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 07:40:01 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 -n 25 -dit -name hf_25 -include hf.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 * 178 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 hf_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(KP833417178, +0.833417178328688677408962550243238843138996060);
|
Chris@10
|
56 DK(KP921177326, +0.921177326965143320250447435415066029359282231);
|
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(KP554608978, +0.554608978404018097464974850792216217022558774);
|
Chris@10
|
61 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
Chris@10
|
62 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
63 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
64 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
Chris@10
|
65 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
66 DK(KP525970792, +0.525970792408939708442463226536226366643874659);
|
Chris@10
|
67 DK(KP726211448, +0.726211448929902658173535992263577167607493062);
|
Chris@10
|
68 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
69 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
Chris@10
|
70 DK(KP871714437, +0.871714437527667770979999223229522602943903653);
|
Chris@10
|
71 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
Chris@10
|
72 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
73 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
Chris@10
|
74 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
Chris@10
|
75 DK(KP851038619, +0.851038619207379630836264138867114231259902550);
|
Chris@10
|
76 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
Chris@10
|
77 DK(KP912018591, +0.912018591466481957908415381764119056233607330);
|
Chris@10
|
78 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
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) * 48); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 48, MAKE_VOLATILE_STRIDE(50, rs)) {
|
Chris@10
|
89 E T7i, T6o, T6m, T7o, T7m, T7h, T6n, T6f, T7j, T7n;
|
Chris@10
|
90 {
|
Chris@10
|
91 E T6W, T5G, T3Y, T3M, T7q, T70, T6V, T7P, Tt, T3L, T5T, T45, T5Q, T4c, T3G;
|
Chris@10
|
92 E T2G, T5P, T49, T5S, T42, T65, T4H, T68, T4A, T2Z, T11, T67, T4x, T64, T4E;
|
Chris@10
|
93 E T5Y, T4W, T61, T4P, T3d, T1z, T60, T4M, T5X, T4T, T3g, T1G, T3q, T4q, T4j;
|
Chris@10
|
94 E T26, T3i, T1M, T3k, T1S;
|
Chris@10
|
95 {
|
Chris@10
|
96 E T3u, T2e, T3E, T44, T4b, T2E, T3w, T2k, T3y, T2q;
|
Chris@10
|
97 {
|
Chris@10
|
98 E T1, T6R, T3P, T7, T3W, Tq, T9, Tc, Tb, T3U, Tk, T3Q, Ta;
|
Chris@10
|
99 {
|
Chris@10
|
100 E T3, T6, T2, T5;
|
Chris@10
|
101 T1 = cr[0];
|
Chris@10
|
102 T6R = ci[0];
|
Chris@10
|
103 T3 = cr[WS(rs, 5)];
|
Chris@10
|
104 T6 = ci[WS(rs, 5)];
|
Chris@10
|
105 T2 = W[8];
|
Chris@10
|
106 T5 = W[9];
|
Chris@10
|
107 {
|
Chris@10
|
108 E Tm, Tp, To, T3V, Tn, T3O, T4, Tl;
|
Chris@10
|
109 Tm = cr[WS(rs, 15)];
|
Chris@10
|
110 Tp = ci[WS(rs, 15)];
|
Chris@10
|
111 T3O = T2 * T6;
|
Chris@10
|
112 T4 = T2 * T3;
|
Chris@10
|
113 Tl = W[28];
|
Chris@10
|
114 To = W[29];
|
Chris@10
|
115 T3P = FNMS(T5, T3, T3O);
|
Chris@10
|
116 T7 = FMA(T5, T6, T4);
|
Chris@10
|
117 T3V = Tl * Tp;
|
Chris@10
|
118 Tn = Tl * Tm;
|
Chris@10
|
119 {
|
Chris@10
|
120 E Tg, Tj, Tf, Ti, T3T, Th, T8;
|
Chris@10
|
121 Tg = cr[WS(rs, 10)];
|
Chris@10
|
122 Tj = ci[WS(rs, 10)];
|
Chris@10
|
123 T3W = FNMS(To, Tm, T3V);
|
Chris@10
|
124 Tq = FMA(To, Tp, Tn);
|
Chris@10
|
125 Tf = W[18];
|
Chris@10
|
126 Ti = W[19];
|
Chris@10
|
127 T9 = cr[WS(rs, 20)];
|
Chris@10
|
128 Tc = ci[WS(rs, 20)];
|
Chris@10
|
129 T3T = Tf * Tj;
|
Chris@10
|
130 Th = Tf * Tg;
|
Chris@10
|
131 T8 = W[38];
|
Chris@10
|
132 Tb = W[39];
|
Chris@10
|
133 T3U = FNMS(Ti, Tg, T3T);
|
Chris@10
|
134 Tk = FMA(Ti, Tj, Th);
|
Chris@10
|
135 T3Q = T8 * Tc;
|
Chris@10
|
136 Ta = T8 * T9;
|
Chris@10
|
137 }
|
Chris@10
|
138 }
|
Chris@10
|
139 }
|
Chris@10
|
140 {
|
Chris@10
|
141 E T6T, T3X, T6Y, Tr, T3R, Td;
|
Chris@10
|
142 T6T = T3U + T3W;
|
Chris@10
|
143 T3X = T3U - T3W;
|
Chris@10
|
144 T6Y = Tk - Tq;
|
Chris@10
|
145 Tr = Tk + Tq;
|
Chris@10
|
146 T3R = FNMS(Tb, T9, T3Q);
|
Chris@10
|
147 Td = FMA(Tb, Tc, Ta);
|
Chris@10
|
148 {
|
Chris@10
|
149 E T3S, T6Z, Te, T6U, T6S, Ts;
|
Chris@10
|
150 T3S = T3P - T3R;
|
Chris@10
|
151 T6S = T3P + T3R;
|
Chris@10
|
152 T6Z = T7 - Td;
|
Chris@10
|
153 Te = T7 + Td;
|
Chris@10
|
154 T6W = T6S - T6T;
|
Chris@10
|
155 T6U = T6S + T6T;
|
Chris@10
|
156 T5G = FNMS(KP618033988, T3S, T3X);
|
Chris@10
|
157 T3Y = FMA(KP618033988, T3X, T3S);
|
Chris@10
|
158 T3M = Te - Tr;
|
Chris@10
|
159 Ts = Te + Tr;
|
Chris@10
|
160 T7q = FMA(KP618033988, T6Y, T6Z);
|
Chris@10
|
161 T70 = FNMS(KP618033988, T6Z, T6Y);
|
Chris@10
|
162 T6V = FNMS(KP250000000, T6U, T6R);
|
Chris@10
|
163 T7P = T6U + T6R;
|
Chris@10
|
164 Tt = T1 + Ts;
|
Chris@10
|
165 T3L = FNMS(KP250000000, Ts, T1);
|
Chris@10
|
166 }
|
Chris@10
|
167 }
|
Chris@10
|
168 }
|
Chris@10
|
169 {
|
Chris@10
|
170 E T2g, T2j, T2m, T3v, T2h, T2p, T2l, T2i, T2o, T3x, T2n;
|
Chris@10
|
171 {
|
Chris@10
|
172 E T2a, T2d, T29, T2c;
|
Chris@10
|
173 T2a = cr[WS(rs, 3)];
|
Chris@10
|
174 T2d = ci[WS(rs, 3)];
|
Chris@10
|
175 T29 = W[4];
|
Chris@10
|
176 T2c = W[5];
|
Chris@10
|
177 {
|
Chris@10
|
178 E T2t, T2w, T2z, T3A, T2u, T2C, T2y, T2v, T2B, T3t, T2b, T2s, T2f;
|
Chris@10
|
179 T2t = cr[WS(rs, 13)];
|
Chris@10
|
180 T2w = ci[WS(rs, 13)];
|
Chris@10
|
181 T3t = T29 * T2d;
|
Chris@10
|
182 T2b = T29 * T2a;
|
Chris@10
|
183 T2s = W[24];
|
Chris@10
|
184 T2z = cr[WS(rs, 18)];
|
Chris@10
|
185 T3u = FNMS(T2c, T2a, T3t);
|
Chris@10
|
186 T2e = FMA(T2c, T2d, T2b);
|
Chris@10
|
187 T3A = T2s * T2w;
|
Chris@10
|
188 T2u = T2s * T2t;
|
Chris@10
|
189 T2C = ci[WS(rs, 18)];
|
Chris@10
|
190 T2y = W[34];
|
Chris@10
|
191 T2v = W[25];
|
Chris@10
|
192 T2B = W[35];
|
Chris@10
|
193 {
|
Chris@10
|
194 E T3B, T2x, T3D, T2D, T3C, T2A;
|
Chris@10
|
195 T2g = cr[WS(rs, 8)];
|
Chris@10
|
196 T3C = T2y * T2C;
|
Chris@10
|
197 T2A = T2y * T2z;
|
Chris@10
|
198 T3B = FNMS(T2v, T2t, T3A);
|
Chris@10
|
199 T2x = FMA(T2v, T2w, T2u);
|
Chris@10
|
200 T3D = FNMS(T2B, T2z, T3C);
|
Chris@10
|
201 T2D = FMA(T2B, T2C, T2A);
|
Chris@10
|
202 T2j = ci[WS(rs, 8)];
|
Chris@10
|
203 T2f = W[14];
|
Chris@10
|
204 T3E = T3B + T3D;
|
Chris@10
|
205 T44 = T3D - T3B;
|
Chris@10
|
206 T4b = T2x - T2D;
|
Chris@10
|
207 T2E = T2x + T2D;
|
Chris@10
|
208 }
|
Chris@10
|
209 T2m = cr[WS(rs, 23)];
|
Chris@10
|
210 T3v = T2f * T2j;
|
Chris@10
|
211 T2h = T2f * T2g;
|
Chris@10
|
212 T2p = ci[WS(rs, 23)];
|
Chris@10
|
213 T2l = W[44];
|
Chris@10
|
214 T2i = W[15];
|
Chris@10
|
215 T2o = W[45];
|
Chris@10
|
216 }
|
Chris@10
|
217 }
|
Chris@10
|
218 T3x = T2l * T2p;
|
Chris@10
|
219 T2n = T2l * T2m;
|
Chris@10
|
220 T3w = FNMS(T2i, T2g, T3v);
|
Chris@10
|
221 T2k = FMA(T2i, T2j, T2h);
|
Chris@10
|
222 T3y = FNMS(T2o, T2m, T3x);
|
Chris@10
|
223 T2q = FMA(T2o, T2p, T2n);
|
Chris@10
|
224 }
|
Chris@10
|
225 {
|
Chris@10
|
226 E T2N, Tz, T2X, T4G, T4z, TZ, T2P, TF, T2R, TL;
|
Chris@10
|
227 {
|
Chris@10
|
228 E TB, TE, TH, T2O, TC, TK, TG, TD, TJ, T2Q, TI;
|
Chris@10
|
229 {
|
Chris@10
|
230 E Tv, Ty, Tu, Tx;
|
Chris@10
|
231 {
|
Chris@10
|
232 E T48, T41, T47, T40, T43, T3z;
|
Chris@10
|
233 Tv = cr[WS(rs, 1)];
|
Chris@10
|
234 T43 = T3y - T3w;
|
Chris@10
|
235 T3z = T3w + T3y;
|
Chris@10
|
236 {
|
Chris@10
|
237 E T4a, T2r, T3F, T2F;
|
Chris@10
|
238 T4a = T2k - T2q;
|
Chris@10
|
239 T2r = T2k + T2q;
|
Chris@10
|
240 T5T = FNMS(KP618033988, T43, T44);
|
Chris@10
|
241 T45 = FMA(KP618033988, T44, T43);
|
Chris@10
|
242 T3F = T3z + T3E;
|
Chris@10
|
243 T48 = T3E - T3z;
|
Chris@10
|
244 T5Q = FNMS(KP618033988, T4a, T4b);
|
Chris@10
|
245 T4c = FMA(KP618033988, T4b, T4a);
|
Chris@10
|
246 T2F = T2r + T2E;
|
Chris@10
|
247 T41 = T2E - T2r;
|
Chris@10
|
248 T3G = T3u + T3F;
|
Chris@10
|
249 T47 = FNMS(KP250000000, T3F, T3u);
|
Chris@10
|
250 T2G = T2e + T2F;
|
Chris@10
|
251 T40 = FNMS(KP250000000, T2F, T2e);
|
Chris@10
|
252 Ty = ci[WS(rs, 1)];
|
Chris@10
|
253 }
|
Chris@10
|
254 T5P = FMA(KP559016994, T48, T47);
|
Chris@10
|
255 T49 = FNMS(KP559016994, T48, T47);
|
Chris@10
|
256 T5S = FMA(KP559016994, T41, T40);
|
Chris@10
|
257 T42 = FNMS(KP559016994, T41, T40);
|
Chris@10
|
258 Tu = W[0];
|
Chris@10
|
259 }
|
Chris@10
|
260 Tx = W[1];
|
Chris@10
|
261 {
|
Chris@10
|
262 E TO, TR, TU, T2T, TP, TX, TT, TQ, TW, T2M, Tw, TN, TA;
|
Chris@10
|
263 TO = cr[WS(rs, 11)];
|
Chris@10
|
264 TR = ci[WS(rs, 11)];
|
Chris@10
|
265 T2M = Tu * Ty;
|
Chris@10
|
266 Tw = Tu * Tv;
|
Chris@10
|
267 TN = W[20];
|
Chris@10
|
268 TU = cr[WS(rs, 16)];
|
Chris@10
|
269 T2N = FNMS(Tx, Tv, T2M);
|
Chris@10
|
270 Tz = FMA(Tx, Ty, Tw);
|
Chris@10
|
271 T2T = TN * TR;
|
Chris@10
|
272 TP = TN * TO;
|
Chris@10
|
273 TX = ci[WS(rs, 16)];
|
Chris@10
|
274 TT = W[30];
|
Chris@10
|
275 TQ = W[21];
|
Chris@10
|
276 TW = W[31];
|
Chris@10
|
277 {
|
Chris@10
|
278 E T2U, TS, T2W, TY, T2V, TV;
|
Chris@10
|
279 TB = cr[WS(rs, 6)];
|
Chris@10
|
280 T2V = TT * TX;
|
Chris@10
|
281 TV = TT * TU;
|
Chris@10
|
282 T2U = FNMS(TQ, TO, T2T);
|
Chris@10
|
283 TS = FMA(TQ, TR, TP);
|
Chris@10
|
284 T2W = FNMS(TW, TU, T2V);
|
Chris@10
|
285 TY = FMA(TW, TX, TV);
|
Chris@10
|
286 TE = ci[WS(rs, 6)];
|
Chris@10
|
287 TA = W[10];
|
Chris@10
|
288 T2X = T2U + T2W;
|
Chris@10
|
289 T4G = T2W - T2U;
|
Chris@10
|
290 T4z = TY - TS;
|
Chris@10
|
291 TZ = TS + TY;
|
Chris@10
|
292 }
|
Chris@10
|
293 TH = cr[WS(rs, 21)];
|
Chris@10
|
294 T2O = TA * TE;
|
Chris@10
|
295 TC = TA * TB;
|
Chris@10
|
296 TK = ci[WS(rs, 21)];
|
Chris@10
|
297 TG = W[40];
|
Chris@10
|
298 TD = W[11];
|
Chris@10
|
299 TJ = W[41];
|
Chris@10
|
300 }
|
Chris@10
|
301 }
|
Chris@10
|
302 T2Q = TG * TK;
|
Chris@10
|
303 TI = TG * TH;
|
Chris@10
|
304 T2P = FNMS(TD, TB, T2O);
|
Chris@10
|
305 TF = FMA(TD, TE, TC);
|
Chris@10
|
306 T2R = FNMS(TJ, TH, T2Q);
|
Chris@10
|
307 TL = FMA(TJ, TK, TI);
|
Chris@10
|
308 }
|
Chris@10
|
309 {
|
Chris@10
|
310 E T31, T17, T3b, T4V, T4O, T1x, T33, T1d, T35, T1j;
|
Chris@10
|
311 {
|
Chris@10
|
312 E T19, T1c, T1f, T32, T1a, T1i, T1e, T1b, T1h, T34, T1g;
|
Chris@10
|
313 {
|
Chris@10
|
314 E T13, T16, T12, T15;
|
Chris@10
|
315 {
|
Chris@10
|
316 E T4w, T4D, T4v, T4C, T4F, T2S;
|
Chris@10
|
317 T13 = cr[WS(rs, 4)];
|
Chris@10
|
318 T4F = T2P - T2R;
|
Chris@10
|
319 T2S = T2P + T2R;
|
Chris@10
|
320 {
|
Chris@10
|
321 E T4y, TM, T2Y, T10;
|
Chris@10
|
322 T4y = TL - TF;
|
Chris@10
|
323 TM = TF + TL;
|
Chris@10
|
324 T65 = FMA(KP618033988, T4F, T4G);
|
Chris@10
|
325 T4H = FNMS(KP618033988, T4G, T4F);
|
Chris@10
|
326 T2Y = T2S + T2X;
|
Chris@10
|
327 T4w = T2S - T2X;
|
Chris@10
|
328 T68 = FNMS(KP618033988, T4y, T4z);
|
Chris@10
|
329 T4A = FMA(KP618033988, T4z, T4y);
|
Chris@10
|
330 T10 = TM + TZ;
|
Chris@10
|
331 T4D = TM - TZ;
|
Chris@10
|
332 T2Z = T2N + T2Y;
|
Chris@10
|
333 T4v = FNMS(KP250000000, T2Y, T2N);
|
Chris@10
|
334 T11 = Tz + T10;
|
Chris@10
|
335 T4C = FNMS(KP250000000, T10, Tz);
|
Chris@10
|
336 T16 = ci[WS(rs, 4)];
|
Chris@10
|
337 }
|
Chris@10
|
338 T67 = FNMS(KP559016994, T4w, T4v);
|
Chris@10
|
339 T4x = FMA(KP559016994, T4w, T4v);
|
Chris@10
|
340 T64 = FNMS(KP559016994, T4D, T4C);
|
Chris@10
|
341 T4E = FMA(KP559016994, T4D, T4C);
|
Chris@10
|
342 T12 = W[6];
|
Chris@10
|
343 }
|
Chris@10
|
344 T15 = W[7];
|
Chris@10
|
345 {
|
Chris@10
|
346 E T1m, T1p, T1s, T37, T1n, T1v, T1r, T1o, T1u, T30, T14, T1l, T18;
|
Chris@10
|
347 T1m = cr[WS(rs, 14)];
|
Chris@10
|
348 T1p = ci[WS(rs, 14)];
|
Chris@10
|
349 T30 = T12 * T16;
|
Chris@10
|
350 T14 = T12 * T13;
|
Chris@10
|
351 T1l = W[26];
|
Chris@10
|
352 T1s = cr[WS(rs, 19)];
|
Chris@10
|
353 T31 = FNMS(T15, T13, T30);
|
Chris@10
|
354 T17 = FMA(T15, T16, T14);
|
Chris@10
|
355 T37 = T1l * T1p;
|
Chris@10
|
356 T1n = T1l * T1m;
|
Chris@10
|
357 T1v = ci[WS(rs, 19)];
|
Chris@10
|
358 T1r = W[36];
|
Chris@10
|
359 T1o = W[27];
|
Chris@10
|
360 T1u = W[37];
|
Chris@10
|
361 {
|
Chris@10
|
362 E T38, T1q, T3a, T1w, T39, T1t;
|
Chris@10
|
363 T19 = cr[WS(rs, 9)];
|
Chris@10
|
364 T39 = T1r * T1v;
|
Chris@10
|
365 T1t = T1r * T1s;
|
Chris@10
|
366 T38 = FNMS(T1o, T1m, T37);
|
Chris@10
|
367 T1q = FMA(T1o, T1p, T1n);
|
Chris@10
|
368 T3a = FNMS(T1u, T1s, T39);
|
Chris@10
|
369 T1w = FMA(T1u, T1v, T1t);
|
Chris@10
|
370 T1c = ci[WS(rs, 9)];
|
Chris@10
|
371 T18 = W[16];
|
Chris@10
|
372 T3b = T38 + T3a;
|
Chris@10
|
373 T4V = T3a - T38;
|
Chris@10
|
374 T4O = T1w - T1q;
|
Chris@10
|
375 T1x = T1q + T1w;
|
Chris@10
|
376 }
|
Chris@10
|
377 T1f = cr[WS(rs, 24)];
|
Chris@10
|
378 T32 = T18 * T1c;
|
Chris@10
|
379 T1a = T18 * T19;
|
Chris@10
|
380 T1i = ci[WS(rs, 24)];
|
Chris@10
|
381 T1e = W[46];
|
Chris@10
|
382 T1b = W[17];
|
Chris@10
|
383 T1h = W[47];
|
Chris@10
|
384 }
|
Chris@10
|
385 }
|
Chris@10
|
386 T34 = T1e * T1i;
|
Chris@10
|
387 T1g = T1e * T1f;
|
Chris@10
|
388 T33 = FNMS(T1b, T19, T32);
|
Chris@10
|
389 T1d = FMA(T1b, T1c, T1a);
|
Chris@10
|
390 T35 = FNMS(T1h, T1f, T34);
|
Chris@10
|
391 T1j = FMA(T1h, T1i, T1g);
|
Chris@10
|
392 }
|
Chris@10
|
393 {
|
Chris@10
|
394 E T1I, T1L, T1O, T3h, T1J, T1R, T1N, T1K, T1Q, T3j, T1P;
|
Chris@10
|
395 {
|
Chris@10
|
396 E T1C, T1F, T1B, T1E;
|
Chris@10
|
397 {
|
Chris@10
|
398 E T4L, T4S, T4K, T4R, T4U, T36;
|
Chris@10
|
399 T1C = cr[WS(rs, 2)];
|
Chris@10
|
400 T4U = T35 - T33;
|
Chris@10
|
401 T36 = T33 + T35;
|
Chris@10
|
402 {
|
Chris@10
|
403 E T4N, T1k, T3c, T1y;
|
Chris@10
|
404 T4N = T1j - T1d;
|
Chris@10
|
405 T1k = T1d + T1j;
|
Chris@10
|
406 T5Y = FNMS(KP618033988, T4U, T4V);
|
Chris@10
|
407 T4W = FMA(KP618033988, T4V, T4U);
|
Chris@10
|
408 T3c = T36 + T3b;
|
Chris@10
|
409 T4L = T3b - T36;
|
Chris@10
|
410 T61 = FNMS(KP618033988, T4N, T4O);
|
Chris@10
|
411 T4P = FMA(KP618033988, T4O, T4N);
|
Chris@10
|
412 T1y = T1k + T1x;
|
Chris@10
|
413 T4S = T1k - T1x;
|
Chris@10
|
414 T3d = T31 + T3c;
|
Chris@10
|
415 T4K = FNMS(KP250000000, T3c, T31);
|
Chris@10
|
416 T1z = T17 + T1y;
|
Chris@10
|
417 T4R = FNMS(KP250000000, T1y, T17);
|
Chris@10
|
418 T1F = ci[WS(rs, 2)];
|
Chris@10
|
419 }
|
Chris@10
|
420 T60 = FMA(KP559016994, T4L, T4K);
|
Chris@10
|
421 T4M = FNMS(KP559016994, T4L, T4K);
|
Chris@10
|
422 T5X = FNMS(KP559016994, T4S, T4R);
|
Chris@10
|
423 T4T = FMA(KP559016994, T4S, T4R);
|
Chris@10
|
424 T1B = W[2];
|
Chris@10
|
425 }
|
Chris@10
|
426 T1E = W[3];
|
Chris@10
|
427 {
|
Chris@10
|
428 E T1V, T1Y, T21, T3m, T1W, T24, T20, T1X, T23, T3f, T1D, T1U, T1H;
|
Chris@10
|
429 T1V = cr[WS(rs, 12)];
|
Chris@10
|
430 T1Y = ci[WS(rs, 12)];
|
Chris@10
|
431 T3f = T1B * T1F;
|
Chris@10
|
432 T1D = T1B * T1C;
|
Chris@10
|
433 T1U = W[22];
|
Chris@10
|
434 T21 = cr[WS(rs, 17)];
|
Chris@10
|
435 T3g = FNMS(T1E, T1C, T3f);
|
Chris@10
|
436 T1G = FMA(T1E, T1F, T1D);
|
Chris@10
|
437 T3m = T1U * T1Y;
|
Chris@10
|
438 T1W = T1U * T1V;
|
Chris@10
|
439 T24 = ci[WS(rs, 17)];
|
Chris@10
|
440 T20 = W[32];
|
Chris@10
|
441 T1X = W[23];
|
Chris@10
|
442 T23 = W[33];
|
Chris@10
|
443 {
|
Chris@10
|
444 E T3n, T1Z, T3p, T25, T3o, T22;
|
Chris@10
|
445 T1I = cr[WS(rs, 7)];
|
Chris@10
|
446 T3o = T20 * T24;
|
Chris@10
|
447 T22 = T20 * T21;
|
Chris@10
|
448 T3n = FNMS(T1X, T1V, T3m);
|
Chris@10
|
449 T1Z = FMA(T1X, T1Y, T1W);
|
Chris@10
|
450 T3p = FNMS(T23, T21, T3o);
|
Chris@10
|
451 T25 = FMA(T23, T24, T22);
|
Chris@10
|
452 T1L = ci[WS(rs, 7)];
|
Chris@10
|
453 T1H = W[12];
|
Chris@10
|
454 T3q = T3n + T3p;
|
Chris@10
|
455 T4q = T3n - T3p;
|
Chris@10
|
456 T4j = T25 - T1Z;
|
Chris@10
|
457 T26 = T1Z + T25;
|
Chris@10
|
458 }
|
Chris@10
|
459 T1O = cr[WS(rs, 22)];
|
Chris@10
|
460 T3h = T1H * T1L;
|
Chris@10
|
461 T1J = T1H * T1I;
|
Chris@10
|
462 T1R = ci[WS(rs, 22)];
|
Chris@10
|
463 T1N = W[42];
|
Chris@10
|
464 T1K = W[13];
|
Chris@10
|
465 T1Q = W[43];
|
Chris@10
|
466 }
|
Chris@10
|
467 }
|
Chris@10
|
468 T3j = T1N * T1R;
|
Chris@10
|
469 T1P = T1N * T1O;
|
Chris@10
|
470 T3i = FNMS(T1K, T1I, T3h);
|
Chris@10
|
471 T1M = FMA(T1K, T1L, T1J);
|
Chris@10
|
472 T3k = FNMS(T1Q, T1O, T3j);
|
Chris@10
|
473 T1S = FMA(T1Q, T1R, T1P);
|
Chris@10
|
474 }
|
Chris@10
|
475 }
|
Chris@10
|
476 }
|
Chris@10
|
477 }
|
Chris@10
|
478 {
|
Chris@10
|
479 E T7Q, T5M, T5J, T7R, T5I, T5L, T7X, T7W, T5F, T6X, T5u, T7M, T7O, T5C, T5E;
|
Chris@10
|
480 E T5t, T7J, T7N;
|
Chris@10
|
481 {
|
Chris@10
|
482 E T4r, T4k, T4h, T4o, T3K, T3I, T1A, T2H, T28;
|
Chris@10
|
483 {
|
Chris@10
|
484 E T3e, T4g, T4n, T4f, T4m, T3H, T4p, T3l;
|
Chris@10
|
485 T7Q = T2Z + T3d;
|
Chris@10
|
486 T3e = T2Z - T3d;
|
Chris@10
|
487 T4p = T3k - T3i;
|
Chris@10
|
488 T3l = T3i + T3k;
|
Chris@10
|
489 {
|
Chris@10
|
490 E T4i, T1T, T3r, T27, T3s;
|
Chris@10
|
491 T4i = T1S - T1M;
|
Chris@10
|
492 T1T = T1M + T1S;
|
Chris@10
|
493 T5M = FMA(KP618033988, T4p, T4q);
|
Chris@10
|
494 T4r = FNMS(KP618033988, T4q, T4p);
|
Chris@10
|
495 T3r = T3l + T3q;
|
Chris@10
|
496 T4g = T3q - T3l;
|
Chris@10
|
497 T5J = FNMS(KP618033988, T4i, T4j);
|
Chris@10
|
498 T4k = FMA(KP618033988, T4j, T4i);
|
Chris@10
|
499 T27 = T1T + T26;
|
Chris@10
|
500 T4n = T26 - T1T;
|
Chris@10
|
501 T3s = T3g + T3r;
|
Chris@10
|
502 T4f = FNMS(KP250000000, T3r, T3g);
|
Chris@10
|
503 T28 = T1G + T27;
|
Chris@10
|
504 T4m = FNMS(KP250000000, T27, T1G);
|
Chris@10
|
505 T3H = T3s - T3G;
|
Chris@10
|
506 T7R = T3s + T3G;
|
Chris@10
|
507 }
|
Chris@10
|
508 T5I = FMA(KP559016994, T4g, T4f);
|
Chris@10
|
509 T4h = FNMS(KP559016994, T4g, T4f);
|
Chris@10
|
510 T5L = FMA(KP559016994, T4n, T4m);
|
Chris@10
|
511 T4o = FNMS(KP559016994, T4n, T4m);
|
Chris@10
|
512 T3K = FNMS(KP618033988, T3e, T3H);
|
Chris@10
|
513 T3I = FMA(KP618033988, T3H, T3e);
|
Chris@10
|
514 }
|
Chris@10
|
515 T1A = T11 + T1z;
|
Chris@10
|
516 T7X = T1z - T11;
|
Chris@10
|
517 T7W = T28 - T2G;
|
Chris@10
|
518 T2H = T28 + T2G;
|
Chris@10
|
519 {
|
Chris@10
|
520 E T3Z, T5d, T7r, T7D, T5h, T5i, T5m, T5l, T59, T7K, T56, T7L, T7I, T7G, T52;
|
Chris@10
|
521 E T50, T5w, T5g, T5q, T5A, T3N, T7p;
|
Chris@10
|
522 T3N = FMA(KP559016994, T3M, T3L);
|
Chris@10
|
523 T5F = FNMS(KP559016994, T3M, T3L);
|
Chris@10
|
524 T6X = FNMS(KP559016994, T6W, T6V);
|
Chris@10
|
525 T7p = FMA(KP559016994, T6W, T6V);
|
Chris@10
|
526 {
|
Chris@10
|
527 E T5o, T5p, T57, T4e, T4Y, T55, T4l, T4s, T4B, T5f, T5e, T4I;
|
Chris@10
|
528 {
|
Chris@10
|
529 E T46, T2K, T2J, T4d, T2I;
|
Chris@10
|
530 T46 = FMA(KP951056516, T45, T42);
|
Chris@10
|
531 T5o = FNMS(KP951056516, T45, T42);
|
Chris@10
|
532 T2I = T1A + T2H;
|
Chris@10
|
533 T2K = T1A - T2H;
|
Chris@10
|
534 T3Z = FNMS(KP951056516, T3Y, T3N);
|
Chris@10
|
535 T5d = FMA(KP951056516, T3Y, T3N);
|
Chris@10
|
536 T7r = FNMS(KP951056516, T7q, T7p);
|
Chris@10
|
537 T7D = FMA(KP951056516, T7q, T7p);
|
Chris@10
|
538 cr[0] = Tt + T2I;
|
Chris@10
|
539 T2J = FNMS(KP250000000, T2I, Tt);
|
Chris@10
|
540 T5p = FNMS(KP951056516, T4c, T49);
|
Chris@10
|
541 T4d = FMA(KP951056516, T4c, T49);
|
Chris@10
|
542 {
|
Chris@10
|
543 E T4Q, T4X, T2L, T3J;
|
Chris@10
|
544 T4Q = FNMS(KP951056516, T4P, T4M);
|
Chris@10
|
545 T5h = FMA(KP951056516, T4P, T4M);
|
Chris@10
|
546 T5i = FNMS(KP951056516, T4W, T4T);
|
Chris@10
|
547 T4X = FMA(KP951056516, T4W, T4T);
|
Chris@10
|
548 T2L = FMA(KP559016994, T2K, T2J);
|
Chris@10
|
549 T3J = FNMS(KP559016994, T2K, T2J);
|
Chris@10
|
550 T57 = FMA(KP126329378, T46, T4d);
|
Chris@10
|
551 T4e = FNMS(KP126329378, T4d, T46);
|
Chris@10
|
552 cr[WS(rs, 5)] = FMA(KP951056516, T3I, T2L);
|
Chris@10
|
553 ci[WS(rs, 4)] = FNMS(KP951056516, T3I, T2L);
|
Chris@10
|
554 ci[WS(rs, 9)] = FMA(KP951056516, T3K, T3J);
|
Chris@10
|
555 cr[WS(rs, 10)] = FNMS(KP951056516, T3K, T3J);
|
Chris@10
|
556 T4Y = FMA(KP827271945, T4X, T4Q);
|
Chris@10
|
557 T55 = FNMS(KP827271945, T4Q, T4X);
|
Chris@10
|
558 }
|
Chris@10
|
559 }
|
Chris@10
|
560 T4l = FNMS(KP951056516, T4k, T4h);
|
Chris@10
|
561 T5m = FMA(KP951056516, T4k, T4h);
|
Chris@10
|
562 T5l = FNMS(KP951056516, T4r, T4o);
|
Chris@10
|
563 T4s = FMA(KP951056516, T4r, T4o);
|
Chris@10
|
564 T4B = FNMS(KP951056516, T4A, T4x);
|
Chris@10
|
565 T5f = FMA(KP951056516, T4A, T4x);
|
Chris@10
|
566 T5e = FMA(KP951056516, T4H, T4E);
|
Chris@10
|
567 T4I = FNMS(KP951056516, T4H, T4E);
|
Chris@10
|
568 {
|
Chris@10
|
569 E T4u, T4Z, T4t, T58;
|
Chris@10
|
570 T4t = FNMS(KP470564281, T4s, T4l);
|
Chris@10
|
571 T58 = FMA(KP470564281, T4l, T4s);
|
Chris@10
|
572 {
|
Chris@10
|
573 E T4J, T54, T7E, T7F;
|
Chris@10
|
574 T4J = FMA(KP634619297, T4I, T4B);
|
Chris@10
|
575 T54 = FNMS(KP634619297, T4B, T4I);
|
Chris@10
|
576 T59 = FNMS(KP912018591, T58, T57);
|
Chris@10
|
577 T7E = FMA(KP912018591, T58, T57);
|
Chris@10
|
578 T7K = FMA(KP912018591, T4t, T4e);
|
Chris@10
|
579 T4u = FNMS(KP912018591, T4t, T4e);
|
Chris@10
|
580 T56 = FMA(KP912575812, T55, T54);
|
Chris@10
|
581 T7F = FNMS(KP912575812, T55, T54);
|
Chris@10
|
582 T7L = FMA(KP912575812, T4Y, T4J);
|
Chris@10
|
583 T4Z = FNMS(KP912575812, T4Y, T4J);
|
Chris@10
|
584 T7I = FNMS(KP851038619, T7F, T7E);
|
Chris@10
|
585 T7G = FMA(KP851038619, T7F, T7E);
|
Chris@10
|
586 }
|
Chris@10
|
587 T52 = FMA(KP851038619, T4Z, T4u);
|
Chris@10
|
588 T50 = FNMS(KP851038619, T4Z, T4u);
|
Chris@10
|
589 }
|
Chris@10
|
590 T5w = FNMS(KP256756360, T5e, T5f);
|
Chris@10
|
591 T5g = FMA(KP256756360, T5f, T5e);
|
Chris@10
|
592 T5q = FMA(KP939062505, T5p, T5o);
|
Chris@10
|
593 T5A = FNMS(KP939062505, T5o, T5p);
|
Chris@10
|
594 }
|
Chris@10
|
595 {
|
Chris@10
|
596 E T5y, T7z, T5B, T7y, T7w, T7u, T5s;
|
Chris@10
|
597 {
|
Chris@10
|
598 E T5k, T5r, T5j, T5x;
|
Chris@10
|
599 cr[WS(rs, 4)] = FNMS(KP992114701, T50, T3Z);
|
Chris@10
|
600 T5j = FMA(KP634619297, T5i, T5h);
|
Chris@10
|
601 T5x = FNMS(KP634619297, T5h, T5i);
|
Chris@10
|
602 {
|
Chris@10
|
603 E T5n, T5z, T7s, T7t;
|
Chris@10
|
604 T5n = FMA(KP549754652, T5m, T5l);
|
Chris@10
|
605 T5z = FNMS(KP549754652, T5l, T5m);
|
Chris@10
|
606 T5y = FMA(KP871714437, T5x, T5w);
|
Chris@10
|
607 T7s = FNMS(KP871714437, T5x, T5w);
|
Chris@10
|
608 T7z = FNMS(KP871714437, T5j, T5g);
|
Chris@10
|
609 T5k = FMA(KP871714437, T5j, T5g);
|
Chris@10
|
610 T5B = FNMS(KP831864738, T5A, T5z);
|
Chris@10
|
611 T7t = FMA(KP831864738, T5A, T5z);
|
Chris@10
|
612 T7y = FNMS(KP831864738, T5q, T5n);
|
Chris@10
|
613 T5r = FMA(KP831864738, T5q, T5n);
|
Chris@10
|
614 T7w = FNMS(KP904730450, T7t, T7s);
|
Chris@10
|
615 T7u = FMA(KP904730450, T7t, T7s);
|
Chris@10
|
616 }
|
Chris@10
|
617 ci[WS(rs, 20)] = FNMS(KP992114701, T7G, T7D);
|
Chris@10
|
618 T5u = FNMS(KP904730450, T5r, T5k);
|
Chris@10
|
619 T5s = FMA(KP904730450, T5r, T5k);
|
Chris@10
|
620 }
|
Chris@10
|
621 {
|
Chris@10
|
622 E T5a, T5c, T7A, T7C, T7v, T53, T5b, T51, T7H, T7x, T7B;
|
Chris@10
|
623 T5a = FNMS(KP726211448, T59, T56);
|
Chris@10
|
624 T5c = FMA(KP525970792, T56, T59);
|
Chris@10
|
625 ci[WS(rs, 23)] = FMA(KP968583161, T7u, T7r);
|
Chris@10
|
626 cr[WS(rs, 1)] = FMA(KP968583161, T5s, T5d);
|
Chris@10
|
627 T51 = FMA(KP248028675, T50, T3Z);
|
Chris@10
|
628 T7A = FNMS(KP683113946, T7z, T7y);
|
Chris@10
|
629 T7C = FMA(KP559154169, T7y, T7z);
|
Chris@10
|
630 T7v = FNMS(KP242145790, T7u, T7r);
|
Chris@10
|
631 T53 = FMA(KP554608978, T52, T51);
|
Chris@10
|
632 T5b = FNMS(KP554608978, T52, T51);
|
Chris@10
|
633 T7M = FNMS(KP525970792, T7L, T7K);
|
Chris@10
|
634 T7O = FMA(KP726211448, T7K, T7L);
|
Chris@10
|
635 ci[WS(rs, 10)] = FNMS(KP943557151, T5c, T5b);
|
Chris@10
|
636 ci[WS(rs, 5)] = FMA(KP943557151, T5c, T5b);
|
Chris@10
|
637 ci[0] = FMA(KP803003575, T5a, T53);
|
Chris@10
|
638 cr[WS(rs, 9)] = FNMS(KP803003575, T5a, T53);
|
Chris@10
|
639 T7x = FNMS(KP541454447, T7w, T7v);
|
Chris@10
|
640 T7B = FMA(KP541454447, T7w, T7v);
|
Chris@10
|
641 T7H = FMA(KP248028675, T7G, T7D);
|
Chris@10
|
642 cr[WS(rs, 21)] = -(FMA(KP921177326, T7C, T7B));
|
Chris@10
|
643 ci[WS(rs, 18)] = FNMS(KP921177326, T7C, T7B);
|
Chris@10
|
644 ci[WS(rs, 13)] = FMA(KP833417178, T7A, T7x);
|
Chris@10
|
645 cr[WS(rs, 16)] = FMS(KP833417178, T7A, T7x);
|
Chris@10
|
646 T5C = FMA(KP559154169, T5B, T5y);
|
Chris@10
|
647 T5E = FNMS(KP683113946, T5y, T5B);
|
Chris@10
|
648 T5t = FNMS(KP242145790, T5s, T5d);
|
Chris@10
|
649 T7J = FNMS(KP554608978, T7I, T7H);
|
Chris@10
|
650 T7N = FMA(KP554608978, T7I, T7H);
|
Chris@10
|
651 }
|
Chris@10
|
652 }
|
Chris@10
|
653 }
|
Chris@10
|
654 }
|
Chris@10
|
655 {
|
Chris@10
|
656 E T7Y, T80, T5v, T5D;
|
Chris@10
|
657 cr[WS(rs, 24)] = -(FMA(KP803003575, T7O, T7N));
|
Chris@10
|
658 ci[WS(rs, 15)] = FNMS(KP803003575, T7O, T7N);
|
Chris@10
|
659 cr[WS(rs, 19)] = FMS(KP943557151, T7M, T7J);
|
Chris@10
|
660 cr[WS(rs, 14)] = -(FMA(KP943557151, T7M, T7J));
|
Chris@10
|
661 T5v = FMA(KP541454447, T5u, T5t);
|
Chris@10
|
662 T5D = FNMS(KP541454447, T5u, T5t);
|
Chris@10
|
663 cr[WS(rs, 11)] = FNMS(KP833417178, T5E, T5D);
|
Chris@10
|
664 ci[WS(rs, 8)] = FMA(KP833417178, T5E, T5D);
|
Chris@10
|
665 cr[WS(rs, 6)] = FMA(KP921177326, T5C, T5v);
|
Chris@10
|
666 ci[WS(rs, 3)] = FNMS(KP921177326, T5C, T5v);
|
Chris@10
|
667 T7Y = FMA(KP618033988, T7X, T7W);
|
Chris@10
|
668 T80 = FNMS(KP618033988, T7W, T7X);
|
Chris@10
|
669 {
|
Chris@10
|
670 E T6t, T6p, T5H, T7d, T71, T6u, T6y, T6x, T6l, T7k, T6i, T7l, T7g, T6c, T6e;
|
Chris@10
|
671 E T6s, T6L, T6J, T6C;
|
Chris@10
|
672 {
|
Chris@10
|
673 E T6A, T6B, T5O, T6j, T6h, T6a, T6q, T5R, T5U, T6r, T5Z, T62;
|
Chris@10
|
674 {
|
Chris@10
|
675 E T5K, T7U, T7T, T5N, T7S;
|
Chris@10
|
676 T6t = FNMS(KP951056516, T5J, T5I);
|
Chris@10
|
677 T5K = FMA(KP951056516, T5J, T5I);
|
Chris@10
|
678 T7U = T7Q - T7R;
|
Chris@10
|
679 T7S = T7Q + T7R;
|
Chris@10
|
680 T6p = FNMS(KP951056516, T5G, T5F);
|
Chris@10
|
681 T5H = FMA(KP951056516, T5G, T5F);
|
Chris@10
|
682 T7d = FNMS(KP951056516, T70, T6X);
|
Chris@10
|
683 T71 = FMA(KP951056516, T70, T6X);
|
Chris@10
|
684 ci[WS(rs, 24)] = T7S + T7P;
|
Chris@10
|
685 T7T = FNMS(KP250000000, T7S, T7P);
|
Chris@10
|
686 T5N = FMA(KP951056516, T5M, T5L);
|
Chris@10
|
687 T6u = FNMS(KP951056516, T5M, T5L);
|
Chris@10
|
688 {
|
Chris@10
|
689 E T66, T69, T7Z, T7V;
|
Chris@10
|
690 T6A = FMA(KP951056516, T65, T64);
|
Chris@10
|
691 T66 = FNMS(KP951056516, T65, T64);
|
Chris@10
|
692 T69 = FMA(KP951056516, T68, T67);
|
Chris@10
|
693 T6B = FNMS(KP951056516, T68, T67);
|
Chris@10
|
694 T7Z = FMA(KP559016994, T7U, T7T);
|
Chris@10
|
695 T7V = FNMS(KP559016994, T7U, T7T);
|
Chris@10
|
696 T5O = FMA(KP062914667, T5N, T5K);
|
Chris@10
|
697 T6j = FNMS(KP062914667, T5K, T5N);
|
Chris@10
|
698 ci[WS(rs, 14)] = FMA(KP951056516, T7Y, T7V);
|
Chris@10
|
699 cr[WS(rs, 15)] = FMS(KP951056516, T7Y, T7V);
|
Chris@10
|
700 ci[WS(rs, 19)] = FMA(KP951056516, T80, T7Z);
|
Chris@10
|
701 cr[WS(rs, 20)] = FMS(KP951056516, T80, T7Z);
|
Chris@10
|
702 T6h = FNMS(KP939062505, T66, T69);
|
Chris@10
|
703 T6a = FMA(KP939062505, T69, T66);
|
Chris@10
|
704 }
|
Chris@10
|
705 }
|
Chris@10
|
706 T6q = FMA(KP951056516, T5Q, T5P);
|
Chris@10
|
707 T5R = FNMS(KP951056516, T5Q, T5P);
|
Chris@10
|
708 T5U = FNMS(KP951056516, T5T, T5S);
|
Chris@10
|
709 T6r = FMA(KP951056516, T5T, T5S);
|
Chris@10
|
710 T6y = FMA(KP951056516, T5Y, T5X);
|
Chris@10
|
711 T5Z = FNMS(KP951056516, T5Y, T5X);
|
Chris@10
|
712 T62 = FMA(KP951056516, T61, T60);
|
Chris@10
|
713 T6x = FNMS(KP951056516, T61, T60);
|
Chris@10
|
714 {
|
Chris@10
|
715 E T5W, T6b, T6k, T5V;
|
Chris@10
|
716 T6k = FMA(KP827271945, T5R, T5U);
|
Chris@10
|
717 T5V = FNMS(KP827271945, T5U, T5R);
|
Chris@10
|
718 {
|
Chris@10
|
719 E T6g, T63, T7e, T7f;
|
Chris@10
|
720 T6g = FMA(KP126329378, T5Z, T62);
|
Chris@10
|
721 T63 = FNMS(KP126329378, T62, T5Z);
|
Chris@10
|
722 T7e = FMA(KP772036680, T6k, T6j);
|
Chris@10
|
723 T6l = FNMS(KP772036680, T6k, T6j);
|
Chris@10
|
724 T5W = FMA(KP772036680, T5V, T5O);
|
Chris@10
|
725 T7k = FNMS(KP772036680, T5V, T5O);
|
Chris@10
|
726 T7f = FNMS(KP734762448, T6h, T6g);
|
Chris@10
|
727 T6i = FMA(KP734762448, T6h, T6g);
|
Chris@10
|
728 T6b = FNMS(KP734762448, T6a, T63);
|
Chris@10
|
729 T7l = FMA(KP734762448, T6a, T63);
|
Chris@10
|
730 T7g = FMA(KP994076283, T7f, T7e);
|
Chris@10
|
731 T7i = FNMS(KP994076283, T7f, T7e);
|
Chris@10
|
732 }
|
Chris@10
|
733 T6c = FNMS(KP994076283, T6b, T5W);
|
Chris@10
|
734 T6e = FMA(KP994076283, T6b, T5W);
|
Chris@10
|
735 }
|
Chris@10
|
736 T6s = FMA(KP062914667, T6r, T6q);
|
Chris@10
|
737 T6L = FNMS(KP062914667, T6q, T6r);
|
Chris@10
|
738 T6J = FNMS(KP549754652, T6A, T6B);
|
Chris@10
|
739 T6C = FMA(KP549754652, T6B, T6A);
|
Chris@10
|
740 }
|
Chris@10
|
741 {
|
Chris@10
|
742 E T6N, T78, T6K, T79, T74, T76, T6E, T6G;
|
Chris@10
|
743 {
|
Chris@10
|
744 E T6w, T6D, T6M, T6v;
|
Chris@10
|
745 cr[WS(rs, 3)] = FMA(KP998026728, T6c, T5H);
|
Chris@10
|
746 T6M = FNMS(KP634619297, T6t, T6u);
|
Chris@10
|
747 T6v = FMA(KP634619297, T6u, T6t);
|
Chris@10
|
748 {
|
Chris@10
|
749 E T6I, T6z, T72, T73;
|
Chris@10
|
750 T6I = FMA(KP470564281, T6x, T6y);
|
Chris@10
|
751 T6z = FNMS(KP470564281, T6y, T6x);
|
Chris@10
|
752 T72 = FMA(KP845997307, T6M, T6L);
|
Chris@10
|
753 T6N = FNMS(KP845997307, T6M, T6L);
|
Chris@10
|
754 T6w = FMA(KP845997307, T6v, T6s);
|
Chris@10
|
755 T78 = FNMS(KP845997307, T6v, T6s);
|
Chris@10
|
756 T73 = FNMS(KP968479752, T6J, T6I);
|
Chris@10
|
757 T6K = FMA(KP968479752, T6J, T6I);
|
Chris@10
|
758 T6D = FMA(KP968479752, T6C, T6z);
|
Chris@10
|
759 T79 = FNMS(KP968479752, T6C, T6z);
|
Chris@10
|
760 T74 = FMA(KP906616052, T73, T72);
|
Chris@10
|
761 T76 = FNMS(KP906616052, T73, T72);
|
Chris@10
|
762 }
|
Chris@10
|
763 ci[WS(rs, 21)] = FNMS(KP998026728, T7g, T7d);
|
Chris@10
|
764 T6E = FMA(KP906616052, T6D, T6w);
|
Chris@10
|
765 T6G = FNMS(KP906616052, T6D, T6w);
|
Chris@10
|
766 }
|
Chris@10
|
767 {
|
Chris@10
|
768 E T7c, T7a, T6Q, T6O, T6F, T7b, T77, T75, T6d, T6P, T6H;
|
Chris@10
|
769 T7c = FMA(KP681693190, T78, T79);
|
Chris@10
|
770 T7a = FNMS(KP560319534, T79, T78);
|
Chris@10
|
771 ci[WS(rs, 22)] = FNMS(KP998026728, T74, T71);
|
Chris@10
|
772 cr[WS(rs, 2)] = FMA(KP998026728, T6E, T6p);
|
Chris@10
|
773 T75 = FMA(KP249506682, T74, T71);
|
Chris@10
|
774 T6Q = FNMS(KP560319534, T6K, T6N);
|
Chris@10
|
775 T6O = FMA(KP681693190, T6N, T6K);
|
Chris@10
|
776 T6F = FNMS(KP249506682, T6E, T6p);
|
Chris@10
|
777 T7b = FMA(KP557913902, T76, T75);
|
Chris@10
|
778 T77 = FNMS(KP557913902, T76, T75);
|
Chris@10
|
779 T6o = FMA(KP614372930, T6i, T6l);
|
Chris@10
|
780 T6m = FNMS(KP621716863, T6l, T6i);
|
Chris@10
|
781 cr[WS(rs, 22)] = FMS(KP860541664, T7c, T7b);
|
Chris@10
|
782 ci[WS(rs, 17)] = FMA(KP860541664, T7c, T7b);
|
Chris@10
|
783 ci[WS(rs, 12)] = FNMS(KP949179823, T7a, T77);
|
Chris@10
|
784 cr[WS(rs, 17)] = -(FMA(KP949179823, T7a, T77));
|
Chris@10
|
785 T6P = FMA(KP557913902, T6G, T6F);
|
Chris@10
|
786 T6H = FNMS(KP557913902, T6G, T6F);
|
Chris@10
|
787 T6d = FNMS(KP249506682, T6c, T5H);
|
Chris@10
|
788 ci[WS(rs, 7)] = FMA(KP949179823, T6Q, T6P);
|
Chris@10
|
789 cr[WS(rs, 12)] = FNMS(KP949179823, T6Q, T6P);
|
Chris@10
|
790 cr[WS(rs, 7)] = FMA(KP860541664, T6O, T6H);
|
Chris@10
|
791 ci[WS(rs, 2)] = FNMS(KP860541664, T6O, T6H);
|
Chris@10
|
792 T7o = FMA(KP621716863, T7k, T7l);
|
Chris@10
|
793 T7m = FNMS(KP614372930, T7l, T7k);
|
Chris@10
|
794 T7h = FMA(KP249506682, T7g, T7d);
|
Chris@10
|
795 T6n = FMA(KP557913902, T6e, T6d);
|
Chris@10
|
796 T6f = FNMS(KP557913902, T6e, T6d);
|
Chris@10
|
797 }
|
Chris@10
|
798 }
|
Chris@10
|
799 }
|
Chris@10
|
800 }
|
Chris@10
|
801 }
|
Chris@10
|
802 }
|
Chris@10
|
803 ci[WS(rs, 6)] = FNMS(KP949179823, T6o, T6n);
|
Chris@10
|
804 ci[WS(rs, 11)] = FMA(KP949179823, T6o, T6n);
|
Chris@10
|
805 cr[WS(rs, 8)] = FMA(KP943557151, T6m, T6f);
|
Chris@10
|
806 ci[WS(rs, 1)] = FNMS(KP943557151, T6m, T6f);
|
Chris@10
|
807 T7j = FNMS(KP557913902, T7i, T7h);
|
Chris@10
|
808 T7n = FMA(KP557913902, T7i, T7h);
|
Chris@10
|
809 cr[WS(rs, 23)] = -(FMA(KP943557151, T7o, T7n));
|
Chris@10
|
810 ci[WS(rs, 16)] = FNMS(KP943557151, T7o, T7n);
|
Chris@10
|
811 cr[WS(rs, 18)] = FMS(KP949179823, T7m, T7j);
|
Chris@10
|
812 cr[WS(rs, 13)] = -(FMA(KP949179823, T7m, T7j));
|
Chris@10
|
813 }
|
Chris@10
|
814 }
|
Chris@10
|
815 }
|
Chris@10
|
816
|
Chris@10
|
817 static const tw_instr twinstr[] = {
|
Chris@10
|
818 {TW_FULL, 1, 25},
|
Chris@10
|
819 {TW_NEXT, 1, 0}
|
Chris@10
|
820 };
|
Chris@10
|
821
|
Chris@10
|
822 static const hc2hc_desc desc = { 25, "hf_25", twinstr, &GENUS, {84, 48, 316, 0} };
|
Chris@10
|
823
|
Chris@10
|
824 void X(codelet_hf_25) (planner *p) {
|
Chris@10
|
825 X(khc2hc_register) (p, hf_25, &desc);
|
Chris@10
|
826 }
|
Chris@10
|
827 #else /* HAVE_FMA */
|
Chris@10
|
828
|
Chris@10
|
829 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 25 -dit -name hf_25 -include hf.h */
|
Chris@10
|
830
|
Chris@10
|
831 /*
|
Chris@10
|
832 * This function contains 400 FP additions, 280 FP multiplications,
|
Chris@10
|
833 * (or, 260 additions, 140 multiplications, 140 fused multiply/add),
|
Chris@10
|
834 * 101 stack variables, 20 constants, and 100 memory accesses
|
Chris@10
|
835 */
|
Chris@10
|
836 #include "hf.h"
|
Chris@10
|
837
|
Chris@10
|
838 static void hf_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
839 {
|
Chris@10
|
840 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
841 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
842 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
843 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
844 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
845 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
846 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
847 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
848 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
Chris@10
|
849 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
Chris@10
|
850 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
851 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
852 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
Chris@10
|
853 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
Chris@10
|
854 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
855 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
856 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
857 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
858 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
859 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
860 {
|
Chris@10
|
861 INT m;
|
Chris@10
|
862 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
|
863 E T1, T6b, T2l, T6g, To, T2m, T6e, T6f, T6a, T6H, T2u, T4I, T2i, T60, T3S;
|
Chris@10
|
864 E T5D, T4r, T58, T3Z, T5C, T4q, T5b, TS, T5W, T2G, T5s, T4g, T4M, T2R, T5t;
|
Chris@10
|
865 E T4h, T4P, T1l, T5X, T37, T5v, T4k, T4T, T3e, T5w, T4j, T4W, T1P, T5Z, T3v;
|
Chris@10
|
866 E T5A, T4o, T54, T3C, T5z, T4n, T51;
|
Chris@10
|
867 {
|
Chris@10
|
868 E T6, T2o, Tb, T2p, Tc, T6c, Th, T2r, Tm, T2s, Tn, T6d;
|
Chris@10
|
869 T1 = cr[0];
|
Chris@10
|
870 T6b = ci[0];
|
Chris@10
|
871 {
|
Chris@10
|
872 E T3, T5, T2, T4;
|
Chris@10
|
873 T3 = cr[WS(rs, 5)];
|
Chris@10
|
874 T5 = ci[WS(rs, 5)];
|
Chris@10
|
875 T2 = W[8];
|
Chris@10
|
876 T4 = W[9];
|
Chris@10
|
877 T6 = FMA(T2, T3, T4 * T5);
|
Chris@10
|
878 T2o = FNMS(T4, T3, T2 * T5);
|
Chris@10
|
879 }
|
Chris@10
|
880 {
|
Chris@10
|
881 E T8, Ta, T7, T9;
|
Chris@10
|
882 T8 = cr[WS(rs, 20)];
|
Chris@10
|
883 Ta = ci[WS(rs, 20)];
|
Chris@10
|
884 T7 = W[38];
|
Chris@10
|
885 T9 = W[39];
|
Chris@10
|
886 Tb = FMA(T7, T8, T9 * Ta);
|
Chris@10
|
887 T2p = FNMS(T9, T8, T7 * Ta);
|
Chris@10
|
888 }
|
Chris@10
|
889 Tc = T6 + Tb;
|
Chris@10
|
890 T6c = T2o + T2p;
|
Chris@10
|
891 {
|
Chris@10
|
892 E Te, Tg, Td, Tf;
|
Chris@10
|
893 Te = cr[WS(rs, 10)];
|
Chris@10
|
894 Tg = ci[WS(rs, 10)];
|
Chris@10
|
895 Td = W[18];
|
Chris@10
|
896 Tf = W[19];
|
Chris@10
|
897 Th = FMA(Td, Te, Tf * Tg);
|
Chris@10
|
898 T2r = FNMS(Tf, Te, Td * Tg);
|
Chris@10
|
899 }
|
Chris@10
|
900 {
|
Chris@10
|
901 E Tj, Tl, Ti, Tk;
|
Chris@10
|
902 Tj = cr[WS(rs, 15)];
|
Chris@10
|
903 Tl = ci[WS(rs, 15)];
|
Chris@10
|
904 Ti = W[28];
|
Chris@10
|
905 Tk = W[29];
|
Chris@10
|
906 Tm = FMA(Ti, Tj, Tk * Tl);
|
Chris@10
|
907 T2s = FNMS(Tk, Tj, Ti * Tl);
|
Chris@10
|
908 }
|
Chris@10
|
909 Tn = Th + Tm;
|
Chris@10
|
910 T6d = T2r + T2s;
|
Chris@10
|
911 T2l = KP559016994 * (Tc - Tn);
|
Chris@10
|
912 T6g = KP559016994 * (T6c - T6d);
|
Chris@10
|
913 To = Tc + Tn;
|
Chris@10
|
914 T2m = FNMS(KP250000000, To, T1);
|
Chris@10
|
915 T6e = T6c + T6d;
|
Chris@10
|
916 T6f = FNMS(KP250000000, T6e, T6b);
|
Chris@10
|
917 {
|
Chris@10
|
918 E T68, T69, T2q, T2t;
|
Chris@10
|
919 T68 = Th - Tm;
|
Chris@10
|
920 T69 = T6 - Tb;
|
Chris@10
|
921 T6a = FNMS(KP587785252, T69, KP951056516 * T68);
|
Chris@10
|
922 T6H = FMA(KP951056516, T69, KP587785252 * T68);
|
Chris@10
|
923 T2q = T2o - T2p;
|
Chris@10
|
924 T2t = T2r - T2s;
|
Chris@10
|
925 T2u = FMA(KP951056516, T2q, KP587785252 * T2t);
|
Chris@10
|
926 T4I = FNMS(KP587785252, T2q, KP951056516 * T2t);
|
Chris@10
|
927 }
|
Chris@10
|
928 }
|
Chris@10
|
929 {
|
Chris@10
|
930 E T1U, T3O, T3E, T3F, T3X, T3W, T3J, T3M, T3P, T25, T2g, T2h;
|
Chris@10
|
931 {
|
Chris@10
|
932 E T1R, T1T, T1Q, T1S;
|
Chris@10
|
933 T1R = cr[WS(rs, 3)];
|
Chris@10
|
934 T1T = ci[WS(rs, 3)];
|
Chris@10
|
935 T1Q = W[4];
|
Chris@10
|
936 T1S = W[5];
|
Chris@10
|
937 T1U = FMA(T1Q, T1R, T1S * T1T);
|
Chris@10
|
938 T3O = FNMS(T1S, T1R, T1Q * T1T);
|
Chris@10
|
939 }
|
Chris@10
|
940 {
|
Chris@10
|
941 E T1Z, T3H, T2f, T3L, T24, T3I, T2a, T3K;
|
Chris@10
|
942 {
|
Chris@10
|
943 E T1W, T1Y, T1V, T1X;
|
Chris@10
|
944 T1W = cr[WS(rs, 8)];
|
Chris@10
|
945 T1Y = ci[WS(rs, 8)];
|
Chris@10
|
946 T1V = W[14];
|
Chris@10
|
947 T1X = W[15];
|
Chris@10
|
948 T1Z = FMA(T1V, T1W, T1X * T1Y);
|
Chris@10
|
949 T3H = FNMS(T1X, T1W, T1V * T1Y);
|
Chris@10
|
950 }
|
Chris@10
|
951 {
|
Chris@10
|
952 E T2c, T2e, T2b, T2d;
|
Chris@10
|
953 T2c = cr[WS(rs, 18)];
|
Chris@10
|
954 T2e = ci[WS(rs, 18)];
|
Chris@10
|
955 T2b = W[34];
|
Chris@10
|
956 T2d = W[35];
|
Chris@10
|
957 T2f = FMA(T2b, T2c, T2d * T2e);
|
Chris@10
|
958 T3L = FNMS(T2d, T2c, T2b * T2e);
|
Chris@10
|
959 }
|
Chris@10
|
960 {
|
Chris@10
|
961 E T21, T23, T20, T22;
|
Chris@10
|
962 T21 = cr[WS(rs, 23)];
|
Chris@10
|
963 T23 = ci[WS(rs, 23)];
|
Chris@10
|
964 T20 = W[44];
|
Chris@10
|
965 T22 = W[45];
|
Chris@10
|
966 T24 = FMA(T20, T21, T22 * T23);
|
Chris@10
|
967 T3I = FNMS(T22, T21, T20 * T23);
|
Chris@10
|
968 }
|
Chris@10
|
969 {
|
Chris@10
|
970 E T27, T29, T26, T28;
|
Chris@10
|
971 T27 = cr[WS(rs, 13)];
|
Chris@10
|
972 T29 = ci[WS(rs, 13)];
|
Chris@10
|
973 T26 = W[24];
|
Chris@10
|
974 T28 = W[25];
|
Chris@10
|
975 T2a = FMA(T26, T27, T28 * T29);
|
Chris@10
|
976 T3K = FNMS(T28, T27, T26 * T29);
|
Chris@10
|
977 }
|
Chris@10
|
978 T3E = T1Z - T24;
|
Chris@10
|
979 T3F = T2a - T2f;
|
Chris@10
|
980 T3X = T3K - T3L;
|
Chris@10
|
981 T3W = T3H - T3I;
|
Chris@10
|
982 T3J = T3H + T3I;
|
Chris@10
|
983 T3M = T3K + T3L;
|
Chris@10
|
984 T3P = T3J + T3M;
|
Chris@10
|
985 T25 = T1Z + T24;
|
Chris@10
|
986 T2g = T2a + T2f;
|
Chris@10
|
987 T2h = T25 + T2g;
|
Chris@10
|
988 }
|
Chris@10
|
989 T2i = T1U + T2h;
|
Chris@10
|
990 T60 = T3O + T3P;
|
Chris@10
|
991 {
|
Chris@10
|
992 E T3G, T57, T3R, T56, T3N, T3Q;
|
Chris@10
|
993 T3G = FMA(KP951056516, T3E, KP587785252 * T3F);
|
Chris@10
|
994 T57 = FNMS(KP587785252, T3E, KP951056516 * T3F);
|
Chris@10
|
995 T3N = KP559016994 * (T3J - T3M);
|
Chris@10
|
996 T3Q = FNMS(KP250000000, T3P, T3O);
|
Chris@10
|
997 T3R = T3N + T3Q;
|
Chris@10
|
998 T56 = T3Q - T3N;
|
Chris@10
|
999 T3S = T3G + T3R;
|
Chris@10
|
1000 T5D = T57 + T56;
|
Chris@10
|
1001 T4r = T3R - T3G;
|
Chris@10
|
1002 T58 = T56 - T57;
|
Chris@10
|
1003 }
|
Chris@10
|
1004 {
|
Chris@10
|
1005 E T3Y, T5a, T3V, T59, T3T, T3U;
|
Chris@10
|
1006 T3Y = FMA(KP951056516, T3W, KP587785252 * T3X);
|
Chris@10
|
1007 T5a = FNMS(KP587785252, T3W, KP951056516 * T3X);
|
Chris@10
|
1008 T3T = KP559016994 * (T25 - T2g);
|
Chris@10
|
1009 T3U = FNMS(KP250000000, T2h, T1U);
|
Chris@10
|
1010 T3V = T3T + T3U;
|
Chris@10
|
1011 T59 = T3U - T3T;
|
Chris@10
|
1012 T3Z = T3V - T3Y;
|
Chris@10
|
1013 T5C = T59 - T5a;
|
Chris@10
|
1014 T4q = T3V + T3Y;
|
Chris@10
|
1015 T5b = T59 + T5a;
|
Chris@10
|
1016 }
|
Chris@10
|
1017 }
|
Chris@10
|
1018 {
|
Chris@10
|
1019 E Tu, T2N, T2B, T2E, T2I, T2H, T2K, T2L, T2O, TF, TQ, TR;
|
Chris@10
|
1020 {
|
Chris@10
|
1021 E Tr, Tt, Tq, Ts;
|
Chris@10
|
1022 Tr = cr[WS(rs, 1)];
|
Chris@10
|
1023 Tt = ci[WS(rs, 1)];
|
Chris@10
|
1024 Tq = W[0];
|
Chris@10
|
1025 Ts = W[1];
|
Chris@10
|
1026 Tu = FMA(Tq, Tr, Ts * Tt);
|
Chris@10
|
1027 T2N = FNMS(Ts, Tr, Tq * Tt);
|
Chris@10
|
1028 }
|
Chris@10
|
1029 {
|
Chris@10
|
1030 E Tz, T2z, TP, T2D, TE, T2A, TK, T2C;
|
Chris@10
|
1031 {
|
Chris@10
|
1032 E Tw, Ty, Tv, Tx;
|
Chris@10
|
1033 Tw = cr[WS(rs, 6)];
|
Chris@10
|
1034 Ty = ci[WS(rs, 6)];
|
Chris@10
|
1035 Tv = W[10];
|
Chris@10
|
1036 Tx = W[11];
|
Chris@10
|
1037 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@10
|
1038 T2z = FNMS(Tx, Tw, Tv * Ty);
|
Chris@10
|
1039 }
|
Chris@10
|
1040 {
|
Chris@10
|
1041 E TM, TO, TL, TN;
|
Chris@10
|
1042 TM = cr[WS(rs, 16)];
|
Chris@10
|
1043 TO = ci[WS(rs, 16)];
|
Chris@10
|
1044 TL = W[30];
|
Chris@10
|
1045 TN = W[31];
|
Chris@10
|
1046 TP = FMA(TL, TM, TN * TO);
|
Chris@10
|
1047 T2D = FNMS(TN, TM, TL * TO);
|
Chris@10
|
1048 }
|
Chris@10
|
1049 {
|
Chris@10
|
1050 E TB, TD, TA, TC;
|
Chris@10
|
1051 TB = cr[WS(rs, 21)];
|
Chris@10
|
1052 TD = ci[WS(rs, 21)];
|
Chris@10
|
1053 TA = W[40];
|
Chris@10
|
1054 TC = W[41];
|
Chris@10
|
1055 TE = FMA(TA, TB, TC * TD);
|
Chris@10
|
1056 T2A = FNMS(TC, TB, TA * TD);
|
Chris@10
|
1057 }
|
Chris@10
|
1058 {
|
Chris@10
|
1059 E TH, TJ, TG, TI;
|
Chris@10
|
1060 TH = cr[WS(rs, 11)];
|
Chris@10
|
1061 TJ = ci[WS(rs, 11)];
|
Chris@10
|
1062 TG = W[20];
|
Chris@10
|
1063 TI = W[21];
|
Chris@10
|
1064 TK = FMA(TG, TH, TI * TJ);
|
Chris@10
|
1065 T2C = FNMS(TI, TH, TG * TJ);
|
Chris@10
|
1066 }
|
Chris@10
|
1067 T2B = T2z - T2A;
|
Chris@10
|
1068 T2E = T2C - T2D;
|
Chris@10
|
1069 T2I = TK - TP;
|
Chris@10
|
1070 T2H = Tz - TE;
|
Chris@10
|
1071 T2K = T2z + T2A;
|
Chris@10
|
1072 T2L = T2C + T2D;
|
Chris@10
|
1073 T2O = T2K + T2L;
|
Chris@10
|
1074 TF = Tz + TE;
|
Chris@10
|
1075 TQ = TK + TP;
|
Chris@10
|
1076 TR = TF + TQ;
|
Chris@10
|
1077 }
|
Chris@10
|
1078 TS = Tu + TR;
|
Chris@10
|
1079 T5W = T2N + T2O;
|
Chris@10
|
1080 {
|
Chris@10
|
1081 E T2F, T4L, T2y, T4K, T2w, T2x;
|
Chris@10
|
1082 T2F = FMA(KP951056516, T2B, KP587785252 * T2E);
|
Chris@10
|
1083 T4L = FNMS(KP587785252, T2B, KP951056516 * T2E);
|
Chris@10
|
1084 T2w = KP559016994 * (TF - TQ);
|
Chris@10
|
1085 T2x = FNMS(KP250000000, TR, Tu);
|
Chris@10
|
1086 T2y = T2w + T2x;
|
Chris@10
|
1087 T4K = T2x - T2w;
|
Chris@10
|
1088 T2G = T2y - T2F;
|
Chris@10
|
1089 T5s = T4K - T4L;
|
Chris@10
|
1090 T4g = T2y + T2F;
|
Chris@10
|
1091 T4M = T4K + T4L;
|
Chris@10
|
1092 }
|
Chris@10
|
1093 {
|
Chris@10
|
1094 E T2J, T4O, T2Q, T4N, T2M, T2P;
|
Chris@10
|
1095 T2J = FMA(KP951056516, T2H, KP587785252 * T2I);
|
Chris@10
|
1096 T4O = FNMS(KP587785252, T2H, KP951056516 * T2I);
|
Chris@10
|
1097 T2M = KP559016994 * (T2K - T2L);
|
Chris@10
|
1098 T2P = FNMS(KP250000000, T2O, T2N);
|
Chris@10
|
1099 T2Q = T2M + T2P;
|
Chris@10
|
1100 T4N = T2P - T2M;
|
Chris@10
|
1101 T2R = T2J + T2Q;
|
Chris@10
|
1102 T5t = T4O + T4N;
|
Chris@10
|
1103 T4h = T2Q - T2J;
|
Chris@10
|
1104 T4P = T4N - T4O;
|
Chris@10
|
1105 }
|
Chris@10
|
1106 }
|
Chris@10
|
1107 {
|
Chris@10
|
1108 E TX, T33, T2T, T2U, T3c, T3b, T2Y, T31, T34, T18, T1j, T1k;
|
Chris@10
|
1109 {
|
Chris@10
|
1110 E TU, TW, TT, TV;
|
Chris@10
|
1111 TU = cr[WS(rs, 4)];
|
Chris@10
|
1112 TW = ci[WS(rs, 4)];
|
Chris@10
|
1113 TT = W[6];
|
Chris@10
|
1114 TV = W[7];
|
Chris@10
|
1115 TX = FMA(TT, TU, TV * TW);
|
Chris@10
|
1116 T33 = FNMS(TV, TU, TT * TW);
|
Chris@10
|
1117 }
|
Chris@10
|
1118 {
|
Chris@10
|
1119 E T12, T2W, T1i, T30, T17, T2X, T1d, T2Z;
|
Chris@10
|
1120 {
|
Chris@10
|
1121 E TZ, T11, TY, T10;
|
Chris@10
|
1122 TZ = cr[WS(rs, 9)];
|
Chris@10
|
1123 T11 = ci[WS(rs, 9)];
|
Chris@10
|
1124 TY = W[16];
|
Chris@10
|
1125 T10 = W[17];
|
Chris@10
|
1126 T12 = FMA(TY, TZ, T10 * T11);
|
Chris@10
|
1127 T2W = FNMS(T10, TZ, TY * T11);
|
Chris@10
|
1128 }
|
Chris@10
|
1129 {
|
Chris@10
|
1130 E T1f, T1h, T1e, T1g;
|
Chris@10
|
1131 T1f = cr[WS(rs, 19)];
|
Chris@10
|
1132 T1h = ci[WS(rs, 19)];
|
Chris@10
|
1133 T1e = W[36];
|
Chris@10
|
1134 T1g = W[37];
|
Chris@10
|
1135 T1i = FMA(T1e, T1f, T1g * T1h);
|
Chris@10
|
1136 T30 = FNMS(T1g, T1f, T1e * T1h);
|
Chris@10
|
1137 }
|
Chris@10
|
1138 {
|
Chris@10
|
1139 E T14, T16, T13, T15;
|
Chris@10
|
1140 T14 = cr[WS(rs, 24)];
|
Chris@10
|
1141 T16 = ci[WS(rs, 24)];
|
Chris@10
|
1142 T13 = W[46];
|
Chris@10
|
1143 T15 = W[47];
|
Chris@10
|
1144 T17 = FMA(T13, T14, T15 * T16);
|
Chris@10
|
1145 T2X = FNMS(T15, T14, T13 * T16);
|
Chris@10
|
1146 }
|
Chris@10
|
1147 {
|
Chris@10
|
1148 E T1a, T1c, T19, T1b;
|
Chris@10
|
1149 T1a = cr[WS(rs, 14)];
|
Chris@10
|
1150 T1c = ci[WS(rs, 14)];
|
Chris@10
|
1151 T19 = W[26];
|
Chris@10
|
1152 T1b = W[27];
|
Chris@10
|
1153 T1d = FMA(T19, T1a, T1b * T1c);
|
Chris@10
|
1154 T2Z = FNMS(T1b, T1a, T19 * T1c);
|
Chris@10
|
1155 }
|
Chris@10
|
1156 T2T = T17 - T12;
|
Chris@10
|
1157 T2U = T1d - T1i;
|
Chris@10
|
1158 T3c = T2Z - T30;
|
Chris@10
|
1159 T3b = T2W - T2X;
|
Chris@10
|
1160 T2Y = T2W + T2X;
|
Chris@10
|
1161 T31 = T2Z + T30;
|
Chris@10
|
1162 T34 = T2Y + T31;
|
Chris@10
|
1163 T18 = T12 + T17;
|
Chris@10
|
1164 T1j = T1d + T1i;
|
Chris@10
|
1165 T1k = T18 + T1j;
|
Chris@10
|
1166 }
|
Chris@10
|
1167 T1l = TX + T1k;
|
Chris@10
|
1168 T5X = T33 + T34;
|
Chris@10
|
1169 {
|
Chris@10
|
1170 E T2V, T4S, T36, T4R, T32, T35;
|
Chris@10
|
1171 T2V = FNMS(KP587785252, T2U, KP951056516 * T2T);
|
Chris@10
|
1172 T4S = FMA(KP587785252, T2T, KP951056516 * T2U);
|
Chris@10
|
1173 T32 = KP559016994 * (T2Y - T31);
|
Chris@10
|
1174 T35 = FNMS(KP250000000, T34, T33);
|
Chris@10
|
1175 T36 = T32 + T35;
|
Chris@10
|
1176 T4R = T35 - T32;
|
Chris@10
|
1177 T37 = T2V - T36;
|
Chris@10
|
1178 T5v = T4S + T4R;
|
Chris@10
|
1179 T4k = T2V + T36;
|
Chris@10
|
1180 T4T = T4R - T4S;
|
Chris@10
|
1181 }
|
Chris@10
|
1182 {
|
Chris@10
|
1183 E T3d, T4V, T3a, T4U, T38, T39;
|
Chris@10
|
1184 T3d = FMA(KP951056516, T3b, KP587785252 * T3c);
|
Chris@10
|
1185 T4V = FNMS(KP587785252, T3b, KP951056516 * T3c);
|
Chris@10
|
1186 T38 = KP559016994 * (T18 - T1j);
|
Chris@10
|
1187 T39 = FNMS(KP250000000, T1k, TX);
|
Chris@10
|
1188 T3a = T38 + T39;
|
Chris@10
|
1189 T4U = T39 - T38;
|
Chris@10
|
1190 T3e = T3a - T3d;
|
Chris@10
|
1191 T5w = T4U - T4V;
|
Chris@10
|
1192 T4j = T3a + T3d;
|
Chris@10
|
1193 T4W = T4U + T4V;
|
Chris@10
|
1194 }
|
Chris@10
|
1195 }
|
Chris@10
|
1196 {
|
Chris@10
|
1197 E T1r, T3r, T3h, T3i, T3A, T3z, T3m, T3p, T3s, T1C, T1N, T1O;
|
Chris@10
|
1198 {
|
Chris@10
|
1199 E T1o, T1q, T1n, T1p;
|
Chris@10
|
1200 T1o = cr[WS(rs, 2)];
|
Chris@10
|
1201 T1q = ci[WS(rs, 2)];
|
Chris@10
|
1202 T1n = W[2];
|
Chris@10
|
1203 T1p = W[3];
|
Chris@10
|
1204 T1r = FMA(T1n, T1o, T1p * T1q);
|
Chris@10
|
1205 T3r = FNMS(T1p, T1o, T1n * T1q);
|
Chris@10
|
1206 }
|
Chris@10
|
1207 {
|
Chris@10
|
1208 E T1w, T3k, T1M, T3o, T1B, T3l, T1H, T3n;
|
Chris@10
|
1209 {
|
Chris@10
|
1210 E T1t, T1v, T1s, T1u;
|
Chris@10
|
1211 T1t = cr[WS(rs, 7)];
|
Chris@10
|
1212 T1v = ci[WS(rs, 7)];
|
Chris@10
|
1213 T1s = W[12];
|
Chris@10
|
1214 T1u = W[13];
|
Chris@10
|
1215 T1w = FMA(T1s, T1t, T1u * T1v);
|
Chris@10
|
1216 T3k = FNMS(T1u, T1t, T1s * T1v);
|
Chris@10
|
1217 }
|
Chris@10
|
1218 {
|
Chris@10
|
1219 E T1J, T1L, T1I, T1K;
|
Chris@10
|
1220 T1J = cr[WS(rs, 17)];
|
Chris@10
|
1221 T1L = ci[WS(rs, 17)];
|
Chris@10
|
1222 T1I = W[32];
|
Chris@10
|
1223 T1K = W[33];
|
Chris@10
|
1224 T1M = FMA(T1I, T1J, T1K * T1L);
|
Chris@10
|
1225 T3o = FNMS(T1K, T1J, T1I * T1L);
|
Chris@10
|
1226 }
|
Chris@10
|
1227 {
|
Chris@10
|
1228 E T1y, T1A, T1x, T1z;
|
Chris@10
|
1229 T1y = cr[WS(rs, 22)];
|
Chris@10
|
1230 T1A = ci[WS(rs, 22)];
|
Chris@10
|
1231 T1x = W[42];
|
Chris@10
|
1232 T1z = W[43];
|
Chris@10
|
1233 T1B = FMA(T1x, T1y, T1z * T1A);
|
Chris@10
|
1234 T3l = FNMS(T1z, T1y, T1x * T1A);
|
Chris@10
|
1235 }
|
Chris@10
|
1236 {
|
Chris@10
|
1237 E T1E, T1G, T1D, T1F;
|
Chris@10
|
1238 T1E = cr[WS(rs, 12)];
|
Chris@10
|
1239 T1G = ci[WS(rs, 12)];
|
Chris@10
|
1240 T1D = W[22];
|
Chris@10
|
1241 T1F = W[23];
|
Chris@10
|
1242 T1H = FMA(T1D, T1E, T1F * T1G);
|
Chris@10
|
1243 T3n = FNMS(T1F, T1E, T1D * T1G);
|
Chris@10
|
1244 }
|
Chris@10
|
1245 T3h = T1w - T1B;
|
Chris@10
|
1246 T3i = T1H - T1M;
|
Chris@10
|
1247 T3A = T3n - T3o;
|
Chris@10
|
1248 T3z = T3k - T3l;
|
Chris@10
|
1249 T3m = T3k + T3l;
|
Chris@10
|
1250 T3p = T3n + T3o;
|
Chris@10
|
1251 T3s = T3m + T3p;
|
Chris@10
|
1252 T1C = T1w + T1B;
|
Chris@10
|
1253 T1N = T1H + T1M;
|
Chris@10
|
1254 T1O = T1C + T1N;
|
Chris@10
|
1255 }
|
Chris@10
|
1256 T1P = T1r + T1O;
|
Chris@10
|
1257 T5Z = T3r + T3s;
|
Chris@10
|
1258 {
|
Chris@10
|
1259 E T3j, T53, T3u, T52, T3q, T3t;
|
Chris@10
|
1260 T3j = FMA(KP951056516, T3h, KP587785252 * T3i);
|
Chris@10
|
1261 T53 = FNMS(KP587785252, T3h, KP951056516 * T3i);
|
Chris@10
|
1262 T3q = KP559016994 * (T3m - T3p);
|
Chris@10
|
1263 T3t = FNMS(KP250000000, T3s, T3r);
|
Chris@10
|
1264 T3u = T3q + T3t;
|
Chris@10
|
1265 T52 = T3t - T3q;
|
Chris@10
|
1266 T3v = T3j + T3u;
|
Chris@10
|
1267 T5A = T53 + T52;
|
Chris@10
|
1268 T4o = T3u - T3j;
|
Chris@10
|
1269 T54 = T52 - T53;
|
Chris@10
|
1270 }
|
Chris@10
|
1271 {
|
Chris@10
|
1272 E T3B, T50, T3y, T4Z, T3w, T3x;
|
Chris@10
|
1273 T3B = FMA(KP951056516, T3z, KP587785252 * T3A);
|
Chris@10
|
1274 T50 = FNMS(KP587785252, T3z, KP951056516 * T3A);
|
Chris@10
|
1275 T3w = KP559016994 * (T1C - T1N);
|
Chris@10
|
1276 T3x = FNMS(KP250000000, T1O, T1r);
|
Chris@10
|
1277 T3y = T3w + T3x;
|
Chris@10
|
1278 T4Z = T3x - T3w;
|
Chris@10
|
1279 T3C = T3y - T3B;
|
Chris@10
|
1280 T5z = T4Z - T50;
|
Chris@10
|
1281 T4n = T3y + T3B;
|
Chris@10
|
1282 T51 = T4Z + T50;
|
Chris@10
|
1283 }
|
Chris@10
|
1284 }
|
Chris@10
|
1285 {
|
Chris@10
|
1286 E T62, T64, Tp, T2k, T5T, T5U, T63, T5V;
|
Chris@10
|
1287 {
|
Chris@10
|
1288 E T5Y, T61, T1m, T2j;
|
Chris@10
|
1289 T5Y = T5W - T5X;
|
Chris@10
|
1290 T61 = T5Z - T60;
|
Chris@10
|
1291 T62 = FMA(KP951056516, T5Y, KP587785252 * T61);
|
Chris@10
|
1292 T64 = FNMS(KP587785252, T5Y, KP951056516 * T61);
|
Chris@10
|
1293 Tp = T1 + To;
|
Chris@10
|
1294 T1m = TS + T1l;
|
Chris@10
|
1295 T2j = T1P + T2i;
|
Chris@10
|
1296 T2k = T1m + T2j;
|
Chris@10
|
1297 T5T = KP559016994 * (T1m - T2j);
|
Chris@10
|
1298 T5U = FNMS(KP250000000, T2k, Tp);
|
Chris@10
|
1299 }
|
Chris@10
|
1300 cr[0] = Tp + T2k;
|
Chris@10
|
1301 T63 = T5U - T5T;
|
Chris@10
|
1302 cr[WS(rs, 10)] = T63 - T64;
|
Chris@10
|
1303 ci[WS(rs, 9)] = T63 + T64;
|
Chris@10
|
1304 T5V = T5T + T5U;
|
Chris@10
|
1305 ci[WS(rs, 4)] = T5V - T62;
|
Chris@10
|
1306 cr[WS(rs, 5)] = T5V + T62;
|
Chris@10
|
1307 }
|
Chris@10
|
1308 {
|
Chris@10
|
1309 E T2v, T4f, T6I, T6U, T42, T6Z, T43, T6Y, T4A, T6N, T4D, T6L, T4u, T6E, T4v;
|
Chris@10
|
1310 E T6D, T48, T6V, T4b, T6T, T2n, T6G;
|
Chris@10
|
1311 T2n = T2l + T2m;
|
Chris@10
|
1312 T2v = T2n - T2u;
|
Chris@10
|
1313 T4f = T2n + T2u;
|
Chris@10
|
1314 T6G = T6g + T6f;
|
Chris@10
|
1315 T6I = T6G - T6H;
|
Chris@10
|
1316 T6U = T6H + T6G;
|
Chris@10
|
1317 {
|
Chris@10
|
1318 E T2S, T3f, T3g, T3D, T40, T41;
|
Chris@10
|
1319 T2S = FMA(KP535826794, T2G, KP844327925 * T2R);
|
Chris@10
|
1320 T3f = FNMS(KP637423989, T3e, KP770513242 * T37);
|
Chris@10
|
1321 T3g = T2S + T3f;
|
Chris@10
|
1322 T3D = FNMS(KP425779291, T3C, KP904827052 * T3v);
|
Chris@10
|
1323 T40 = FNMS(KP992114701, T3Z, KP125333233 * T3S);
|
Chris@10
|
1324 T41 = T3D + T40;
|
Chris@10
|
1325 T42 = T3g + T41;
|
Chris@10
|
1326 T6Z = T3D - T40;
|
Chris@10
|
1327 T43 = KP559016994 * (T3g - T41);
|
Chris@10
|
1328 T6Y = T3f - T2S;
|
Chris@10
|
1329 }
|
Chris@10
|
1330 {
|
Chris@10
|
1331 E T4y, T4z, T6J, T4B, T4C, T6K;
|
Chris@10
|
1332 T4y = FNMS(KP248689887, T4g, KP968583161 * T4h);
|
Chris@10
|
1333 T4z = FNMS(KP844327925, T4j, KP535826794 * T4k);
|
Chris@10
|
1334 T6J = T4y + T4z;
|
Chris@10
|
1335 T4B = FNMS(KP481753674, T4n, KP876306680 * T4o);
|
Chris@10
|
1336 T4C = FNMS(KP684547105, T4q, KP728968627 * T4r);
|
Chris@10
|
1337 T6K = T4B + T4C;
|
Chris@10
|
1338 T4A = T4y - T4z;
|
Chris@10
|
1339 T6N = KP559016994 * (T6J - T6K);
|
Chris@10
|
1340 T4D = T4B - T4C;
|
Chris@10
|
1341 T6L = T6J + T6K;
|
Chris@10
|
1342 }
|
Chris@10
|
1343 {
|
Chris@10
|
1344 E T4i, T4l, T4m, T4p, T4s, T4t;
|
Chris@10
|
1345 T4i = FMA(KP968583161, T4g, KP248689887 * T4h);
|
Chris@10
|
1346 T4l = FMA(KP535826794, T4j, KP844327925 * T4k);
|
Chris@10
|
1347 T4m = T4i + T4l;
|
Chris@10
|
1348 T4p = FMA(KP876306680, T4n, KP481753674 * T4o);
|
Chris@10
|
1349 T4s = FMA(KP728968627, T4q, KP684547105 * T4r);
|
Chris@10
|
1350 T4t = T4p + T4s;
|
Chris@10
|
1351 T4u = T4m + T4t;
|
Chris@10
|
1352 T6E = T4p - T4s;
|
Chris@10
|
1353 T4v = KP559016994 * (T4m - T4t);
|
Chris@10
|
1354 T6D = T4l - T4i;
|
Chris@10
|
1355 }
|
Chris@10
|
1356 {
|
Chris@10
|
1357 E T46, T47, T6R, T49, T4a, T6S;
|
Chris@10
|
1358 T46 = FNMS(KP844327925, T2G, KP535826794 * T2R);
|
Chris@10
|
1359 T47 = FMA(KP770513242, T3e, KP637423989 * T37);
|
Chris@10
|
1360 T6R = T46 + T47;
|
Chris@10
|
1361 T49 = FMA(KP125333233, T3Z, KP992114701 * T3S);
|
Chris@10
|
1362 T4a = FMA(KP904827052, T3C, KP425779291 * T3v);
|
Chris@10
|
1363 T6S = T4a + T49;
|
Chris@10
|
1364 T48 = T46 - T47;
|
Chris@10
|
1365 T6V = T6R - T6S;
|
Chris@10
|
1366 T4b = T49 - T4a;
|
Chris@10
|
1367 T6T = KP559016994 * (T6R + T6S);
|
Chris@10
|
1368 }
|
Chris@10
|
1369 cr[WS(rs, 4)] = T2v + T42;
|
Chris@10
|
1370 ci[WS(rs, 23)] = T6L + T6I;
|
Chris@10
|
1371 ci[WS(rs, 20)] = T6V + T6U;
|
Chris@10
|
1372 cr[WS(rs, 1)] = T4f + T4u;
|
Chris@10
|
1373 {
|
Chris@10
|
1374 E T4c, T4e, T45, T4d, T44;
|
Chris@10
|
1375 T4c = FMA(KP951056516, T48, KP587785252 * T4b);
|
Chris@10
|
1376 T4e = FNMS(KP587785252, T48, KP951056516 * T4b);
|
Chris@10
|
1377 T44 = FNMS(KP250000000, T42, T2v);
|
Chris@10
|
1378 T45 = T43 + T44;
|
Chris@10
|
1379 T4d = T44 - T43;
|
Chris@10
|
1380 ci[0] = T45 - T4c;
|
Chris@10
|
1381 ci[WS(rs, 5)] = T4d + T4e;
|
Chris@10
|
1382 cr[WS(rs, 9)] = T45 + T4c;
|
Chris@10
|
1383 ci[WS(rs, 10)] = T4d - T4e;
|
Chris@10
|
1384 }
|
Chris@10
|
1385 {
|
Chris@10
|
1386 E T6F, T6P, T6O, T6Q, T6M;
|
Chris@10
|
1387 T6F = FMA(KP587785252, T6D, KP951056516 * T6E);
|
Chris@10
|
1388 T6P = FNMS(KP587785252, T6E, KP951056516 * T6D);
|
Chris@10
|
1389 T6M = FNMS(KP250000000, T6L, T6I);
|
Chris@10
|
1390 T6O = T6M - T6N;
|
Chris@10
|
1391 T6Q = T6N + T6M;
|
Chris@10
|
1392 cr[WS(rs, 16)] = T6F - T6O;
|
Chris@10
|
1393 ci[WS(rs, 18)] = T6P + T6Q;
|
Chris@10
|
1394 ci[WS(rs, 13)] = T6F + T6O;
|
Chris@10
|
1395 cr[WS(rs, 21)] = T6P - T6Q;
|
Chris@10
|
1396 }
|
Chris@10
|
1397 {
|
Chris@10
|
1398 E T70, T71, T6X, T72, T6W;
|
Chris@10
|
1399 T70 = FMA(KP587785252, T6Y, KP951056516 * T6Z);
|
Chris@10
|
1400 T71 = FNMS(KP587785252, T6Z, KP951056516 * T6Y);
|
Chris@10
|
1401 T6W = FNMS(KP250000000, T6V, T6U);
|
Chris@10
|
1402 T6X = T6T - T6W;
|
Chris@10
|
1403 T72 = T6T + T6W;
|
Chris@10
|
1404 cr[WS(rs, 14)] = T6X - T70;
|
Chris@10
|
1405 ci[WS(rs, 15)] = T71 + T72;
|
Chris@10
|
1406 cr[WS(rs, 19)] = T70 + T6X;
|
Chris@10
|
1407 cr[WS(rs, 24)] = T71 - T72;
|
Chris@10
|
1408 }
|
Chris@10
|
1409 {
|
Chris@10
|
1410 E T4E, T4G, T4x, T4F, T4w;
|
Chris@10
|
1411 T4E = FMA(KP951056516, T4A, KP587785252 * T4D);
|
Chris@10
|
1412 T4G = FNMS(KP587785252, T4A, KP951056516 * T4D);
|
Chris@10
|
1413 T4w = FNMS(KP250000000, T4u, T4f);
|
Chris@10
|
1414 T4x = T4v + T4w;
|
Chris@10
|
1415 T4F = T4w - T4v;
|
Chris@10
|
1416 ci[WS(rs, 3)] = T4x - T4E;
|
Chris@10
|
1417 ci[WS(rs, 8)] = T4F + T4G;
|
Chris@10
|
1418 cr[WS(rs, 6)] = T4x + T4E;
|
Chris@10
|
1419 cr[WS(rs, 11)] = T4F - T4G;
|
Chris@10
|
1420 }
|
Chris@10
|
1421 }
|
Chris@10
|
1422 {
|
Chris@10
|
1423 E T75, T7d, T76, T79, T7a, T7b, T7e, T7c;
|
Chris@10
|
1424 {
|
Chris@10
|
1425 E T73, T74, T77, T78;
|
Chris@10
|
1426 T73 = T1l - TS;
|
Chris@10
|
1427 T74 = T1P - T2i;
|
Chris@10
|
1428 T75 = FMA(KP587785252, T73, KP951056516 * T74);
|
Chris@10
|
1429 T7d = FNMS(KP587785252, T74, KP951056516 * T73);
|
Chris@10
|
1430 T76 = T6e + T6b;
|
Chris@10
|
1431 T77 = T5W + T5X;
|
Chris@10
|
1432 T78 = T5Z + T60;
|
Chris@10
|
1433 T79 = T77 + T78;
|
Chris@10
|
1434 T7a = FNMS(KP250000000, T79, T76);
|
Chris@10
|
1435 T7b = KP559016994 * (T77 - T78);
|
Chris@10
|
1436 }
|
Chris@10
|
1437 ci[WS(rs, 24)] = T79 + T76;
|
Chris@10
|
1438 T7e = T7b + T7a;
|
Chris@10
|
1439 cr[WS(rs, 20)] = T7d - T7e;
|
Chris@10
|
1440 ci[WS(rs, 19)] = T7d + T7e;
|
Chris@10
|
1441 T7c = T7a - T7b;
|
Chris@10
|
1442 cr[WS(rs, 15)] = T75 - T7c;
|
Chris@10
|
1443 ci[WS(rs, 14)] = T75 + T7c;
|
Chris@10
|
1444 }
|
Chris@10
|
1445 {
|
Chris@10
|
1446 E T4J, T5r, T6i, T6u, T5e, T6z, T5f, T6y, T5M, T6n, T5P, T6l, T5G, T66, T5H;
|
Chris@10
|
1447 E T65, T5k, T6v, T5n, T6t, T4H, T6h;
|
Chris@10
|
1448 T4H = T2m - T2l;
|
Chris@10
|
1449 T4J = T4H + T4I;
|
Chris@10
|
1450 T5r = T4H - T4I;
|
Chris@10
|
1451 T6h = T6f - T6g;
|
Chris@10
|
1452 T6i = T6a + T6h;
|
Chris@10
|
1453 T6u = T6h - T6a;
|
Chris@10
|
1454 {
|
Chris@10
|
1455 E T4Q, T4X, T4Y, T55, T5c, T5d;
|
Chris@10
|
1456 T4Q = FMA(KP728968627, T4M, KP684547105 * T4P);
|
Chris@10
|
1457 T4X = FNMS(KP992114701, T4W, KP125333233 * T4T);
|
Chris@10
|
1458 T4Y = T4Q + T4X;
|
Chris@10
|
1459 T55 = FMA(KP062790519, T51, KP998026728 * T54);
|
Chris@10
|
1460 T5c = FNMS(KP637423989, T5b, KP770513242 * T58);
|
Chris@10
|
1461 T5d = T55 + T5c;
|
Chris@10
|
1462 T5e = T4Y + T5d;
|
Chris@10
|
1463 T6z = T55 - T5c;
|
Chris@10
|
1464 T5f = KP559016994 * (T4Y - T5d);
|
Chris@10
|
1465 T6y = T4X - T4Q;
|
Chris@10
|
1466 }
|
Chris@10
|
1467 {
|
Chris@10
|
1468 E T5K, T5L, T6j, T5N, T5O, T6k;
|
Chris@10
|
1469 T5K = FNMS(KP481753674, T5s, KP876306680 * T5t);
|
Chris@10
|
1470 T5L = FMA(KP904827052, T5w, KP425779291 * T5v);
|
Chris@10
|
1471 T6j = T5K - T5L;
|
Chris@10
|
1472 T5N = FNMS(KP844327925, T5z, KP535826794 * T5A);
|
Chris@10
|
1473 T5O = FNMS(KP998026728, T5C, KP062790519 * T5D);
|
Chris@10
|
1474 T6k = T5N + T5O;
|
Chris@10
|
1475 T5M = T5K + T5L;
|
Chris@10
|
1476 T6n = KP559016994 * (T6j - T6k);
|
Chris@10
|
1477 T5P = T5N - T5O;
|
Chris@10
|
1478 T6l = T6j + T6k;
|
Chris@10
|
1479 }
|
Chris@10
|
1480 {
|
Chris@10
|
1481 E T5u, T5x, T5y, T5B, T5E, T5F;
|
Chris@10
|
1482 T5u = FMA(KP876306680, T5s, KP481753674 * T5t);
|
Chris@10
|
1483 T5x = FNMS(KP425779291, T5w, KP904827052 * T5v);
|
Chris@10
|
1484 T5y = T5u + T5x;
|
Chris@10
|
1485 T5B = FMA(KP535826794, T5z, KP844327925 * T5A);
|
Chris@10
|
1486 T5E = FMA(KP062790519, T5C, KP998026728 * T5D);
|
Chris@10
|
1487 T5F = T5B + T5E;
|
Chris@10
|
1488 T5G = T5y + T5F;
|
Chris@10
|
1489 T66 = T5B - T5E;
|
Chris@10
|
1490 T5H = KP559016994 * (T5y - T5F);
|
Chris@10
|
1491 T65 = T5x - T5u;
|
Chris@10
|
1492 }
|
Chris@10
|
1493 {
|
Chris@10
|
1494 E T5i, T5j, T6r, T5l, T5m, T6s;
|
Chris@10
|
1495 T5i = FNMS(KP684547105, T4M, KP728968627 * T4P);
|
Chris@10
|
1496 T5j = FMA(KP125333233, T4W, KP992114701 * T4T);
|
Chris@10
|
1497 T6r = T5i - T5j;
|
Chris@10
|
1498 T5l = FNMS(KP998026728, T51, KP062790519 * T54);
|
Chris@10
|
1499 T5m = FMA(KP770513242, T5b, KP637423989 * T58);
|
Chris@10
|
1500 T6s = T5l - T5m;
|
Chris@10
|
1501 T5k = T5i + T5j;
|
Chris@10
|
1502 T6v = T6r + T6s;
|
Chris@10
|
1503 T5n = T5l + T5m;
|
Chris@10
|
1504 T6t = KP559016994 * (T6r - T6s);
|
Chris@10
|
1505 }
|
Chris@10
|
1506 cr[WS(rs, 3)] = T4J + T5e;
|
Chris@10
|
1507 ci[WS(rs, 22)] = T6l + T6i;
|
Chris@10
|
1508 ci[WS(rs, 21)] = T6v + T6u;
|
Chris@10
|
1509 cr[WS(rs, 2)] = T5r + T5G;
|
Chris@10
|
1510 {
|
Chris@10
|
1511 E T67, T6p, T6o, T6q, T6m;
|
Chris@10
|
1512 T67 = FMA(KP587785252, T65, KP951056516 * T66);
|
Chris@10
|
1513 T6p = FNMS(KP587785252, T66, KP951056516 * T65);
|
Chris@10
|
1514 T6m = FNMS(KP250000000, T6l, T6i);
|
Chris@10
|
1515 T6o = T6m - T6n;
|
Chris@10
|
1516 T6q = T6n + T6m;
|
Chris@10
|
1517 cr[WS(rs, 17)] = T67 - T6o;
|
Chris@10
|
1518 ci[WS(rs, 17)] = T6p + T6q;
|
Chris@10
|
1519 ci[WS(rs, 12)] = T67 + T6o;
|
Chris@10
|
1520 cr[WS(rs, 22)] = T6p - T6q;
|
Chris@10
|
1521 }
|
Chris@10
|
1522 {
|
Chris@10
|
1523 E T5Q, T5S, T5J, T5R, T5I;
|
Chris@10
|
1524 T5Q = FMA(KP951056516, T5M, KP587785252 * T5P);
|
Chris@10
|
1525 T5S = FNMS(KP587785252, T5M, KP951056516 * T5P);
|
Chris@10
|
1526 T5I = FNMS(KP250000000, T5G, T5r);
|
Chris@10
|
1527 T5J = T5H + T5I;
|
Chris@10
|
1528 T5R = T5I - T5H;
|
Chris@10
|
1529 ci[WS(rs, 2)] = T5J - T5Q;
|
Chris@10
|
1530 ci[WS(rs, 7)] = T5R + T5S;
|
Chris@10
|
1531 cr[WS(rs, 7)] = T5J + T5Q;
|
Chris@10
|
1532 cr[WS(rs, 12)] = T5R - T5S;
|
Chris@10
|
1533 }
|
Chris@10
|
1534 {
|
Chris@10
|
1535 E T5o, T5q, T5h, T5p, T5g;
|
Chris@10
|
1536 T5o = FMA(KP951056516, T5k, KP587785252 * T5n);
|
Chris@10
|
1537 T5q = FNMS(KP587785252, T5k, KP951056516 * T5n);
|
Chris@10
|
1538 T5g = FNMS(KP250000000, T5e, T4J);
|
Chris@10
|
1539 T5h = T5f + T5g;
|
Chris@10
|
1540 T5p = T5g - T5f;
|
Chris@10
|
1541 ci[WS(rs, 1)] = T5h - T5o;
|
Chris@10
|
1542 ci[WS(rs, 6)] = T5p + T5q;
|
Chris@10
|
1543 cr[WS(rs, 8)] = T5h + T5o;
|
Chris@10
|
1544 ci[WS(rs, 11)] = T5p - T5q;
|
Chris@10
|
1545 }
|
Chris@10
|
1546 {
|
Chris@10
|
1547 E T6A, T6B, T6x, T6C, T6w;
|
Chris@10
|
1548 T6A = FMA(KP587785252, T6y, KP951056516 * T6z);
|
Chris@10
|
1549 T6B = FNMS(KP587785252, T6z, KP951056516 * T6y);
|
Chris@10
|
1550 T6w = FNMS(KP250000000, T6v, T6u);
|
Chris@10
|
1551 T6x = T6t - T6w;
|
Chris@10
|
1552 T6C = T6t + T6w;
|
Chris@10
|
1553 cr[WS(rs, 13)] = T6x - T6A;
|
Chris@10
|
1554 ci[WS(rs, 16)] = T6B + T6C;
|
Chris@10
|
1555 cr[WS(rs, 18)] = T6A + T6x;
|
Chris@10
|
1556 cr[WS(rs, 23)] = T6B - T6C;
|
Chris@10
|
1557 }
|
Chris@10
|
1558 }
|
Chris@10
|
1559 }
|
Chris@10
|
1560 }
|
Chris@10
|
1561 }
|
Chris@10
|
1562
|
Chris@10
|
1563 static const tw_instr twinstr[] = {
|
Chris@10
|
1564 {TW_FULL, 1, 25},
|
Chris@10
|
1565 {TW_NEXT, 1, 0}
|
Chris@10
|
1566 };
|
Chris@10
|
1567
|
Chris@10
|
1568 static const hc2hc_desc desc = { 25, "hf_25", twinstr, &GENUS, {260, 140, 140, 0} };
|
Chris@10
|
1569
|
Chris@10
|
1570 void X(codelet_hf_25) (planner *p) {
|
Chris@10
|
1571 X(khc2hc_register) (p, hf_25, &desc);
|
Chris@10
|
1572 }
|
Chris@10
|
1573 #endif /* HAVE_FMA */
|