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 08:06:38 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 -twiddle-log3 -precompute-twiddles -n 25 -dit -name hf2_25 -include rdft/scalar/hf.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 * 186 stack variables, 47 constants, and 100 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hf2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP614372930, +0.614372930789563808870829930444362096004872855);
|
cannam@167
|
40 DK(KP621716863, +0.621716863012209892444754556304102309693593202);
|
cannam@167
|
41 DK(KP860541664, +0.860541664367944677098261680920518816412804187);
|
cannam@167
|
42 DK(KP949179823, +0.949179823508441261575555465843363271711583843);
|
cannam@167
|
43 DK(KP557913902, +0.557913902031834264187699648465567037992437152);
|
cannam@167
|
44 DK(KP249506682, +0.249506682107067890488084201715862638334226305);
|
cannam@167
|
45 DK(KP681693190, +0.681693190061530575150324149145440022633095390);
|
cannam@167
|
46 DK(KP560319534, +0.560319534973832390111614715371676131169633784);
|
cannam@167
|
47 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
cannam@167
|
48 DK(KP906616052, +0.906616052148196230441134447086066874408359177);
|
cannam@167
|
49 DK(KP845997307, +0.845997307939530944175097360758058292389769300);
|
cannam@167
|
50 DK(KP968479752, +0.968479752739016373193524836781420152702090879);
|
cannam@167
|
51 DK(KP994076283, +0.994076283785401014123185814696322018529298887);
|
cannam@167
|
52 DK(KP772036680, +0.772036680810363904029489473607579825330539880);
|
cannam@167
|
53 DK(KP734762448, +0.734762448793050413546343770063151342619912334);
|
cannam@167
|
54 DK(KP062914667, +0.062914667253649757225485955897349402364686947);
|
cannam@167
|
55 DK(KP921177326, +0.921177326965143320250447435415066029359282231);
|
cannam@167
|
56 DK(KP833417178, +0.833417178328688677408962550243238843138996060);
|
cannam@167
|
57 DK(KP541454447, +0.541454447536312777046285590082819509052033189);
|
cannam@167
|
58 DK(KP242145790, +0.242145790282157779872542093866183953459003101);
|
cannam@167
|
59 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
cannam@167
|
60 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
cannam@167
|
61 DK(KP943557151, +0.943557151597354104399655195398983005179443399);
|
cannam@167
|
62 DK(KP803003575, +0.803003575438660414833440593570376004635464850);
|
cannam@167
|
63 DK(KP554608978, +0.554608978404018097464974850792216217022558774);
|
cannam@167
|
64 DK(KP248028675, +0.248028675328619457762448260696444630363259177);
|
cannam@167
|
65 DK(KP525970792, +0.525970792408939708442463226536226366643874659);
|
cannam@167
|
66 DK(KP726211448, +0.726211448929902658173535992263577167607493062);
|
cannam@167
|
67 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
cannam@167
|
68 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
cannam@167
|
69 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
cannam@167
|
70 DK(KP831864738, +0.831864738706457140726048799369896829771167132);
|
cannam@167
|
71 DK(KP871714437, +0.871714437527667770979999223229522602943903653);
|
cannam@167
|
72 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
cannam@167
|
73 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
cannam@167
|
74 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
cannam@167
|
75 DK(KP851038619, +0.851038619207379630836264138867114231259902550);
|
cannam@167
|
76 DK(KP912018591, +0.912018591466481957908415381764119056233607330);
|
cannam@167
|
77 DK(KP912575812, +0.912575812670962425556968549836277086778922727);
|
cannam@167
|
78 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
cannam@167
|
79 DK(KP470564281, +0.470564281212251493087595091036643380879947982);
|
cannam@167
|
80 DK(KP827271945, +0.827271945972475634034355757144307982555673741);
|
cannam@167
|
81 DK(KP126329378, +0.126329378446108174786050455341811215027378105);
|
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 T2, T8, T3, T6, Tk, Tm, T5, T7, T19, Tb, T1b, Tc, Tw, TT, T1j;
|
cannam@167
|
90 E TE, T2p, T1c, T2U, TI, T11, T15, T2Q, T2M, T2m, T2i, T2e, Tn, Tr, TX;
|
cannam@167
|
91 E T31, T35, T1l, T1m, T1q, TA, T1K, T1O, T2a, T27, T1g, T2x, T2t, Th, Td;
|
cannam@167
|
92 E T1S, T2X, T1W;
|
cannam@167
|
93 {
|
cannam@167
|
94 E TS, TD, T2L, T10, TH, T2P, T14, T9, T1a, Tz, TW, T4, Ta, Tv, T1J;
|
cannam@167
|
95 E T1N;
|
cannam@167
|
96 T2 = W[0];
|
cannam@167
|
97 T8 = W[4];
|
cannam@167
|
98 T3 = W[2];
|
cannam@167
|
99 T6 = W[3];
|
cannam@167
|
100 T4 = T2 * T3;
|
cannam@167
|
101 TS = T3 * T8;
|
cannam@167
|
102 Ta = T2 * T6;
|
cannam@167
|
103 Tv = T2 * T8;
|
cannam@167
|
104 Tk = W[6];
|
cannam@167
|
105 TD = T8 * Tk;
|
cannam@167
|
106 T2L = T2 * Tk;
|
cannam@167
|
107 T10 = T3 * Tk;
|
cannam@167
|
108 Tm = W[7];
|
cannam@167
|
109 TH = T8 * Tm;
|
cannam@167
|
110 T2P = T2 * Tm;
|
cannam@167
|
111 T14 = T3 * Tm;
|
cannam@167
|
112 T5 = W[1];
|
cannam@167
|
113 T7 = FNMS(T5, T6, T4);
|
cannam@167
|
114 T19 = FMA(T5, T6, T4);
|
cannam@167
|
115 T9 = T7 * T8;
|
cannam@167
|
116 T1a = T19 * T8;
|
cannam@167
|
117 Tb = FMA(T5, T3, Ta);
|
cannam@167
|
118 T1b = FNMS(T5, T3, Ta);
|
cannam@167
|
119 Tc = W[5];
|
cannam@167
|
120 Tz = T2 * Tc;
|
cannam@167
|
121 TW = T3 * Tc;
|
cannam@167
|
122 Tw = FNMS(T5, Tc, Tv);
|
cannam@167
|
123 TT = FMA(T6, Tc, TS);
|
cannam@167
|
124 T1j = FMA(T5, Tc, Tv);
|
cannam@167
|
125 TE = FMA(Tc, Tm, TD);
|
cannam@167
|
126 T2p = FMA(T6, T8, TW);
|
cannam@167
|
127 T1c = FNMS(T1b, Tc, T1a);
|
cannam@167
|
128 T2U = FNMS(Tb, Tc, T9);
|
cannam@167
|
129 TI = FNMS(Tc, Tk, TH);
|
cannam@167
|
130 T11 = FMA(T6, Tm, T10);
|
cannam@167
|
131 T15 = FNMS(T6, Tk, T14);
|
cannam@167
|
132 T2Q = FNMS(T5, Tk, T2P);
|
cannam@167
|
133 T2M = FMA(T5, Tm, T2L);
|
cannam@167
|
134 {
|
cannam@167
|
135 E T2h, T2d, Tl, Tq;
|
cannam@167
|
136 T2m = FNMS(T6, Tc, TS);
|
cannam@167
|
137 T2h = T19 * Tm;
|
cannam@167
|
138 T2i = FNMS(T1b, Tk, T2h);
|
cannam@167
|
139 T2d = T19 * Tk;
|
cannam@167
|
140 T2e = FMA(T1b, Tm, T2d);
|
cannam@167
|
141 Tl = T7 * Tk;
|
cannam@167
|
142 Tn = FMA(Tb, Tm, Tl);
|
cannam@167
|
143 Tq = T7 * Tm;
|
cannam@167
|
144 Tr = FNMS(Tb, Tk, Tq);
|
cannam@167
|
145 }
|
cannam@167
|
146 {
|
cannam@167
|
147 E T30, T34, T1k, T1p;
|
cannam@167
|
148 T30 = TT * Tk;
|
cannam@167
|
149 T34 = TT * Tm;
|
cannam@167
|
150 TX = FNMS(T6, T8, TW);
|
cannam@167
|
151 T31 = FMA(TX, Tm, T30);
|
cannam@167
|
152 T35 = FNMS(TX, Tk, T34);
|
cannam@167
|
153 T1k = T1j * Tk;
|
cannam@167
|
154 T1p = T1j * Tm;
|
cannam@167
|
155 T1l = FNMS(T5, T8, Tz);
|
cannam@167
|
156 T1m = FMA(T1l, Tm, T1k);
|
cannam@167
|
157 T1q = FNMS(T1l, Tk, T1p);
|
cannam@167
|
158 }
|
cannam@167
|
159 T1J = Tw * Tk;
|
cannam@167
|
160 T1N = Tw * Tm;
|
cannam@167
|
161 TA = FMA(T5, T8, Tz);
|
cannam@167
|
162 T1K = FMA(TA, Tm, T1J);
|
cannam@167
|
163 T1O = FNMS(TA, Tk, T1N);
|
cannam@167
|
164 {
|
cannam@167
|
165 E T1f, T2s, T2w, Tg, T1R, T1V;
|
cannam@167
|
166 T1f = T19 * Tc;
|
cannam@167
|
167 T2a = FNMS(T1b, T8, T1f);
|
cannam@167
|
168 T27 = FMA(T1b, Tc, T1a);
|
cannam@167
|
169 T2s = T27 * Tk;
|
cannam@167
|
170 T2w = T27 * Tm;
|
cannam@167
|
171 T1g = FMA(T1b, T8, T1f);
|
cannam@167
|
172 T2x = FNMS(T2a, Tk, T2w);
|
cannam@167
|
173 T2t = FMA(T2a, Tm, T2s);
|
cannam@167
|
174 Tg = T7 * Tc;
|
cannam@167
|
175 Th = FNMS(Tb, T8, Tg);
|
cannam@167
|
176 Td = FMA(Tb, Tc, T9);
|
cannam@167
|
177 T1R = Td * Tk;
|
cannam@167
|
178 T1V = Td * Tm;
|
cannam@167
|
179 T1S = FMA(Th, Tm, T1R);
|
cannam@167
|
180 T2X = FMA(Tb, T8, Tg);
|
cannam@167
|
181 T1W = FNMS(Th, Tk, T1V);
|
cannam@167
|
182 }
|
cannam@167
|
183 }
|
cannam@167
|
184 {
|
cannam@167
|
185 E T1, T7l, T4s, T6a, T7u, T7U, TM, T4f, T4g, T7o, T7p, T7q, T4z, T6n, T4G;
|
cannam@167
|
186 E T6k, T3a, T6m, T4w, T4a, T6j, T4D, T54, T6C, T5b, T6z, T1v, T6y, T58, T3t;
|
cannam@167
|
187 E T6B, T51, T5j, T6v, T5q, T6s, T21, T6r, T5n, T3H, T6u, T5g, T4O, T6d, T4V;
|
cannam@167
|
188 E T6g, T2C, T6f, T4S, T3W, T6c, T4L;
|
cannam@167
|
189 {
|
cannam@167
|
190 E Tj, T4j, TK, T4q, TC, T4o, Tt, T4l;
|
cannam@167
|
191 T1 = cr[0];
|
cannam@167
|
192 T7l = ci[0];
|
cannam@167
|
193 {
|
cannam@167
|
194 E Te, Tf, Ti, T4i;
|
cannam@167
|
195 Te = cr[WS(rs, 5)];
|
cannam@167
|
196 Tf = Td * Te;
|
cannam@167
|
197 Ti = ci[WS(rs, 5)];
|
cannam@167
|
198 T4i = Td * Ti;
|
cannam@167
|
199 Tj = FMA(Th, Ti, Tf);
|
cannam@167
|
200 T4j = FNMS(Th, Te, T4i);
|
cannam@167
|
201 }
|
cannam@167
|
202 {
|
cannam@167
|
203 E TF, TG, TJ, T4p;
|
cannam@167
|
204 TF = cr[WS(rs, 15)];
|
cannam@167
|
205 TG = TE * TF;
|
cannam@167
|
206 TJ = ci[WS(rs, 15)];
|
cannam@167
|
207 T4p = TE * TJ;
|
cannam@167
|
208 TK = FMA(TI, TJ, TG);
|
cannam@167
|
209 T4q = FNMS(TI, TF, T4p);
|
cannam@167
|
210 }
|
cannam@167
|
211 {
|
cannam@167
|
212 E Tx, Ty, TB, T4n;
|
cannam@167
|
213 Tx = cr[WS(rs, 10)];
|
cannam@167
|
214 Ty = Tw * Tx;
|
cannam@167
|
215 TB = ci[WS(rs, 10)];
|
cannam@167
|
216 T4n = Tw * TB;
|
cannam@167
|
217 TC = FMA(TA, TB, Ty);
|
cannam@167
|
218 T4o = FNMS(TA, Tx, T4n);
|
cannam@167
|
219 }
|
cannam@167
|
220 {
|
cannam@167
|
221 E To, Tp, Ts, T4k;
|
cannam@167
|
222 To = cr[WS(rs, 20)];
|
cannam@167
|
223 Tp = Tn * To;
|
cannam@167
|
224 Ts = ci[WS(rs, 20)];
|
cannam@167
|
225 T4k = Tn * Ts;
|
cannam@167
|
226 Tt = FMA(Tr, Ts, Tp);
|
cannam@167
|
227 T4l = FNMS(Tr, To, T4k);
|
cannam@167
|
228 }
|
cannam@167
|
229 {
|
cannam@167
|
230 E T4m, T4r, T7s, T7t;
|
cannam@167
|
231 T4m = T4j - T4l;
|
cannam@167
|
232 T4r = T4o - T4q;
|
cannam@167
|
233 T4s = FMA(KP618033988, T4r, T4m);
|
cannam@167
|
234 T6a = FNMS(KP618033988, T4m, T4r);
|
cannam@167
|
235 T7s = TC - TK;
|
cannam@167
|
236 T7t = Tj - Tt;
|
cannam@167
|
237 T7u = FNMS(KP618033988, T7t, T7s);
|
cannam@167
|
238 T7U = FMA(KP618033988, T7s, T7t);
|
cannam@167
|
239 }
|
cannam@167
|
240 {
|
cannam@167
|
241 E Tu, TL, T7m, T7n;
|
cannam@167
|
242 Tu = Tj + Tt;
|
cannam@167
|
243 TL = TC + TK;
|
cannam@167
|
244 TM = Tu + TL;
|
cannam@167
|
245 T4f = FNMS(KP250000000, TM, T1);
|
cannam@167
|
246 T4g = Tu - TL;
|
cannam@167
|
247 T7m = T4j + T4l;
|
cannam@167
|
248 T7n = T4o + T4q;
|
cannam@167
|
249 T7o = T7m + T7n;
|
cannam@167
|
250 T7p = FNMS(KP250000000, T7o, T7l);
|
cannam@167
|
251 T7q = T7m - T7n;
|
cannam@167
|
252 }
|
cannam@167
|
253 }
|
cannam@167
|
254 {
|
cannam@167
|
255 E T2G, T3Y, T2Z, T37, T38, T45, T47, T48, T2K, T2S, T2T, T40, T42, T43;
|
cannam@167
|
256 {
|
cannam@167
|
257 E T2D, T2E, T2F, T3X;
|
cannam@167
|
258 T2D = cr[WS(rs, 3)];
|
cannam@167
|
259 T2E = T3 * T2D;
|
cannam@167
|
260 T2F = ci[WS(rs, 3)];
|
cannam@167
|
261 T3X = T3 * T2F;
|
cannam@167
|
262 T2G = FMA(T6, T2F, T2E);
|
cannam@167
|
263 T3Y = FNMS(T6, T2D, T3X);
|
cannam@167
|
264 }
|
cannam@167
|
265 {
|
cannam@167
|
266 E T2V, T2W, T2Y, T44, T32, T33, T36, T46;
|
cannam@167
|
267 T2V = cr[WS(rs, 13)];
|
cannam@167
|
268 T2W = T2U * T2V;
|
cannam@167
|
269 T2Y = ci[WS(rs, 13)];
|
cannam@167
|
270 T44 = T2U * T2Y;
|
cannam@167
|
271 T32 = cr[WS(rs, 18)];
|
cannam@167
|
272 T33 = T31 * T32;
|
cannam@167
|
273 T36 = ci[WS(rs, 18)];
|
cannam@167
|
274 T46 = T31 * T36;
|
cannam@167
|
275 T2Z = FMA(T2X, T2Y, T2W);
|
cannam@167
|
276 T37 = FMA(T35, T36, T33);
|
cannam@167
|
277 T38 = T2Z + T37;
|
cannam@167
|
278 T45 = FNMS(T2X, T2V, T44);
|
cannam@167
|
279 T47 = FNMS(T35, T32, T46);
|
cannam@167
|
280 T48 = T45 + T47;
|
cannam@167
|
281 }
|
cannam@167
|
282 {
|
cannam@167
|
283 E T2H, T2I, T2J, T3Z, T2N, T2O, T2R, T41;
|
cannam@167
|
284 T2H = cr[WS(rs, 8)];
|
cannam@167
|
285 T2I = T1j * T2H;
|
cannam@167
|
286 T2J = ci[WS(rs, 8)];
|
cannam@167
|
287 T3Z = T1j * T2J;
|
cannam@167
|
288 T2N = cr[WS(rs, 23)];
|
cannam@167
|
289 T2O = T2M * T2N;
|
cannam@167
|
290 T2R = ci[WS(rs, 23)];
|
cannam@167
|
291 T41 = T2M * T2R;
|
cannam@167
|
292 T2K = FMA(T1l, T2J, T2I);
|
cannam@167
|
293 T2S = FMA(T2Q, T2R, T2O);
|
cannam@167
|
294 T2T = T2K + T2S;
|
cannam@167
|
295 T40 = FNMS(T1l, T2H, T3Z);
|
cannam@167
|
296 T42 = FNMS(T2Q, T2N, T41);
|
cannam@167
|
297 T43 = T40 + T42;
|
cannam@167
|
298 }
|
cannam@167
|
299 {
|
cannam@167
|
300 E T4x, T4y, T4E, T4F;
|
cannam@167
|
301 T4x = T42 - T40;
|
cannam@167
|
302 T4y = T47 - T45;
|
cannam@167
|
303 T4z = FMA(KP618033988, T4y, T4x);
|
cannam@167
|
304 T6n = FNMS(KP618033988, T4x, T4y);
|
cannam@167
|
305 T4E = T2K - T2S;
|
cannam@167
|
306 T4F = T2Z - T37;
|
cannam@167
|
307 T4G = FMA(KP618033988, T4F, T4E);
|
cannam@167
|
308 T6k = FNMS(KP618033988, T4E, T4F);
|
cannam@167
|
309 }
|
cannam@167
|
310 {
|
cannam@167
|
311 E T4v, T39, T4u, T4C, T49, T4B;
|
cannam@167
|
312 T4v = T38 - T2T;
|
cannam@167
|
313 T39 = T2T + T38;
|
cannam@167
|
314 T4u = FNMS(KP250000000, T39, T2G);
|
cannam@167
|
315 T3a = T2G + T39;
|
cannam@167
|
316 T6m = FMA(KP559016994, T4v, T4u);
|
cannam@167
|
317 T4w = FNMS(KP559016994, T4v, T4u);
|
cannam@167
|
318 T4C = T48 - T43;
|
cannam@167
|
319 T49 = T43 + T48;
|
cannam@167
|
320 T4B = FNMS(KP250000000, T49, T3Y);
|
cannam@167
|
321 T4a = T3Y + T49;
|
cannam@167
|
322 T6j = FMA(KP559016994, T4C, T4B);
|
cannam@167
|
323 T4D = FNMS(KP559016994, T4C, T4B);
|
cannam@167
|
324 }
|
cannam@167
|
325 }
|
cannam@167
|
326 {
|
cannam@167
|
327 E TR, T3h, T1i, T1s, T1t, T3o, T3q, T3r, TZ, T17, T18, T3j, T3l, T3m;
|
cannam@167
|
328 {
|
cannam@167
|
329 E TO, TP, TQ, T3g;
|
cannam@167
|
330 TO = cr[WS(rs, 1)];
|
cannam@167
|
331 TP = T2 * TO;
|
cannam@167
|
332 TQ = ci[WS(rs, 1)];
|
cannam@167
|
333 T3g = T2 * TQ;
|
cannam@167
|
334 TR = FMA(T5, TQ, TP);
|
cannam@167
|
335 T3h = FNMS(T5, TO, T3g);
|
cannam@167
|
336 }
|
cannam@167
|
337 {
|
cannam@167
|
338 E T1d, T1e, T1h, T3n, T1n, T1o, T1r, T3p;
|
cannam@167
|
339 T1d = cr[WS(rs, 11)];
|
cannam@167
|
340 T1e = T1c * T1d;
|
cannam@167
|
341 T1h = ci[WS(rs, 11)];
|
cannam@167
|
342 T3n = T1c * T1h;
|
cannam@167
|
343 T1n = cr[WS(rs, 16)];
|
cannam@167
|
344 T1o = T1m * T1n;
|
cannam@167
|
345 T1r = ci[WS(rs, 16)];
|
cannam@167
|
346 T3p = T1m * T1r;
|
cannam@167
|
347 T1i = FMA(T1g, T1h, T1e);
|
cannam@167
|
348 T1s = FMA(T1q, T1r, T1o);
|
cannam@167
|
349 T1t = T1i + T1s;
|
cannam@167
|
350 T3o = FNMS(T1g, T1d, T3n);
|
cannam@167
|
351 T3q = FNMS(T1q, T1n, T3p);
|
cannam@167
|
352 T3r = T3o + T3q;
|
cannam@167
|
353 }
|
cannam@167
|
354 {
|
cannam@167
|
355 E TU, TV, TY, T3i, T12, T13, T16, T3k;
|
cannam@167
|
356 TU = cr[WS(rs, 6)];
|
cannam@167
|
357 TV = TT * TU;
|
cannam@167
|
358 TY = ci[WS(rs, 6)];
|
cannam@167
|
359 T3i = TT * TY;
|
cannam@167
|
360 T12 = cr[WS(rs, 21)];
|
cannam@167
|
361 T13 = T11 * T12;
|
cannam@167
|
362 T16 = ci[WS(rs, 21)];
|
cannam@167
|
363 T3k = T11 * T16;
|
cannam@167
|
364 TZ = FMA(TX, TY, TV);
|
cannam@167
|
365 T17 = FMA(T15, T16, T13);
|
cannam@167
|
366 T18 = TZ + T17;
|
cannam@167
|
367 T3j = FNMS(TX, TU, T3i);
|
cannam@167
|
368 T3l = FNMS(T15, T12, T3k);
|
cannam@167
|
369 T3m = T3j + T3l;
|
cannam@167
|
370 }
|
cannam@167
|
371 {
|
cannam@167
|
372 E T52, T53, T59, T5a;
|
cannam@167
|
373 T52 = T17 - TZ;
|
cannam@167
|
374 T53 = T1s - T1i;
|
cannam@167
|
375 T54 = FMA(KP618033988, T53, T52);
|
cannam@167
|
376 T6C = FNMS(KP618033988, T52, T53);
|
cannam@167
|
377 T59 = T3j - T3l;
|
cannam@167
|
378 T5a = T3q - T3o;
|
cannam@167
|
379 T5b = FNMS(KP618033988, T5a, T59);
|
cannam@167
|
380 T6z = FMA(KP618033988, T59, T5a);
|
cannam@167
|
381 }
|
cannam@167
|
382 {
|
cannam@167
|
383 E T57, T1u, T56, T50, T3s, T4Z;
|
cannam@167
|
384 T57 = T18 - T1t;
|
cannam@167
|
385 T1u = T18 + T1t;
|
cannam@167
|
386 T56 = FNMS(KP250000000, T1u, TR);
|
cannam@167
|
387 T1v = TR + T1u;
|
cannam@167
|
388 T6y = FNMS(KP559016994, T57, T56);
|
cannam@167
|
389 T58 = FMA(KP559016994, T57, T56);
|
cannam@167
|
390 T50 = T3m - T3r;
|
cannam@167
|
391 T3s = T3m + T3r;
|
cannam@167
|
392 T4Z = FNMS(KP250000000, T3s, T3h);
|
cannam@167
|
393 T3t = T3h + T3s;
|
cannam@167
|
394 T6B = FNMS(KP559016994, T50, T4Z);
|
cannam@167
|
395 T51 = FMA(KP559016994, T50, T4Z);
|
cannam@167
|
396 }
|
cannam@167
|
397 }
|
cannam@167
|
398 {
|
cannam@167
|
399 E T1z, T3v, T1Q, T1Y, T1Z, T3C, T3E, T3F, T1D, T1H, T1I, T3x, T3z, T3A;
|
cannam@167
|
400 {
|
cannam@167
|
401 E T1w, T1x, T1y, T3u;
|
cannam@167
|
402 T1w = cr[WS(rs, 4)];
|
cannam@167
|
403 T1x = T7 * T1w;
|
cannam@167
|
404 T1y = ci[WS(rs, 4)];
|
cannam@167
|
405 T3u = T7 * T1y;
|
cannam@167
|
406 T1z = FMA(Tb, T1y, T1x);
|
cannam@167
|
407 T3v = FNMS(Tb, T1w, T3u);
|
cannam@167
|
408 }
|
cannam@167
|
409 {
|
cannam@167
|
410 E T1L, T1M, T1P, T3B, T1T, T1U, T1X, T3D;
|
cannam@167
|
411 T1L = cr[WS(rs, 14)];
|
cannam@167
|
412 T1M = T1K * T1L;
|
cannam@167
|
413 T1P = ci[WS(rs, 14)];
|
cannam@167
|
414 T3B = T1K * T1P;
|
cannam@167
|
415 T1T = cr[WS(rs, 19)];
|
cannam@167
|
416 T1U = T1S * T1T;
|
cannam@167
|
417 T1X = ci[WS(rs, 19)];
|
cannam@167
|
418 T3D = T1S * T1X;
|
cannam@167
|
419 T1Q = FMA(T1O, T1P, T1M);
|
cannam@167
|
420 T1Y = FMA(T1W, T1X, T1U);
|
cannam@167
|
421 T1Z = T1Q + T1Y;
|
cannam@167
|
422 T3C = FNMS(T1O, T1L, T3B);
|
cannam@167
|
423 T3E = FNMS(T1W, T1T, T3D);
|
cannam@167
|
424 T3F = T3C + T3E;
|
cannam@167
|
425 }
|
cannam@167
|
426 {
|
cannam@167
|
427 E T1A, T1B, T1C, T3w, T1E, T1F, T1G, T3y;
|
cannam@167
|
428 T1A = cr[WS(rs, 9)];
|
cannam@167
|
429 T1B = T8 * T1A;
|
cannam@167
|
430 T1C = ci[WS(rs, 9)];
|
cannam@167
|
431 T3w = T8 * T1C;
|
cannam@167
|
432 T1E = cr[WS(rs, 24)];
|
cannam@167
|
433 T1F = Tk * T1E;
|
cannam@167
|
434 T1G = ci[WS(rs, 24)];
|
cannam@167
|
435 T3y = Tk * T1G;
|
cannam@167
|
436 T1D = FMA(Tc, T1C, T1B);
|
cannam@167
|
437 T1H = FMA(Tm, T1G, T1F);
|
cannam@167
|
438 T1I = T1D + T1H;
|
cannam@167
|
439 T3x = FNMS(Tc, T1A, T3w);
|
cannam@167
|
440 T3z = FNMS(Tm, T1E, T3y);
|
cannam@167
|
441 T3A = T3x + T3z;
|
cannam@167
|
442 }
|
cannam@167
|
443 {
|
cannam@167
|
444 E T5h, T5i, T5o, T5p;
|
cannam@167
|
445 T5h = T1H - T1D;
|
cannam@167
|
446 T5i = T1Y - T1Q;
|
cannam@167
|
447 T5j = FMA(KP618033988, T5i, T5h);
|
cannam@167
|
448 T6v = FNMS(KP618033988, T5h, T5i);
|
cannam@167
|
449 T5o = T3z - T3x;
|
cannam@167
|
450 T5p = T3E - T3C;
|
cannam@167
|
451 T5q = FMA(KP618033988, T5p, T5o);
|
cannam@167
|
452 T6s = FNMS(KP618033988, T5o, T5p);
|
cannam@167
|
453 }
|
cannam@167
|
454 {
|
cannam@167
|
455 E T5m, T20, T5l, T5f, T3G, T5e;
|
cannam@167
|
456 T5m = T1I - T1Z;
|
cannam@167
|
457 T20 = T1I + T1Z;
|
cannam@167
|
458 T5l = FNMS(KP250000000, T20, T1z);
|
cannam@167
|
459 T21 = T1z + T20;
|
cannam@167
|
460 T6r = FNMS(KP559016994, T5m, T5l);
|
cannam@167
|
461 T5n = FMA(KP559016994, T5m, T5l);
|
cannam@167
|
462 T5f = T3F - T3A;
|
cannam@167
|
463 T3G = T3A + T3F;
|
cannam@167
|
464 T5e = FNMS(KP250000000, T3G, T3v);
|
cannam@167
|
465 T3H = T3v + T3G;
|
cannam@167
|
466 T6u = FMA(KP559016994, T5f, T5e);
|
cannam@167
|
467 T5g = FNMS(KP559016994, T5f, T5e);
|
cannam@167
|
468 }
|
cannam@167
|
469 }
|
cannam@167
|
470 {
|
cannam@167
|
471 E T26, T3K, T2r, T2z, T2A, T3R, T3T, T3U, T2c, T2k, T2l, T3M, T3O, T3P;
|
cannam@167
|
472 {
|
cannam@167
|
473 E T23, T24, T25, T3J;
|
cannam@167
|
474 T23 = cr[WS(rs, 2)];
|
cannam@167
|
475 T24 = T19 * T23;
|
cannam@167
|
476 T25 = ci[WS(rs, 2)];
|
cannam@167
|
477 T3J = T19 * T25;
|
cannam@167
|
478 T26 = FMA(T1b, T25, T24);
|
cannam@167
|
479 T3K = FNMS(T1b, T23, T3J);
|
cannam@167
|
480 }
|
cannam@167
|
481 {
|
cannam@167
|
482 E T2n, T2o, T2q, T3Q, T2u, T2v, T2y, T3S;
|
cannam@167
|
483 T2n = cr[WS(rs, 12)];
|
cannam@167
|
484 T2o = T2m * T2n;
|
cannam@167
|
485 T2q = ci[WS(rs, 12)];
|
cannam@167
|
486 T3Q = T2m * T2q;
|
cannam@167
|
487 T2u = cr[WS(rs, 17)];
|
cannam@167
|
488 T2v = T2t * T2u;
|
cannam@167
|
489 T2y = ci[WS(rs, 17)];
|
cannam@167
|
490 T3S = T2t * T2y;
|
cannam@167
|
491 T2r = FMA(T2p, T2q, T2o);
|
cannam@167
|
492 T2z = FMA(T2x, T2y, T2v);
|
cannam@167
|
493 T2A = T2r + T2z;
|
cannam@167
|
494 T3R = FNMS(T2p, T2n, T3Q);
|
cannam@167
|
495 T3T = FNMS(T2x, T2u, T3S);
|
cannam@167
|
496 T3U = T3R + T3T;
|
cannam@167
|
497 }
|
cannam@167
|
498 {
|
cannam@167
|
499 E T28, T29, T2b, T3L, T2f, T2g, T2j, T3N;
|
cannam@167
|
500 T28 = cr[WS(rs, 7)];
|
cannam@167
|
501 T29 = T27 * T28;
|
cannam@167
|
502 T2b = ci[WS(rs, 7)];
|
cannam@167
|
503 T3L = T27 * T2b;
|
cannam@167
|
504 T2f = cr[WS(rs, 22)];
|
cannam@167
|
505 T2g = T2e * T2f;
|
cannam@167
|
506 T2j = ci[WS(rs, 22)];
|
cannam@167
|
507 T3N = T2e * T2j;
|
cannam@167
|
508 T2c = FMA(T2a, T2b, T29);
|
cannam@167
|
509 T2k = FMA(T2i, T2j, T2g);
|
cannam@167
|
510 T2l = T2c + T2k;
|
cannam@167
|
511 T3M = FNMS(T2a, T28, T3L);
|
cannam@167
|
512 T3O = FNMS(T2i, T2f, T3N);
|
cannam@167
|
513 T3P = T3M + T3O;
|
cannam@167
|
514 }
|
cannam@167
|
515 {
|
cannam@167
|
516 E T4M, T4N, T4T, T4U;
|
cannam@167
|
517 T4M = T2k - T2c;
|
cannam@167
|
518 T4N = T2z - T2r;
|
cannam@167
|
519 T4O = FMA(KP618033988, T4N, T4M);
|
cannam@167
|
520 T6d = FNMS(KP618033988, T4M, T4N);
|
cannam@167
|
521 T4T = T3O - T3M;
|
cannam@167
|
522 T4U = T3R - T3T;
|
cannam@167
|
523 T4V = FNMS(KP618033988, T4U, T4T);
|
cannam@167
|
524 T6g = FMA(KP618033988, T4T, T4U);
|
cannam@167
|
525 }
|
cannam@167
|
526 {
|
cannam@167
|
527 E T4R, T2B, T4Q, T4K, T3V, T4J;
|
cannam@167
|
528 T4R = T2A - T2l;
|
cannam@167
|
529 T2B = T2l + T2A;
|
cannam@167
|
530 T4Q = FNMS(KP250000000, T2B, T26);
|
cannam@167
|
531 T2C = T26 + T2B;
|
cannam@167
|
532 T6f = FMA(KP559016994, T4R, T4Q);
|
cannam@167
|
533 T4S = FNMS(KP559016994, T4R, T4Q);
|
cannam@167
|
534 T4K = T3U - T3P;
|
cannam@167
|
535 T3V = T3P + T3U;
|
cannam@167
|
536 T4J = FNMS(KP250000000, T3V, T3K);
|
cannam@167
|
537 T3W = T3K + T3V;
|
cannam@167
|
538 T6c = FMA(KP559016994, T4K, T4J);
|
cannam@167
|
539 T4L = FNMS(KP559016994, T4K, T4J);
|
cannam@167
|
540 }
|
cannam@167
|
541 }
|
cannam@167
|
542 {
|
cannam@167
|
543 E T4c, T4e, TN, T3c, T3d, T3e, T4d, T3f;
|
cannam@167
|
544 {
|
cannam@167
|
545 E T3I, T4b, T22, T3b;
|
cannam@167
|
546 T3I = T3t - T3H;
|
cannam@167
|
547 T4b = T3W - T4a;
|
cannam@167
|
548 T4c = FMA(KP618033988, T4b, T3I);
|
cannam@167
|
549 T4e = FNMS(KP618033988, T3I, T4b);
|
cannam@167
|
550 TN = T1 + TM;
|
cannam@167
|
551 T22 = T1v + T21;
|
cannam@167
|
552 T3b = T2C + T3a;
|
cannam@167
|
553 T3c = T22 + T3b;
|
cannam@167
|
554 T3d = FNMS(KP250000000, T3c, TN);
|
cannam@167
|
555 T3e = T22 - T3b;
|
cannam@167
|
556 }
|
cannam@167
|
557 cr[0] = TN + T3c;
|
cannam@167
|
558 T4d = FNMS(KP559016994, T3e, T3d);
|
cannam@167
|
559 cr[WS(rs, 10)] = FNMS(KP951056516, T4e, T4d);
|
cannam@167
|
560 ci[WS(rs, 9)] = FMA(KP951056516, T4e, T4d);
|
cannam@167
|
561 T3f = FMA(KP559016994, T3e, T3d);
|
cannam@167
|
562 ci[WS(rs, 4)] = FNMS(KP951056516, T4c, T3f);
|
cannam@167
|
563 cr[WS(rs, 5)] = FMA(KP951056516, T4c, T3f);
|
cannam@167
|
564 }
|
cannam@167
|
565 {
|
cannam@167
|
566 E T4t, T5H, T7V, T87, T5A, T5D, T8f, T8e, T88, T89, T8a, T4Y, T5t, T5u, T62;
|
cannam@167
|
567 E T65, T83, T82, T7W, T7X, T7Y, T5O, T5V, T5W, T4h, T7T;
|
cannam@167
|
568 T4h = FMA(KP559016994, T4g, T4f);
|
cannam@167
|
569 T4t = FNMS(KP951056516, T4s, T4h);
|
cannam@167
|
570 T5H = FMA(KP951056516, T4s, T4h);
|
cannam@167
|
571 T7T = FMA(KP559016994, T7q, T7p);
|
cannam@167
|
572 T7V = FNMS(KP951056516, T7U, T7T);
|
cannam@167
|
573 T87 = FMA(KP951056516, T7U, T7T);
|
cannam@167
|
574 {
|
cannam@167
|
575 E T4I, T5B, T5s, T5z, T4X, T5C, T5d, T5y;
|
cannam@167
|
576 {
|
cannam@167
|
577 E T4A, T4H, T5k, T5r;
|
cannam@167
|
578 T4A = FMA(KP951056516, T4z, T4w);
|
cannam@167
|
579 T4H = FMA(KP951056516, T4G, T4D);
|
cannam@167
|
580 T4I = FNMS(KP126329378, T4H, T4A);
|
cannam@167
|
581 T5B = FMA(KP126329378, T4A, T4H);
|
cannam@167
|
582 T5k = FNMS(KP951056516, T5j, T5g);
|
cannam@167
|
583 T5r = FMA(KP951056516, T5q, T5n);
|
cannam@167
|
584 T5s = FMA(KP827271945, T5r, T5k);
|
cannam@167
|
585 T5z = FNMS(KP827271945, T5k, T5r);
|
cannam@167
|
586 }
|
cannam@167
|
587 {
|
cannam@167
|
588 E T4P, T4W, T55, T5c;
|
cannam@167
|
589 T4P = FNMS(KP951056516, T4O, T4L);
|
cannam@167
|
590 T4W = FMA(KP951056516, T4V, T4S);
|
cannam@167
|
591 T4X = FNMS(KP470564281, T4W, T4P);
|
cannam@167
|
592 T5C = FMA(KP470564281, T4P, T4W);
|
cannam@167
|
593 T55 = FNMS(KP951056516, T54, T51);
|
cannam@167
|
594 T5c = FNMS(KP951056516, T5b, T58);
|
cannam@167
|
595 T5d = FMA(KP634619297, T5c, T55);
|
cannam@167
|
596 T5y = FNMS(KP634619297, T55, T5c);
|
cannam@167
|
597 }
|
cannam@167
|
598 T5A = FMA(KP912575812, T5z, T5y);
|
cannam@167
|
599 T5D = FNMS(KP912018591, T5C, T5B);
|
cannam@167
|
600 T8f = FMA(KP912575812, T5s, T5d);
|
cannam@167
|
601 T8e = FMA(KP912018591, T4X, T4I);
|
cannam@167
|
602 T88 = FMA(KP912018591, T5C, T5B);
|
cannam@167
|
603 T89 = FNMS(KP912575812, T5z, T5y);
|
cannam@167
|
604 T8a = FMA(KP851038619, T89, T88);
|
cannam@167
|
605 T4Y = FNMS(KP912018591, T4X, T4I);
|
cannam@167
|
606 T5t = FNMS(KP912575812, T5s, T5d);
|
cannam@167
|
607 T5u = FNMS(KP851038619, T5t, T4Y);
|
cannam@167
|
608 }
|
cannam@167
|
609 {
|
cannam@167
|
610 E T5K, T60, T5U, T64, T5N, T61, T5R, T63;
|
cannam@167
|
611 {
|
cannam@167
|
612 E T5I, T5J, T5S, T5T;
|
cannam@167
|
613 T5I = FMA(KP951056516, T5b, T58);
|
cannam@167
|
614 T5J = FMA(KP951056516, T54, T51);
|
cannam@167
|
615 T5K = FMA(KP256756360, T5J, T5I);
|
cannam@167
|
616 T60 = FNMS(KP256756360, T5I, T5J);
|
cannam@167
|
617 T5S = FNMS(KP951056516, T4z, T4w);
|
cannam@167
|
618 T5T = FNMS(KP951056516, T4G, T4D);
|
cannam@167
|
619 T5U = FMA(KP939062505, T5T, T5S);
|
cannam@167
|
620 T64 = FNMS(KP939062505, T5S, T5T);
|
cannam@167
|
621 }
|
cannam@167
|
622 {
|
cannam@167
|
623 E T5L, T5M, T5P, T5Q;
|
cannam@167
|
624 T5L = FMA(KP951056516, T5j, T5g);
|
cannam@167
|
625 T5M = FNMS(KP951056516, T5q, T5n);
|
cannam@167
|
626 T5N = FMA(KP634619297, T5M, T5L);
|
cannam@167
|
627 T61 = FNMS(KP634619297, T5L, T5M);
|
cannam@167
|
628 T5P = FNMS(KP951056516, T4V, T4S);
|
cannam@167
|
629 T5Q = FMA(KP951056516, T4O, T4L);
|
cannam@167
|
630 T5R = FMA(KP549754652, T5Q, T5P);
|
cannam@167
|
631 T63 = FNMS(KP549754652, T5P, T5Q);
|
cannam@167
|
632 }
|
cannam@167
|
633 T62 = FMA(KP871714437, T61, T60);
|
cannam@167
|
634 T65 = FNMS(KP831864738, T64, T63);
|
cannam@167
|
635 T83 = FNMS(KP871714437, T5N, T5K);
|
cannam@167
|
636 T82 = FNMS(KP831864738, T5U, T5R);
|
cannam@167
|
637 T7W = FNMS(KP871714437, T61, T60);
|
cannam@167
|
638 T7X = FMA(KP831864738, T64, T63);
|
cannam@167
|
639 T7Y = FMA(KP904730450, T7X, T7W);
|
cannam@167
|
640 T5O = FMA(KP871714437, T5N, T5K);
|
cannam@167
|
641 T5V = FMA(KP831864738, T5U, T5R);
|
cannam@167
|
642 T5W = FMA(KP904730450, T5V, T5O);
|
cannam@167
|
643 }
|
cannam@167
|
644 cr[WS(rs, 4)] = FNMS(KP992114701, T5u, T4t);
|
cannam@167
|
645 ci[WS(rs, 23)] = FMA(KP968583161, T7Y, T7V);
|
cannam@167
|
646 ci[WS(rs, 20)] = FNMS(KP992114701, T8a, T87);
|
cannam@167
|
647 cr[WS(rs, 1)] = FMA(KP968583161, T5W, T5H);
|
cannam@167
|
648 {
|
cannam@167
|
649 E T5E, T5G, T5x, T5F, T5v, T5w;
|
cannam@167
|
650 T5E = FNMS(KP726211448, T5D, T5A);
|
cannam@167
|
651 T5G = FMA(KP525970792, T5A, T5D);
|
cannam@167
|
652 T5v = FMA(KP248028675, T5u, T4t);
|
cannam@167
|
653 T5w = FMA(KP851038619, T5t, T4Y);
|
cannam@167
|
654 T5x = FMA(KP554608978, T5w, T5v);
|
cannam@167
|
655 T5F = FNMS(KP554608978, T5w, T5v);
|
cannam@167
|
656 cr[WS(rs, 9)] = FNMS(KP803003575, T5E, T5x);
|
cannam@167
|
657 ci[WS(rs, 5)] = FMA(KP943557151, T5G, T5F);
|
cannam@167
|
658 ci[0] = FMA(KP803003575, T5E, T5x);
|
cannam@167
|
659 ci[WS(rs, 10)] = FNMS(KP943557151, T5G, T5F);
|
cannam@167
|
660 }
|
cannam@167
|
661 {
|
cannam@167
|
662 E T84, T86, T81, T85, T7Z, T80;
|
cannam@167
|
663 T84 = FNMS(KP683113946, T83, T82);
|
cannam@167
|
664 T86 = FMA(KP559154169, T82, T83);
|
cannam@167
|
665 T7Z = FNMS(KP242145790, T7Y, T7V);
|
cannam@167
|
666 T80 = FNMS(KP904730450, T7X, T7W);
|
cannam@167
|
667 T81 = FNMS(KP541454447, T80, T7Z);
|
cannam@167
|
668 T85 = FMA(KP541454447, T80, T7Z);
|
cannam@167
|
669 cr[WS(rs, 16)] = FMS(KP833417178, T84, T81);
|
cannam@167
|
670 ci[WS(rs, 18)] = FNMS(KP921177326, T86, T85);
|
cannam@167
|
671 ci[WS(rs, 13)] = FMA(KP833417178, T84, T81);
|
cannam@167
|
672 cr[WS(rs, 21)] = -(FMA(KP921177326, T86, T85));
|
cannam@167
|
673 }
|
cannam@167
|
674 {
|
cannam@167
|
675 E T8g, T8i, T8d, T8h, T8b, T8c;
|
cannam@167
|
676 T8g = FNMS(KP525970792, T8f, T8e);
|
cannam@167
|
677 T8i = FMA(KP726211448, T8e, T8f);
|
cannam@167
|
678 T8b = FMA(KP248028675, T8a, T87);
|
cannam@167
|
679 T8c = FNMS(KP851038619, T89, T88);
|
cannam@167
|
680 T8d = FNMS(KP554608978, T8c, T8b);
|
cannam@167
|
681 T8h = FMA(KP554608978, T8c, T8b);
|
cannam@167
|
682 cr[WS(rs, 14)] = -(FMA(KP943557151, T8g, T8d));
|
cannam@167
|
683 ci[WS(rs, 15)] = FNMS(KP803003575, T8i, T8h);
|
cannam@167
|
684 cr[WS(rs, 19)] = FMS(KP943557151, T8g, T8d);
|
cannam@167
|
685 cr[WS(rs, 24)] = -(FMA(KP803003575, T8i, T8h));
|
cannam@167
|
686 }
|
cannam@167
|
687 {
|
cannam@167
|
688 E T66, T68, T5Z, T67, T5X, T5Y;
|
cannam@167
|
689 T66 = FMA(KP559154169, T65, T62);
|
cannam@167
|
690 T68 = FNMS(KP683113946, T62, T65);
|
cannam@167
|
691 T5X = FNMS(KP242145790, T5W, T5H);
|
cannam@167
|
692 T5Y = FNMS(KP904730450, T5V, T5O);
|
cannam@167
|
693 T5Z = FMA(KP541454447, T5Y, T5X);
|
cannam@167
|
694 T67 = FNMS(KP541454447, T5Y, T5X);
|
cannam@167
|
695 ci[WS(rs, 3)] = FNMS(KP921177326, T66, T5Z);
|
cannam@167
|
696 ci[WS(rs, 8)] = FMA(KP833417178, T68, T67);
|
cannam@167
|
697 cr[WS(rs, 6)] = FMA(KP921177326, T66, T5Z);
|
cannam@167
|
698 cr[WS(rs, 11)] = FNMS(KP833417178, T68, T67);
|
cannam@167
|
699 }
|
cannam@167
|
700 }
|
cannam@167
|
701 {
|
cannam@167
|
702 E T8s, T8u, T8j, T8m, T8n, T8o, T8t, T8p;
|
cannam@167
|
703 {
|
cannam@167
|
704 E T8q, T8r, T8k, T8l;
|
cannam@167
|
705 T8q = T2C - T3a;
|
cannam@167
|
706 T8r = T21 - T1v;
|
cannam@167
|
707 T8s = FMA(KP618033988, T8r, T8q);
|
cannam@167
|
708 T8u = FNMS(KP618033988, T8q, T8r);
|
cannam@167
|
709 T8j = T7o + T7l;
|
cannam@167
|
710 T8k = T3t + T3H;
|
cannam@167
|
711 T8l = T3W + T4a;
|
cannam@167
|
712 T8m = T8k + T8l;
|
cannam@167
|
713 T8n = FNMS(KP250000000, T8m, T8j);
|
cannam@167
|
714 T8o = T8k - T8l;
|
cannam@167
|
715 }
|
cannam@167
|
716 ci[WS(rs, 24)] = T8m + T8j;
|
cannam@167
|
717 T8t = FMA(KP559016994, T8o, T8n);
|
cannam@167
|
718 cr[WS(rs, 20)] = FMS(KP951056516, T8u, T8t);
|
cannam@167
|
719 ci[WS(rs, 19)] = FMA(KP951056516, T8u, T8t);
|
cannam@167
|
720 T8p = FNMS(KP559016994, T8o, T8n);
|
cannam@167
|
721 cr[WS(rs, 15)] = FMS(KP951056516, T8s, T8p);
|
cannam@167
|
722 ci[WS(rs, 14)] = FMA(KP951056516, T8s, T8p);
|
cannam@167
|
723 }
|
cannam@167
|
724 {
|
cannam@167
|
725 E T6b, T6T, T7v, T7H, T6M, T6P, T7P, T7O, T7I, T7J, T7K, T6q, T6F, T6G, T7e;
|
cannam@167
|
726 E T7h, T7D, T7C, T7w, T7x, T7y, T70, T77, T78, T69, T7r;
|
cannam@167
|
727 T69 = FNMS(KP559016994, T4g, T4f);
|
cannam@167
|
728 T6b = FMA(KP951056516, T6a, T69);
|
cannam@167
|
729 T6T = FNMS(KP951056516, T6a, T69);
|
cannam@167
|
730 T7r = FNMS(KP559016994, T7q, T7p);
|
cannam@167
|
731 T7v = FMA(KP951056516, T7u, T7r);
|
cannam@167
|
732 T7H = FNMS(KP951056516, T7u, T7r);
|
cannam@167
|
733 {
|
cannam@167
|
734 E T6i, T6N, T6E, T6L, T6p, T6O, T6x, T6K;
|
cannam@167
|
735 {
|
cannam@167
|
736 E T6e, T6h, T6A, T6D;
|
cannam@167
|
737 T6e = FMA(KP951056516, T6d, T6c);
|
cannam@167
|
738 T6h = FMA(KP951056516, T6g, T6f);
|
cannam@167
|
739 T6i = FMA(KP062914667, T6h, T6e);
|
cannam@167
|
740 T6N = FNMS(KP062914667, T6e, T6h);
|
cannam@167
|
741 T6A = FNMS(KP951056516, T6z, T6y);
|
cannam@167
|
742 T6D = FMA(KP951056516, T6C, T6B);
|
cannam@167
|
743 T6E = FMA(KP939062505, T6D, T6A);
|
cannam@167
|
744 T6L = FNMS(KP939062505, T6A, T6D);
|
cannam@167
|
745 }
|
cannam@167
|
746 {
|
cannam@167
|
747 E T6l, T6o, T6t, T6w;
|
cannam@167
|
748 T6l = FNMS(KP951056516, T6k, T6j);
|
cannam@167
|
749 T6o = FNMS(KP951056516, T6n, T6m);
|
cannam@167
|
750 T6p = FNMS(KP827271945, T6o, T6l);
|
cannam@167
|
751 T6O = FMA(KP827271945, T6l, T6o);
|
cannam@167
|
752 T6t = FNMS(KP951056516, T6s, T6r);
|
cannam@167
|
753 T6w = FMA(KP951056516, T6v, T6u);
|
cannam@167
|
754 T6x = FNMS(KP126329378, T6w, T6t);
|
cannam@167
|
755 T6K = FMA(KP126329378, T6t, T6w);
|
cannam@167
|
756 }
|
cannam@167
|
757 T6M = FMA(KP734762448, T6L, T6K);
|
cannam@167
|
758 T6P = FNMS(KP772036680, T6O, T6N);
|
cannam@167
|
759 T7P = FMA(KP734762448, T6E, T6x);
|
cannam@167
|
760 T7O = FNMS(KP772036680, T6p, T6i);
|
cannam@167
|
761 T7I = FMA(KP772036680, T6O, T6N);
|
cannam@167
|
762 T7J = FNMS(KP734762448, T6L, T6K);
|
cannam@167
|
763 T7K = FMA(KP994076283, T7J, T7I);
|
cannam@167
|
764 T6q = FMA(KP772036680, T6p, T6i);
|
cannam@167
|
765 T6F = FNMS(KP734762448, T6E, T6x);
|
cannam@167
|
766 T6G = FNMS(KP994076283, T6F, T6q);
|
cannam@167
|
767 }
|
cannam@167
|
768 {
|
cannam@167
|
769 E T6W, T7f, T76, T7d, T6Z, T7g, T73, T7c;
|
cannam@167
|
770 {
|
cannam@167
|
771 E T6U, T6V, T74, T75;
|
cannam@167
|
772 T6U = FMA(KP951056516, T6k, T6j);
|
cannam@167
|
773 T6V = FMA(KP951056516, T6n, T6m);
|
cannam@167
|
774 T6W = FMA(KP062914667, T6V, T6U);
|
cannam@167
|
775 T7f = FNMS(KP062914667, T6U, T6V);
|
cannam@167
|
776 T74 = FMA(KP951056516, T6z, T6y);
|
cannam@167
|
777 T75 = FNMS(KP951056516, T6C, T6B);
|
cannam@167
|
778 T76 = FMA(KP549754652, T75, T74);
|
cannam@167
|
779 T7d = FNMS(KP549754652, T74, T75);
|
cannam@167
|
780 }
|
cannam@167
|
781 {
|
cannam@167
|
782 E T6X, T6Y, T71, T72;
|
cannam@167
|
783 T6X = FNMS(KP951056516, T6d, T6c);
|
cannam@167
|
784 T6Y = FNMS(KP951056516, T6g, T6f);
|
cannam@167
|
785 T6Z = FMA(KP634619297, T6Y, T6X);
|
cannam@167
|
786 T7g = FNMS(KP634619297, T6X, T6Y);
|
cannam@167
|
787 T71 = FNMS(KP951056516, T6v, T6u);
|
cannam@167
|
788 T72 = FMA(KP951056516, T6s, T6r);
|
cannam@167
|
789 T73 = FNMS(KP470564281, T72, T71);
|
cannam@167
|
790 T7c = FMA(KP470564281, T71, T72);
|
cannam@167
|
791 }
|
cannam@167
|
792 T7e = FMA(KP968479752, T7d, T7c);
|
cannam@167
|
793 T7h = FNMS(KP845997307, T7g, T7f);
|
cannam@167
|
794 T7D = FNMS(KP968479752, T76, T73);
|
cannam@167
|
795 T7C = FNMS(KP845997307, T6Z, T6W);
|
cannam@167
|
796 T7w = FMA(KP845997307, T7g, T7f);
|
cannam@167
|
797 T7x = FNMS(KP968479752, T7d, T7c);
|
cannam@167
|
798 T7y = FMA(KP906616052, T7x, T7w);
|
cannam@167
|
799 T70 = FMA(KP845997307, T6Z, T6W);
|
cannam@167
|
800 T77 = FMA(KP968479752, T76, T73);
|
cannam@167
|
801 T78 = FMA(KP906616052, T77, T70);
|
cannam@167
|
802 }
|
cannam@167
|
803 cr[WS(rs, 3)] = FMA(KP998026728, T6G, T6b);
|
cannam@167
|
804 ci[WS(rs, 22)] = FNMS(KP998026728, T7y, T7v);
|
cannam@167
|
805 ci[WS(rs, 21)] = FNMS(KP998026728, T7K, T7H);
|
cannam@167
|
806 cr[WS(rs, 2)] = FMA(KP998026728, T78, T6T);
|
cannam@167
|
807 {
|
cannam@167
|
808 E T7E, T7G, T7B, T7F, T7z, T7A;
|
cannam@167
|
809 T7E = FNMS(KP560319534, T7D, T7C);
|
cannam@167
|
810 T7G = FMA(KP681693190, T7C, T7D);
|
cannam@167
|
811 T7z = FMA(KP249506682, T7y, T7v);
|
cannam@167
|
812 T7A = FNMS(KP906616052, T7x, T7w);
|
cannam@167
|
813 T7B = FNMS(KP557913902, T7A, T7z);
|
cannam@167
|
814 T7F = FMA(KP557913902, T7A, T7z);
|
cannam@167
|
815 cr[WS(rs, 17)] = -(FMA(KP949179823, T7E, T7B));
|
cannam@167
|
816 ci[WS(rs, 17)] = FMA(KP860541664, T7G, T7F);
|
cannam@167
|
817 ci[WS(rs, 12)] = FNMS(KP949179823, T7E, T7B);
|
cannam@167
|
818 cr[WS(rs, 22)] = FMS(KP860541664, T7G, T7F);
|
cannam@167
|
819 }
|
cannam@167
|
820 {
|
cannam@167
|
821 E T7i, T7k, T7b, T7j, T79, T7a;
|
cannam@167
|
822 T7i = FMA(KP681693190, T7h, T7e);
|
cannam@167
|
823 T7k = FNMS(KP560319534, T7e, T7h);
|
cannam@167
|
824 T79 = FNMS(KP249506682, T78, T6T);
|
cannam@167
|
825 T7a = FNMS(KP906616052, T77, T70);
|
cannam@167
|
826 T7b = FNMS(KP557913902, T7a, T79);
|
cannam@167
|
827 T7j = FMA(KP557913902, T7a, T79);
|
cannam@167
|
828 ci[WS(rs, 2)] = FNMS(KP860541664, T7i, T7b);
|
cannam@167
|
829 cr[WS(rs, 12)] = FNMS(KP949179823, T7k, T7j);
|
cannam@167
|
830 cr[WS(rs, 7)] = FMA(KP860541664, T7i, T7b);
|
cannam@167
|
831 ci[WS(rs, 7)] = FMA(KP949179823, T7k, T7j);
|
cannam@167
|
832 }
|
cannam@167
|
833 {
|
cannam@167
|
834 E T6Q, T6S, T6J, T6R, T6H, T6I;
|
cannam@167
|
835 T6Q = FNMS(KP621716863, T6P, T6M);
|
cannam@167
|
836 T6S = FMA(KP614372930, T6M, T6P);
|
cannam@167
|
837 T6H = FNMS(KP249506682, T6G, T6b);
|
cannam@167
|
838 T6I = FMA(KP994076283, T6F, T6q);
|
cannam@167
|
839 T6J = FNMS(KP557913902, T6I, T6H);
|
cannam@167
|
840 T6R = FMA(KP557913902, T6I, T6H);
|
cannam@167
|
841 ci[WS(rs, 1)] = FNMS(KP943557151, T6Q, T6J);
|
cannam@167
|
842 ci[WS(rs, 11)] = FMA(KP949179823, T6S, T6R);
|
cannam@167
|
843 cr[WS(rs, 8)] = FMA(KP943557151, T6Q, T6J);
|
cannam@167
|
844 ci[WS(rs, 6)] = FNMS(KP949179823, T6S, T6R);
|
cannam@167
|
845 }
|
cannam@167
|
846 {
|
cannam@167
|
847 E T7Q, T7S, T7N, T7R, T7L, T7M;
|
cannam@167
|
848 T7Q = FNMS(KP614372930, T7P, T7O);
|
cannam@167
|
849 T7S = FMA(KP621716863, T7O, T7P);
|
cannam@167
|
850 T7L = FMA(KP249506682, T7K, T7H);
|
cannam@167
|
851 T7M = FNMS(KP994076283, T7J, T7I);
|
cannam@167
|
852 T7N = FNMS(KP557913902, T7M, T7L);
|
cannam@167
|
853 T7R = FMA(KP557913902, T7M, T7L);
|
cannam@167
|
854 cr[WS(rs, 13)] = -(FMA(KP949179823, T7Q, T7N));
|
cannam@167
|
855 ci[WS(rs, 16)] = FNMS(KP943557151, T7S, T7R);
|
cannam@167
|
856 cr[WS(rs, 18)] = FMS(KP949179823, T7Q, T7N);
|
cannam@167
|
857 cr[WS(rs, 23)] = -(FMA(KP943557151, T7S, T7R));
|
cannam@167
|
858 }
|
cannam@167
|
859 }
|
cannam@167
|
860 }
|
cannam@167
|
861 }
|
cannam@167
|
862 }
|
cannam@167
|
863 }
|
cannam@167
|
864
|
cannam@167
|
865 static const tw_instr twinstr[] = {
|
cannam@167
|
866 {TW_CEXP, 1, 1},
|
cannam@167
|
867 {TW_CEXP, 1, 3},
|
cannam@167
|
868 {TW_CEXP, 1, 9},
|
cannam@167
|
869 {TW_CEXP, 1, 24},
|
cannam@167
|
870 {TW_NEXT, 1, 0}
|
cannam@167
|
871 };
|
cannam@167
|
872
|
cannam@167
|
873 static const hc2hc_desc desc = { 25, "hf2_25", twinstr, &GENUS, {84, 78, 356, 0} };
|
cannam@167
|
874
|
cannam@167
|
875 void X(codelet_hf2_25) (planner *p) {
|
cannam@167
|
876 X(khc2hc_register) (p, hf2_25, &desc);
|
cannam@167
|
877 }
|
cannam@167
|
878 #else
|
cannam@167
|
879
|
cannam@167
|
880 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 25 -dit -name hf2_25 -include rdft/scalar/hf.h */
|
cannam@167
|
881
|
cannam@167
|
882 /*
|
cannam@167
|
883 * This function contains 440 FP additions, 340 FP multiplications,
|
cannam@167
|
884 * (or, 280 additions, 180 multiplications, 160 fused multiply/add),
|
cannam@167
|
885 * 149 stack variables, 20 constants, and 100 memory accesses
|
cannam@167
|
886 */
|
cannam@167
|
887 #include "rdft/scalar/hf.h"
|
cannam@167
|
888
|
cannam@167
|
889 static void hf2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
890 {
|
cannam@167
|
891 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
cannam@167
|
892 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
cannam@167
|
893 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
cannam@167
|
894 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
cannam@167
|
895 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
cannam@167
|
896 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
cannam@167
|
897 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
cannam@167
|
898 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
cannam@167
|
899 DK(KP992114701, +0.992114701314477831049793042785778521453036709);
|
cannam@167
|
900 DK(KP125333233, +0.125333233564304245373118759816508793942918247);
|
cannam@167
|
901 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
cannam@167
|
902 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
cannam@167
|
903 DK(KP637423989, +0.637423989748689710176712811676016195434917298);
|
cannam@167
|
904 DK(KP770513242, +0.770513242775789230803009636396177847271667672);
|
cannam@167
|
905 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
cannam@167
|
906 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
cannam@167
|
907 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
cannam@167
|
908 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
909 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
910 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
911 {
|
cannam@167
|
912 INT m;
|
cannam@167
|
913 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
|
914 E T2, T5, T3, T6, T8, Td, T16, T14, Te, T9, T21, T23, Tx, TR, T1g;
|
cannam@167
|
915 E TB, T1f, TV, T1Q, Tg, T1S, Tk, T18, T2s, T1c, T2q, Tn, To, Tp, Tr;
|
cannam@167
|
916 E T28, T2x, TY, T2k, T2m, T2v, TG, TE, T10, T1h, T1E, T26, T1B, T1G, T1V;
|
cannam@167
|
917 E T1X, T1z, T1j;
|
cannam@167
|
918 {
|
cannam@167
|
919 E Tw, TT, Tz, TQ, Tv, TU, TA, TP;
|
cannam@167
|
920 {
|
cannam@167
|
921 E T4, Tc, T7, Tb;
|
cannam@167
|
922 T2 = W[0];
|
cannam@167
|
923 T5 = W[1];
|
cannam@167
|
924 T3 = W[2];
|
cannam@167
|
925 T6 = W[3];
|
cannam@167
|
926 T4 = T2 * T3;
|
cannam@167
|
927 Tc = T5 * T3;
|
cannam@167
|
928 T7 = T5 * T6;
|
cannam@167
|
929 Tb = T2 * T6;
|
cannam@167
|
930 T8 = T4 - T7;
|
cannam@167
|
931 Td = Tb + Tc;
|
cannam@167
|
932 T16 = Tb - Tc;
|
cannam@167
|
933 T14 = T4 + T7;
|
cannam@167
|
934 Te = W[5];
|
cannam@167
|
935 Tw = T5 * Te;
|
cannam@167
|
936 TT = T3 * Te;
|
cannam@167
|
937 Tz = T2 * Te;
|
cannam@167
|
938 TQ = T6 * Te;
|
cannam@167
|
939 T9 = W[4];
|
cannam@167
|
940 Tv = T2 * T9;
|
cannam@167
|
941 TU = T6 * T9;
|
cannam@167
|
942 TA = T5 * T9;
|
cannam@167
|
943 TP = T3 * T9;
|
cannam@167
|
944 }
|
cannam@167
|
945 T21 = TP - TQ;
|
cannam@167
|
946 T23 = TT + TU;
|
cannam@167
|
947 {
|
cannam@167
|
948 E T15, T17, Ta, Tf, T1a, T1b, Ti, Tj;
|
cannam@167
|
949 Tx = Tv - Tw;
|
cannam@167
|
950 TR = TP + TQ;
|
cannam@167
|
951 T1g = Tz - TA;
|
cannam@167
|
952 TB = Tz + TA;
|
cannam@167
|
953 T1f = Tv + Tw;
|
cannam@167
|
954 TV = TT - TU;
|
cannam@167
|
955 T15 = T14 * T9;
|
cannam@167
|
956 T17 = T16 * Te;
|
cannam@167
|
957 T1Q = T15 + T17;
|
cannam@167
|
958 Ta = T8 * T9;
|
cannam@167
|
959 Tf = Td * Te;
|
cannam@167
|
960 Tg = Ta + Tf;
|
cannam@167
|
961 T1a = T14 * Te;
|
cannam@167
|
962 T1b = T16 * T9;
|
cannam@167
|
963 T1S = T1a - T1b;
|
cannam@167
|
964 Ti = T8 * Te;
|
cannam@167
|
965 Tj = Td * T9;
|
cannam@167
|
966 Tk = Ti - Tj;
|
cannam@167
|
967 T18 = T15 - T17;
|
cannam@167
|
968 T2s = Ti + Tj;
|
cannam@167
|
969 T1c = T1a + T1b;
|
cannam@167
|
970 T2q = Ta - Tf;
|
cannam@167
|
971 Tn = W[6];
|
cannam@167
|
972 To = W[7];
|
cannam@167
|
973 Tp = FMA(T8, Tn, Td * To);
|
cannam@167
|
974 Tr = FNMS(Td, Tn, T8 * To);
|
cannam@167
|
975 T28 = FNMS(T1S, Tn, T1Q * To);
|
cannam@167
|
976 T2x = FNMS(TV, Tn, TR * To);
|
cannam@167
|
977 TY = FMA(T3, Tn, T6 * To);
|
cannam@167
|
978 T2k = FMA(T2, Tn, T5 * To);
|
cannam@167
|
979 T2m = FNMS(T5, Tn, T2 * To);
|
cannam@167
|
980 T2v = FMA(TR, Tn, TV * To);
|
cannam@167
|
981 TG = FNMS(Te, Tn, T9 * To);
|
cannam@167
|
982 TE = FMA(T9, Tn, Te * To);
|
cannam@167
|
983 T10 = FNMS(T6, Tn, T3 * To);
|
cannam@167
|
984 T1h = FMA(T1f, Tn, T1g * To);
|
cannam@167
|
985 T1E = FMA(Tg, Tn, Tk * To);
|
cannam@167
|
986 T26 = FMA(T1Q, Tn, T1S * To);
|
cannam@167
|
987 T1B = FNMS(TB, Tn, Tx * To);
|
cannam@167
|
988 T1G = FNMS(Tk, Tn, Tg * To);
|
cannam@167
|
989 T1V = FMA(T14, Tn, T16 * To);
|
cannam@167
|
990 T1X = FNMS(T16, Tn, T14 * To);
|
cannam@167
|
991 T1z = FMA(Tx, Tn, TB * To);
|
cannam@167
|
992 T1j = FNMS(T1g, Tn, T1f * To);
|
cannam@167
|
993 }
|
cannam@167
|
994 }
|
cannam@167
|
995 {
|
cannam@167
|
996 E T1, T6v, T2F, T6A, TK, T2G, T6y, T6z, T6u, T71, T2O, T52, T2C, T6k, T4c;
|
cannam@167
|
997 E T5X, T4L, T5s, T4j, T5W, T4K, T5v, T1o, T6g, T30, T5M, T4A, T56, T3b, T5N;
|
cannam@167
|
998 E T4B, T59, T1L, T6h, T3r, T5P, T4E, T5d, T3y, T5Q, T4D, T5g, T2d, T6j, T3P;
|
cannam@167
|
999 E T5U, T4I, T5o, T3W, T5T, T4H, T5l;
|
cannam@167
|
1000 {
|
cannam@167
|
1001 E Tm, T2I, Tt, T2J, Tu, T6w, TD, T2L, TI, T2M, TJ, T6x;
|
cannam@167
|
1002 T1 = cr[0];
|
cannam@167
|
1003 T6v = ci[0];
|
cannam@167
|
1004 {
|
cannam@167
|
1005 E Th, Tl, Tq, Ts;
|
cannam@167
|
1006 Th = cr[WS(rs, 5)];
|
cannam@167
|
1007 Tl = ci[WS(rs, 5)];
|
cannam@167
|
1008 Tm = FMA(Tg, Th, Tk * Tl);
|
cannam@167
|
1009 T2I = FNMS(Tk, Th, Tg * Tl);
|
cannam@167
|
1010 Tq = cr[WS(rs, 20)];
|
cannam@167
|
1011 Ts = ci[WS(rs, 20)];
|
cannam@167
|
1012 Tt = FMA(Tp, Tq, Tr * Ts);
|
cannam@167
|
1013 T2J = FNMS(Tr, Tq, Tp * Ts);
|
cannam@167
|
1014 }
|
cannam@167
|
1015 Tu = Tm + Tt;
|
cannam@167
|
1016 T6w = T2I + T2J;
|
cannam@167
|
1017 {
|
cannam@167
|
1018 E Ty, TC, TF, TH;
|
cannam@167
|
1019 Ty = cr[WS(rs, 10)];
|
cannam@167
|
1020 TC = ci[WS(rs, 10)];
|
cannam@167
|
1021 TD = FMA(Tx, Ty, TB * TC);
|
cannam@167
|
1022 T2L = FNMS(TB, Ty, Tx * TC);
|
cannam@167
|
1023 TF = cr[WS(rs, 15)];
|
cannam@167
|
1024 TH = ci[WS(rs, 15)];
|
cannam@167
|
1025 TI = FMA(TE, TF, TG * TH);
|
cannam@167
|
1026 T2M = FNMS(TG, TF, TE * TH);
|
cannam@167
|
1027 }
|
cannam@167
|
1028 TJ = TD + TI;
|
cannam@167
|
1029 T6x = T2L + T2M;
|
cannam@167
|
1030 T2F = KP559016994 * (Tu - TJ);
|
cannam@167
|
1031 T6A = KP559016994 * (T6w - T6x);
|
cannam@167
|
1032 TK = Tu + TJ;
|
cannam@167
|
1033 T2G = FNMS(KP250000000, TK, T1);
|
cannam@167
|
1034 T6y = T6w + T6x;
|
cannam@167
|
1035 T6z = FNMS(KP250000000, T6y, T6v);
|
cannam@167
|
1036 {
|
cannam@167
|
1037 E T6s, T6t, T2K, T2N;
|
cannam@167
|
1038 T6s = TD - TI;
|
cannam@167
|
1039 T6t = Tm - Tt;
|
cannam@167
|
1040 T6u = FNMS(KP587785252, T6t, KP951056516 * T6s);
|
cannam@167
|
1041 T71 = FMA(KP951056516, T6t, KP587785252 * T6s);
|
cannam@167
|
1042 T2K = T2I - T2J;
|
cannam@167
|
1043 T2N = T2L - T2M;
|
cannam@167
|
1044 T2O = FMA(KP951056516, T2K, KP587785252 * T2N);
|
cannam@167
|
1045 T52 = FNMS(KP587785252, T2K, KP951056516 * T2N);
|
cannam@167
|
1046 }
|
cannam@167
|
1047 }
|
cannam@167
|
1048 {
|
cannam@167
|
1049 E T2g, T48, T3Y, T3Z, T4h, T4g, T43, T46, T49, T2p, T2A, T2B, T2e, T2f;
|
cannam@167
|
1050 T2e = cr[WS(rs, 3)];
|
cannam@167
|
1051 T2f = ci[WS(rs, 3)];
|
cannam@167
|
1052 T2g = FMA(T3, T2e, T6 * T2f);
|
cannam@167
|
1053 T48 = FNMS(T6, T2e, T3 * T2f);
|
cannam@167
|
1054 {
|
cannam@167
|
1055 E T2j, T41, T2z, T45, T2o, T42, T2u, T44;
|
cannam@167
|
1056 {
|
cannam@167
|
1057 E T2h, T2i, T2w, T2y;
|
cannam@167
|
1058 T2h = cr[WS(rs, 8)];
|
cannam@167
|
1059 T2i = ci[WS(rs, 8)];
|
cannam@167
|
1060 T2j = FMA(T1f, T2h, T1g * T2i);
|
cannam@167
|
1061 T41 = FNMS(T1g, T2h, T1f * T2i);
|
cannam@167
|
1062 T2w = cr[WS(rs, 18)];
|
cannam@167
|
1063 T2y = ci[WS(rs, 18)];
|
cannam@167
|
1064 T2z = FMA(T2v, T2w, T2x * T2y);
|
cannam@167
|
1065 T45 = FNMS(T2x, T2w, T2v * T2y);
|
cannam@167
|
1066 }
|
cannam@167
|
1067 {
|
cannam@167
|
1068 E T2l, T2n, T2r, T2t;
|
cannam@167
|
1069 T2l = cr[WS(rs, 23)];
|
cannam@167
|
1070 T2n = ci[WS(rs, 23)];
|
cannam@167
|
1071 T2o = FMA(T2k, T2l, T2m * T2n);
|
cannam@167
|
1072 T42 = FNMS(T2m, T2l, T2k * T2n);
|
cannam@167
|
1073 T2r = cr[WS(rs, 13)];
|
cannam@167
|
1074 T2t = ci[WS(rs, 13)];
|
cannam@167
|
1075 T2u = FMA(T2q, T2r, T2s * T2t);
|
cannam@167
|
1076 T44 = FNMS(T2s, T2r, T2q * T2t);
|
cannam@167
|
1077 }
|
cannam@167
|
1078 T3Y = T2j - T2o;
|
cannam@167
|
1079 T3Z = T2u - T2z;
|
cannam@167
|
1080 T4h = T44 - T45;
|
cannam@167
|
1081 T4g = T41 - T42;
|
cannam@167
|
1082 T43 = T41 + T42;
|
cannam@167
|
1083 T46 = T44 + T45;
|
cannam@167
|
1084 T49 = T43 + T46;
|
cannam@167
|
1085 T2p = T2j + T2o;
|
cannam@167
|
1086 T2A = T2u + T2z;
|
cannam@167
|
1087 T2B = T2p + T2A;
|
cannam@167
|
1088 }
|
cannam@167
|
1089 T2C = T2g + T2B;
|
cannam@167
|
1090 T6k = T48 + T49;
|
cannam@167
|
1091 {
|
cannam@167
|
1092 E T40, T5r, T4b, T5q, T47, T4a;
|
cannam@167
|
1093 T40 = FMA(KP951056516, T3Y, KP587785252 * T3Z);
|
cannam@167
|
1094 T5r = FNMS(KP587785252, T3Y, KP951056516 * T3Z);
|
cannam@167
|
1095 T47 = KP559016994 * (T43 - T46);
|
cannam@167
|
1096 T4a = FNMS(KP250000000, T49, T48);
|
cannam@167
|
1097 T4b = T47 + T4a;
|
cannam@167
|
1098 T5q = T4a - T47;
|
cannam@167
|
1099 T4c = T40 + T4b;
|
cannam@167
|
1100 T5X = T5r + T5q;
|
cannam@167
|
1101 T4L = T4b - T40;
|
cannam@167
|
1102 T5s = T5q - T5r;
|
cannam@167
|
1103 }
|
cannam@167
|
1104 {
|
cannam@167
|
1105 E T4i, T5u, T4f, T5t, T4d, T4e;
|
cannam@167
|
1106 T4i = FMA(KP951056516, T4g, KP587785252 * T4h);
|
cannam@167
|
1107 T5u = FNMS(KP587785252, T4g, KP951056516 * T4h);
|
cannam@167
|
1108 T4d = KP559016994 * (T2p - T2A);
|
cannam@167
|
1109 T4e = FNMS(KP250000000, T2B, T2g);
|
cannam@167
|
1110 T4f = T4d + T4e;
|
cannam@167
|
1111 T5t = T4e - T4d;
|
cannam@167
|
1112 T4j = T4f - T4i;
|
cannam@167
|
1113 T5W = T5t - T5u;
|
cannam@167
|
1114 T4K = T4f + T4i;
|
cannam@167
|
1115 T5v = T5t + T5u;
|
cannam@167
|
1116 }
|
cannam@167
|
1117 }
|
cannam@167
|
1118 {
|
cannam@167
|
1119 E TO, T37, T2V, T2Y, T32, T31, T34, T35, T38, T13, T1m, T1n, TM, TN;
|
cannam@167
|
1120 TM = cr[WS(rs, 1)];
|
cannam@167
|
1121 TN = ci[WS(rs, 1)];
|
cannam@167
|
1122 TO = FMA(T2, TM, T5 * TN);
|
cannam@167
|
1123 T37 = FNMS(T5, TM, T2 * TN);
|
cannam@167
|
1124 {
|
cannam@167
|
1125 E TX, T2T, T1l, T2X, T12, T2U, T1e, T2W;
|
cannam@167
|
1126 {
|
cannam@167
|
1127 E TS, TW, T1i, T1k;
|
cannam@167
|
1128 TS = cr[WS(rs, 6)];
|
cannam@167
|
1129 TW = ci[WS(rs, 6)];
|
cannam@167
|
1130 TX = FMA(TR, TS, TV * TW);
|
cannam@167
|
1131 T2T = FNMS(TV, TS, TR * TW);
|
cannam@167
|
1132 T1i = cr[WS(rs, 16)];
|
cannam@167
|
1133 T1k = ci[WS(rs, 16)];
|
cannam@167
|
1134 T1l = FMA(T1h, T1i, T1j * T1k);
|
cannam@167
|
1135 T2X = FNMS(T1j, T1i, T1h * T1k);
|
cannam@167
|
1136 }
|
cannam@167
|
1137 {
|
cannam@167
|
1138 E TZ, T11, T19, T1d;
|
cannam@167
|
1139 TZ = cr[WS(rs, 21)];
|
cannam@167
|
1140 T11 = ci[WS(rs, 21)];
|
cannam@167
|
1141 T12 = FMA(TY, TZ, T10 * T11);
|
cannam@167
|
1142 T2U = FNMS(T10, TZ, TY * T11);
|
cannam@167
|
1143 T19 = cr[WS(rs, 11)];
|
cannam@167
|
1144 T1d = ci[WS(rs, 11)];
|
cannam@167
|
1145 T1e = FMA(T18, T19, T1c * T1d);
|
cannam@167
|
1146 T2W = FNMS(T1c, T19, T18 * T1d);
|
cannam@167
|
1147 }
|
cannam@167
|
1148 T2V = T2T - T2U;
|
cannam@167
|
1149 T2Y = T2W - T2X;
|
cannam@167
|
1150 T32 = T1e - T1l;
|
cannam@167
|
1151 T31 = TX - T12;
|
cannam@167
|
1152 T34 = T2T + T2U;
|
cannam@167
|
1153 T35 = T2W + T2X;
|
cannam@167
|
1154 T38 = T34 + T35;
|
cannam@167
|
1155 T13 = TX + T12;
|
cannam@167
|
1156 T1m = T1e + T1l;
|
cannam@167
|
1157 T1n = T13 + T1m;
|
cannam@167
|
1158 }
|
cannam@167
|
1159 T1o = TO + T1n;
|
cannam@167
|
1160 T6g = T37 + T38;
|
cannam@167
|
1161 {
|
cannam@167
|
1162 E T2Z, T55, T2S, T54, T2Q, T2R;
|
cannam@167
|
1163 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
|
cannam@167
|
1164 T55 = FNMS(KP587785252, T2V, KP951056516 * T2Y);
|
cannam@167
|
1165 T2Q = KP559016994 * (T13 - T1m);
|
cannam@167
|
1166 T2R = FNMS(KP250000000, T1n, TO);
|
cannam@167
|
1167 T2S = T2Q + T2R;
|
cannam@167
|
1168 T54 = T2R - T2Q;
|
cannam@167
|
1169 T30 = T2S - T2Z;
|
cannam@167
|
1170 T5M = T54 - T55;
|
cannam@167
|
1171 T4A = T2S + T2Z;
|
cannam@167
|
1172 T56 = T54 + T55;
|
cannam@167
|
1173 }
|
cannam@167
|
1174 {
|
cannam@167
|
1175 E T33, T58, T3a, T57, T36, T39;
|
cannam@167
|
1176 T33 = FMA(KP951056516, T31, KP587785252 * T32);
|
cannam@167
|
1177 T58 = FNMS(KP587785252, T31, KP951056516 * T32);
|
cannam@167
|
1178 T36 = KP559016994 * (T34 - T35);
|
cannam@167
|
1179 T39 = FNMS(KP250000000, T38, T37);
|
cannam@167
|
1180 T3a = T36 + T39;
|
cannam@167
|
1181 T57 = T39 - T36;
|
cannam@167
|
1182 T3b = T33 + T3a;
|
cannam@167
|
1183 T5N = T58 + T57;
|
cannam@167
|
1184 T4B = T3a - T33;
|
cannam@167
|
1185 T59 = T57 - T58;
|
cannam@167
|
1186 }
|
cannam@167
|
1187 }
|
cannam@167
|
1188 {
|
cannam@167
|
1189 E T1r, T3n, T3d, T3e, T3w, T3v, T3i, T3l, T3o, T1y, T1J, T1K, T1p, T1q;
|
cannam@167
|
1190 T1p = cr[WS(rs, 4)];
|
cannam@167
|
1191 T1q = ci[WS(rs, 4)];
|
cannam@167
|
1192 T1r = FMA(T8, T1p, Td * T1q);
|
cannam@167
|
1193 T3n = FNMS(Td, T1p, T8 * T1q);
|
cannam@167
|
1194 {
|
cannam@167
|
1195 E T1u, T3g, T1I, T3k, T1x, T3h, T1D, T3j;
|
cannam@167
|
1196 {
|
cannam@167
|
1197 E T1s, T1t, T1F, T1H;
|
cannam@167
|
1198 T1s = cr[WS(rs, 9)];
|
cannam@167
|
1199 T1t = ci[WS(rs, 9)];
|
cannam@167
|
1200 T1u = FMA(T9, T1s, Te * T1t);
|
cannam@167
|
1201 T3g = FNMS(Te, T1s, T9 * T1t);
|
cannam@167
|
1202 T1F = cr[WS(rs, 19)];
|
cannam@167
|
1203 T1H = ci[WS(rs, 19)];
|
cannam@167
|
1204 T1I = FMA(T1E, T1F, T1G * T1H);
|
cannam@167
|
1205 T3k = FNMS(T1G, T1F, T1E * T1H);
|
cannam@167
|
1206 }
|
cannam@167
|
1207 {
|
cannam@167
|
1208 E T1v, T1w, T1A, T1C;
|
cannam@167
|
1209 T1v = cr[WS(rs, 24)];
|
cannam@167
|
1210 T1w = ci[WS(rs, 24)];
|
cannam@167
|
1211 T1x = FMA(Tn, T1v, To * T1w);
|
cannam@167
|
1212 T3h = FNMS(To, T1v, Tn * T1w);
|
cannam@167
|
1213 T1A = cr[WS(rs, 14)];
|
cannam@167
|
1214 T1C = ci[WS(rs, 14)];
|
cannam@167
|
1215 T1D = FMA(T1z, T1A, T1B * T1C);
|
cannam@167
|
1216 T3j = FNMS(T1B, T1A, T1z * T1C);
|
cannam@167
|
1217 }
|
cannam@167
|
1218 T3d = T1x - T1u;
|
cannam@167
|
1219 T3e = T1D - T1I;
|
cannam@167
|
1220 T3w = T3j - T3k;
|
cannam@167
|
1221 T3v = T3g - T3h;
|
cannam@167
|
1222 T3i = T3g + T3h;
|
cannam@167
|
1223 T3l = T3j + T3k;
|
cannam@167
|
1224 T3o = T3i + T3l;
|
cannam@167
|
1225 T1y = T1u + T1x;
|
cannam@167
|
1226 T1J = T1D + T1I;
|
cannam@167
|
1227 T1K = T1y + T1J;
|
cannam@167
|
1228 }
|
cannam@167
|
1229 T1L = T1r + T1K;
|
cannam@167
|
1230 T6h = T3n + T3o;
|
cannam@167
|
1231 {
|
cannam@167
|
1232 E T3f, T5c, T3q, T5b, T3m, T3p;
|
cannam@167
|
1233 T3f = FNMS(KP587785252, T3e, KP951056516 * T3d);
|
cannam@167
|
1234 T5c = FMA(KP587785252, T3d, KP951056516 * T3e);
|
cannam@167
|
1235 T3m = KP559016994 * (T3i - T3l);
|
cannam@167
|
1236 T3p = FNMS(KP250000000, T3o, T3n);
|
cannam@167
|
1237 T3q = T3m + T3p;
|
cannam@167
|
1238 T5b = T3p - T3m;
|
cannam@167
|
1239 T3r = T3f - T3q;
|
cannam@167
|
1240 T5P = T5c + T5b;
|
cannam@167
|
1241 T4E = T3f + T3q;
|
cannam@167
|
1242 T5d = T5b - T5c;
|
cannam@167
|
1243 }
|
cannam@167
|
1244 {
|
cannam@167
|
1245 E T3x, T5f, T3u, T5e, T3s, T3t;
|
cannam@167
|
1246 T3x = FMA(KP951056516, T3v, KP587785252 * T3w);
|
cannam@167
|
1247 T5f = FNMS(KP587785252, T3v, KP951056516 * T3w);
|
cannam@167
|
1248 T3s = KP559016994 * (T1y - T1J);
|
cannam@167
|
1249 T3t = FNMS(KP250000000, T1K, T1r);
|
cannam@167
|
1250 T3u = T3s + T3t;
|
cannam@167
|
1251 T5e = T3t - T3s;
|
cannam@167
|
1252 T3y = T3u - T3x;
|
cannam@167
|
1253 T5Q = T5e - T5f;
|
cannam@167
|
1254 T4D = T3u + T3x;
|
cannam@167
|
1255 T5g = T5e + T5f;
|
cannam@167
|
1256 }
|
cannam@167
|
1257 }
|
cannam@167
|
1258 {
|
cannam@167
|
1259 E T1P, T3L, T3B, T3C, T3U, T3T, T3G, T3J, T3M, T20, T2b, T2c, T1N, T1O;
|
cannam@167
|
1260 T1N = cr[WS(rs, 2)];
|
cannam@167
|
1261 T1O = ci[WS(rs, 2)];
|
cannam@167
|
1262 T1P = FMA(T14, T1N, T16 * T1O);
|
cannam@167
|
1263 T3L = FNMS(T16, T1N, T14 * T1O);
|
cannam@167
|
1264 {
|
cannam@167
|
1265 E T1U, T3E, T2a, T3I, T1Z, T3F, T25, T3H;
|
cannam@167
|
1266 {
|
cannam@167
|
1267 E T1R, T1T, T27, T29;
|
cannam@167
|
1268 T1R = cr[WS(rs, 7)];
|
cannam@167
|
1269 T1T = ci[WS(rs, 7)];
|
cannam@167
|
1270 T1U = FMA(T1Q, T1R, T1S * T1T);
|
cannam@167
|
1271 T3E = FNMS(T1S, T1R, T1Q * T1T);
|
cannam@167
|
1272 T27 = cr[WS(rs, 17)];
|
cannam@167
|
1273 T29 = ci[WS(rs, 17)];
|
cannam@167
|
1274 T2a = FMA(T26, T27, T28 * T29);
|
cannam@167
|
1275 T3I = FNMS(T28, T27, T26 * T29);
|
cannam@167
|
1276 }
|
cannam@167
|
1277 {
|
cannam@167
|
1278 E T1W, T1Y, T22, T24;
|
cannam@167
|
1279 T1W = cr[WS(rs, 22)];
|
cannam@167
|
1280 T1Y = ci[WS(rs, 22)];
|
cannam@167
|
1281 T1Z = FMA(T1V, T1W, T1X * T1Y);
|
cannam@167
|
1282 T3F = FNMS(T1X, T1W, T1V * T1Y);
|
cannam@167
|
1283 T22 = cr[WS(rs, 12)];
|
cannam@167
|
1284 T24 = ci[WS(rs, 12)];
|
cannam@167
|
1285 T25 = FMA(T21, T22, T23 * T24);
|
cannam@167
|
1286 T3H = FNMS(T23, T22, T21 * T24);
|
cannam@167
|
1287 }
|
cannam@167
|
1288 T3B = T1U - T1Z;
|
cannam@167
|
1289 T3C = T25 - T2a;
|
cannam@167
|
1290 T3U = T3H - T3I;
|
cannam@167
|
1291 T3T = T3E - T3F;
|
cannam@167
|
1292 T3G = T3E + T3F;
|
cannam@167
|
1293 T3J = T3H + T3I;
|
cannam@167
|
1294 T3M = T3G + T3J;
|
cannam@167
|
1295 T20 = T1U + T1Z;
|
cannam@167
|
1296 T2b = T25 + T2a;
|
cannam@167
|
1297 T2c = T20 + T2b;
|
cannam@167
|
1298 }
|
cannam@167
|
1299 T2d = T1P + T2c;
|
cannam@167
|
1300 T6j = T3L + T3M;
|
cannam@167
|
1301 {
|
cannam@167
|
1302 E T3D, T5n, T3O, T5m, T3K, T3N;
|
cannam@167
|
1303 T3D = FMA(KP951056516, T3B, KP587785252 * T3C);
|
cannam@167
|
1304 T5n = FNMS(KP587785252, T3B, KP951056516 * T3C);
|
cannam@167
|
1305 T3K = KP559016994 * (T3G - T3J);
|
cannam@167
|
1306 T3N = FNMS(KP250000000, T3M, T3L);
|
cannam@167
|
1307 T3O = T3K + T3N;
|
cannam@167
|
1308 T5m = T3N - T3K;
|
cannam@167
|
1309 T3P = T3D + T3O;
|
cannam@167
|
1310 T5U = T5n + T5m;
|
cannam@167
|
1311 T4I = T3O - T3D;
|
cannam@167
|
1312 T5o = T5m - T5n;
|
cannam@167
|
1313 }
|
cannam@167
|
1314 {
|
cannam@167
|
1315 E T3V, T5k, T3S, T5j, T3Q, T3R;
|
cannam@167
|
1316 T3V = FMA(KP951056516, T3T, KP587785252 * T3U);
|
cannam@167
|
1317 T5k = FNMS(KP587785252, T3T, KP951056516 * T3U);
|
cannam@167
|
1318 T3Q = KP559016994 * (T20 - T2b);
|
cannam@167
|
1319 T3R = FNMS(KP250000000, T2c, T1P);
|
cannam@167
|
1320 T3S = T3Q + T3R;
|
cannam@167
|
1321 T5j = T3R - T3Q;
|
cannam@167
|
1322 T3W = T3S - T3V;
|
cannam@167
|
1323 T5T = T5j - T5k;
|
cannam@167
|
1324 T4H = T3S + T3V;
|
cannam@167
|
1325 T5l = T5j + T5k;
|
cannam@167
|
1326 }
|
cannam@167
|
1327 }
|
cannam@167
|
1328 {
|
cannam@167
|
1329 E T6m, T6o, TL, T2E, T6d, T6e, T6n, T6f;
|
cannam@167
|
1330 {
|
cannam@167
|
1331 E T6i, T6l, T1M, T2D;
|
cannam@167
|
1332 T6i = T6g - T6h;
|
cannam@167
|
1333 T6l = T6j - T6k;
|
cannam@167
|
1334 T6m = FMA(KP951056516, T6i, KP587785252 * T6l);
|
cannam@167
|
1335 T6o = FNMS(KP587785252, T6i, KP951056516 * T6l);
|
cannam@167
|
1336 TL = T1 + TK;
|
cannam@167
|
1337 T1M = T1o + T1L;
|
cannam@167
|
1338 T2D = T2d + T2C;
|
cannam@167
|
1339 T2E = T1M + T2D;
|
cannam@167
|
1340 T6d = KP559016994 * (T1M - T2D);
|
cannam@167
|
1341 T6e = FNMS(KP250000000, T2E, TL);
|
cannam@167
|
1342 }
|
cannam@167
|
1343 cr[0] = TL + T2E;
|
cannam@167
|
1344 T6n = T6e - T6d;
|
cannam@167
|
1345 cr[WS(rs, 10)] = T6n - T6o;
|
cannam@167
|
1346 ci[WS(rs, 9)] = T6n + T6o;
|
cannam@167
|
1347 T6f = T6d + T6e;
|
cannam@167
|
1348 ci[WS(rs, 4)] = T6f - T6m;
|
cannam@167
|
1349 cr[WS(rs, 5)] = T6f + T6m;
|
cannam@167
|
1350 }
|
cannam@167
|
1351 {
|
cannam@167
|
1352 E T2P, T4z, T72, T7e, T4m, T7j, T4n, T7i, T4U, T77, T4X, T75, T4O, T6Y, T4P;
|
cannam@167
|
1353 E T6X, T4s, T7f, T4v, T7d, T2H, T70;
|
cannam@167
|
1354 T2H = T2F + T2G;
|
cannam@167
|
1355 T2P = T2H - T2O;
|
cannam@167
|
1356 T4z = T2H + T2O;
|
cannam@167
|
1357 T70 = T6A + T6z;
|
cannam@167
|
1358 T72 = T70 - T71;
|
cannam@167
|
1359 T7e = T71 + T70;
|
cannam@167
|
1360 {
|
cannam@167
|
1361 E T3c, T3z, T3A, T3X, T4k, T4l;
|
cannam@167
|
1362 T3c = FMA(KP535826794, T30, KP844327925 * T3b);
|
cannam@167
|
1363 T3z = FNMS(KP637423989, T3y, KP770513242 * T3r);
|
cannam@167
|
1364 T3A = T3c + T3z;
|
cannam@167
|
1365 T3X = FNMS(KP425779291, T3W, KP904827052 * T3P);
|
cannam@167
|
1366 T4k = FNMS(KP992114701, T4j, KP125333233 * T4c);
|
cannam@167
|
1367 T4l = T3X + T4k;
|
cannam@167
|
1368 T4m = T3A + T4l;
|
cannam@167
|
1369 T7j = T3X - T4k;
|
cannam@167
|
1370 T4n = KP559016994 * (T3A - T4l);
|
cannam@167
|
1371 T7i = T3z - T3c;
|
cannam@167
|
1372 }
|
cannam@167
|
1373 {
|
cannam@167
|
1374 E T4S, T4T, T73, T4V, T4W, T74;
|
cannam@167
|
1375 T4S = FNMS(KP248689887, T4A, KP968583161 * T4B);
|
cannam@167
|
1376 T4T = FNMS(KP844327925, T4D, KP535826794 * T4E);
|
cannam@167
|
1377 T73 = T4S + T4T;
|
cannam@167
|
1378 T4V = FNMS(KP481753674, T4H, KP876306680 * T4I);
|
cannam@167
|
1379 T4W = FNMS(KP684547105, T4K, KP728968627 * T4L);
|
cannam@167
|
1380 T74 = T4V + T4W;
|
cannam@167
|
1381 T4U = T4S - T4T;
|
cannam@167
|
1382 T77 = KP559016994 * (T73 - T74);
|
cannam@167
|
1383 T4X = T4V - T4W;
|
cannam@167
|
1384 T75 = T73 + T74;
|
cannam@167
|
1385 }
|
cannam@167
|
1386 {
|
cannam@167
|
1387 E T4C, T4F, T4G, T4J, T4M, T4N;
|
cannam@167
|
1388 T4C = FMA(KP968583161, T4A, KP248689887 * T4B);
|
cannam@167
|
1389 T4F = FMA(KP535826794, T4D, KP844327925 * T4E);
|
cannam@167
|
1390 T4G = T4C + T4F;
|
cannam@167
|
1391 T4J = FMA(KP876306680, T4H, KP481753674 * T4I);
|
cannam@167
|
1392 T4M = FMA(KP728968627, T4K, KP684547105 * T4L);
|
cannam@167
|
1393 T4N = T4J + T4M;
|
cannam@167
|
1394 T4O = T4G + T4N;
|
cannam@167
|
1395 T6Y = T4J - T4M;
|
cannam@167
|
1396 T4P = KP559016994 * (T4G - T4N);
|
cannam@167
|
1397 T6X = T4F - T4C;
|
cannam@167
|
1398 }
|
cannam@167
|
1399 {
|
cannam@167
|
1400 E T4q, T4r, T7b, T4t, T4u, T7c;
|
cannam@167
|
1401 T4q = FNMS(KP844327925, T30, KP535826794 * T3b);
|
cannam@167
|
1402 T4r = FMA(KP770513242, T3y, KP637423989 * T3r);
|
cannam@167
|
1403 T7b = T4q + T4r;
|
cannam@167
|
1404 T4t = FMA(KP125333233, T4j, KP992114701 * T4c);
|
cannam@167
|
1405 T4u = FMA(KP904827052, T3W, KP425779291 * T3P);
|
cannam@167
|
1406 T7c = T4u + T4t;
|
cannam@167
|
1407 T4s = T4q - T4r;
|
cannam@167
|
1408 T7f = T7b - T7c;
|
cannam@167
|
1409 T4v = T4t - T4u;
|
cannam@167
|
1410 T7d = KP559016994 * (T7b + T7c);
|
cannam@167
|
1411 }
|
cannam@167
|
1412 cr[WS(rs, 4)] = T2P + T4m;
|
cannam@167
|
1413 ci[WS(rs, 23)] = T75 + T72;
|
cannam@167
|
1414 ci[WS(rs, 20)] = T7f + T7e;
|
cannam@167
|
1415 cr[WS(rs, 1)] = T4z + T4O;
|
cannam@167
|
1416 {
|
cannam@167
|
1417 E T4w, T4y, T4p, T4x, T4o;
|
cannam@167
|
1418 T4w = FMA(KP951056516, T4s, KP587785252 * T4v);
|
cannam@167
|
1419 T4y = FNMS(KP587785252, T4s, KP951056516 * T4v);
|
cannam@167
|
1420 T4o = FNMS(KP250000000, T4m, T2P);
|
cannam@167
|
1421 T4p = T4n + T4o;
|
cannam@167
|
1422 T4x = T4o - T4n;
|
cannam@167
|
1423 ci[0] = T4p - T4w;
|
cannam@167
|
1424 ci[WS(rs, 5)] = T4x + T4y;
|
cannam@167
|
1425 cr[WS(rs, 9)] = T4p + T4w;
|
cannam@167
|
1426 ci[WS(rs, 10)] = T4x - T4y;
|
cannam@167
|
1427 }
|
cannam@167
|
1428 {
|
cannam@167
|
1429 E T6Z, T79, T78, T7a, T76;
|
cannam@167
|
1430 T6Z = FMA(KP587785252, T6X, KP951056516 * T6Y);
|
cannam@167
|
1431 T79 = FNMS(KP587785252, T6Y, KP951056516 * T6X);
|
cannam@167
|
1432 T76 = FNMS(KP250000000, T75, T72);
|
cannam@167
|
1433 T78 = T76 - T77;
|
cannam@167
|
1434 T7a = T77 + T76;
|
cannam@167
|
1435 cr[WS(rs, 16)] = T6Z - T78;
|
cannam@167
|
1436 ci[WS(rs, 18)] = T79 + T7a;
|
cannam@167
|
1437 ci[WS(rs, 13)] = T6Z + T78;
|
cannam@167
|
1438 cr[WS(rs, 21)] = T79 - T7a;
|
cannam@167
|
1439 }
|
cannam@167
|
1440 {
|
cannam@167
|
1441 E T7k, T7l, T7h, T7m, T7g;
|
cannam@167
|
1442 T7k = FMA(KP587785252, T7i, KP951056516 * T7j);
|
cannam@167
|
1443 T7l = FNMS(KP587785252, T7j, KP951056516 * T7i);
|
cannam@167
|
1444 T7g = FNMS(KP250000000, T7f, T7e);
|
cannam@167
|
1445 T7h = T7d - T7g;
|
cannam@167
|
1446 T7m = T7d + T7g;
|
cannam@167
|
1447 cr[WS(rs, 14)] = T7h - T7k;
|
cannam@167
|
1448 ci[WS(rs, 15)] = T7l + T7m;
|
cannam@167
|
1449 cr[WS(rs, 19)] = T7k + T7h;
|
cannam@167
|
1450 cr[WS(rs, 24)] = T7l - T7m;
|
cannam@167
|
1451 }
|
cannam@167
|
1452 {
|
cannam@167
|
1453 E T4Y, T50, T4R, T4Z, T4Q;
|
cannam@167
|
1454 T4Y = FMA(KP951056516, T4U, KP587785252 * T4X);
|
cannam@167
|
1455 T50 = FNMS(KP587785252, T4U, KP951056516 * T4X);
|
cannam@167
|
1456 T4Q = FNMS(KP250000000, T4O, T4z);
|
cannam@167
|
1457 T4R = T4P + T4Q;
|
cannam@167
|
1458 T4Z = T4Q - T4P;
|
cannam@167
|
1459 ci[WS(rs, 3)] = T4R - T4Y;
|
cannam@167
|
1460 ci[WS(rs, 8)] = T4Z + T50;
|
cannam@167
|
1461 cr[WS(rs, 6)] = T4R + T4Y;
|
cannam@167
|
1462 cr[WS(rs, 11)] = T4Z - T50;
|
cannam@167
|
1463 }
|
cannam@167
|
1464 }
|
cannam@167
|
1465 {
|
cannam@167
|
1466 E T7p, T7x, T7q, T7t, T7u, T7v, T7y, T7w;
|
cannam@167
|
1467 {
|
cannam@167
|
1468 E T7n, T7o, T7r, T7s;
|
cannam@167
|
1469 T7n = T1L - T1o;
|
cannam@167
|
1470 T7o = T2d - T2C;
|
cannam@167
|
1471 T7p = FMA(KP587785252, T7n, KP951056516 * T7o);
|
cannam@167
|
1472 T7x = FNMS(KP587785252, T7o, KP951056516 * T7n);
|
cannam@167
|
1473 T7q = T6y + T6v;
|
cannam@167
|
1474 T7r = T6g + T6h;
|
cannam@167
|
1475 T7s = T6j + T6k;
|
cannam@167
|
1476 T7t = T7r + T7s;
|
cannam@167
|
1477 T7u = FNMS(KP250000000, T7t, T7q);
|
cannam@167
|
1478 T7v = KP559016994 * (T7r - T7s);
|
cannam@167
|
1479 }
|
cannam@167
|
1480 ci[WS(rs, 24)] = T7t + T7q;
|
cannam@167
|
1481 T7y = T7v + T7u;
|
cannam@167
|
1482 cr[WS(rs, 20)] = T7x - T7y;
|
cannam@167
|
1483 ci[WS(rs, 19)] = T7x + T7y;
|
cannam@167
|
1484 T7w = T7u - T7v;
|
cannam@167
|
1485 cr[WS(rs, 15)] = T7p - T7w;
|
cannam@167
|
1486 ci[WS(rs, 14)] = T7p + T7w;
|
cannam@167
|
1487 }
|
cannam@167
|
1488 {
|
cannam@167
|
1489 E T53, T5L, T6C, T6O, T5y, T6T, T5z, T6S, T66, T6H, T69, T6F, T60, T6q, T61;
|
cannam@167
|
1490 E T6p, T5E, T6P, T5H, T6N, T51, T6B;
|
cannam@167
|
1491 T51 = T2G - T2F;
|
cannam@167
|
1492 T53 = T51 + T52;
|
cannam@167
|
1493 T5L = T51 - T52;
|
cannam@167
|
1494 T6B = T6z - T6A;
|
cannam@167
|
1495 T6C = T6u + T6B;
|
cannam@167
|
1496 T6O = T6B - T6u;
|
cannam@167
|
1497 {
|
cannam@167
|
1498 E T5a, T5h, T5i, T5p, T5w, T5x;
|
cannam@167
|
1499 T5a = FMA(KP728968627, T56, KP684547105 * T59);
|
cannam@167
|
1500 T5h = FNMS(KP992114701, T5g, KP125333233 * T5d);
|
cannam@167
|
1501 T5i = T5a + T5h;
|
cannam@167
|
1502 T5p = FMA(KP062790519, T5l, KP998026728 * T5o);
|
cannam@167
|
1503 T5w = FNMS(KP637423989, T5v, KP770513242 * T5s);
|
cannam@167
|
1504 T5x = T5p + T5w;
|
cannam@167
|
1505 T5y = T5i + T5x;
|
cannam@167
|
1506 T6T = T5p - T5w;
|
cannam@167
|
1507 T5z = KP559016994 * (T5i - T5x);
|
cannam@167
|
1508 T6S = T5h - T5a;
|
cannam@167
|
1509 }
|
cannam@167
|
1510 {
|
cannam@167
|
1511 E T64, T65, T6D, T67, T68, T6E;
|
cannam@167
|
1512 T64 = FNMS(KP481753674, T5M, KP876306680 * T5N);
|
cannam@167
|
1513 T65 = FMA(KP904827052, T5Q, KP425779291 * T5P);
|
cannam@167
|
1514 T6D = T64 - T65;
|
cannam@167
|
1515 T67 = FNMS(KP844327925, T5T, KP535826794 * T5U);
|
cannam@167
|
1516 T68 = FNMS(KP998026728, T5W, KP062790519 * T5X);
|
cannam@167
|
1517 T6E = T67 + T68;
|
cannam@167
|
1518 T66 = T64 + T65;
|
cannam@167
|
1519 T6H = KP559016994 * (T6D - T6E);
|
cannam@167
|
1520 T69 = T67 - T68;
|
cannam@167
|
1521 T6F = T6D + T6E;
|
cannam@167
|
1522 }
|
cannam@167
|
1523 {
|
cannam@167
|
1524 E T5O, T5R, T5S, T5V, T5Y, T5Z;
|
cannam@167
|
1525 T5O = FMA(KP876306680, T5M, KP481753674 * T5N);
|
cannam@167
|
1526 T5R = FNMS(KP425779291, T5Q, KP904827052 * T5P);
|
cannam@167
|
1527 T5S = T5O + T5R;
|
cannam@167
|
1528 T5V = FMA(KP535826794, T5T, KP844327925 * T5U);
|
cannam@167
|
1529 T5Y = FMA(KP062790519, T5W, KP998026728 * T5X);
|
cannam@167
|
1530 T5Z = T5V + T5Y;
|
cannam@167
|
1531 T60 = T5S + T5Z;
|
cannam@167
|
1532 T6q = T5V - T5Y;
|
cannam@167
|
1533 T61 = KP559016994 * (T5S - T5Z);
|
cannam@167
|
1534 T6p = T5R - T5O;
|
cannam@167
|
1535 }
|
cannam@167
|
1536 {
|
cannam@167
|
1537 E T5C, T5D, T6L, T5F, T5G, T6M;
|
cannam@167
|
1538 T5C = FNMS(KP684547105, T56, KP728968627 * T59);
|
cannam@167
|
1539 T5D = FMA(KP125333233, T5g, KP992114701 * T5d);
|
cannam@167
|
1540 T6L = T5C - T5D;
|
cannam@167
|
1541 T5F = FNMS(KP998026728, T5l, KP062790519 * T5o);
|
cannam@167
|
1542 T5G = FMA(KP770513242, T5v, KP637423989 * T5s);
|
cannam@167
|
1543 T6M = T5F - T5G;
|
cannam@167
|
1544 T5E = T5C + T5D;
|
cannam@167
|
1545 T6P = T6L + T6M;
|
cannam@167
|
1546 T5H = T5F + T5G;
|
cannam@167
|
1547 T6N = KP559016994 * (T6L - T6M);
|
cannam@167
|
1548 }
|
cannam@167
|
1549 cr[WS(rs, 3)] = T53 + T5y;
|
cannam@167
|
1550 ci[WS(rs, 22)] = T6F + T6C;
|
cannam@167
|
1551 ci[WS(rs, 21)] = T6P + T6O;
|
cannam@167
|
1552 cr[WS(rs, 2)] = T5L + T60;
|
cannam@167
|
1553 {
|
cannam@167
|
1554 E T6r, T6J, T6I, T6K, T6G;
|
cannam@167
|
1555 T6r = FMA(KP587785252, T6p, KP951056516 * T6q);
|
cannam@167
|
1556 T6J = FNMS(KP587785252, T6q, KP951056516 * T6p);
|
cannam@167
|
1557 T6G = FNMS(KP250000000, T6F, T6C);
|
cannam@167
|
1558 T6I = T6G - T6H;
|
cannam@167
|
1559 T6K = T6H + T6G;
|
cannam@167
|
1560 cr[WS(rs, 17)] = T6r - T6I;
|
cannam@167
|
1561 ci[WS(rs, 17)] = T6J + T6K;
|
cannam@167
|
1562 ci[WS(rs, 12)] = T6r + T6I;
|
cannam@167
|
1563 cr[WS(rs, 22)] = T6J - T6K;
|
cannam@167
|
1564 }
|
cannam@167
|
1565 {
|
cannam@167
|
1566 E T6a, T6c, T63, T6b, T62;
|
cannam@167
|
1567 T6a = FMA(KP951056516, T66, KP587785252 * T69);
|
cannam@167
|
1568 T6c = FNMS(KP587785252, T66, KP951056516 * T69);
|
cannam@167
|
1569 T62 = FNMS(KP250000000, T60, T5L);
|
cannam@167
|
1570 T63 = T61 + T62;
|
cannam@167
|
1571 T6b = T62 - T61;
|
cannam@167
|
1572 ci[WS(rs, 2)] = T63 - T6a;
|
cannam@167
|
1573 ci[WS(rs, 7)] = T6b + T6c;
|
cannam@167
|
1574 cr[WS(rs, 7)] = T63 + T6a;
|
cannam@167
|
1575 cr[WS(rs, 12)] = T6b - T6c;
|
cannam@167
|
1576 }
|
cannam@167
|
1577 {
|
cannam@167
|
1578 E T5I, T5K, T5B, T5J, T5A;
|
cannam@167
|
1579 T5I = FMA(KP951056516, T5E, KP587785252 * T5H);
|
cannam@167
|
1580 T5K = FNMS(KP587785252, T5E, KP951056516 * T5H);
|
cannam@167
|
1581 T5A = FNMS(KP250000000, T5y, T53);
|
cannam@167
|
1582 T5B = T5z + T5A;
|
cannam@167
|
1583 T5J = T5A - T5z;
|
cannam@167
|
1584 ci[WS(rs, 1)] = T5B - T5I;
|
cannam@167
|
1585 ci[WS(rs, 6)] = T5J + T5K;
|
cannam@167
|
1586 cr[WS(rs, 8)] = T5B + T5I;
|
cannam@167
|
1587 ci[WS(rs, 11)] = T5J - T5K;
|
cannam@167
|
1588 }
|
cannam@167
|
1589 {
|
cannam@167
|
1590 E T6U, T6V, T6R, T6W, T6Q;
|
cannam@167
|
1591 T6U = FMA(KP587785252, T6S, KP951056516 * T6T);
|
cannam@167
|
1592 T6V = FNMS(KP587785252, T6T, KP951056516 * T6S);
|
cannam@167
|
1593 T6Q = FNMS(KP250000000, T6P, T6O);
|
cannam@167
|
1594 T6R = T6N - T6Q;
|
cannam@167
|
1595 T6W = T6N + T6Q;
|
cannam@167
|
1596 cr[WS(rs, 13)] = T6R - T6U;
|
cannam@167
|
1597 ci[WS(rs, 16)] = T6V + T6W;
|
cannam@167
|
1598 cr[WS(rs, 18)] = T6U + T6R;
|
cannam@167
|
1599 cr[WS(rs, 23)] = T6V - T6W;
|
cannam@167
|
1600 }
|
cannam@167
|
1601 }
|
cannam@167
|
1602 }
|
cannam@167
|
1603 }
|
cannam@167
|
1604 }
|
cannam@167
|
1605 }
|
cannam@167
|
1606
|
cannam@167
|
1607 static const tw_instr twinstr[] = {
|
cannam@167
|
1608 {TW_CEXP, 1, 1},
|
cannam@167
|
1609 {TW_CEXP, 1, 3},
|
cannam@167
|
1610 {TW_CEXP, 1, 9},
|
cannam@167
|
1611 {TW_CEXP, 1, 24},
|
cannam@167
|
1612 {TW_NEXT, 1, 0}
|
cannam@167
|
1613 };
|
cannam@167
|
1614
|
cannam@167
|
1615 static const hc2hc_desc desc = { 25, "hf2_25", twinstr, &GENUS, {280, 180, 160, 0} };
|
cannam@167
|
1616
|
cannam@167
|
1617 void X(codelet_hf2_25) (planner *p) {
|
cannam@167
|
1618 X(khc2hc_register) (p, hf2_25, &desc);
|
cannam@167
|
1619 }
|
cannam@167
|
1620 #endif
|