6a;
|
Chris@10
|
1578 ci[WS(rs, 7)] = T6b + T6c;
|
Chris@10
|
1579 cr[WS(rs, 7)] = T63 + T6a;
|
Chris@10
|
1580 cr[WS(rs, 12)] = T6b - T6c;
|
Chris@10
|
1581 }
|
Chris@10
|
1582 {
|
Chris@10
|
1583 E T5I, T5K, T5B, T5J, T5A;
|
Chris@10
|
1584 T5I = FMA(KP951056516, T5E, KP587785252 * T5H);
|
Chris@10
|
1585 T5K = FNMS(KP587785252, T5E, KP951056516 * T5H);
|
Chris@10
|
1586 T5A = FNMS(KP250000000, T5y, T53);
|
Chris@10
|
1587 T5B = T5z + T5A;
|
Chris@10
|
1588 T5J = T5A - T5z;
|
Chris@10
|
1589 ci[WS(rs, 1)] = T5B - T5I;
|
Chris@10
|
1590 ci[WS(rs, 6)] = T5J + T5K;
|
Chris@10
|
1591 cr[WS(rs, 8)] = T5B + T5I;
|
Chris@10
|
1592 ci[WS(rs, 11)] = T5J - T5K;
|
Chris@10
|
1593 }
|
Chris@10
|
1594 {
|
Chris@10
|
1595 E T6U, T6V, T6R, T6W, T6Q;
|
Chris@10
|
1596 T6U = FMA(KP587785252, T6S, KP951056516 * T6T);
|
Chris@10
|
1597 T6V = FNMS(KP587785252, T6T, KP951056516 * T6S);
|
Chris@10
|
1598 T6Q = FNMS(KP250000000, T6P, T6O);
|
Chris@10
|
1599 T6R = T6N - T6Q;
|
Chris@10
|
1600 T6W = T6N + T6Q;
|
Chris@10
|
1601 cr[WS(rs, 13)] = T6R - T6U;
|
Chris@10
|
1602 ci[WS(rs, 16)] = T6V + T6W;
|
Chris@10
|
1603 cr[WS(rs, 18)] = T6U + T6R;
|
Chris@10
|
1604 cr[WS(rs, 23)] = T6V - T6W;
|
Chris@10
|
1605 }
|
Chris@10
|
1606 }
|
Chris@10
|
1607 }
|
Chris@10
|
1608 }
|
Chris@10
|
1609 }
|
Chris@10
|
1610 }
|
Chris@10
|
1611
|
Chris@10
|
1612 static const tw_instr twinstr[] = {
|
Chris@10
|
1613 {TW_CEXP, 1, 1},
|
Chris@10
|
1614 {TW_CEXP, 1, 3},
|
Chris@10
|
1615 {TW_CEXP, 1, 9},
|
Chris@10
|
1616 {TW_CEXP, 1, 24},
|
Chris@10
|
1617 {TW_NEXT, 1, 0}
|
Chris@10
|
1618 };
|
Chris@10
|
1619
|
Chris@10
|
1620 static const hc2hc_desc desc = { 25, "hf2_25", twinstr, &GENUS, {280, 180, 160, 0} };
|
Chris@10
|
1621
|
Chris@10
|
1622 void X(codelet_hf2_25) (planner *p) {
|
Chris@10
|
1623 X(khc2hc_register) (p, hf2_25, &desc);
|
Chris@10
|
1624 }
|
Chris@10
|
1625 #endif /* HAVE_FMA */
|