cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:07:41 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 25 -dif -name hb2_25 -include rdft/scalar/hb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 440 FP additions, 434 FP multiplications,
|
cannam@167
|
32 * (or, 84 additions, 78 multiplications, 356 fused multiply/add),
|
cannam@167
|
33 * 206 stack variables, 47 constants, and 100 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hb2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP921177326, +0.921177326965143320250447435415066029359282231);
|
cannam@167
|
40 DK(KP833417178, +0.833417178328688677408962550243238843138996060);
|
cannam@167
|
41 DK(KP541454447, +0.541454447536312777046285590082819509052033189);
|
cannam@167
|
42 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
cannam@167
|
43 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
cannam@167
|
44 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
cannam@167
|
45 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
cannam@167
|
46 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
cannam@167
|
47 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
cannam@167
|
48 DK(KP871714437, +0.871714437527667770979999223229522602943903653);
|
cannam@167
|
49 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
cannam@167
|
50 DK(KP554608978, +0.554608978404018097464974850792216217022558774);
|
cannam@167
|
51 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
cannam@167
|
52 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
cannam@167
|
53 DK(KP851038619, +0.851038619207379630836264138867114231259902550);
|
cannam@167
|
54 DK(KP525970792, +0.525970792408939708442463226536226366643874659);
|
cannam@167
|
55 DK(KP726211448, +0.726211448929902658173535992263577167607493062);
|
cannam@167
|
56 DK(KP912018591, +0.912018591466481957908415381764119056233607330);
|
cannam@167
|
57 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
cannam@167
|
58 DK(KP943557151, +0.943557151597354104399655195398983005179443399);
|
cannam@167
|
59 DK(KP994076283, +0.994076283785401014123185814696322018529298887);
|
cannam@167
|
60 DK(KP614372930, +0.614372930789563808870829930444362096004872855);
|
cannam@167
|
61 DK(KP621716863, +0.621716863012209892444754556304102309693593202);
|
cannam@167
|
62 DK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
cannam@167
|
63 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
cannam@167
|
64 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
cannam@167
|
65 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
cannam@167
|
66 DK(KP557913902, +0.557913902031834264187699648465567037992437152);
|
cannam@167
|
67 DK(KP249506682, +0.249506682107067890488084201715862638334226305);
|
cannam@167
|
68 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
cannam@167
|
69 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
cannam@167
|
70 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
cannam@167
|
71 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
cannam@167
|
72 DK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
cannam@167
|
73 DK(KP968479752, +0.968479752739016373193524836781420152702090879);
|
cannam@167
|
74 DK(KP062914667, +0.062914667253649757225485955897349402364686947);
|
cannam@167
|
75 DK(KP827271945, +0.827271945972475634034355757144307982555673741);
|
cannam@167
|
76 DK(KP126329378, +0.126329378446108174786050455341811215027378105);
|
cannam@167
|
77 DK(KP470564281, +0.470564281212251493087595091036643380879947982);
|
cannam@167
|
78 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
cannam@167
|
79 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
cannam@167
|
80 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
cannam@167
|
81 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
cannam@167
|
82 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
83 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
84 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
85 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
cannam@167
|
86 {
|
cannam@167
|
87 INT m;
|
cannam@167
|
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)) {
|
cannam@167
|
89 E TN, TT, TO, TR, T23, T25, TQ, TS, T4l, TW, T4n, TX, T2e, T2y, T4z;
|
cannam@167
|
90 E T2q, T76, T4o, T8d, T2u, T4e, T4i, T8a, T86, T71, T6Y, T6U, T26, T2a, T3U;
|
cannam@167
|
91 E T8o, T8s, T4B, T4C, T4G, T2k, T5w, T5C, T6E, T5T, T4u, T7g, T7c, T1I, TY;
|
cannam@167
|
92 E T5I, T8i, T5M;
|
cannam@167
|
93 {
|
cannam@167
|
94 E T2x, T2p, T85, T4d, T2t, T89, T4h, TU, T4m, T2j, T3T, TP, TV, T2d, T5v;
|
cannam@167
|
95 E T5B;
|
cannam@167
|
96 TN = W[0];
|
cannam@167
|
97 TT = W[4];
|
cannam@167
|
98 TO = W[2];
|
cannam@167
|
99 TR = W[3];
|
cannam@167
|
100 TP = TN * TO;
|
cannam@167
|
101 T2x = TO * TT;
|
cannam@167
|
102 TV = TN * TR;
|
cannam@167
|
103 T2d = TN * TT;
|
cannam@167
|
104 T23 = W[6];
|
cannam@167
|
105 T2p = TT * T23;
|
cannam@167
|
106 T85 = TN * T23;
|
cannam@167
|
107 T4d = TO * T23;
|
cannam@167
|
108 T25 = W[7];
|
cannam@167
|
109 T2t = TT * T25;
|
cannam@167
|
110 T89 = TN * T25;
|
cannam@167
|
111 T4h = TO * T25;
|
cannam@167
|
112 TQ = W[1];
|
cannam@167
|
113 TS = FNMS(TQ, TR, TP);
|
cannam@167
|
114 T4l = FMA(TQ, TR, TP);
|
cannam@167
|
115 TU = TS * TT;
|
cannam@167
|
116 T4m = T4l * TT;
|
cannam@167
|
117 TW = FMA(TQ, TO, TV);
|
cannam@167
|
118 T4n = FNMS(TQ, TO, TV);
|
cannam@167
|
119 TX = W[5];
|
cannam@167
|
120 T2j = TN * TX;
|
cannam@167
|
121 T3T = TO * TX;
|
cannam@167
|
122 T2e = FNMS(TQ, TX, T2d);
|
cannam@167
|
123 T2y = FMA(TR, TX, T2x);
|
cannam@167
|
124 T4z = FMA(TQ, TX, T2d);
|
cannam@167
|
125 T2q = FMA(TX, T25, T2p);
|
cannam@167
|
126 T76 = FMA(TR, TT, T3T);
|
cannam@167
|
127 T4o = FNMS(T4n, TX, T4m);
|
cannam@167
|
128 T8d = FNMS(TW, TX, TU);
|
cannam@167
|
129 T2u = FNMS(TX, T23, T2t);
|
cannam@167
|
130 T4e = FMA(TR, T25, T4d);
|
cannam@167
|
131 T4i = FNMS(TR, T23, T4h);
|
cannam@167
|
132 T8a = FNMS(TQ, T23, T89);
|
cannam@167
|
133 T86 = FMA(TQ, T25, T85);
|
cannam@167
|
134 {
|
cannam@167
|
135 E T6X, T6T, T24, T29;
|
cannam@167
|
136 T71 = FNMS(TR, TX, T2x);
|
cannam@167
|
137 T6X = T4l * T25;
|
cannam@167
|
138 T6Y = FNMS(T4n, T23, T6X);
|
cannam@167
|
139 T6T = T4l * T23;
|
cannam@167
|
140 T6U = FMA(T4n, T25, T6T);
|
cannam@167
|
141 T24 = TS * T23;
|
cannam@167
|
142 T26 = FMA(TW, T25, T24);
|
cannam@167
|
143 T29 = TS * T25;
|
cannam@167
|
144 T2a = FNMS(TW, T23, T29);
|
cannam@167
|
145 }
|
cannam@167
|
146 {
|
cannam@167
|
147 E T8n, T8r, T4A, T4F;
|
cannam@167
|
148 T8n = T2y * T23;
|
cannam@167
|
149 T8r = T2y * T25;
|
cannam@167
|
150 T3U = FNMS(TR, TT, T3T);
|
cannam@167
|
151 T8o = FMA(T3U, T25, T8n);
|
cannam@167
|
152 T8s = FNMS(T3U, T23, T8r);
|
cannam@167
|
153 T4A = T4z * T23;
|
cannam@167
|
154 T4F = T4z * T25;
|
cannam@167
|
155 T4B = FNMS(TQ, TT, T2j);
|
cannam@167
|
156 T4C = FMA(T4B, T25, T4A);
|
cannam@167
|
157 T4G = FNMS(T4B, T23, T4F);
|
cannam@167
|
158 }
|
cannam@167
|
159 T5v = T2e * T23;
|
cannam@167
|
160 T5B = T2e * T25;
|
cannam@167
|
161 T2k = FMA(TQ, TT, T2j);
|
cannam@167
|
162 T5w = FMA(T2k, T25, T5v);
|
cannam@167
|
163 T5C = FNMS(T2k, T23, T5B);
|
cannam@167
|
164 {
|
cannam@167
|
165 E T4t, T7b, T7f, T1H, T5H, T5L;
|
cannam@167
|
166 T4t = T4l * TX;
|
cannam@167
|
167 T6E = FNMS(T4n, TT, T4t);
|
cannam@167
|
168 T5T = FMA(T4n, TX, T4m);
|
cannam@167
|
169 T7b = T5T * T23;
|
cannam@167
|
170 T7f = T5T * T25;
|
cannam@167
|
171 T4u = FMA(T4n, TT, T4t);
|
cannam@167
|
172 T7g = FNMS(T6E, T23, T7f);
|
cannam@167
|
173 T7c = FMA(T6E, T25, T7b);
|
cannam@167
|
174 T1H = TS * TX;
|
cannam@167
|
175 T1I = FNMS(TW, TT, T1H);
|
cannam@167
|
176 TY = FMA(TW, TX, TU);
|
cannam@167
|
177 T5H = TY * T23;
|
cannam@167
|
178 T5L = TY * T25;
|
cannam@167
|
179 T5I = FMA(T1I, T25, T5H);
|
cannam@167
|
180 T8i = FMA(TW, TT, T1H);
|
cannam@167
|
181 T5M = FNMS(T1I, T23, T5L);
|
cannam@167
|
182 }
|
cannam@167
|
183 }
|
cannam@167
|
184 {
|
cannam@167
|
185 E T9, T40, T1R, T6G, T6F, T3X, T6H, T2F, T7n, T4N, T5W, T1k, T1S, T1D, T1T;
|
cannam@167
|
186 E Ti, Tr, Ts, TB, TK, TL, TM, T6p, T7K, T6w, T7A, T2U, T56, T3K, T4X;
|
cannam@167
|
187 E T6i, T7J, T6v, T7x, T39, T57, T3L, T50, T3E, T59, T3O, T4Q, T63, T7H, T6y;
|
cannam@167
|
188 E T7t, T3p, T5a, T3N, T4T, T6a, T7G, T6z, T7q;
|
cannam@167
|
189 {
|
cannam@167
|
190 E T1, T1J, T8, T3Z, T2A, T3Y, T1Q, T3W, T2C, T2D, T3V;
|
cannam@167
|
191 T1 = cr[0];
|
cannam@167
|
192 T1J = ci[WS(rs, 24)];
|
cannam@167
|
193 {
|
cannam@167
|
194 E T2, T3, T4, T5, T6, T7;
|
cannam@167
|
195 T2 = cr[WS(rs, 5)];
|
cannam@167
|
196 T3 = ci[WS(rs, 4)];
|
cannam@167
|
197 T4 = T2 + T3;
|
cannam@167
|
198 T5 = cr[WS(rs, 10)];
|
cannam@167
|
199 T6 = ci[WS(rs, 9)];
|
cannam@167
|
200 T7 = T5 + T6;
|
cannam@167
|
201 T8 = T4 + T7;
|
cannam@167
|
202 T3Z = T5 - T6;
|
cannam@167
|
203 T2A = T4 - T7;
|
cannam@167
|
204 T3Y = T2 - T3;
|
cannam@167
|
205 }
|
cannam@167
|
206 {
|
cannam@167
|
207 E T1K, T1L, T1M, T1N, T1O, T1P;
|
cannam@167
|
208 T1K = ci[WS(rs, 19)];
|
cannam@167
|
209 T1L = cr[WS(rs, 20)];
|
cannam@167
|
210 T1M = T1K - T1L;
|
cannam@167
|
211 T1N = ci[WS(rs, 14)];
|
cannam@167
|
212 T1O = cr[WS(rs, 15)];
|
cannam@167
|
213 T1P = T1N - T1O;
|
cannam@167
|
214 T1Q = T1M + T1P;
|
cannam@167
|
215 T3W = T1M - T1P;
|
cannam@167
|
216 T2C = T1K + T1L;
|
cannam@167
|
217 T2D = T1N + T1O;
|
cannam@167
|
218 }
|
cannam@167
|
219 T9 = T1 + T8;
|
cannam@167
|
220 T40 = FMA(KP618033988, T3Z, T3Y);
|
cannam@167
|
221 T1R = T1J + T1Q;
|
cannam@167
|
222 T6G = FNMS(KP618033988, T3Y, T3Z);
|
cannam@167
|
223 T3V = FNMS(KP250000000, T1Q, T1J);
|
cannam@167
|
224 T6F = FNMS(KP559016994, T3W, T3V);
|
cannam@167
|
225 T3X = FMA(KP559016994, T3W, T3V);
|
cannam@167
|
226 T6H = FNMS(KP951056516, T6G, T6F);
|
cannam@167
|
227 {
|
cannam@167
|
228 E T2E, T5V, T2B, T5U, T2z;
|
cannam@167
|
229 T2E = FMA(KP618033988, T2D, T2C);
|
cannam@167
|
230 T5V = FNMS(KP618033988, T2C, T2D);
|
cannam@167
|
231 T2z = FNMS(KP250000000, T8, T1);
|
cannam@167
|
232 T2B = FMA(KP559016994, T2A, T2z);
|
cannam@167
|
233 T5U = FNMS(KP559016994, T2A, T2z);
|
cannam@167
|
234 T2F = FNMS(KP951056516, T2E, T2B);
|
cannam@167
|
235 T7n = FNMS(KP951056516, T5V, T5U);
|
cannam@167
|
236 T4N = FMA(KP951056516, T2E, T2B);
|
cannam@167
|
237 T5W = FMA(KP951056516, T5V, T5U);
|
cannam@167
|
238 }
|
cannam@167
|
239 }
|
cannam@167
|
240 {
|
cannam@167
|
241 E Ta, T2H, T6n, T2S, Th, T2G, TC, T3r, T5Y, T3C, TJ, T3q, Tj, T30, T6d;
|
cannam@167
|
242 E T33, Tq, T32, T1u, T3v, T61, T3y, T1B, T3x, T12, T2L, T6k, T2O, T19, T2N;
|
cannam@167
|
243 E T1b, T2W, T6g, T37, T1i, T2V, T1l, T3g, T68, T3j, T1s, T3i, Tt, T3c, T65;
|
cannam@167
|
244 E T3n, TA, T3b;
|
cannam@167
|
245 {
|
cannam@167
|
246 E Tg, T2R, Td, T2Q;
|
cannam@167
|
247 Ta = cr[WS(rs, 1)];
|
cannam@167
|
248 {
|
cannam@167
|
249 E Te, Tf, Tb, Tc;
|
cannam@167
|
250 Te = cr[WS(rs, 11)];
|
cannam@167
|
251 Tf = ci[WS(rs, 8)];
|
cannam@167
|
252 Tg = Te + Tf;
|
cannam@167
|
253 T2R = Tf - Te;
|
cannam@167
|
254 Tb = cr[WS(rs, 6)];
|
cannam@167
|
255 Tc = ci[WS(rs, 3)];
|
cannam@167
|
256 Td = Tb + Tc;
|
cannam@167
|
257 T2Q = Tb - Tc;
|
cannam@167
|
258 }
|
cannam@167
|
259 T2H = Td - Tg;
|
cannam@167
|
260 T6n = FMA(KP618033988, T2Q, T2R);
|
cannam@167
|
261 T2S = FNMS(KP618033988, T2R, T2Q);
|
cannam@167
|
262 Th = Td + Tg;
|
cannam@167
|
263 T2G = FNMS(KP250000000, Th, Ta);
|
cannam@167
|
264 }
|
cannam@167
|
265 {
|
cannam@167
|
266 E TI, T3B, TF, T3A;
|
cannam@167
|
267 TC = cr[WS(rs, 3)];
|
cannam@167
|
268 {
|
cannam@167
|
269 E TG, TH, TD, TE;
|
cannam@167
|
270 TG = ci[WS(rs, 11)];
|
cannam@167
|
271 TH = ci[WS(rs, 6)];
|
cannam@167
|
272 TI = TG + TH;
|
cannam@167
|
273 T3B = TG - TH;
|
cannam@167
|
274 TD = cr[WS(rs, 8)];
|
cannam@167
|
275 TE = ci[WS(rs, 1)];
|
cannam@167
|
276 TF = TD + TE;
|
cannam@167
|
277 T3A = TD - TE;
|
cannam@167
|
278 }
|
cannam@167
|
279 T3r = TI - TF;
|
cannam@167
|
280 T5Y = FNMS(KP618033988, T3A, T3B);
|
cannam@167
|
281 T3C = FMA(KP618033988, T3B, T3A);
|
cannam@167
|
282 TJ = TF + TI;
|
cannam@167
|
283 T3q = FNMS(KP250000000, TJ, TC);
|
cannam@167
|
284 }
|
cannam@167
|
285 {
|
cannam@167
|
286 E Tp, T2Z, Tm, T2Y;
|
cannam@167
|
287 Tj = cr[WS(rs, 4)];
|
cannam@167
|
288 {
|
cannam@167
|
289 E Tn, To, Tk, Tl;
|
cannam@167
|
290 Tn = ci[WS(rs, 10)];
|
cannam@167
|
291 To = ci[WS(rs, 5)];
|
cannam@167
|
292 Tp = Tn + To;
|
cannam@167
|
293 T2Z = To - Tn;
|
cannam@167
|
294 Tk = cr[WS(rs, 9)];
|
cannam@167
|
295 Tl = ci[0];
|
cannam@167
|
296 Tm = Tk + Tl;
|
cannam@167
|
297 T2Y = Tl - Tk;
|
cannam@167
|
298 }
|
cannam@167
|
299 T30 = FMA(KP618033988, T2Z, T2Y);
|
cannam@167
|
300 T6d = FNMS(KP618033988, T2Y, T2Z);
|
cannam@167
|
301 T33 = Tm - Tp;
|
cannam@167
|
302 Tq = Tm + Tp;
|
cannam@167
|
303 T32 = FMS(KP250000000, Tq, Tj);
|
cannam@167
|
304 }
|
cannam@167
|
305 {
|
cannam@167
|
306 E T1A, T3u, T1x, T3t;
|
cannam@167
|
307 T1u = ci[WS(rs, 21)];
|
cannam@167
|
308 {
|
cannam@167
|
309 E T1y, T1z, T1v, T1w;
|
cannam@167
|
310 T1y = cr[WS(rs, 13)];
|
cannam@167
|
311 T1z = cr[WS(rs, 18)];
|
cannam@167
|
312 T1A = T1y + T1z;
|
cannam@167
|
313 T3u = T1z - T1y;
|
cannam@167
|
314 T1v = ci[WS(rs, 16)];
|
cannam@167
|
315 T1w = cr[WS(rs, 23)];
|
cannam@167
|
316 T1x = T1v - T1w;
|
cannam@167
|
317 T3t = T1v + T1w;
|
cannam@167
|
318 }
|
cannam@167
|
319 T3v = FMA(KP618033988, T3u, T3t);
|
cannam@167
|
320 T61 = FNMS(KP618033988, T3t, T3u);
|
cannam@167
|
321 T3y = T1x + T1A;
|
cannam@167
|
322 T1B = T1x - T1A;
|
cannam@167
|
323 T3x = FMS(KP250000000, T1B, T1u);
|
cannam@167
|
324 }
|
cannam@167
|
325 {
|
cannam@167
|
326 E T18, T2K, T15, T2J;
|
cannam@167
|
327 T12 = ci[WS(rs, 23)];
|
cannam@167
|
328 {
|
cannam@167
|
329 E T16, T17, T13, T14;
|
cannam@167
|
330 T16 = ci[WS(rs, 13)];
|
cannam@167
|
331 T17 = cr[WS(rs, 16)];
|
cannam@167
|
332 T18 = T16 - T17;
|
cannam@167
|
333 T2K = T16 + T17;
|
cannam@167
|
334 T13 = ci[WS(rs, 18)];
|
cannam@167
|
335 T14 = cr[WS(rs, 21)];
|
cannam@167
|
336 T15 = T13 - T14;
|
cannam@167
|
337 T2J = T13 + T14;
|
cannam@167
|
338 }
|
cannam@167
|
339 T2L = FMA(KP618033988, T2K, T2J);
|
cannam@167
|
340 T6k = FNMS(KP618033988, T2J, T2K);
|
cannam@167
|
341 T2O = T15 - T18;
|
cannam@167
|
342 T19 = T15 + T18;
|
cannam@167
|
343 T2N = FNMS(KP250000000, T19, T12);
|
cannam@167
|
344 }
|
cannam@167
|
345 {
|
cannam@167
|
346 E T1h, T36, T1e, T35;
|
cannam@167
|
347 T1b = ci[WS(rs, 20)];
|
cannam@167
|
348 {
|
cannam@167
|
349 E T1f, T1g, T1c, T1d;
|
cannam@167
|
350 T1f = cr[WS(rs, 14)];
|
cannam@167
|
351 T1g = cr[WS(rs, 19)];
|
cannam@167
|
352 T1h = T1f + T1g;
|
cannam@167
|
353 T36 = T1g - T1f;
|
cannam@167
|
354 T1c = ci[WS(rs, 15)];
|
cannam@167
|
355 T1d = cr[WS(rs, 24)];
|
cannam@167
|
356 T1e = T1c - T1d;
|
cannam@167
|
357 T35 = T1c + T1d;
|
cannam@167
|
358 }
|
cannam@167
|
359 T2W = T1e + T1h;
|
cannam@167
|
360 T6g = FNMS(KP618033988, T35, T36);
|
cannam@167
|
361 T37 = FMA(KP618033988, T36, T35);
|
cannam@167
|
362 T1i = T1e - T1h;
|
cannam@167
|
363 T2V = FMS(KP250000000, T1i, T1b);
|
cannam@167
|
364 }
|
cannam@167
|
365 {
|
cannam@167
|
366 E T1o, T3e, T1r, T3f;
|
cannam@167
|
367 T1l = ci[WS(rs, 22)];
|
cannam@167
|
368 {
|
cannam@167
|
369 E T1m, T1n, T1p, T1q;
|
cannam@167
|
370 T1m = ci[WS(rs, 17)];
|
cannam@167
|
371 T1n = cr[WS(rs, 22)];
|
cannam@167
|
372 T1o = T1m - T1n;
|
cannam@167
|
373 T3e = T1m + T1n;
|
cannam@167
|
374 T1p = ci[WS(rs, 12)];
|
cannam@167
|
375 T1q = cr[WS(rs, 17)];
|
cannam@167
|
376 T1r = T1p - T1q;
|
cannam@167
|
377 T3f = T1p + T1q;
|
cannam@167
|
378 }
|
cannam@167
|
379 T3g = FMA(KP618033988, T3f, T3e);
|
cannam@167
|
380 T68 = FNMS(KP618033988, T3e, T3f);
|
cannam@167
|
381 T3j = T1o - T1r;
|
cannam@167
|
382 T1s = T1o + T1r;
|
cannam@167
|
383 T3i = FMS(KP250000000, T1s, T1l);
|
cannam@167
|
384 }
|
cannam@167
|
385 {
|
cannam@167
|
386 E Tw, T3l, Tz, T3m;
|
cannam@167
|
387 Tt = cr[WS(rs, 2)];
|
cannam@167
|
388 {
|
cannam@167
|
389 E Tu, Tv, Tx, Ty;
|
cannam@167
|
390 Tu = cr[WS(rs, 7)];
|
cannam@167
|
391 Tv = ci[WS(rs, 2)];
|
cannam@167
|
392 Tw = Tu + Tv;
|
cannam@167
|
393 T3l = Tu - Tv;
|
cannam@167
|
394 Tx = cr[WS(rs, 12)];
|
cannam@167
|
395 Ty = ci[WS(rs, 7)];
|
cannam@167
|
396 Tz = Tx + Ty;
|
cannam@167
|
397 T3m = Ty - Tx;
|
cannam@167
|
398 }
|
cannam@167
|
399 T3c = Tz - Tw;
|
cannam@167
|
400 T65 = FMA(KP618033988, T3l, T3m);
|
cannam@167
|
401 T3n = FNMS(KP618033988, T3m, T3l);
|
cannam@167
|
402 TA = Tw + Tz;
|
cannam@167
|
403 T3b = FNMS(KP250000000, TA, Tt);
|
cannam@167
|
404 }
|
cannam@167
|
405 {
|
cannam@167
|
406 E T1a, T1j, T1t, T1C;
|
cannam@167
|
407 T1a = T12 + T19;
|
cannam@167
|
408 T1j = T1b + T1i;
|
cannam@167
|
409 T1k = T1a - T1j;
|
cannam@167
|
410 T1S = T1a + T1j;
|
cannam@167
|
411 T1t = T1l + T1s;
|
cannam@167
|
412 T1C = T1u + T1B;
|
cannam@167
|
413 T1D = T1t - T1C;
|
cannam@167
|
414 T1T = T1t + T1C;
|
cannam@167
|
415 }
|
cannam@167
|
416 Ti = Ta + Th;
|
cannam@167
|
417 Tr = Tj + Tq;
|
cannam@167
|
418 Ts = Ti + Tr;
|
cannam@167
|
419 TB = Tt + TA;
|
cannam@167
|
420 TK = TC + TJ;
|
cannam@167
|
421 TL = TB + TK;
|
cannam@167
|
422 TM = Ts + TL;
|
cannam@167
|
423 {
|
cannam@167
|
424 E T6l, T7y, T6o, T7z, T6j, T6m;
|
cannam@167
|
425 T6j = FNMS(KP559016994, T2H, T2G);
|
cannam@167
|
426 T6l = FMA(KP951056516, T6k, T6j);
|
cannam@167
|
427 T7y = FNMS(KP951056516, T6k, T6j);
|
cannam@167
|
428 T6m = FNMS(KP559016994, T2O, T2N);
|
cannam@167
|
429 T6o = FMA(KP951056516, T6n, T6m);
|
cannam@167
|
430 T7z = FNMS(KP951056516, T6n, T6m);
|
cannam@167
|
431 T6p = FNMS(KP549754652, T6o, T6l);
|
cannam@167
|
432 T7K = FMA(KP939062505, T7y, T7z);
|
cannam@167
|
433 T6w = FMA(KP549754652, T6l, T6o);
|
cannam@167
|
434 T7A = FNMS(KP939062505, T7z, T7y);
|
cannam@167
|
435 }
|
cannam@167
|
436 {
|
cannam@167
|
437 E T2M, T4W, T2T, T4V, T2I, T2P;
|
cannam@167
|
438 T2I = FMA(KP559016994, T2H, T2G);
|
cannam@167
|
439 T2M = FNMS(KP951056516, T2L, T2I);
|
cannam@167
|
440 T4W = FMA(KP951056516, T2L, T2I);
|
cannam@167
|
441 T2P = FMA(KP559016994, T2O, T2N);
|
cannam@167
|
442 T2T = FMA(KP951056516, T2S, T2P);
|
cannam@167
|
443 T4V = FNMS(KP951056516, T2S, T2P);
|
cannam@167
|
444 T2U = FNMS(KP256756360, T2T, T2M);
|
cannam@167
|
445 T56 = FMA(KP634619297, T4V, T4W);
|
cannam@167
|
446 T3K = FMA(KP256756360, T2M, T2T);
|
cannam@167
|
447 T4X = FNMS(KP634619297, T4W, T4V);
|
cannam@167
|
448 }
|
cannam@167
|
449 {
|
cannam@167
|
450 E T6e, T7w, T6h, T7v, T6c, T6f;
|
cannam@167
|
451 T6c = FMA(KP559016994, T2W, T2V);
|
cannam@167
|
452 T6e = FNMS(KP951056516, T6d, T6c);
|
cannam@167
|
453 T7w = FMA(KP951056516, T6d, T6c);
|
cannam@167
|
454 T6f = FMA(KP559016994, T33, T32);
|
cannam@167
|
455 T6h = FNMS(KP951056516, T6g, T6f);
|
cannam@167
|
456 T7v = FMA(KP951056516, T6g, T6f);
|
cannam@167
|
457 T6i = FMA(KP470564281, T6h, T6e);
|
cannam@167
|
458 T7J = FNMS(KP126329378, T7v, T7w);
|
cannam@167
|
459 T6v = FNMS(KP470564281, T6e, T6h);
|
cannam@167
|
460 T7x = FMA(KP126329378, T7w, T7v);
|
cannam@167
|
461 }
|
cannam@167
|
462 {
|
cannam@167
|
463 E T31, T4Y, T38, T4Z, T2X, T34;
|
cannam@167
|
464 T2X = FNMS(KP559016994, T2W, T2V);
|
cannam@167
|
465 T31 = FMA(KP951056516, T30, T2X);
|
cannam@167
|
466 T4Y = FNMS(KP951056516, T30, T2X);
|
cannam@167
|
467 T34 = FNMS(KP559016994, T33, T32);
|
cannam@167
|
468 T38 = FMA(KP951056516, T37, T34);
|
cannam@167
|
469 T4Z = FNMS(KP951056516, T37, T34);
|
cannam@167
|
470 T39 = FNMS(KP634619297, T38, T31);
|
cannam@167
|
471 T57 = FMA(KP827271945, T4Y, T4Z);
|
cannam@167
|
472 T3L = FMA(KP634619297, T31, T38);
|
cannam@167
|
473 T50 = FNMS(KP827271945, T4Z, T4Y);
|
cannam@167
|
474 }
|
cannam@167
|
475 {
|
cannam@167
|
476 E T3w, T4O, T3D, T4P, T3s, T3z;
|
cannam@167
|
477 T3s = FNMS(KP559016994, T3r, T3q);
|
cannam@167
|
478 T3w = FNMS(KP951056516, T3v, T3s);
|
cannam@167
|
479 T4O = FMA(KP951056516, T3v, T3s);
|
cannam@167
|
480 T3z = FNMS(KP559016994, T3y, T3x);
|
cannam@167
|
481 T3D = FNMS(KP951056516, T3C, T3z);
|
cannam@167
|
482 T4P = FMA(KP951056516, T3C, T3z);
|
cannam@167
|
483 T3E = FMA(KP939062505, T3D, T3w);
|
cannam@167
|
484 T59 = FMA(KP126329378, T4O, T4P);
|
cannam@167
|
485 T3O = FNMS(KP939062505, T3w, T3D);
|
cannam@167
|
486 T4Q = FNMS(KP126329378, T4P, T4O);
|
cannam@167
|
487 }
|
cannam@167
|
488 {
|
cannam@167
|
489 E T5Z, T7r, T62, T7s, T5X, T60;
|
cannam@167
|
490 T5X = FMA(KP559016994, T3y, T3x);
|
cannam@167
|
491 T5Z = FMA(KP951056516, T5Y, T5X);
|
cannam@167
|
492 T7r = FNMS(KP951056516, T5Y, T5X);
|
cannam@167
|
493 T60 = FMA(KP559016994, T3r, T3q);
|
cannam@167
|
494 T62 = FMA(KP951056516, T61, T60);
|
cannam@167
|
495 T7s = FNMS(KP951056516, T61, T60);
|
cannam@167
|
496 T63 = FMA(KP062914667, T62, T5Z);
|
cannam@167
|
497 T7H = FMA(KP827271945, T7r, T7s);
|
cannam@167
|
498 T6y = FNMS(KP062914667, T5Z, T62);
|
cannam@167
|
499 T7t = FNMS(KP827271945, T7s, T7r);
|
cannam@167
|
500 }
|
cannam@167
|
501 {
|
cannam@167
|
502 E T3h, T4S, T3o, T4R, T3d, T3k;
|
cannam@167
|
503 T3d = FNMS(KP559016994, T3c, T3b);
|
cannam@167
|
504 T3h = FNMS(KP951056516, T3g, T3d);
|
cannam@167
|
505 T4S = FMA(KP951056516, T3g, T3d);
|
cannam@167
|
506 T3k = FNMS(KP559016994, T3j, T3i);
|
cannam@167
|
507 T3o = FNMS(KP951056516, T3n, T3k);
|
cannam@167
|
508 T4R = FMA(KP951056516, T3n, T3k);
|
cannam@167
|
509 T3p = FMA(KP549754652, T3o, T3h);
|
cannam@167
|
510 T5a = FMA(KP470564281, T4R, T4S);
|
cannam@167
|
511 T3N = FNMS(KP549754652, T3h, T3o);
|
cannam@167
|
512 T4T = FNMS(KP470564281, T4S, T4R);
|
cannam@167
|
513 }
|
cannam@167
|
514 {
|
cannam@167
|
515 E T66, T7o, T69, T7p, T64, T67;
|
cannam@167
|
516 T64 = FMA(KP559016994, T3j, T3i);
|
cannam@167
|
517 T66 = FNMS(KP951056516, T65, T64);
|
cannam@167
|
518 T7o = FMA(KP951056516, T65, T64);
|
cannam@167
|
519 T67 = FMA(KP559016994, T3c, T3b);
|
cannam@167
|
520 T69 = FMA(KP951056516, T68, T67);
|
cannam@167
|
521 T7p = FNMS(KP951056516, T68, T67);
|
cannam@167
|
522 T6a = FMA(KP634619297, T69, T66);
|
cannam@167
|
523 T7G = FNMS(KP062914667, T7o, T7p);
|
cannam@167
|
524 T6z = FNMS(KP634619297, T66, T69);
|
cannam@167
|
525 T7q = FMA(KP062914667, T7p, T7o);
|
cannam@167
|
526 }
|
cannam@167
|
527 }
|
cannam@167
|
528 cr[0] = T9 + TM;
|
cannam@167
|
529 {
|
cannam@167
|
530 E T1U, T1X, T2l, T20, T2m, T1F, T2r, T27, T2h;
|
cannam@167
|
531 {
|
cannam@167
|
532 E T1W, T1V, T1Y, T1Z;
|
cannam@167
|
533 T1W = T1S - T1T;
|
cannam@167
|
534 T1U = T1S + T1T;
|
cannam@167
|
535 T1V = FNMS(KP250000000, T1U, T1R);
|
cannam@167
|
536 T1X = FMA(KP559016994, T1W, T1V);
|
cannam@167
|
537 T2l = FNMS(KP559016994, T1W, T1V);
|
cannam@167
|
538 T1Y = Ti - Tr;
|
cannam@167
|
539 T1Z = TB - TK;
|
cannam@167
|
540 T20 = FMA(KP618033988, T1Z, T1Y);
|
cannam@167
|
541 T2m = FNMS(KP618033988, T1Y, T1Z);
|
cannam@167
|
542 {
|
cannam@167
|
543 E T1E, T2g, T11, T2f, TZ, T10;
|
cannam@167
|
544 T1E = FMA(KP618033988, T1D, T1k);
|
cannam@167
|
545 T2g = FNMS(KP618033988, T1k, T1D);
|
cannam@167
|
546 TZ = FNMS(KP250000000, TM, T9);
|
cannam@167
|
547 T10 = Ts - TL;
|
cannam@167
|
548 T11 = FMA(KP559016994, T10, TZ);
|
cannam@167
|
549 T2f = FNMS(KP559016994, T10, TZ);
|
cannam@167
|
550 T1F = FNMS(KP951056516, T1E, T11);
|
cannam@167
|
551 T2r = FNMS(KP951056516, T2g, T2f);
|
cannam@167
|
552 T27 = FMA(KP951056516, T1E, T11);
|
cannam@167
|
553 T2h = FMA(KP951056516, T2g, T2f);
|
cannam@167
|
554 }
|
cannam@167
|
555 }
|
cannam@167
|
556 {
|
cannam@167
|
557 E T2s, T2w, T2v, T1G, T22, T21;
|
cannam@167
|
558 ci[0] = T1R + T1U;
|
cannam@167
|
559 T2s = T2q * T2r;
|
cannam@167
|
560 T2w = T2u * T2r;
|
cannam@167
|
561 T2v = FMA(KP951056516, T2m, T2l);
|
cannam@167
|
562 cr[WS(rs, 15)] = FNMS(T2u, T2v, T2s);
|
cannam@167
|
563 ci[WS(rs, 15)] = FMA(T2q, T2v, T2w);
|
cannam@167
|
564 T1G = TY * T1F;
|
cannam@167
|
565 T22 = T1I * T1F;
|
cannam@167
|
566 T21 = FMA(KP951056516, T20, T1X);
|
cannam@167
|
567 cr[WS(rs, 5)] = FNMS(T1I, T21, T1G);
|
cannam@167
|
568 ci[WS(rs, 5)] = FMA(TY, T21, T22);
|
cannam@167
|
569 {
|
cannam@167
|
570 E T28, T2c, T2b, T2i, T2o, T2n;
|
cannam@167
|
571 T28 = T26 * T27;
|
cannam@167
|
572 T2c = T2a * T27;
|
cannam@167
|
573 T2b = FNMS(KP951056516, T20, T1X);
|
cannam@167
|
574 cr[WS(rs, 20)] = FNMS(T2a, T2b, T28);
|
cannam@167
|
575 ci[WS(rs, 20)] = FMA(T26, T2b, T2c);
|
cannam@167
|
576 T2i = T2e * T2h;
|
cannam@167
|
577 T2o = T2k * T2h;
|
cannam@167
|
578 T2n = FNMS(KP951056516, T2m, T2l);
|
cannam@167
|
579 cr[WS(rs, 10)] = FNMS(T2k, T2n, T2i);
|
cannam@167
|
580 ci[WS(rs, 10)] = FMA(T2e, T2n, T2o);
|
cannam@167
|
581 }
|
cannam@167
|
582 }
|
cannam@167
|
583 }
|
cannam@167
|
584 {
|
cannam@167
|
585 E T6B, T73, T6Q, T78, T7j, T6u, T72, T7l, T6N, T77, T7k, T7m;
|
cannam@167
|
586 {
|
cannam@167
|
587 E T6x, T6A, T6O, T6P;
|
cannam@167
|
588 T6x = FMA(KP968479752, T6w, T6v);
|
cannam@167
|
589 T6A = FNMS(KP845997307, T6z, T6y);
|
cannam@167
|
590 T6B = FNMS(KP681693190, T6A, T6x);
|
cannam@167
|
591 T73 = FMA(KP560319534, T6x, T6A);
|
cannam@167
|
592 T6O = FNMS(KP968479752, T6p, T6i);
|
cannam@167
|
593 T6P = FNMS(KP845997307, T6a, T63);
|
cannam@167
|
594 T6Q = FMA(KP681693190, T6P, T6O);
|
cannam@167
|
595 T78 = FNMS(KP560319534, T6O, T6P);
|
cannam@167
|
596 }
|
cannam@167
|
597 {
|
cannam@167
|
598 E T6r, T6t, T6b, T6q, T6s;
|
cannam@167
|
599 T6b = FMA(KP845997307, T6a, T63);
|
cannam@167
|
600 T6q = FMA(KP968479752, T6p, T6i);
|
cannam@167
|
601 T6r = FMA(KP906616052, T6q, T6b);
|
cannam@167
|
602 T6t = FNMS(KP906616052, T6q, T6b);
|
cannam@167
|
603 T7j = FMA(KP998026728, T6r, T5W);
|
cannam@167
|
604 T6s = FNMS(KP249506682, T6r, T5W);
|
cannam@167
|
605 T6u = FNMS(KP557913902, T6t, T6s);
|
cannam@167
|
606 T72 = FMA(KP557913902, T6t, T6s);
|
cannam@167
|
607 }
|
cannam@167
|
608 {
|
cannam@167
|
609 E T6K, T6M, T6I, T6J, T6L;
|
cannam@167
|
610 T6I = FMA(KP845997307, T6z, T6y);
|
cannam@167
|
611 T6J = FNMS(KP968479752, T6w, T6v);
|
cannam@167
|
612 T6K = FNMS(KP906616052, T6J, T6I);
|
cannam@167
|
613 T6M = FMA(KP906616052, T6J, T6I);
|
cannam@167
|
614 T7l = FMA(KP998026728, T6K, T6H);
|
cannam@167
|
615 T6L = FNMS(KP249506682, T6K, T6H);
|
cannam@167
|
616 T6N = FNMS(KP557913902, T6M, T6L);
|
cannam@167
|
617 T77 = FMA(KP557913902, T6M, T6L);
|
cannam@167
|
618 }
|
cannam@167
|
619 T7k = T4l * T7j;
|
cannam@167
|
620 cr[WS(rs, 2)] = FNMS(T4n, T7l, T7k);
|
cannam@167
|
621 T7m = T4l * T7l;
|
cannam@167
|
622 ci[WS(rs, 2)] = FMA(T4n, T7j, T7m);
|
cannam@167
|
623 {
|
cannam@167
|
624 E T6C, T6D, T6R, T6S;
|
cannam@167
|
625 T6C = FNMS(KP860541664, T6B, T6u);
|
cannam@167
|
626 T6D = T5T * T6C;
|
cannam@167
|
627 T6R = FNMS(KP860541664, T6Q, T6N);
|
cannam@167
|
628 T6S = T5T * T6R;
|
cannam@167
|
629 cr[WS(rs, 7)] = FNMS(T6E, T6R, T6D);
|
cannam@167
|
630 ci[WS(rs, 7)] = FMA(T6E, T6C, T6S);
|
cannam@167
|
631 }
|
cannam@167
|
632 {
|
cannam@167
|
633 E T7d, T7e, T7h, T7i;
|
cannam@167
|
634 T7d = FMA(KP949179823, T73, T72);
|
cannam@167
|
635 T7e = T7c * T7d;
|
cannam@167
|
636 T7h = FNMS(KP949179823, T78, T77);
|
cannam@167
|
637 T7i = T7c * T7h;
|
cannam@167
|
638 cr[WS(rs, 17)] = FNMS(T7g, T7h, T7e);
|
cannam@167
|
639 ci[WS(rs, 17)] = FMA(T7g, T7d, T7i);
|
cannam@167
|
640 }
|
cannam@167
|
641 {
|
cannam@167
|
642 E T74, T75, T79, T7a;
|
cannam@167
|
643 T74 = FNMS(KP949179823, T73, T72);
|
cannam@167
|
644 T75 = T71 * T74;
|
cannam@167
|
645 T79 = FMA(KP949179823, T78, T77);
|
cannam@167
|
646 T7a = T71 * T79;
|
cannam@167
|
647 cr[WS(rs, 12)] = FNMS(T76, T79, T75);
|
cannam@167
|
648 ci[WS(rs, 12)] = FMA(T76, T74, T7a);
|
cannam@167
|
649 }
|
cannam@167
|
650 {
|
cannam@167
|
651 E T6V, T6W, T6Z, T70;
|
cannam@167
|
652 T6V = FMA(KP860541664, T6B, T6u);
|
cannam@167
|
653 T6W = T6U * T6V;
|
cannam@167
|
654 T6Z = FMA(KP860541664, T6Q, T6N);
|
cannam@167
|
655 T70 = T6U * T6Z;
|
cannam@167
|
656 cr[WS(rs, 22)] = FNMS(T6Y, T6Z, T6W);
|
cannam@167
|
657 ci[WS(rs, 22)] = FMA(T6Y, T6V, T70);
|
cannam@167
|
658 }
|
cannam@167
|
659 }
|
cannam@167
|
660 {
|
cannam@167
|
661 E T7U, T8f, T82, T8k, T7F, T7M, T7X, T7Y, T7D, T7R, T8e, T7I, T7L, T7E, T7O;
|
cannam@167
|
662 E T7N;
|
cannam@167
|
663 {
|
cannam@167
|
664 E T7S, T7T, T80, T81;
|
cannam@167
|
665 T7S = FNMS(KP734762448, T7K, T7J);
|
cannam@167
|
666 T7T = FNMS(KP772036680, T7H, T7G);
|
cannam@167
|
667 T7U = FNMS(KP621716863, T7T, T7S);
|
cannam@167
|
668 T8f = FMA(KP614372930, T7S, T7T);
|
cannam@167
|
669 T80 = FNMS(KP734762448, T7A, T7x);
|
cannam@167
|
670 T81 = FNMS(KP772036680, T7t, T7q);
|
cannam@167
|
671 T82 = FNMS(KP621716863, T81, T80);
|
cannam@167
|
672 T8k = FMA(KP614372930, T80, T81);
|
cannam@167
|
673 }
|
cannam@167
|
674 T7F = FMA(KP951056516, T6G, T6F);
|
cannam@167
|
675 T7I = FMA(KP772036680, T7H, T7G);
|
cannam@167
|
676 T7L = FMA(KP734762448, T7K, T7J);
|
cannam@167
|
677 T7M = FMA(KP994076283, T7L, T7I);
|
cannam@167
|
678 T7X = FNMS(KP249506682, T7M, T7F);
|
cannam@167
|
679 T7Y = FNMS(KP994076283, T7L, T7I);
|
cannam@167
|
680 {
|
cannam@167
|
681 E T7C, T7Q, T7u, T7B, T7P;
|
cannam@167
|
682 T7u = FMA(KP772036680, T7t, T7q);
|
cannam@167
|
683 T7B = FMA(KP734762448, T7A, T7x);
|
cannam@167
|
684 T7C = FMA(KP994076283, T7B, T7u);
|
cannam@167
|
685 T7Q = FNMS(KP994076283, T7B, T7u);
|
cannam@167
|
686 T7D = FMA(KP998026728, T7C, T7n);
|
cannam@167
|
687 T7P = FNMS(KP249506682, T7C, T7n);
|
cannam@167
|
688 T7R = FNMS(KP557913902, T7Q, T7P);
|
cannam@167
|
689 T8e = FMA(KP557913902, T7Q, T7P);
|
cannam@167
|
690 }
|
cannam@167
|
691 T7E = TO * T7D;
|
cannam@167
|
692 T7O = TR * T7D;
|
cannam@167
|
693 T7N = FMA(KP998026728, T7M, T7F);
|
cannam@167
|
694 cr[WS(rs, 3)] = FNMS(TR, T7N, T7E);
|
cannam@167
|
695 ci[WS(rs, 3)] = FMA(TO, T7N, T7O);
|
cannam@167
|
696 {
|
cannam@167
|
697 E T8l, T8t, T8q, T8u, T8h, T8m, T8j, T8p, T8g;
|
cannam@167
|
698 T8j = FMA(KP557913902, T7Y, T7X);
|
cannam@167
|
699 T8l = FNMS(KP949179823, T8k, T8j);
|
cannam@167
|
700 T8t = FMA(KP949179823, T8k, T8j);
|
cannam@167
|
701 T8p = FNMS(KP949179823, T8f, T8e);
|
cannam@167
|
702 T8q = T8o * T8p;
|
cannam@167
|
703 T8u = T8s * T8p;
|
cannam@167
|
704 T8g = FMA(KP949179823, T8f, T8e);
|
cannam@167
|
705 T8h = T8d * T8g;
|
cannam@167
|
706 T8m = T8i * T8g;
|
cannam@167
|
707 cr[WS(rs, 13)] = FNMS(T8i, T8l, T8h);
|
cannam@167
|
708 ci[WS(rs, 13)] = FMA(T8d, T8l, T8m);
|
cannam@167
|
709 cr[WS(rs, 18)] = FNMS(T8s, T8t, T8q);
|
cannam@167
|
710 ci[WS(rs, 18)] = FMA(T8o, T8t, T8u);
|
cannam@167
|
711 }
|
cannam@167
|
712 {
|
cannam@167
|
713 E T83, T8b, T88, T8c, T7W, T84, T7Z, T87, T7V;
|
cannam@167
|
714 T7Z = FNMS(KP557913902, T7Y, T7X);
|
cannam@167
|
715 T83 = FNMS(KP943557151, T82, T7Z);
|
cannam@167
|
716 T8b = FMA(KP943557151, T82, T7Z);
|
cannam@167
|
717 T87 = FNMS(KP943557151, T7U, T7R);
|
cannam@167
|
718 T88 = T86 * T87;
|
cannam@167
|
719 T8c = T8a * T87;
|
cannam@167
|
720 T7V = FMA(KP943557151, T7U, T7R);
|
cannam@167
|
721 T7W = T4z * T7V;
|
cannam@167
|
722 T84 = T4B * T7V;
|
cannam@167
|
723 cr[WS(rs, 8)] = FNMS(T4B, T83, T7W);
|
cannam@167
|
724 ci[WS(rs, 8)] = FMA(T4z, T83, T84);
|
cannam@167
|
725 cr[WS(rs, 23)] = FNMS(T8a, T8b, T88);
|
cannam@167
|
726 ci[WS(rs, 23)] = FMA(T86, T8b, T8c);
|
cannam@167
|
727 }
|
cannam@167
|
728 }
|
cannam@167
|
729 {
|
cannam@167
|
730 E T5c, T5y, T5o, T5E, T5f, T5i, T5j, T5k, T5P, T55, T5x, T5g, T5h, T5Q, T5S;
|
cannam@167
|
731 E T5R;
|
cannam@167
|
732 {
|
cannam@167
|
733 E T58, T5b, T5m, T5n;
|
cannam@167
|
734 T58 = FNMS(KP912575812, T57, T56);
|
cannam@167
|
735 T5b = FNMS(KP912018591, T5a, T59);
|
cannam@167
|
736 T5c = FNMS(KP726211448, T5b, T58);
|
cannam@167
|
737 T5y = FMA(KP525970792, T58, T5b);
|
cannam@167
|
738 T5m = FNMS(KP912575812, T50, T4X);
|
cannam@167
|
739 T5n = FMA(KP912018591, T4T, T4Q);
|
cannam@167
|
740 T5o = FNMS(KP726211448, T5n, T5m);
|
cannam@167
|
741 T5E = FMA(KP525970792, T5m, T5n);
|
cannam@167
|
742 }
|
cannam@167
|
743 T5f = FNMS(KP951056516, T40, T3X);
|
cannam@167
|
744 T5g = FMA(KP912018591, T5a, T59);
|
cannam@167
|
745 T5h = FMA(KP912575812, T57, T56);
|
cannam@167
|
746 T5i = FMA(KP851038619, T5h, T5g);
|
cannam@167
|
747 T5j = FNMS(KP248028675, T5i, T5f);
|
cannam@167
|
748 T5k = FNMS(KP851038619, T5h, T5g);
|
cannam@167
|
749 {
|
cannam@167
|
750 E T52, T54, T4U, T51, T53;
|
cannam@167
|
751 T4U = FNMS(KP912018591, T4T, T4Q);
|
cannam@167
|
752 T51 = FMA(KP912575812, T50, T4X);
|
cannam@167
|
753 T52 = FMA(KP851038619, T51, T4U);
|
cannam@167
|
754 T54 = FNMS(KP851038619, T51, T4U);
|
cannam@167
|
755 T5P = FNMS(KP992114701, T52, T4N);
|
cannam@167
|
756 T53 = FMA(KP248028675, T52, T4N);
|
cannam@167
|
757 T55 = FMA(KP554608978, T54, T53);
|
cannam@167
|
758 T5x = FNMS(KP554608978, T54, T53);
|
cannam@167
|
759 }
|
cannam@167
|
760 T5Q = TS * T5P;
|
cannam@167
|
761 T5S = TW * T5P;
|
cannam@167
|
762 T5R = FMA(KP992114701, T5i, T5f);
|
cannam@167
|
763 cr[WS(rs, 4)] = FNMS(TW, T5R, T5Q);
|
cannam@167
|
764 ci[WS(rs, 4)] = FMA(TS, T5R, T5S);
|
cannam@167
|
765 {
|
cannam@167
|
766 E T5F, T5N, T5K, T5O, T5A, T5G, T5D, T5J, T5z;
|
cannam@167
|
767 T5D = FMA(KP554608978, T5k, T5j);
|
cannam@167
|
768 T5F = FNMS(KP943557151, T5E, T5D);
|
cannam@167
|
769 T5N = FMA(KP943557151, T5E, T5D);
|
cannam@167
|
770 T5J = FMA(KP943557151, T5y, T5x);
|
cannam@167
|
771 T5K = T5I * T5J;
|
cannam@167
|
772 T5O = T5M * T5J;
|
cannam@167
|
773 T5z = FNMS(KP943557151, T5y, T5x);
|
cannam@167
|
774 T5A = T5w * T5z;
|
cannam@167
|
775 T5G = T5C * T5z;
|
cannam@167
|
776 cr[WS(rs, 14)] = FNMS(T5C, T5F, T5A);
|
cannam@167
|
777 ci[WS(rs, 14)] = FMA(T5w, T5F, T5G);
|
cannam@167
|
778 cr[WS(rs, 19)] = FNMS(T5M, T5N, T5K);
|
cannam@167
|
779 ci[WS(rs, 19)] = FMA(T5I, T5N, T5O);
|
cannam@167
|
780 }
|
cannam@167
|
781 {
|
cannam@167
|
782 E T5p, T5t, T5s, T5u, T5e, T5q, T5l, T5r, T5d;
|
cannam@167
|
783 T5l = FNMS(KP554608978, T5k, T5j);
|
cannam@167
|
784 T5p = FNMS(KP803003575, T5o, T5l);
|
cannam@167
|
785 T5t = FMA(KP803003575, T5o, T5l);
|
cannam@167
|
786 T5r = FMA(KP803003575, T5c, T55);
|
cannam@167
|
787 T5s = T23 * T5r;
|
cannam@167
|
788 T5u = T25 * T5r;
|
cannam@167
|
789 T5d = FNMS(KP803003575, T5c, T55);
|
cannam@167
|
790 T5e = TT * T5d;
|
cannam@167
|
791 T5q = TX * T5d;
|
cannam@167
|
792 cr[WS(rs, 9)] = FNMS(TX, T5p, T5e);
|
cannam@167
|
793 ci[WS(rs, 9)] = FMA(TT, T5p, T5q);
|
cannam@167
|
794 cr[WS(rs, 24)] = FNMS(T25, T5t, T5s);
|
cannam@167
|
795 ci[WS(rs, 24)] = FMA(T23, T5t, T5u);
|
cannam@167
|
796 }
|
cannam@167
|
797 }
|
cannam@167
|
798 {
|
cannam@167
|
799 E T3Q, T4q, T4a, T4w, T41, T44, T45, T46, T4J, T3J, T4p, T42, T43, T4K, T4M;
|
cannam@167
|
800 E T4L;
|
cannam@167
|
801 {
|
cannam@167
|
802 E T3M, T3P, T48, T49;
|
cannam@167
|
803 T3M = FMA(KP871714437, T3L, T3K);
|
cannam@167
|
804 T3P = FNMS(KP831864738, T3O, T3N);
|
cannam@167
|
805 T3Q = FNMS(KP559154169, T3P, T3M);
|
cannam@167
|
806 T4q = FMA(KP683113946, T3M, T3P);
|
cannam@167
|
807 T48 = FNMS(KP871714437, T39, T2U);
|
cannam@167
|
808 T49 = FNMS(KP831864738, T3E, T3p);
|
cannam@167
|
809 T4a = FMA(KP559154169, T49, T48);
|
cannam@167
|
810 T4w = FNMS(KP683113946, T48, T49);
|
cannam@167
|
811 }
|
cannam@167
|
812 T41 = FMA(KP951056516, T40, T3X);
|
cannam@167
|
813 T42 = FNMS(KP871714437, T3L, T3K);
|
cannam@167
|
814 T43 = FMA(KP831864738, T3O, T3N);
|
cannam@167
|
815 T44 = FNMS(KP904730450, T43, T42);
|
cannam@167
|
816 T45 = FNMS(KP242145790, T44, T41);
|
cannam@167
|
817 T46 = FMA(KP904730450, T43, T42);
|
cannam@167
|
818 {
|
cannam@167
|
819 E T3G, T3I, T3a, T3F, T3H;
|
cannam@167
|
820 T3a = FMA(KP871714437, T39, T2U);
|
cannam@167
|
821 T3F = FMA(KP831864738, T3E, T3p);
|
cannam@167
|
822 T3G = FMA(KP904730450, T3F, T3a);
|
cannam@167
|
823 T3I = FNMS(KP904730450, T3F, T3a);
|
cannam@167
|
824 T4J = FMA(KP968583161, T3G, T2F);
|
cannam@167
|
825 T3H = FNMS(KP242145790, T3G, T2F);
|
cannam@167
|
826 T3J = FMA(KP541454447, T3I, T3H);
|
cannam@167
|
827 T4p = FNMS(KP541454447, T3I, T3H);
|
cannam@167
|
828 }
|
cannam@167
|
829 T4K = TN * T4J;
|
cannam@167
|
830 T4M = TQ * T4J;
|
cannam@167
|
831 T4L = FMA(KP968583161, T44, T41);
|
cannam@167
|
832 cr[WS(rs, 1)] = FNMS(TQ, T4L, T4K);
|
cannam@167
|
833 ci[WS(rs, 1)] = FMA(TN, T4L, T4M);
|
cannam@167
|
834 {
|
cannam@167
|
835 E T4x, T4H, T4E, T4I, T4s, T4y, T4v, T4D, T4r;
|
cannam@167
|
836 T4v = FNMS(KP541454447, T46, T45);
|
cannam@167
|
837 T4x = FNMS(KP833417178, T4w, T4v);
|
cannam@167
|
838 T4H = FMA(KP833417178, T4w, T4v);
|
cannam@167
|
839 T4D = FMA(KP833417178, T4q, T4p);
|
cannam@167
|
840 T4E = T4C * T4D;
|
cannam@167
|
841 T4I = T4G * T4D;
|
cannam@167
|
842 T4r = FNMS(KP833417178, T4q, T4p);
|
cannam@167
|
843 T4s = T4o * T4r;
|
cannam@167
|
844 T4y = T4u * T4r;
|
cannam@167
|
845 cr[WS(rs, 11)] = FNMS(T4u, T4x, T4s);
|
cannam@167
|
846 ci[WS(rs, 11)] = FMA(T4o, T4x, T4y);
|
cannam@167
|
847 cr[WS(rs, 16)] = FNMS(T4G, T4H, T4E);
|
cannam@167
|
848 ci[WS(rs, 16)] = FMA(T4C, T4H, T4I);
|
cannam@167
|
849 }
|
cannam@167
|
850 {
|
cannam@167
|
851 E T4b, T4j, T4g, T4k, T3S, T4c, T47, T4f, T3R;
|
cannam@167
|
852 T47 = FMA(KP541454447, T46, T45);
|
cannam@167
|
853 T4b = FMA(KP921177326, T4a, T47);
|
cannam@167
|
854 T4j = FNMS(KP921177326, T4a, T47);
|
cannam@167
|
855 T4f = FMA(KP921177326, T3Q, T3J);
|
cannam@167
|
856 T4g = T4e * T4f;
|
cannam@167
|
857 T4k = T4i * T4f;
|
cannam@167
|
858 T3R = FNMS(KP921177326, T3Q, T3J);
|
cannam@167
|
859 T3S = T2y * T3R;
|
cannam@167
|
860 T4c = T3U * T3R;
|
cannam@167
|
861 cr[WS(rs, 6)] = FNMS(T3U, T4b, T3S);
|
cannam@167
|
862 ci[WS(rs, 6)] = FMA(T2y, T4b, T4c);
|
cannam@167
|
863 cr[WS(rs, 21)] = FNMS(T4i, T4j, T4g);
|
cannam@167
|
864 ci[WS(rs, 21)] = FMA(T4e, T4j, T4k);
|
cannam@167
|
865 }
|
cannam@167
|
866 }
|
cannam@167
|
867 }
|
cannam@167
|
868 }
|
cannam@167
|
869 }
|
cannam@167
|
870 }
|
cannam@167
|
871
|
cannam@167
|
872 static const tw_instr twinstr[] = {
|
cannam@167
|
873 {TW_CEXP, 1, 1},
|
cannam@167
|
874 {TW_CEXP, 1, 3},
|
cannam@167
|
875 {TW_CEXP, 1, 9},
|
cannam@167
|
876 {TW_CEXP, 1, 24},
|
cannam@167
|
877 {TW_NEXT, 1, 0}
|
cannam@167
|
878 };
|
cannam@167
|
879
|
cannam@167
|
880 static const hc2hc_desc desc = { 25, "hb2_25", twinstr, &GENUS, {84, 78, 356, 0} };
|
cannam@167
|
881
|
cannam@167
|
882 void X(codelet_hb2_25) (planner *p) {
|
cannam@167
|
883 X(khc2hc_register) (p, hb2_25, &desc);
|
cannam@167
|
884 }
|
cannam@167
|
885 #else
|
cannam@167
|
886
|
cannam@167
|
887 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 25 -dif -name hb2_25 -include rdft/scalar/hb.h */
|
cannam@167
|
888
|
cannam@167
|
889 /*
|
cannam@167
|
890 * This function contains 440 FP additions, 340 FP multiplications,
|
cannam@167
|
891 * (or, 280 additions, 180 multiplications, 160 fused multiply/add),
|
cannam@167
|
892 * 155 stack variables, 20 constants, and 100 memory accesses
|
cannam@167
|
893 */
|
cannam@167
|
894 #include "rdft/scalar/hb.h"
|
cannam@167
|
895
|
cannam@167
|
896 static void hb2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
897 {
|
cannam@167
|
898 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
cannam@167
|
899 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
cannam@167
|
900 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
cannam@167
|
901 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
cannam@167
|
902 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
cannam@167
|
903 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
cannam@167
|
904 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
cannam@167
|
905 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
cannam@167
|
906 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
cannam@167
|
907 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
cannam@167
|
908 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
cannam@167
|
909 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
cannam@167
|
910 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
cannam@167
|
911 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
cannam@167
|
912 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
cannam@167
|
913 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
cannam@167
|
914 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
915 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
916 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
cannam@167
|
917 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
918 {
|
cannam@167
|
919 INT m;
|
cannam@167
|
920 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)) {
|
cannam@167
|
921 E TN, TQ, TO, TR, TT, TY, T2t, T2r, TZ, TU, T4f, T4l, T2d, T4v, T5m;
|
cannam@167
|
922 E T2j, T5l, T4X, T2v, T11, T3R, T1L, T5d, T6x, T5h, T6t, T25, T26, T27, T29;
|
cannam@167
|
923 E T6D, T7v, T49, T7l, T7p, T7t, T2p, T2n, T4b, T4p, T5n, T6B, T5b, T5p, T6p;
|
cannam@167
|
924 E T6r, T59, T4r;
|
cannam@167
|
925 {
|
cannam@167
|
926 E T2c, T4j, T2h, T4e, T2b, T4k, T2i, T4d;
|
cannam@167
|
927 {
|
cannam@167
|
928 E TP, TX, TS, TW;
|
cannam@167
|
929 TN = W[0];
|
cannam@167
|
930 TQ = W[1];
|
cannam@167
|
931 TO = W[2];
|
cannam@167
|
932 TR = W[3];
|
cannam@167
|
933 TP = TN * TO;
|
cannam@167
|
934 TX = TQ * TO;
|
cannam@167
|
935 TS = TQ * TR;
|
cannam@167
|
936 TW = TN * TR;
|
cannam@167
|
937 TT = TP - TS;
|
cannam@167
|
938 TY = TW + TX;
|
cannam@167
|
939 T2t = TW - TX;
|
cannam@167
|
940 T2r = TP + TS;
|
cannam@167
|
941 TZ = W[5];
|
cannam@167
|
942 T2c = TQ * TZ;
|
cannam@167
|
943 T4j = TO * TZ;
|
cannam@167
|
944 T2h = TN * TZ;
|
cannam@167
|
945 T4e = TR * TZ;
|
cannam@167
|
946 TU = W[4];
|
cannam@167
|
947 T2b = TN * TU;
|
cannam@167
|
948 T4k = TR * TU;
|
cannam@167
|
949 T2i = TQ * TU;
|
cannam@167
|
950 T4d = TO * TU;
|
cannam@167
|
951 }
|
cannam@167
|
952 T4f = T4d - T4e;
|
cannam@167
|
953 T4l = T4j + T4k;
|
cannam@167
|
954 {
|
cannam@167
|
955 E T2s, T2u, TV, T10, T3P, T3Q, T1J, T1K;
|
cannam@167
|
956 T2d = T2b - T2c;
|
cannam@167
|
957 T4v = T2b + T2c;
|
cannam@167
|
958 T5m = T4j - T4k;
|
cannam@167
|
959 T2j = T2h + T2i;
|
cannam@167
|
960 T5l = T4d + T4e;
|
cannam@167
|
961 T4X = T2h - T2i;
|
cannam@167
|
962 T2s = T2r * TU;
|
cannam@167
|
963 T2u = T2t * TZ;
|
cannam@167
|
964 T2v = T2s + T2u;
|
cannam@167
|
965 TV = TT * TU;
|
cannam@167
|
966 T10 = TY * TZ;
|
cannam@167
|
967 T11 = TV + T10;
|
cannam@167
|
968 T3P = T2r * TZ;
|
cannam@167
|
969 T3Q = T2t * TU;
|
cannam@167
|
970 T3R = T3P - T3Q;
|
cannam@167
|
971 T1J = TT * TZ;
|
cannam@167
|
972 T1K = TY * TU;
|
cannam@167
|
973 T1L = T1J - T1K;
|
cannam@167
|
974 T5d = TV - T10;
|
cannam@167
|
975 T6x = T3P + T3Q;
|
cannam@167
|
976 T5h = T1J + T1K;
|
cannam@167
|
977 T6t = T2s - T2u;
|
cannam@167
|
978 T25 = W[6];
|
cannam@167
|
979 T26 = W[7];
|
cannam@167
|
980 T27 = FMA(TT, T25, TY * T26);
|
cannam@167
|
981 T29 = FNMS(TY, T25, TT * T26);
|
cannam@167
|
982 T6D = FNMS(T4X, T25, T4v * T26);
|
cannam@167
|
983 T7v = FNMS(T1L, T25, T11 * T26);
|
cannam@167
|
984 T49 = FMA(T2r, T25, T2t * T26);
|
cannam@167
|
985 T7l = FMA(T2d, T25, T2j * T26);
|
cannam@167
|
986 T7p = FNMS(T2j, T25, T2d * T26);
|
cannam@167
|
987 T7t = FMA(T11, T25, T1L * T26);
|
cannam@167
|
988 T2p = FNMS(TZ, T25, TU * T26);
|
cannam@167
|
989 T2n = FMA(TU, T25, TZ * T26);
|
cannam@167
|
990 T4b = FNMS(T2t, T25, T2r * T26);
|
cannam@167
|
991 T4p = FMA(T2v, T25, T3R * T26);
|
cannam@167
|
992 T5n = FMA(T5l, T25, T5m * T26);
|
cannam@167
|
993 T6B = FMA(T4v, T25, T4X * T26);
|
cannam@167
|
994 T5b = FNMS(TQ, T25, TN * T26);
|
cannam@167
|
995 T5p = FNMS(T5m, T25, T5l * T26);
|
cannam@167
|
996 T6p = FMA(TO, T25, TR * T26);
|
cannam@167
|
997 T6r = FNMS(TR, T25, TO * T26);
|
cannam@167
|
998 T59 = FMA(TN, T25, TQ * T26);
|
cannam@167
|
999 T4r = FNMS(T3R, T25, T2v * T26);
|
cannam@167
|
1000 }
|
cannam@167
|
1001 }
|
cannam@167
|
1002 {
|
cannam@167
|
1003 E T9, T6i, T40, T3z, T5Y, Ti, Tr, Ts, T1d, T1m, T1P, T2K, T4P, T3H, T4y;
|
cannam@167
|
1004 E T5G, T71, T65, T6N, T5z, T70, T64, T6K, T2Z, T4Q, T3I, T4B, T20, T5Z, T3C;
|
cannam@167
|
1005 E T43, T6j, TB, TK, TL, T1w, T1F, T1Q, T3f, T4S, T3K, T4F, T5V, T74, T68;
|
cannam@167
|
1006 E T6U, T5O, T73, T67, T6R, T3u, T4T, T3L, T4I;
|
cannam@167
|
1007 {
|
cannam@167
|
1008 E T1, T4, T7, T8, T3Z, T3Y, T3x, T3y;
|
cannam@167
|
1009 T1 = cr[0];
|
cannam@167
|
1010 {
|
cannam@167
|
1011 E T2, T3, T5, T6;
|
cannam@167
|
1012 T2 = cr[WS(rs, 5)];
|
cannam@167
|
1013 T3 = ci[WS(rs, 4)];
|
cannam@167
|
1014 T4 = T2 + T3;
|
cannam@167
|
1015 T5 = cr[WS(rs, 10)];
|
cannam@167
|
1016 T6 = ci[WS(rs, 9)];
|
cannam@167
|
1017 T7 = T5 + T6;
|
cannam@167
|
1018 T8 = T4 + T7;
|
cannam@167
|
1019 T3Z = T5 - T6;
|
cannam@167
|
1020 T3Y = T2 - T3;
|
cannam@167
|
1021 }
|
cannam@167
|
1022 T9 = T1 + T8;
|
cannam@167
|
1023 T6i = FMA(KP951056516, T3Y, KP587785252 * T3Z);
|
cannam@167
|
1024 T40 = FNMS(KP951056516, T3Z, KP587785252 * T3Y);
|
cannam@167
|
1025 T3x = FNMS(KP250000000, T8, T1);
|
cannam@167
|
1026 T3y = KP559016994 * (T4 - T7);
|
cannam@167
|
1027 T3z = T3x - T3y;
|
cannam@167
|
1028 T5Y = T3y + T3x;
|
cannam@167
|
1029 }
|
cannam@167
|
1030 {
|
cannam@167
|
1031 E Ta, T2x, T5w, T2F, Th, T2w, T1e, T2P, T5B, T2X, T1l, T2O, Tj, T2N, T5D;
|
cannam@167
|
1032 E T2T, Tq, T2S, T15, T2B, T5u, T2H, T1c, T2G;
|
cannam@167
|
1033 {
|
cannam@167
|
1034 E Tg, T2E, Td, T2D;
|
cannam@167
|
1035 Ta = cr[WS(rs, 1)];
|
cannam@167
|
1036 {
|
cannam@167
|
1037 E Te, Tf, Tb, Tc;
|
cannam@167
|
1038 Te = cr[WS(rs, 11)];
|
cannam@167
|
1039 Tf = ci[WS(rs, 8)];
|
cannam@167
|
1040 Tg = Te + Tf;
|
cannam@167
|
1041 T2E = Te - Tf;
|
cannam@167
|
1042 Tb = cr[WS(rs, 6)];
|
cannam@167
|
1043 Tc = ci[WS(rs, 3)];
|
cannam@167
|
1044 Td = Tb + Tc;
|
cannam@167
|
1045 T2D = Tb - Tc;
|
cannam@167
|
1046 }
|
cannam@167
|
1047 T2x = KP559016994 * (Td - Tg);
|
cannam@167
|
1048 T5w = FMA(KP951056516, T2D, KP587785252 * T2E);
|
cannam@167
|
1049 T2F = FNMS(KP951056516, T2E, KP587785252 * T2D);
|
cannam@167
|
1050 Th = Td + Tg;
|
cannam@167
|
1051 T2w = FNMS(KP250000000, Th, Ta);
|
cannam@167
|
1052 }
|
cannam@167
|
1053 {
|
cannam@167
|
1054 E T1k, T2W, T1h, T2V;
|
cannam@167
|
1055 T1e = ci[WS(rs, 20)];
|
cannam@167
|
1056 {
|
cannam@167
|
1057 E T1i, T1j, T1f, T1g;
|
cannam@167
|
1058 T1i = cr[WS(rs, 14)];
|
cannam@167
|
1059 T1j = cr[WS(rs, 19)];
|
cannam@167
|
1060 T1k = T1i + T1j;
|
cannam@167
|
1061 T2W = T1j - T1i;
|
cannam@167
|
1062 T1f = ci[WS(rs, 15)];
|
cannam@167
|
1063 T1g = cr[WS(rs, 24)];
|
cannam@167
|
1064 T1h = T1f - T1g;
|
cannam@167
|
1065 T2V = T1f + T1g;
|
cannam@167
|
1066 }
|
cannam@167
|
1067 T2P = KP559016994 * (T1h + T1k);
|
cannam@167
|
1068 T5B = FMA(KP951056516, T2V, KP587785252 * T2W);
|
cannam@167
|
1069 T2X = FNMS(KP951056516, T2W, KP587785252 * T2V);
|
cannam@167
|
1070 T1l = T1h - T1k;
|
cannam@167
|
1071 T2O = FNMS(KP250000000, T1l, T1e);
|
cannam@167
|
1072 }
|
cannam@167
|
1073 {
|
cannam@167
|
1074 E Tp, T2M, Tm, T2L;
|
cannam@167
|
1075 Tj = cr[WS(rs, 4)];
|
cannam@167
|
1076 {
|
cannam@167
|
1077 E Tn, To, Tk, Tl;
|
cannam@167
|
1078 Tn = ci[WS(rs, 10)];
|
cannam@167
|
1079 To = ci[WS(rs, 5)];
|
cannam@167
|
1080 Tp = Tn + To;
|
cannam@167
|
1081 T2M = Tn - To;
|
cannam@167
|
1082 Tk = cr[WS(rs, 9)];
|
cannam@167
|
1083 Tl = ci[0];
|
cannam@167
|
1084 Tm = Tk + Tl;
|
cannam@167
|
1085 T2L = Tk - Tl;
|
cannam@167
|
1086 }
|
cannam@167
|
1087 T2N = FNMS(KP951056516, T2M, KP587785252 * T2L);
|
cannam@167
|
1088 T5D = FMA(KP951056516, T2L, KP587785252 * T2M);
|
cannam@167
|
1089 T2T = KP559016994 * (Tm - Tp);
|
cannam@167
|
1090 Tq = Tm + Tp;
|
cannam@167
|
1091 T2S = FNMS(KP250000000, Tq, Tj);
|
cannam@167
|
1092 }
|
cannam@167
|
1093 {
|
cannam@167
|
1094 E T1b, T2A, T18, T2z;
|
cannam@167
|
1095 T15 = ci[WS(rs, 23)];
|
cannam@167
|
1096 {
|
cannam@167
|
1097 E T19, T1a, T16, T17;
|
cannam@167
|
1098 T19 = ci[WS(rs, 13)];
|
cannam@167
|
1099 T1a = cr[WS(rs, 16)];
|
cannam@167
|
1100 T1b = T19 - T1a;
|
cannam@167
|
1101 T2A = T19 + T1a;
|
cannam@167
|
1102 T16 = ci[WS(rs, 18)];
|
cannam@167
|
1103 T17 = cr[WS(rs, 21)];
|
cannam@167
|
1104 T18 = T16 - T17;
|
cannam@167
|
1105 T2z = T16 + T17;
|
cannam@167
|
1106 }
|
cannam@167
|
1107 T2B = FNMS(KP951056516, T2A, KP587785252 * T2z);
|
cannam@167
|
1108 T5u = FMA(KP951056516, T2z, KP587785252 * T2A);
|
cannam@167
|
1109 T2H = KP559016994 * (T18 - T1b);
|
cannam@167
|
1110 T1c = T18 + T1b;
|
cannam@167
|
1111 T2G = FNMS(KP250000000, T1c, T15);
|
cannam@167
|
1112 }
|
cannam@167
|
1113 Ti = Ta + Th;
|
cannam@167
|
1114 Tr = Tj + Tq;
|
cannam@167
|
1115 Ts = Ti + Tr;
|
cannam@167
|
1116 T1d = T15 + T1c;
|
cannam@167
|
1117 T1m = T1e + T1l;
|
cannam@167
|
1118 T1P = T1d + T1m;
|
cannam@167
|
1119 {
|
cannam@167
|
1120 E T2C, T4w, T2J, T4x, T2y, T2I;
|
cannam@167
|
1121 T2y = T2w - T2x;
|
cannam@167
|
1122 T2C = T2y - T2B;
|
cannam@167
|
1123 T4w = T2y + T2B;
|
cannam@167
|
1124 T2I = T2G - T2H;
|
cannam@167
|
1125 T2J = T2F + T2I;
|
cannam@167
|
1126 T4x = T2I - T2F;
|
cannam@167
|
1127 T2K = FNMS(KP481753674, T2J, KP876306680 * T2C);
|
cannam@167
|
1128 T4P = FMA(KP728968627, T4x, KP684547105 * T4w);
|
cannam@167
|
1129 T3H = FMA(KP876306680, T2J, KP481753674 * T2C);
|
cannam@167
|
1130 T4y = FNMS(KP684547105, T4x, KP728968627 * T4w);
|
cannam@167
|
1131 }
|
cannam@167
|
1132 {
|
cannam@167
|
1133 E T5C, T6M, T5F, T6L, T5A, T5E;
|
cannam@167
|
1134 T5A = T2T + T2S;
|
cannam@167
|
1135 T5C = T5A - T5B;
|
cannam@167
|
1136 T6M = T5A + T5B;
|
cannam@167
|
1137 T5E = T2O + T2P;
|
cannam@167
|
1138 T5F = T5D + T5E;
|
cannam@167
|
1139 T6L = T5E - T5D;
|
cannam@167
|
1140 T5G = FNMS(KP844327925, T5F, KP535826794 * T5C);
|
cannam@167
|
1141 T71 = FMA(KP637423989, T6L, KP770513242 * T6M);
|
cannam@167
|
1142 T65 = FMA(KP535826794, T5F, KP844327925 * T5C);
|
cannam@167
|
1143 T6N = FNMS(KP637423989, T6M, KP770513242 * T6L);
|
cannam@167
|
1144 }
|
cannam@167
|
1145 {
|
cannam@167
|
1146 E T5v, T6I, T5y, T6J, T5t, T5x;
|
cannam@167
|
1147 T5t = T2x + T2w;
|
cannam@167
|
1148 T5v = T5t - T5u;
|
cannam@167
|
1149 T6I = T5t + T5u;
|
cannam@167
|
1150 T5x = T2H + T2G;
|
cannam@167
|
1151 T5y = T5w + T5x;
|
cannam@167
|
1152 T6J = T5x - T5w;
|
cannam@167
|
1153 T5z = FNMS(KP248689887, T5y, KP968583161 * T5v);
|
cannam@167
|
1154 T70 = FMA(KP535826794, T6J, KP844327925 * T6I);
|
cannam@167
|
1155 T64 = FMA(KP968583161, T5y, KP248689887 * T5v);
|
cannam@167
|
1156 T6K = FNMS(KP844327925, T6J, KP535826794 * T6I);
|
cannam@167
|
1157 }
|
cannam@167
|
1158 {
|
cannam@167
|
1159 E T2R, T4z, T2Y, T4A, T2Q, T2U;
|
cannam@167
|
1160 T2Q = T2O - T2P;
|
cannam@167
|
1161 T2R = T2N + T2Q;
|
cannam@167
|
1162 T4z = T2Q - T2N;
|
cannam@167
|
1163 T2U = T2S - T2T;
|
cannam@167
|
1164 T2Y = T2U - T2X;
|
cannam@167
|
1165 T4A = T2U + T2X;
|
cannam@167
|
1166 T2Z = FMA(KP904827052, T2R, KP425779291 * T2Y);
|
cannam@167
|
1167 T4Q = FNMS(KP992114701, T4z, KP125333233 * T4A);
|
cannam@167
|
1168 T3I = FNMS(KP425779291, T2R, KP904827052 * T2Y);
|
cannam@167
|
1169 T4B = FMA(KP125333233, T4z, KP992114701 * T4A);
|
cannam@167
|
1170 }
|
cannam@167
|
1171 }
|
cannam@167
|
1172 {
|
cannam@167
|
1173 E T1S, T1V, T1Y, T1Z, T3B, T3A, T41, T42;
|
cannam@167
|
1174 T1S = ci[WS(rs, 24)];
|
cannam@167
|
1175 {
|
cannam@167
|
1176 E T1T, T1U, T1W, T1X;
|
cannam@167
|
1177 T1T = ci[WS(rs, 19)];
|
cannam@167
|
1178 T1U = cr[WS(rs, 20)];
|
cannam@167
|
1179 T1V = T1T - T1U;
|
cannam@167
|
1180 T1W = ci[WS(rs, 14)];
|
cannam@167
|
1181 T1X = cr[WS(rs, 15)];
|
cannam@167
|
1182 T1Y = T1W - T1X;
|
cannam@167
|
1183 T1Z = T1V + T1Y;
|
cannam@167
|
1184 T3B = T1W + T1X;
|
cannam@167
|
1185 T3A = T1T + T1U;
|
cannam@167
|
1186 }
|
cannam@167
|
1187 T20 = T1S + T1Z;
|
cannam@167
|
1188 T5Z = FMA(KP951056516, T3A, KP587785252 * T3B);
|
cannam@167
|
1189 T3C = FNMS(KP951056516, T3B, KP587785252 * T3A);
|
cannam@167
|
1190 T41 = FNMS(KP250000000, T1Z, T1S);
|
cannam@167
|
1191 T42 = KP559016994 * (T1V - T1Y);
|
cannam@167
|
1192 T43 = T41 - T42;
|
cannam@167
|
1193 T6j = T42 + T41;
|
cannam@167
|
1194 }
|
cannam@167
|
1195 {
|
cannam@167
|
1196 E Tt, T32, T5L, T3a, TA, T31, T1o, T36, T5J, T3c, T1v, T3b, TC, T3h, T5S;
|
cannam@167
|
1197 E T3p, TJ, T3g, T1x, T3l, T5Q, T3r, T1E, T3q;
|
cannam@167
|
1198 {
|
cannam@167
|
1199 E Tw, T38, Tz, T39;
|
cannam@167
|
1200 Tt = cr[WS(rs, 2)];
|
cannam@167
|
1201 {
|
cannam@167
|
1202 E Tu, Tv, Tx, Ty;
|
cannam@167
|
1203 Tu = cr[WS(rs, 7)];
|
cannam@167
|
1204 Tv = ci[WS(rs, 2)];
|
cannam@167
|
1205 Tw = Tu + Tv;
|
cannam@167
|
1206 T38 = Tu - Tv;
|
cannam@167
|
1207 Tx = cr[WS(rs, 12)];
|
cannam@167
|
1208 Ty = ci[WS(rs, 7)];
|
cannam@167
|
1209 Tz = Tx + Ty;
|
cannam@167
|
1210 T39 = Tx - Ty;
|
cannam@167
|
1211 }
|
cannam@167
|
1212 T32 = KP559016994 * (Tw - Tz);
|
cannam@167
|
1213 T5L = FMA(KP951056516, T38, KP587785252 * T39);
|
cannam@167
|
1214 T3a = FNMS(KP951056516, T39, KP587785252 * T38);
|
cannam@167
|
1215 TA = Tw + Tz;
|
cannam@167
|
1216 T31 = FNMS(KP250000000, TA, Tt);
|
cannam@167
|
1217 }
|
cannam@167
|
1218 {
|
cannam@167
|
1219 E T1r, T34, T1u, T35;
|
cannam@167
|
1220 T1o = ci[WS(rs, 22)];
|
cannam@167
|
1221 {
|
cannam@167
|
1222 E T1p, T1q, T1s, T1t;
|
cannam@167
|
1223 T1p = ci[WS(rs, 17)];
|
cannam@167
|
1224 T1q = cr[WS(rs, 22)];
|
cannam@167
|
1225 T1r = T1p - T1q;
|
cannam@167
|
1226 T34 = T1p + T1q;
|
cannam@167
|
1227 T1s = ci[WS(rs, 12)];
|
cannam@167
|
1228 T1t = cr[WS(rs, 17)];
|
cannam@167
|
1229 T1u = T1s - T1t;
|
cannam@167
|
1230 T35 = T1s + T1t;
|
cannam@167
|
1231 }
|
cannam@167
|
1232 T36 = FNMS(KP951056516, T35, KP587785252 * T34);
|
cannam@167
|
1233 T5J = FMA(KP951056516, T34, KP587785252 * T35);
|
cannam@167
|
1234 T3c = KP559016994 * (T1r - T1u);
|
cannam@167
|
1235 T1v = T1r + T1u;
|
cannam@167
|
1236 T3b = FNMS(KP250000000, T1v, T1o);
|
cannam@167
|
1237 }
|
cannam@167
|
1238 {
|
cannam@167
|
1239 E TI, T3o, TF, T3n;
|
cannam@167
|
1240 TC = cr[WS(rs, 3)];
|
cannam@167
|
1241 {
|
cannam@167
|
1242 E TG, TH, TD, TE;
|
cannam@167
|
1243 TG = ci[WS(rs, 11)];
|
cannam@167
|
1244 TH = ci[WS(rs, 6)];
|
cannam@167
|
1245 TI = TG + TH;
|
cannam@167
|
1246 T3o = TG - TH;
|
cannam@167
|
1247 TD = cr[WS(rs, 8)];
|
cannam@167
|
1248 TE = ci[WS(rs, 1)];
|
cannam@167
|
1249 TF = TD + TE;
|
cannam@167
|
1250 T3n = TD - TE;
|
cannam@167
|
1251 }
|
cannam@167
|
1252 T3h = KP559016994 * (TF - TI);
|
cannam@167
|
1253 T5S = FMA(KP951056516, T3n, KP587785252 * T3o);
|
cannam@167
|
1254 T3p = FNMS(KP951056516, T3o, KP587785252 * T3n);
|
cannam@167
|
1255 TJ = TF + TI;
|
cannam@167
|
1256 T3g = FNMS(KP250000000, TJ, TC);
|
cannam@167
|
1257 }
|
cannam@167
|
1258 {
|
cannam@167
|
1259 E T1D, T3k, T1A, T3j;
|
cannam@167
|
1260 T1x = ci[WS(rs, 21)];
|
cannam@167
|
1261 {
|
cannam@167
|
1262 E T1B, T1C, T1y, T1z;
|
cannam@167
|
1263 T1B = cr[WS(rs, 13)];
|
cannam@167
|
1264 T1C = cr[WS(rs, 18)];
|
cannam@167
|
1265 T1D = T1B + T1C;
|
cannam@167
|
1266 T3k = T1C - T1B;
|
cannam@167
|
1267 T1y = ci[WS(rs, 16)];
|
cannam@167
|
1268 T1z = cr[WS(rs, 23)];
|
cannam@167
|
1269 T1A = T1y - T1z;
|
cannam@167
|
1270 T3j = T1y + T1z;
|
cannam@167
|
1271 }
|
cannam@167
|
1272 T3l = FNMS(KP951056516, T3k, KP587785252 * T3j);
|
cannam@167
|
1273 T5Q = FMA(KP951056516, T3j, KP587785252 * T3k);
|
cannam@167
|
1274 T3r = KP559016994 * (T1A + T1D);
|
cannam@167
|
1275 T1E = T1A - T1D;
|
cannam@167
|
1276 T3q = FNMS(KP250000000, T1E, T1x);
|
cannam@167
|
1277 }
|
cannam@167
|
1278 TB = Tt + TA;
|
cannam@167
|
1279 TK = TC + TJ;
|
cannam@167
|
1280 TL = TB + TK;
|
cannam@167
|
1281 T1w = T1o + T1v;
|
cannam@167
|
1282 T1F = T1x + T1E;
|
cannam@167
|
1283 T1Q = T1w + T1F;
|
cannam@167
|
1284 {
|
cannam@167
|
1285 E T37, T4D, T3e, T4E, T33, T3d;
|
cannam@167
|
1286 T33 = T31 - T32;
|
cannam@167
|
1287 T37 = T33 - T36;
|
cannam@167
|
1288 T4D = T33 + T36;
|
cannam@167
|
1289 T3d = T3b - T3c;
|
cannam@167
|
1290 T3e = T3a + T3d;
|
cannam@167
|
1291 T4E = T3d - T3a;
|
cannam@167
|
1292 T3f = FNMS(KP844327925, T3e, KP535826794 * T37);
|
cannam@167
|
1293 T4S = FMA(KP062790519, T4E, KP998026728 * T4D);
|
cannam@167
|
1294 T3K = FMA(KP535826794, T3e, KP844327925 * T37);
|
cannam@167
|
1295 T4F = FNMS(KP998026728, T4E, KP062790519 * T4D);
|
cannam@167
|
1296 }
|
cannam@167
|
1297 {
|
cannam@167
|
1298 E T5R, T6T, T5U, T6S, T5P, T5T;
|
cannam@167
|
1299 T5P = T3h + T3g;
|
cannam@167
|
1300 T5R = T5P - T5Q;
|
cannam@167
|
1301 T6T = T5P + T5Q;
|
cannam@167
|
1302 T5T = T3q + T3r;
|
cannam@167
|
1303 T5U = T5S + T5T;
|
cannam@167
|
1304 T6S = T5T - T5S;
|
cannam@167
|
1305 T5V = FNMS(KP684547105, T5U, KP728968627 * T5R);
|
cannam@167
|
1306 T74 = FNMS(KP992114701, T6S, KP125333233 * T6T);
|
cannam@167
|
1307 T68 = FMA(KP728968627, T5U, KP684547105 * T5R);
|
cannam@167
|
1308 T6U = FMA(KP125333233, T6S, KP992114701 * T6T);
|
cannam@167
|
1309 }
|
cannam@167
|
1310 {
|
cannam@167
|
1311 E T5K, T6Q, T5N, T6P, T5I, T5M;
|
cannam@167
|
1312 T5I = T32 + T31;
|
cannam@167
|
1313 T5K = T5I - T5J;
|
cannam@167
|
1314 T6Q = T5I + T5J;
|
cannam@167
|
1315 T5M = T3c + T3b;
|
cannam@167
|
1316 T5N = T5L + T5M;
|
cannam@167
|
1317 T6P = T5M - T5L;
|
cannam@167
|
1318 T5O = FNMS(KP481753674, T5N, KP876306680 * T5K);
|
cannam@167
|
1319 T73 = FNMS(KP425779291, T6P, KP904827052 * T6Q);
|
cannam@167
|
1320 T67 = FMA(KP876306680, T5N, KP481753674 * T5K);
|
cannam@167
|
1321 T6R = FMA(KP904827052, T6P, KP425779291 * T6Q);
|
cannam@167
|
1322 }
|
cannam@167
|
1323 {
|
cannam@167
|
1324 E T3m, T4H, T3t, T4G, T3i, T3s;
|
cannam@167
|
1325 T3i = T3g - T3h;
|
cannam@167
|
1326 T3m = T3i - T3l;
|
cannam@167
|
1327 T4H = T3i + T3l;
|
cannam@167
|
1328 T3s = T3q - T3r;
|
cannam@167
|
1329 T3t = T3p + T3s;
|
cannam@167
|
1330 T4G = T3s - T3p;
|
cannam@167
|
1331 T3u = FNMS(KP998026728, T3t, KP062790519 * T3m);
|
cannam@167
|
1332 T4T = FNMS(KP637423989, T4G, KP770513242 * T4H);
|
cannam@167
|
1333 T3L = FMA(KP062790519, T3t, KP998026728 * T3m);
|
cannam@167
|
1334 T4I = FMA(KP770513242, T4G, KP637423989 * T4H);
|
cannam@167
|
1335 }
|
cannam@167
|
1336 }
|
cannam@167
|
1337 {
|
cannam@167
|
1338 E TM, T14, T2e, T21, T23, T2l, T1H, T2f, T1O, T2k;
|
cannam@167
|
1339 {
|
cannam@167
|
1340 E T12, T13, T1R, T22;
|
cannam@167
|
1341 T12 = KP559016994 * (Ts - TL);
|
cannam@167
|
1342 TM = Ts + TL;
|
cannam@167
|
1343 T13 = FNMS(KP250000000, TM, T9);
|
cannam@167
|
1344 T14 = T12 + T13;
|
cannam@167
|
1345 T2e = T13 - T12;
|
cannam@167
|
1346 T1R = KP559016994 * (T1P - T1Q);
|
cannam@167
|
1347 T21 = T1P + T1Q;
|
cannam@167
|
1348 T22 = FNMS(KP250000000, T21, T20);
|
cannam@167
|
1349 T23 = T1R + T22;
|
cannam@167
|
1350 T2l = T22 - T1R;
|
cannam@167
|
1351 }
|
cannam@167
|
1352 {
|
cannam@167
|
1353 E T1n, T1G, T1M, T1N;
|
cannam@167
|
1354 T1n = T1d - T1m;
|
cannam@167
|
1355 T1G = T1w - T1F;
|
cannam@167
|
1356 T1H = FMA(KP951056516, T1n, KP587785252 * T1G);
|
cannam@167
|
1357 T2f = FNMS(KP951056516, T1G, KP587785252 * T1n);
|
cannam@167
|
1358 T1M = Ti - Tr;
|
cannam@167
|
1359 T1N = TB - TK;
|
cannam@167
|
1360 T1O = FMA(KP951056516, T1M, KP587785252 * T1N);
|
cannam@167
|
1361 T2k = FNMS(KP951056516, T1N, KP587785252 * T1M);
|
cannam@167
|
1362 }
|
cannam@167
|
1363 {
|
cannam@167
|
1364 E T1I, T24, T2o, T2q;
|
cannam@167
|
1365 cr[0] = T9 + TM;
|
cannam@167
|
1366 ci[0] = T20 + T21;
|
cannam@167
|
1367 T1I = T14 - T1H;
|
cannam@167
|
1368 T24 = T1O + T23;
|
cannam@167
|
1369 cr[WS(rs, 5)] = FNMS(T1L, T24, T11 * T1I);
|
cannam@167
|
1370 ci[WS(rs, 5)] = FMA(T1L, T1I, T11 * T24);
|
cannam@167
|
1371 T2o = T2e + T2f;
|
cannam@167
|
1372 T2q = T2l - T2k;
|
cannam@167
|
1373 cr[WS(rs, 15)] = FNMS(T2p, T2q, T2n * T2o);
|
cannam@167
|
1374 ci[WS(rs, 15)] = FMA(T2p, T2o, T2n * T2q);
|
cannam@167
|
1375 {
|
cannam@167
|
1376 E T2g, T2m, T28, T2a;
|
cannam@167
|
1377 T2g = T2e - T2f;
|
cannam@167
|
1378 T2m = T2k + T2l;
|
cannam@167
|
1379 cr[WS(rs, 10)] = FNMS(T2j, T2m, T2d * T2g);
|
cannam@167
|
1380 ci[WS(rs, 10)] = FMA(T2j, T2g, T2d * T2m);
|
cannam@167
|
1381 T28 = T14 + T1H;
|
cannam@167
|
1382 T2a = T23 - T1O;
|
cannam@167
|
1383 cr[WS(rs, 20)] = FNMS(T29, T2a, T27 * T28);
|
cannam@167
|
1384 ci[WS(rs, 20)] = FMA(T29, T28, T27 * T2a);
|
cannam@167
|
1385 }
|
cannam@167
|
1386 }
|
cannam@167
|
1387 }
|
cannam@167
|
1388 {
|
cannam@167
|
1389 E T76, T7n, T7a, T7q, T6H, T6W, T6X, T6Y, T7e, T7f, T7d, T7g, T7x, T7y;
|
cannam@167
|
1390 {
|
cannam@167
|
1391 E T72, T75, T78, T79;
|
cannam@167
|
1392 T72 = T70 + T71;
|
cannam@167
|
1393 T75 = T73 - T74;
|
cannam@167
|
1394 T76 = FMA(KP951056516, T72, KP587785252 * T75);
|
cannam@167
|
1395 T7n = FNMS(KP951056516, T75, KP587785252 * T72);
|
cannam@167
|
1396 T78 = T6K - T6N;
|
cannam@167
|
1397 T79 = T6U - T6R;
|
cannam@167
|
1398 T7a = FMA(KP951056516, T78, KP587785252 * T79);
|
cannam@167
|
1399 T7q = FNMS(KP951056516, T79, KP587785252 * T78);
|
cannam@167
|
1400 }
|
cannam@167
|
1401 {
|
cannam@167
|
1402 E T6O, T6V, T7b, T7c;
|
cannam@167
|
1403 T6H = T5Y + T5Z;
|
cannam@167
|
1404 T6O = T6K + T6N;
|
cannam@167
|
1405 T6V = T6R + T6U;
|
cannam@167
|
1406 T6W = T6O - T6V;
|
cannam@167
|
1407 T6X = FNMS(KP250000000, T6W, T6H);
|
cannam@167
|
1408 T6Y = KP559016994 * (T6O + T6V);
|
cannam@167
|
1409 T7e = T6j - T6i;
|
cannam@167
|
1410 T7b = T70 - T71;
|
cannam@167
|
1411 T7c = T73 + T74;
|
cannam@167
|
1412 T7f = T7b + T7c;
|
cannam@167
|
1413 T7d = KP559016994 * (T7b - T7c);
|
cannam@167
|
1414 T7g = FNMS(KP250000000, T7f, T7e);
|
cannam@167
|
1415 }
|
cannam@167
|
1416 T7x = T6H + T6W;
|
cannam@167
|
1417 T7y = T7e + T7f;
|
cannam@167
|
1418 cr[WS(rs, 4)] = FNMS(TY, T7y, TT * T7x);
|
cannam@167
|
1419 ci[WS(rs, 4)] = FMA(TY, T7x, TT * T7y);
|
cannam@167
|
1420 {
|
cannam@167
|
1421 E T7o, T7u, T7s, T7w, T7m, T7r;
|
cannam@167
|
1422 T7m = T6X - T6Y;
|
cannam@167
|
1423 T7o = T7m - T7n;
|
cannam@167
|
1424 T7u = T7m + T7n;
|
cannam@167
|
1425 T7r = T7g - T7d;
|
cannam@167
|
1426 T7s = T7q + T7r;
|
cannam@167
|
1427 T7w = T7r - T7q;
|
cannam@167
|
1428 cr[WS(rs, 14)] = FNMS(T7p, T7s, T7l * T7o);
|
cannam@167
|
1429 ci[WS(rs, 14)] = FMA(T7p, T7o, T7l * T7s);
|
cannam@167
|
1430 cr[WS(rs, 19)] = FNMS(T7v, T7w, T7t * T7u);
|
cannam@167
|
1431 ci[WS(rs, 19)] = FMA(T7v, T7u, T7t * T7w);
|
cannam@167
|
1432 }
|
cannam@167
|
1433 {
|
cannam@167
|
1434 E T77, T7j, T7i, T7k, T6Z, T7h;
|
cannam@167
|
1435 T6Z = T6X + T6Y;
|
cannam@167
|
1436 T77 = T6Z - T76;
|
cannam@167
|
1437 T7j = T6Z + T76;
|
cannam@167
|
1438 T7h = T7d + T7g;
|
cannam@167
|
1439 T7i = T7a + T7h;
|
cannam@167
|
1440 T7k = T7h - T7a;
|
cannam@167
|
1441 cr[WS(rs, 9)] = FNMS(TZ, T7i, TU * T77);
|
cannam@167
|
1442 ci[WS(rs, 9)] = FMA(TZ, T77, TU * T7i);
|
cannam@167
|
1443 cr[WS(rs, 24)] = FNMS(T26, T7k, T25 * T7j);
|
cannam@167
|
1444 ci[WS(rs, 24)] = FMA(T26, T7j, T25 * T7k);
|
cannam@167
|
1445 }
|
cannam@167
|
1446 }
|
cannam@167
|
1447 {
|
cannam@167
|
1448 E T3N, T4h, T3U, T4m, T3D, T3E, T3w, T3F, T44, T45, T3X, T46, T4t, T4u;
|
cannam@167
|
1449 {
|
cannam@167
|
1450 E T3J, T3M, T3S, T3T;
|
cannam@167
|
1451 T3J = T3H - T3I;
|
cannam@167
|
1452 T3M = T3K - T3L;
|
cannam@167
|
1453 T3N = FMA(KP951056516, T3J, KP587785252 * T3M);
|
cannam@167
|
1454 T4h = FNMS(KP951056516, T3M, KP587785252 * T3J);
|
cannam@167
|
1455 T3S = T2K + T2Z;
|
cannam@167
|
1456 T3T = T3f - T3u;
|
cannam@167
|
1457 T3U = FMA(KP951056516, T3S, KP587785252 * T3T);
|
cannam@167
|
1458 T4m = FNMS(KP951056516, T3T, KP587785252 * T3S);
|
cannam@167
|
1459 }
|
cannam@167
|
1460 {
|
cannam@167
|
1461 E T30, T3v, T3V, T3W;
|
cannam@167
|
1462 T3D = T3z - T3C;
|
cannam@167
|
1463 T30 = T2K - T2Z;
|
cannam@167
|
1464 T3v = T3f + T3u;
|
cannam@167
|
1465 T3E = T30 + T3v;
|
cannam@167
|
1466 T3w = KP559016994 * (T30 - T3v);
|
cannam@167
|
1467 T3F = FNMS(KP250000000, T3E, T3D);
|
cannam@167
|
1468 T44 = T40 + T43;
|
cannam@167
|
1469 T3V = T3H + T3I;
|
cannam@167
|
1470 T3W = T3K + T3L;
|
cannam@167
|
1471 T45 = T3V + T3W;
|
cannam@167
|
1472 T3X = KP559016994 * (T3V - T3W);
|
cannam@167
|
1473 T46 = FNMS(KP250000000, T45, T44);
|
cannam@167
|
1474 }
|
cannam@167
|
1475 T4t = T3D + T3E;
|
cannam@167
|
1476 T4u = T44 + T45;
|
cannam@167
|
1477 cr[WS(rs, 2)] = FNMS(T2t, T4u, T2r * T4t);
|
cannam@167
|
1478 ci[WS(rs, 2)] = FMA(T2t, T4t, T2r * T4u);
|
cannam@167
|
1479 {
|
cannam@167
|
1480 E T4i, T4q, T4o, T4s, T4g, T4n;
|
cannam@167
|
1481 T4g = T3F - T3w;
|
cannam@167
|
1482 T4i = T4g - T4h;
|
cannam@167
|
1483 T4q = T4g + T4h;
|
cannam@167
|
1484 T4n = T46 - T3X;
|
cannam@167
|
1485 T4o = T4m + T4n;
|
cannam@167
|
1486 T4s = T4n - T4m;
|
cannam@167
|
1487 cr[WS(rs, 12)] = FNMS(T4l, T4o, T4f * T4i);
|
cannam@167
|
1488 ci[WS(rs, 12)] = FMA(T4l, T4i, T4f * T4o);
|
cannam@167
|
1489 cr[WS(rs, 17)] = FNMS(T4r, T4s, T4p * T4q);
|
cannam@167
|
1490 ci[WS(rs, 17)] = FMA(T4r, T4q, T4p * T4s);
|
cannam@167
|
1491 }
|
cannam@167
|
1492 {
|
cannam@167
|
1493 E T3O, T4a, T48, T4c, T3G, T47;
|
cannam@167
|
1494 T3G = T3w + T3F;
|
cannam@167
|
1495 T3O = T3G - T3N;
|
cannam@167
|
1496 T4a = T3G + T3N;
|
cannam@167
|
1497 T47 = T3X + T46;
|
cannam@167
|
1498 T48 = T3U + T47;
|
cannam@167
|
1499 T4c = T47 - T3U;
|
cannam@167
|
1500 cr[WS(rs, 7)] = FNMS(T3R, T48, T2v * T3O);
|
cannam@167
|
1501 ci[WS(rs, 7)] = FMA(T3R, T3O, T2v * T48);
|
cannam@167
|
1502 cr[WS(rs, 22)] = FNMS(T4b, T4c, T49 * T4a);
|
cannam@167
|
1503 ci[WS(rs, 22)] = FMA(T4b, T4a, T49 * T4c);
|
cannam@167
|
1504 }
|
cannam@167
|
1505 }
|
cannam@167
|
1506 {
|
cannam@167
|
1507 E T4V, T5f, T50, T5i, T4L, T4M, T4K, T4N, T54, T55, T53, T56, T5r, T5s;
|
cannam@167
|
1508 {
|
cannam@167
|
1509 E T4R, T4U, T4Y, T4Z;
|
cannam@167
|
1510 T4R = T4P - T4Q;
|
cannam@167
|
1511 T4U = T4S - T4T;
|
cannam@167
|
1512 T4V = FMA(KP951056516, T4R, KP587785252 * T4U);
|
cannam@167
|
1513 T5f = FNMS(KP951056516, T4U, KP587785252 * T4R);
|
cannam@167
|
1514 T4Y = T4y + T4B;
|
cannam@167
|
1515 T4Z = T4F + T4I;
|
cannam@167
|
1516 T50 = FMA(KP951056516, T4Y, KP587785252 * T4Z);
|
cannam@167
|
1517 T5i = FNMS(KP951056516, T4Z, KP587785252 * T4Y);
|
cannam@167
|
1518 }
|
cannam@167
|
1519 {
|
cannam@167
|
1520 E T4C, T4J, T51, T52;
|
cannam@167
|
1521 T4L = T3z + T3C;
|
cannam@167
|
1522 T4C = T4y - T4B;
|
cannam@167
|
1523 T4J = T4F - T4I;
|
cannam@167
|
1524 T4M = T4C + T4J;
|
cannam@167
|
1525 T4K = KP559016994 * (T4C - T4J);
|
cannam@167
|
1526 T4N = FNMS(KP250000000, T4M, T4L);
|
cannam@167
|
1527 T54 = T43 - T40;
|
cannam@167
|
1528 T51 = T4P + T4Q;
|
cannam@167
|
1529 T52 = T4S + T4T;
|
cannam@167
|
1530 T55 = T51 + T52;
|
cannam@167
|
1531 T53 = KP559016994 * (T51 - T52);
|
cannam@167
|
1532 T56 = FNMS(KP250000000, T55, T54);
|
cannam@167
|
1533 }
|
cannam@167
|
1534 T5r = T4L + T4M;
|
cannam@167
|
1535 T5s = T54 + T55;
|
cannam@167
|
1536 cr[WS(rs, 3)] = FNMS(TR, T5s, TO * T5r);
|
cannam@167
|
1537 ci[WS(rs, 3)] = FMA(TR, T5r, TO * T5s);
|
cannam@167
|
1538 {
|
cannam@167
|
1539 E T5g, T5o, T5k, T5q, T5e, T5j;
|
cannam@167
|
1540 T5e = T4N - T4K;
|
cannam@167
|
1541 T5g = T5e - T5f;
|
cannam@167
|
1542 T5o = T5e + T5f;
|
cannam@167
|
1543 T5j = T56 - T53;
|
cannam@167
|
1544 T5k = T5i + T5j;
|
cannam@167
|
1545 T5q = T5j - T5i;
|
cannam@167
|
1546 cr[WS(rs, 13)] = FNMS(T5h, T5k, T5d * T5g);
|
cannam@167
|
1547 ci[WS(rs, 13)] = FMA(T5h, T5g, T5d * T5k);
|
cannam@167
|
1548 cr[WS(rs, 18)] = FNMS(T5p, T5q, T5n * T5o);
|
cannam@167
|
1549 ci[WS(rs, 18)] = FMA(T5p, T5o, T5n * T5q);
|
cannam@167
|
1550 }
|
cannam@167
|
1551 {
|
cannam@167
|
1552 E T4W, T5a, T58, T5c, T4O, T57;
|
cannam@167
|
1553 T4O = T4K + T4N;
|
cannam@167
|
1554 T4W = T4O - T4V;
|
cannam@167
|
1555 T5a = T4O + T4V;
|
cannam@167
|
1556 T57 = T53 + T56;
|
cannam@167
|
1557 T58 = T50 + T57;
|
cannam@167
|
1558 T5c = T57 - T50;
|
cannam@167
|
1559 cr[WS(rs, 8)] = FNMS(T4X, T58, T4v * T4W);
|
cannam@167
|
1560 ci[WS(rs, 8)] = FMA(T4X, T4W, T4v * T58);
|
cannam@167
|
1561 cr[WS(rs, 23)] = FNMS(T5b, T5c, T59 * T5a);
|
cannam@167
|
1562 ci[WS(rs, 23)] = FMA(T5b, T5a, T59 * T5c);
|
cannam@167
|
1563 }
|
cannam@167
|
1564 }
|
cannam@167
|
1565 {
|
cannam@167
|
1566 E T6a, T6v, T6e, T6y, T60, T61, T5X, T62, T6k, T6l, T6h, T6m, T6F, T6G;
|
cannam@167
|
1567 {
|
cannam@167
|
1568 E T66, T69, T6c, T6d;
|
cannam@167
|
1569 T66 = T64 - T65;
|
cannam@167
|
1570 T69 = T67 - T68;
|
cannam@167
|
1571 T6a = FMA(KP951056516, T66, KP587785252 * T69);
|
cannam@167
|
1572 T6v = FNMS(KP951056516, T69, KP587785252 * T66);
|
cannam@167
|
1573 T6c = T5z - T5G;
|
cannam@167
|
1574 T6d = T5O - T5V;
|
cannam@167
|
1575 T6e = FMA(KP951056516, T6c, KP587785252 * T6d);
|
cannam@167
|
1576 T6y = FNMS(KP951056516, T6d, KP587785252 * T6c);
|
cannam@167
|
1577 }
|
cannam@167
|
1578 {
|
cannam@167
|
1579 E T5H, T5W, T6f, T6g;
|
cannam@167
|
1580 T60 = T5Y - T5Z;
|
cannam@167
|
1581 T5H = T5z + T5G;
|
cannam@167
|
1582 T5W = T5O + T5V;
|
cannam@167
|
1583 T61 = T5H + T5W;
|
cannam@167
|
1584 T5X = KP559016994 * (T5H - T5W);
|
cannam@167
|
1585 T62 = FNMS(KP250000000, T61, T60);
|
cannam@167
|
1586 T6k = T6i + T6j;
|
cannam@167
|
1587 T6f = T64 + T65;
|
cannam@167
|
1588 T6g = T67 + T68;
|
cannam@167
|
1589 T6l = T6f + T6g;
|
cannam@167
|
1590 T6h = KP559016994 * (T6f - T6g);
|
cannam@167
|
1591 T6m = FNMS(KP250000000, T6l, T6k);
|
cannam@167
|
1592 }
|
cannam@167
|
1593 T6F = T60 + T61;
|
cannam@167
|
1594 T6G = T6k + T6l;
|
cannam@167
|
1595 cr[WS(rs, 1)] = FNMS(TQ, T6G, TN * T6F);
|
cannam@167
|
1596 ci[WS(rs, 1)] = FMA(TQ, T6F, TN * T6G);
|
cannam@167
|
1597 {
|
cannam@167
|
1598 E T6w, T6C, T6A, T6E, T6u, T6z;
|
cannam@167
|
1599 T6u = T62 - T5X;
|
cannam@167
|
1600 T6w = T6u - T6v;
|
cannam@167
|
1601 T6C = T6u + T6v;
|
cannam@167
|
1602 T6z = T6m - T6h;
|
cannam@167
|
1603 T6A = T6y + T6z;
|
cannam@167
|
1604 T6E = T6z - T6y;
|
cannam@167
|
1605 cr[WS(rs, 11)] = FNMS(T6x, T6A, T6t * T6w);
|
cannam@167
|
1606 ci[WS(rs, 11)] = FMA(T6x, T6w, T6t * T6A);
|
cannam@167
|
1607 cr[WS(rs, 16)] = FNMS(T6D, T6E, T6B * T6C);
|
cannam@167
|
1608 ci[WS(rs, 16)] = FMA(T6D, T6C, T6B * T6E);
|
cannam@167
|
1609 }
|
cannam@167
|
1610 {
|
cannam@167
|
1611 E T6b, T6q, T6o, T6s, T63, T6n;
|
cannam@167
|
1612 T63 = T5X + T62;
|
cannam@167
|
1613 T6b = T63 - T6a;
|
cannam@167
|
1614 T6q = T63 + T6a;
|
cannam@167
|
1615 T6n = T6h + T6m;
|
cannam@167
|
1616 T6o = T6e + T6n;
|
cannam@167
|
1617 T6s = T6n - T6e;
|
cannam@167
|
1618 cr[WS(rs, 6)] = FNMS(T5m, T6o, T5l * T6b);
|
cannam@167
|
1619 ci[WS(rs, 6)] = FMA(T5m, T6b, T5l * T6o);
|
cannam@167
|
1620 cr[WS(rs, 21)] = FNMS(T6r, T6s, T6p * T6q);
|
cannam@167
|
1621 ci[WS(rs, 21)] = FMA(T6r, T6q, T6p * T6s);
|
cannam@167
|
1622 }
|
cannam@167
|
1623 }
|
cannam@167
|
1624 }
|
cannam@167
|
1625 }
|
cannam@167
|
1626 }
|
cannam@167
|
1627 }
|
cannam@167
|
1628
|
cannam@167
|
1629 static const tw_instr twinstr[] = {
|
cannam@167
|
1630 {TW_CEXP, 1, 1},
|
cannam@167
|
1631 {TW_CEXP, 1, 3},
|
cannam@167
|
1632 {TW_CEXP, 1, 9},
|
cannam@167
|
1633 {TW_CEXP, 1, 24},
|
cannam@167
|
1634 {TW_NEXT, 1, 0}
|
cannam@167
|
1635 };
|
cannam@167
|
1636
|
cannam@167
|
1637 static const hc2hc_desc desc = { 25, "hb2_25", twinstr, &GENUS, {280, 180, 160, 0} };
|
cannam@167
|
1638
|
cannam@167
|
1639 void X(codelet_hb2_25) (planner *p) {
|
cannam@167
|
1640 X(khc2hc_register) (p, hb2_25, &desc);
|
cannam@167
|
1641 }
|
cannam@167
|
1642 #endif
|