Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/rdft/scalar/r2cb/hb_25.c @ 10:37bf6b4a2645
Add FFTW3
author | Chris Cannam |
---|---|
date | Wed, 20 Mar 2013 15:35:50 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
9:c0fb53affa76 | 10:37bf6b4a2645 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-11 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Sun Nov 25 07:41:23 EST 2012 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -dif -name hb_25 -include hb.h */ | |
29 | |
30 /* | |
31 * This function contains 400 FP additions, 364 FP multiplications, | |
32 * (or, 84 additions, 48 multiplications, 316 fused multiply/add), | |
33 * 176 stack variables, 47 constants, and 100 memory accesses | |
34 */ | |
35 #include "hb.h" | |
36 | |
37 static void hb_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DK(KP921177326, +0.921177326965143320250447435415066029359282231); | |
40 DK(KP833417178, +0.833417178328688677408962550243238843138996060); | |
41 DK(KP541454447, +0.541454447536312777046285590082819509052033189); | |
42 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
43 DK(KP242145790, +0.242145790282157779872542093866183953459003101); | |
44 DK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
45 DK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
46 DK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
47 DK(KP831864738, +0.831864738706457140726048799369896829771167132); | |
48 DK(KP871714437, +0.871714437527667770979999223229522602943903653); | |
49 DK(KP803003575, +0.803003575438660414833440593570376004635464850); | |
50 DK(KP554608978, +0.554608978404018097464974850792216217022558774); | |
51 DK(KP248028675, +0.248028675328619457762448260696444630363259177); | |
52 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
53 DK(KP851038619, +0.851038619207379630836264138867114231259902550); | |
54 DK(KP525970792, +0.525970792408939708442463226536226366643874659); | |
55 DK(KP726211448, +0.726211448929902658173535992263577167607493062); | |
56 DK(KP912018591, +0.912018591466481957908415381764119056233607330); | |
57 DK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
58 DK(KP943557151, +0.943557151597354104399655195398983005179443399); | |
59 DK(KP994076283, +0.994076283785401014123185814696322018529298887); | |
60 DK(KP614372930, +0.614372930789563808870829930444362096004872855); | |
61 DK(KP621716863, +0.621716863012209892444754556304102309693593202); | |
62 DK(KP772036680, +0.772036680810363904029489473607579825330539880); | |
63 DK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
64 DK(KP860541664, +0.860541664367944677098261680920518816412804187); | |
65 DK(KP949179823, +0.949179823508441261575555465843363271711583843); | |
66 DK(KP557913902, +0.557913902031834264187699648465567037992437152); | |
67 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
68 DK(KP249506682, +0.249506682107067890488084201715862638334226305); | |
69 DK(KP906616052, +0.906616052148196230441134447086066874408359177); | |
70 DK(KP681693190, +0.681693190061530575150324149145440022633095390); | |
71 DK(KP560319534, +0.560319534973832390111614715371676131169633784); | |
72 DK(KP845997307, +0.845997307939530944175097360758058292389769300); | |
73 DK(KP968479752, +0.968479752739016373193524836781420152702090879); | |
74 DK(KP062914667, +0.062914667253649757225485955897349402364686947); | |
75 DK(KP827271945, +0.827271945972475634034355757144307982555673741); | |
76 DK(KP470564281, +0.470564281212251493087595091036643380879947982); | |
77 DK(KP126329378, +0.126329378446108174786050455341811215027378105); | |
78 DK(KP256756360, +0.256756360367726783319498520922669048172391148); | |
79 DK(KP634619297, +0.634619297544148100711287640319130485732531031); | |
80 DK(KP549754652, +0.549754652192770074288023275540779861653779767); | |
81 DK(KP939062505, +0.939062505817492352556001843133229685779824606); | |
82 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
83 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
84 DK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
85 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
86 { | |
87 INT m; | |
88 for (m = mb, W = W + ((mb - 1) * 48); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 48, MAKE_VOLATILE_STRIDE(50, rs)) { | |
89 E T3w, T3P, T2d, T3y, T3x, T3Q; | |
90 { | |
91 E T9, T3E, T1F, T3B, T6f, T7d, T5u, T6U, T4k, T2k, T5G, T1G, T19, T1H, T1s; | |
92 E T1M, T1N, TP, TM, T7i, T77, T5X, T64, T4u, T4D, T3p, T2z, T74, T7h, T63; | |
93 E T5Q, T4x, T4E, T3q, T2O, T4n, T4G, T3t, T3j, T5F, T70, T7f, T66, T5B, T4q; | |
94 E T4H, T3s, T34, T5E, T6V; | |
95 { | |
96 E T2f, T2e, T6e, T2j, T5t, T6d; | |
97 { | |
98 E T1, T1x, T3C, T3D, T8, T2h, T1A, T1D, T2i, T3A, T1E, T3z; | |
99 T1 = cr[0]; | |
100 T1x = ci[WS(rs, 24)]; | |
101 { | |
102 E T2, T3, T5, T6; | |
103 T2 = cr[WS(rs, 5)]; | |
104 T3 = ci[WS(rs, 4)]; | |
105 T5 = cr[WS(rs, 10)]; | |
106 T6 = ci[WS(rs, 9)]; | |
107 { | |
108 E T1y, T4, T7, T1z, T1B, T1C; | |
109 T1y = ci[WS(rs, 19)]; | |
110 T3C = T2 - T3; | |
111 T4 = T2 + T3; | |
112 T3D = T5 - T6; | |
113 T7 = T5 + T6; | |
114 T1z = cr[WS(rs, 20)]; | |
115 T1B = ci[WS(rs, 14)]; | |
116 T1C = cr[WS(rs, 15)]; | |
117 T8 = T4 + T7; | |
118 T2f = T4 - T7; | |
119 T2h = T1y + T1z; | |
120 T1A = T1y - T1z; | |
121 T1D = T1B - T1C; | |
122 T2i = T1B + T1C; | |
123 } | |
124 } | |
125 T2e = FNMS(KP250000000, T8, T1); | |
126 T9 = T1 + T8; | |
127 T3A = T1A - T1D; | |
128 T1E = T1A + T1D; | |
129 T3E = FMA(KP618033988, T3D, T3C); | |
130 T6e = FNMS(KP618033988, T3C, T3D); | |
131 T2j = FMA(KP618033988, T2i, T2h); | |
132 T5t = FNMS(KP618033988, T2h, T2i); | |
133 T1F = T1x + T1E; | |
134 T3z = FNMS(KP250000000, T1E, T1x); | |
135 T6d = FNMS(KP559016994, T3A, T3z); | |
136 T3B = FMA(KP559016994, T3A, T3z); | |
137 } | |
138 { | |
139 E T2x, T5V, T2m, T2l, Ti, T5w, T3h, T36, TK, T35, T2F, T5L, T2I, Tr, T2H; | |
140 E T3a, T5z, T3d, T1r, T3c, T2q, T5S, T2t, TZ, T2s, T5O, T2M, T2B, Tt, T18; | |
141 E T2A, Tx, T2V, T2Y, Tw, T30, T1i, T2X, Ty; | |
142 { | |
143 E T1j, T1k, T1p, T39, T1l; | |
144 { | |
145 E TC, TI, T3g, TD, TE; | |
146 { | |
147 E Ta, Te, Tf, Tb, Tc, T5s, T2g, T2w, Tg; | |
148 Ta = cr[WS(rs, 1)]; | |
149 T5s = FNMS(KP559016994, T2f, T2e); | |
150 T2g = FMA(KP559016994, T2f, T2e); | |
151 T6f = FNMS(KP951056516, T6e, T6d); | |
152 T7d = FMA(KP951056516, T6e, T6d); | |
153 Te = cr[WS(rs, 11)]; | |
154 T5u = FMA(KP951056516, T5t, T5s); | |
155 T6U = FNMS(KP951056516, T5t, T5s); | |
156 T4k = FMA(KP951056516, T2j, T2g); | |
157 T2k = FNMS(KP951056516, T2j, T2g); | |
158 Tf = ci[WS(rs, 8)]; | |
159 Tb = cr[WS(rs, 6)]; | |
160 Tc = ci[WS(rs, 3)]; | |
161 TC = cr[WS(rs, 3)]; | |
162 T2w = Tf - Te; | |
163 Tg = Te + Tf; | |
164 { | |
165 E T2v, Td, Th, TG, TH; | |
166 T2v = Tb - Tc; | |
167 Td = Tb + Tc; | |
168 TG = ci[WS(rs, 11)]; | |
169 TH = ci[WS(rs, 6)]; | |
170 T2x = FNMS(KP618033988, T2w, T2v); | |
171 T5V = FMA(KP618033988, T2v, T2w); | |
172 Th = Td + Tg; | |
173 T2m = Td - Tg; | |
174 TI = TG + TH; | |
175 T3g = TG - TH; | |
176 T2l = FNMS(KP250000000, Th, Ta); | |
177 Ti = Ta + Th; | |
178 TD = cr[WS(rs, 8)]; | |
179 TE = ci[WS(rs, 1)]; | |
180 } | |
181 } | |
182 { | |
183 E Tj, Tk, Tp, T2E, TJ, Tl; | |
184 Tj = cr[WS(rs, 4)]; | |
185 { | |
186 E Tn, To, T3f, TF; | |
187 Tn = ci[WS(rs, 10)]; | |
188 To = ci[WS(rs, 5)]; | |
189 T3f = TD - TE; | |
190 TF = TD + TE; | |
191 Tk = cr[WS(rs, 9)]; | |
192 Tp = Tn + To; | |
193 T2E = To - Tn; | |
194 T5w = FNMS(KP618033988, T3f, T3g); | |
195 T3h = FMA(KP618033988, T3g, T3f); | |
196 T36 = TI - TF; | |
197 TJ = TF + TI; | |
198 Tl = ci[0]; | |
199 } | |
200 T1j = ci[WS(rs, 21)]; | |
201 TK = TC + TJ; | |
202 T35 = FNMS(KP250000000, TJ, TC); | |
203 { | |
204 E T1n, Tm, T2D, T1o, Tq; | |
205 T1n = cr[WS(rs, 13)]; | |
206 Tm = Tk + Tl; | |
207 T2D = Tl - Tk; | |
208 T1o = cr[WS(rs, 18)]; | |
209 T1k = ci[WS(rs, 16)]; | |
210 T2F = FMA(KP618033988, T2E, T2D); | |
211 T5L = FNMS(KP618033988, T2D, T2E); | |
212 T2I = Tm - Tp; | |
213 Tq = Tm + Tp; | |
214 T1p = T1n + T1o; | |
215 T39 = T1o - T1n; | |
216 Tr = Tj + Tq; | |
217 T2H = FMS(KP250000000, Tq, Tj); | |
218 T1l = cr[WS(rs, 23)]; | |
219 } | |
220 } | |
221 } | |
222 { | |
223 E T10, T11, T16, T2L, T12; | |
224 { | |
225 E TR, TS, TX, T2p, T1q, TT; | |
226 TR = ci[WS(rs, 23)]; | |
227 { | |
228 E TV, TW, T38, T1m; | |
229 TV = ci[WS(rs, 13)]; | |
230 TW = cr[WS(rs, 16)]; | |
231 T38 = T1k + T1l; | |
232 T1m = T1k - T1l; | |
233 TS = ci[WS(rs, 18)]; | |
234 TX = TV - TW; | |
235 T2p = TV + TW; | |
236 T3a = FMA(KP618033988, T39, T38); | |
237 T5z = FNMS(KP618033988, T38, T39); | |
238 T3d = T1m + T1p; | |
239 T1q = T1m - T1p; | |
240 TT = cr[WS(rs, 21)]; | |
241 } | |
242 T10 = ci[WS(rs, 20)]; | |
243 T1r = T1j + T1q; | |
244 T3c = FMS(KP250000000, T1q, T1j); | |
245 { | |
246 E T14, TU, T2o, T15, TY; | |
247 T14 = cr[WS(rs, 14)]; | |
248 TU = TS - TT; | |
249 T2o = TS + TT; | |
250 T15 = cr[WS(rs, 19)]; | |
251 T11 = ci[WS(rs, 15)]; | |
252 T2q = FMA(KP618033988, T2p, T2o); | |
253 T5S = FNMS(KP618033988, T2o, T2p); | |
254 T2t = TU - TX; | |
255 TY = TU + TX; | |
256 T16 = T14 + T15; | |
257 T2L = T15 - T14; | |
258 TZ = TR + TY; | |
259 T2s = FNMS(KP250000000, TY, TR); | |
260 T12 = cr[WS(rs, 24)]; | |
261 } | |
262 } | |
263 { | |
264 E T1a, T1e, T1d, T2T, T17, T1f; | |
265 T1a = ci[WS(rs, 22)]; | |
266 { | |
267 E T1b, T1c, T2K, T13; | |
268 T1b = ci[WS(rs, 17)]; | |
269 T1c = cr[WS(rs, 22)]; | |
270 T2K = T11 + T12; | |
271 T13 = T11 - T12; | |
272 T1e = ci[WS(rs, 12)]; | |
273 T1d = T1b - T1c; | |
274 T2T = T1b + T1c; | |
275 T5O = FNMS(KP618033988, T2K, T2L); | |
276 T2M = FMA(KP618033988, T2L, T2K); | |
277 T2B = T13 + T16; | |
278 T17 = T13 - T16; | |
279 T1f = cr[WS(rs, 17)]; | |
280 } | |
281 Tt = cr[WS(rs, 2)]; | |
282 T18 = T10 + T17; | |
283 T2A = FMS(KP250000000, T17, T10); | |
284 { | |
285 E Tu, T1g, T2U, Tv, T1h; | |
286 Tu = cr[WS(rs, 7)]; | |
287 T1g = T1e - T1f; | |
288 T2U = T1e + T1f; | |
289 Tv = ci[WS(rs, 2)]; | |
290 Tx = cr[WS(rs, 12)]; | |
291 T2V = FMA(KP618033988, T2U, T2T); | |
292 T5G = FNMS(KP618033988, T2T, T2U); | |
293 T2Y = T1d - T1g; | |
294 T1h = T1d + T1g; | |
295 Tw = Tu + Tv; | |
296 T30 = Tu - Tv; | |
297 T1i = T1a + T1h; | |
298 T2X = FMS(KP250000000, T1h, T1a); | |
299 Ty = ci[WS(rs, 7)]; | |
300 } | |
301 } | |
302 } | |
303 } | |
304 { | |
305 E T32, T5D, T2R, T2Q, T2u, T2r, T4t; | |
306 { | |
307 E TA, T31, Tz, TB, Ts; | |
308 T31 = Ty - Tx; | |
309 Tz = Tx + Ty; | |
310 T1G = TZ + T18; | |
311 T19 = TZ - T18; | |
312 T32 = FNMS(KP618033988, T31, T30); | |
313 T5D = FMA(KP618033988, T30, T31); | |
314 TA = Tw + Tz; | |
315 T2R = Tz - Tw; | |
316 T2Q = FNMS(KP250000000, TA, Tt); | |
317 TB = Tt + TA; | |
318 T1H = T1i + T1r; | |
319 T1s = T1i - T1r; | |
320 T1M = Ti - Tr; | |
321 Ts = Ti + Tr; | |
322 { | |
323 E T2n, T5R, T5U, TL; | |
324 T2n = FMA(KP559016994, T2m, T2l); | |
325 T5R = FNMS(KP559016994, T2m, T2l); | |
326 T5U = FNMS(KP559016994, T2t, T2s); | |
327 T2u = FMA(KP559016994, T2t, T2s); | |
328 TL = TB + TK; | |
329 T1N = TB - TK; | |
330 { | |
331 E T5T, T75, T5W, T76; | |
332 T5T = FMA(KP951056516, T5S, T5R); | |
333 T75 = FNMS(KP951056516, T5S, T5R); | |
334 T5W = FMA(KP951056516, T5V, T5U); | |
335 T76 = FNMS(KP951056516, T5V, T5U); | |
336 TP = Ts - TL; | |
337 TM = Ts + TL; | |
338 T2r = FNMS(KP951056516, T2q, T2n); | |
339 T4t = FMA(KP951056516, T2q, T2n); | |
340 T7i = FMA(KP939062505, T75, T76); | |
341 T77 = FNMS(KP939062505, T76, T75); | |
342 T5X = FNMS(KP549754652, T5W, T5T); | |
343 T64 = FMA(KP549754652, T5T, T5W); | |
344 } | |
345 } | |
346 } | |
347 { | |
348 E T2J, T2G, T4v, T5y, T37, T3e, T5v; | |
349 { | |
350 E T2C, T5K, T5N, T4s, T2y; | |
351 T2C = FNMS(KP559016994, T2B, T2A); | |
352 T5K = FMA(KP559016994, T2B, T2A); | |
353 T5N = FMA(KP559016994, T2I, T2H); | |
354 T2J = FNMS(KP559016994, T2I, T2H); | |
355 T4s = FNMS(KP951056516, T2x, T2u); | |
356 T2y = FMA(KP951056516, T2x, T2u); | |
357 { | |
358 E T73, T5M, T72, T5P; | |
359 T73 = FMA(KP951056516, T5L, T5K); | |
360 T5M = FNMS(KP951056516, T5L, T5K); | |
361 T72 = FMA(KP951056516, T5O, T5N); | |
362 T5P = FNMS(KP951056516, T5O, T5N); | |
363 T4u = FNMS(KP634619297, T4t, T4s); | |
364 T4D = FMA(KP634619297, T4s, T4t); | |
365 T3p = FMA(KP256756360, T2r, T2y); | |
366 T2z = FNMS(KP256756360, T2y, T2r); | |
367 T74 = FMA(KP126329378, T73, T72); | |
368 T7h = FNMS(KP126329378, T72, T73); | |
369 T63 = FNMS(KP470564281, T5M, T5P); | |
370 T5Q = FMA(KP470564281, T5P, T5M); | |
371 T2G = FMA(KP951056516, T2F, T2C); | |
372 T4v = FNMS(KP951056516, T2F, T2C); | |
373 } | |
374 T5y = FMA(KP559016994, T36, T35); | |
375 T37 = FNMS(KP559016994, T36, T35); | |
376 T3e = FNMS(KP559016994, T3d, T3c); | |
377 T5v = FMA(KP559016994, T3d, T3c); | |
378 } | |
379 { | |
380 E T5x, T6Y, T4w, T2N; | |
381 T4w = FNMS(KP951056516, T2M, T2J); | |
382 T2N = FMA(KP951056516, T2M, T2J); | |
383 { | |
384 E T4l, T3b, T4m, T3i; | |
385 T4l = FMA(KP951056516, T3a, T37); | |
386 T3b = FNMS(KP951056516, T3a, T37); | |
387 T4m = FMA(KP951056516, T3h, T3e); | |
388 T3i = FNMS(KP951056516, T3h, T3e); | |
389 T4x = FNMS(KP827271945, T4w, T4v); | |
390 T4E = FMA(KP827271945, T4v, T4w); | |
391 T3q = FMA(KP634619297, T2G, T2N); | |
392 T2O = FNMS(KP634619297, T2N, T2G); | |
393 T4n = FNMS(KP126329378, T4m, T4l); | |
394 T4G = FMA(KP126329378, T4l, T4m); | |
395 T3t = FNMS(KP939062505, T3b, T3i); | |
396 T3j = FMA(KP939062505, T3i, T3b); | |
397 T5x = FMA(KP951056516, T5w, T5v); | |
398 T6Y = FNMS(KP951056516, T5w, T5v); | |
399 } | |
400 { | |
401 E T2S, T2Z, T5C, T6Z, T5A; | |
402 T5F = FMA(KP559016994, T2R, T2Q); | |
403 T2S = FNMS(KP559016994, T2R, T2Q); | |
404 T2Z = FNMS(KP559016994, T2Y, T2X); | |
405 T5C = FMA(KP559016994, T2Y, T2X); | |
406 T6Z = FNMS(KP951056516, T5z, T5y); | |
407 T5A = FMA(KP951056516, T5z, T5y); | |
408 { | |
409 E T4p, T2W, T4o, T33; | |
410 T4p = FMA(KP951056516, T2V, T2S); | |
411 T2W = FNMS(KP951056516, T2V, T2S); | |
412 T4o = FMA(KP951056516, T32, T2Z); | |
413 T33 = FNMS(KP951056516, T32, T2Z); | |
414 T70 = FNMS(KP827271945, T6Z, T6Y); | |
415 T7f = FMA(KP827271945, T6Y, T6Z); | |
416 T66 = FNMS(KP062914667, T5x, T5A); | |
417 T5B = FMA(KP062914667, T5A, T5x); | |
418 T4q = FNMS(KP470564281, T4p, T4o); | |
419 T4H = FMA(KP470564281, T4o, T4p); | |
420 T3s = FNMS(KP549754652, T2W, T33); | |
421 T34 = FMA(KP549754652, T33, T2W); | |
422 T5E = FNMS(KP951056516, T5D, T5C); | |
423 T6V = FMA(KP951056516, T5D, T5C); | |
424 } | |
425 } | |
426 } | |
427 } | |
428 } | |
429 } | |
430 } | |
431 { | |
432 E T6X, T7e, T6A, T6F, T6C, T6G, T6B; | |
433 cr[0] = T9 + TM; | |
434 { | |
435 E T67, T5I, T25, T22, T1X, T26, T21; | |
436 { | |
437 E T1I, T23, T1L, T1Z, T1t, TO, T24, T1O; | |
438 { | |
439 E T1K, T6W, T5H, T1J; | |
440 T1K = T1G - T1H; | |
441 T1I = T1G + T1H; | |
442 T6W = FNMS(KP951056516, T5G, T5F); | |
443 T5H = FMA(KP951056516, T5G, T5F); | |
444 T1J = FNMS(KP250000000, T1I, T1F); | |
445 T6X = FMA(KP062914667, T6W, T6V); | |
446 T7e = FNMS(KP062914667, T6V, T6W); | |
447 T67 = FNMS(KP634619297, T5E, T5H); | |
448 T5I = FMA(KP634619297, T5H, T5E); | |
449 T23 = FNMS(KP559016994, T1K, T1J); | |
450 T1L = FMA(KP559016994, T1K, T1J); | |
451 T1Z = FNMS(KP618033988, T19, T1s); | |
452 T1t = FMA(KP618033988, T1s, T19); | |
453 TO = FNMS(KP250000000, TM, T9); | |
454 T24 = FNMS(KP618033988, T1M, T1N); | |
455 T1O = FMA(KP618033988, T1N, T1M); | |
456 } | |
457 { | |
458 E T2b, T2a, T1Y, TQ, T27; | |
459 ci[0] = T1F + T1I; | |
460 T2b = FMA(KP951056516, T24, T23); | |
461 T25 = FNMS(KP951056516, T24, T23); | |
462 T2a = W[29]; | |
463 T1Y = FNMS(KP559016994, TP, TO); | |
464 TQ = FMA(KP559016994, TP, TO); | |
465 T27 = W[28]; | |
466 { | |
467 E T1V, T1P, T20, T1S, T1w, T1v, TN, T1Q; | |
468 T1V = FNMS(KP951056516, T1O, T1L); | |
469 T1P = FMA(KP951056516, T1O, T1L); | |
470 { | |
471 E T28, T1u, T29, T2c; | |
472 T20 = FMA(KP951056516, T1Z, T1Y); | |
473 T28 = FNMS(KP951056516, T1Z, T1Y); | |
474 T1S = FMA(KP951056516, T1t, TQ); | |
475 T1u = FNMS(KP951056516, T1t, TQ); | |
476 T1w = W[9]; | |
477 T29 = T27 * T28; | |
478 T2c = T2a * T28; | |
479 TN = W[8]; | |
480 T1Q = T1w * T1u; | |
481 cr[WS(rs, 15)] = FNMS(T2a, T2b, T29); | |
482 ci[WS(rs, 15)] = FMA(T27, T2b, T2c); | |
483 T1v = TN * T1u; | |
484 } | |
485 ci[WS(rs, 5)] = FMA(TN, T1P, T1Q); | |
486 { | |
487 E T1U, T1R, T1W, T1T; | |
488 T1U = W[39]; | |
489 cr[WS(rs, 5)] = FNMS(T1w, T1P, T1v); | |
490 T1R = W[38]; | |
491 T1W = T1U * T1S; | |
492 T22 = W[19]; | |
493 T1T = T1R * T1S; | |
494 T1X = W[18]; | |
495 ci[WS(rs, 20)] = FMA(T1R, T1V, T1W); | |
496 T26 = T22 * T20; | |
497 cr[WS(rs, 20)] = FNMS(T1U, T1V, T1T); | |
498 T21 = T1X * T20; | |
499 } | |
500 } | |
501 } | |
502 } | |
503 { | |
504 E T6h, T6g, T5Y, T5J, T6z, T69, T6o, T6E; | |
505 { | |
506 E T6m, T6n, T65, T68; | |
507 T65 = FMA(KP968479752, T64, T63); | |
508 T6h = FNMS(KP968479752, T64, T63); | |
509 ci[WS(rs, 10)] = FMA(T1X, T25, T26); | |
510 T68 = FNMS(KP845997307, T67, T66); | |
511 T6g = FMA(KP845997307, T67, T66); | |
512 cr[WS(rs, 10)] = FNMS(T22, T25, T21); | |
513 T6m = FNMS(KP968479752, T5X, T5Q); | |
514 T5Y = FMA(KP968479752, T5X, T5Q); | |
515 T5J = FMA(KP845997307, T5I, T5B); | |
516 T6n = FNMS(KP845997307, T5I, T5B); | |
517 T6z = FMA(KP560319534, T65, T68); | |
518 T69 = FNMS(KP681693190, T68, T65); | |
519 T6o = FMA(KP681693190, T6n, T6m); | |
520 T6E = FNMS(KP560319534, T6m, T6n); | |
521 } | |
522 { | |
523 E T62, T6l, T6I, T6L, T6H, T6K; | |
524 { | |
525 E T6Q, T6O, T6y, T6D, T6S; | |
526 { | |
527 E T6N, T5Z, T61, T6i, T6k; | |
528 T6N = W[2]; | |
529 T5Z = FMA(KP906616052, T5Y, T5J); | |
530 T61 = FNMS(KP906616052, T5Y, T5J); | |
531 T6i = FNMS(KP906616052, T6h, T6g); | |
532 T6k = FMA(KP906616052, T6h, T6g); | |
533 T6Q = W[3]; | |
534 { | |
535 E T60, T6j, T6R, T6P; | |
536 T60 = FNMS(KP249506682, T5Z, T5u); | |
537 T6O = FMA(KP998026728, T5Z, T5u); | |
538 T6j = FNMS(KP249506682, T6i, T6f); | |
539 T6R = FMA(KP998026728, T6i, T6f); | |
540 T6y = FMA(KP557913902, T61, T60); | |
541 T62 = FNMS(KP557913902, T61, T60); | |
542 T6P = T6N * T6O; | |
543 T6l = FNMS(KP557913902, T6k, T6j); | |
544 T6D = FMA(KP557913902, T6k, T6j); | |
545 T6S = T6N * T6R; | |
546 cr[WS(rs, 2)] = FNMS(T6Q, T6R, T6P); | |
547 } | |
548 } | |
549 T6A = FNMS(KP949179823, T6z, T6y); | |
550 T6I = FMA(KP949179823, T6z, T6y); | |
551 T6L = FNMS(KP949179823, T6E, T6D); | |
552 T6F = FMA(KP949179823, T6E, T6D); | |
553 ci[WS(rs, 2)] = FMA(T6Q, T6O, T6S); | |
554 T6H = W[32]; | |
555 T6K = W[33]; | |
556 } | |
557 { | |
558 E T6a, T6s, T6v, T6p, T6c, T6q, T6b, T6M, T6J, T5r; | |
559 T6a = FNMS(KP860541664, T69, T62); | |
560 T6s = FMA(KP860541664, T69, T62); | |
561 T6v = FMA(KP860541664, T6o, T6l); | |
562 T6p = FNMS(KP860541664, T6o, T6l); | |
563 T6M = T6H * T6L; | |
564 T6J = T6H * T6I; | |
565 T5r = W[12]; | |
566 T6c = W[13]; | |
567 ci[WS(rs, 17)] = FMA(T6K, T6I, T6M); | |
568 cr[WS(rs, 17)] = FNMS(T6K, T6L, T6J); | |
569 T6q = T5r * T6p; | |
570 T6b = T5r * T6a; | |
571 { | |
572 E T6r, T6u, T6w, T6t, T6x; | |
573 ci[WS(rs, 7)] = FMA(T6c, T6a, T6q); | |
574 cr[WS(rs, 7)] = FNMS(T6c, T6p, T6b); | |
575 T6r = W[42]; | |
576 T6u = W[43]; | |
577 T6w = T6r * T6v; | |
578 T6t = T6r * T6s; | |
579 T6x = W[22]; | |
580 T6C = W[23]; | |
581 ci[WS(rs, 22)] = FMA(T6u, T6s, T6w); | |
582 cr[WS(rs, 22)] = FNMS(T6u, T6v, T6t); | |
583 T6G = T6x * T6F; | |
584 T6B = T6x * T6A; | |
585 } | |
586 } | |
587 } | |
588 } | |
589 } | |
590 { | |
591 E T7u, T7D, T7n, T7w, T7v, T7E; | |
592 { | |
593 E T78, T7t, T7N, T71, T7C, T7S, T7y, T7k; | |
594 { | |
595 E T7j, T7g, T7A, T7B, T7r, T7s; | |
596 T7r = FNMS(KP734762448, T7i, T7h); | |
597 T7j = FMA(KP734762448, T7i, T7h); | |
598 T7g = FMA(KP772036680, T7f, T7e); | |
599 T7s = FNMS(KP772036680, T7f, T7e); | |
600 ci[WS(rs, 12)] = FMA(T6C, T6A, T6G); | |
601 cr[WS(rs, 12)] = FNMS(T6C, T6F, T6B); | |
602 T7A = FNMS(KP734762448, T77, T74); | |
603 T78 = FMA(KP734762448, T77, T74); | |
604 T7t = FNMS(KP621716863, T7s, T7r); | |
605 T7N = FMA(KP614372930, T7r, T7s); | |
606 T71 = FMA(KP772036680, T70, T6X); | |
607 T7B = FNMS(KP772036680, T70, T6X); | |
608 T7C = FNMS(KP621716863, T7B, T7A); | |
609 T7S = FMA(KP614372930, T7A, T7B); | |
610 T7y = FNMS(KP994076283, T7j, T7g); | |
611 T7k = FMA(KP994076283, T7j, T7g); | |
612 } | |
613 { | |
614 E T7c, T6T, T7x, T7l, T79, T7p; | |
615 T7c = W[5]; | |
616 T6T = W[4]; | |
617 T7x = FNMS(KP249506682, T7k, T7d); | |
618 T7l = FMA(KP998026728, T7k, T7d); | |
619 T79 = FMA(KP994076283, T78, T71); | |
620 T7p = FNMS(KP994076283, T78, T71); | |
621 { | |
622 E T7z, T7Y, T7Z, T7T, T7q, T7O, T7X, T7L, T7Q, T7P, T7U; | |
623 { | |
624 E T7V, T80, T7b, T7m, T7W; | |
625 { | |
626 E T7R, T7o, T7a, T7M; | |
627 T7V = W[34]; | |
628 T7R = FMA(KP557913902, T7y, T7x); | |
629 T7z = FNMS(KP557913902, T7y, T7x); | |
630 T7Y = W[35]; | |
631 T7o = FNMS(KP249506682, T79, T6U); | |
632 T7a = FMA(KP998026728, T79, T6U); | |
633 T7Z = FMA(KP949179823, T7S, T7R); | |
634 T7T = FNMS(KP949179823, T7S, T7R); | |
635 T7M = FMA(KP557913902, T7p, T7o); | |
636 T7q = FNMS(KP557913902, T7p, T7o); | |
637 T7b = T6T * T7a; | |
638 T7m = T7c * T7a; | |
639 T7W = FNMS(KP949179823, T7N, T7M); | |
640 T7O = FMA(KP949179823, T7N, T7M); | |
641 } | |
642 cr[WS(rs, 3)] = FNMS(T7c, T7l, T7b); | |
643 ci[WS(rs, 3)] = FMA(T6T, T7l, T7m); | |
644 T80 = T7Y * T7W; | |
645 T7X = T7V * T7W; | |
646 T7L = W[24]; | |
647 T7Q = W[25]; | |
648 ci[WS(rs, 18)] = FMA(T7V, T7Z, T80); | |
649 } | |
650 cr[WS(rs, 18)] = FNMS(T7Y, T7Z, T7X); | |
651 T7P = T7L * T7O; | |
652 T7U = T7Q * T7O; | |
653 { | |
654 E T7J, T7F, T7I, T7H, T7K, T7G; | |
655 T7u = FMA(KP943557151, T7t, T7q); | |
656 T7G = FNMS(KP943557151, T7t, T7q); | |
657 cr[WS(rs, 13)] = FNMS(T7Q, T7T, T7P); | |
658 ci[WS(rs, 13)] = FMA(T7L, T7T, T7U); | |
659 T7J = FMA(KP943557151, T7C, T7z); | |
660 T7D = FNMS(KP943557151, T7C, T7z); | |
661 T7F = W[44]; | |
662 T7I = W[45]; | |
663 T7n = W[14]; | |
664 T7H = T7F * T7G; | |
665 T7K = T7I * T7G; | |
666 T7w = W[15]; | |
667 T7v = T7n * T7u; | |
668 cr[WS(rs, 23)] = FNMS(T7I, T7J, T7H); | |
669 ci[WS(rs, 23)] = FMA(T7F, T7J, T7K); | |
670 } | |
671 } | |
672 } | |
673 } | |
674 T7E = T7w * T7u; | |
675 cr[WS(rs, 8)] = FNMS(T7w, T7D, T7v); | |
676 { | |
677 E T3F, T4K, T4X, T4j, T4M, T4L, T4Y; | |
678 { | |
679 E T4P, T4O, T4y, T4r, T4J, T57, T4N, T5c, T4W; | |
680 { | |
681 E T4U, T4V, T4F, T4I; | |
682 T4F = FNMS(KP912575812, T4E, T4D); | |
683 T4P = FMA(KP912575812, T4E, T4D); | |
684 T4O = FMA(KP912018591, T4H, T4G); | |
685 T4I = FNMS(KP912018591, T4H, T4G); | |
686 ci[WS(rs, 8)] = FMA(T7n, T7D, T7E); | |
687 T4y = FMA(KP912575812, T4x, T4u); | |
688 T4U = FNMS(KP912575812, T4x, T4u); | |
689 T4V = FMA(KP912018591, T4q, T4n); | |
690 T4r = FNMS(KP912018591, T4q, T4n); | |
691 T4J = FNMS(KP726211448, T4I, T4F); | |
692 T57 = FMA(KP525970792, T4F, T4I); | |
693 T3F = FMA(KP951056516, T3E, T3B); | |
694 T4N = FNMS(KP951056516, T3E, T3B); | |
695 T5c = FMA(KP525970792, T4U, T4V); | |
696 T4W = FNMS(KP726211448, T4V, T4U); | |
697 } | |
698 { | |
699 E T5o, T4S, T4B, T5l, T5p, T4R, T4A, T5m, T4Q, T4z; | |
700 T5o = W[7]; | |
701 T4Q = FMA(KP851038619, T4P, T4O); | |
702 T4S = FNMS(KP851038619, T4P, T4O); | |
703 T4z = FMA(KP851038619, T4y, T4r); | |
704 T4B = FNMS(KP851038619, T4y, T4r); | |
705 T5l = W[6]; | |
706 T5p = FMA(KP992114701, T4Q, T4N); | |
707 T4R = FNMS(KP248028675, T4Q, T4N); | |
708 T4A = FMA(KP248028675, T4z, T4k); | |
709 T5m = FNMS(KP992114701, T4z, T4k); | |
710 { | |
711 E T4T, T4C, T5d, T58, T55, T5a, T59, T5e; | |
712 { | |
713 E T5f, T5j, T5i, T5h, T5k, T5g; | |
714 T5f = W[36]; | |
715 { | |
716 E T5b, T56, T5n, T5q; | |
717 T4T = FNMS(KP554608978, T4S, T4R); | |
718 T5b = FMA(KP554608978, T4S, T4R); | |
719 T56 = FNMS(KP554608978, T4B, T4A); | |
720 T4C = FMA(KP554608978, T4B, T4A); | |
721 T5n = T5l * T5m; | |
722 T5q = T5o * T5m; | |
723 T5j = FMA(KP943557151, T5c, T5b); | |
724 T5d = FNMS(KP943557151, T5c, T5b); | |
725 T5g = FMA(KP943557151, T57, T56); | |
726 T58 = FNMS(KP943557151, T57, T56); | |
727 cr[WS(rs, 4)] = FNMS(T5o, T5p, T5n); | |
728 ci[WS(rs, 4)] = FMA(T5l, T5p, T5q); | |
729 } | |
730 T5i = W[37]; | |
731 T5h = T5f * T5g; | |
732 T55 = W[26]; | |
733 T5k = T5i * T5g; | |
734 T5a = W[27]; | |
735 cr[WS(rs, 19)] = FNMS(T5i, T5j, T5h); | |
736 T59 = T55 * T58; | |
737 ci[WS(rs, 19)] = FMA(T5f, T5j, T5k); | |
738 } | |
739 T5e = T5a * T58; | |
740 { | |
741 E T53, T4Z, T52, T51, T54, T50; | |
742 cr[WS(rs, 14)] = FNMS(T5a, T5d, T59); | |
743 T4K = FNMS(KP803003575, T4J, T4C); | |
744 T50 = FMA(KP803003575, T4J, T4C); | |
745 ci[WS(rs, 14)] = FMA(T55, T5d, T5e); | |
746 T4X = FNMS(KP803003575, T4W, T4T); | |
747 T53 = FMA(KP803003575, T4W, T4T); | |
748 T4Z = W[46]; | |
749 T52 = W[47]; | |
750 T4j = W[16]; | |
751 T51 = T4Z * T50; | |
752 T54 = T52 * T50; | |
753 T4M = W[17]; | |
754 T4L = T4j * T4K; | |
755 cr[WS(rs, 24)] = FNMS(T52, T53, T51); | |
756 ci[WS(rs, 24)] = FMA(T4Z, T53, T54); | |
757 } | |
758 } | |
759 } | |
760 } | |
761 T4Y = T4M * T4K; | |
762 cr[WS(rs, 9)] = FNMS(T4M, T4X, T4L); | |
763 { | |
764 E T3G, T3H, T2P, T3k, T3Z, T3v, T3O, T44; | |
765 { | |
766 E T3M, T3N, T3r, T3u; | |
767 T3G = FNMS(KP871714437, T3q, T3p); | |
768 T3r = FMA(KP871714437, T3q, T3p); | |
769 T3u = FNMS(KP831864738, T3t, T3s); | |
770 T3H = FMA(KP831864738, T3t, T3s); | |
771 ci[WS(rs, 9)] = FMA(T4j, T4X, T4Y); | |
772 T3M = FNMS(KP871714437, T2O, T2z); | |
773 T2P = FMA(KP871714437, T2O, T2z); | |
774 T3k = FMA(KP831864738, T3j, T34); | |
775 T3N = FNMS(KP831864738, T3j, T34); | |
776 T3Z = FMA(KP683113946, T3r, T3u); | |
777 T3v = FNMS(KP559154169, T3u, T3r); | |
778 T3O = FMA(KP559154169, T3N, T3M); | |
779 T44 = FNMS(KP683113946, T3M, T3N); | |
780 } | |
781 { | |
782 E T4g, T3K, T3n, T4d, T3J, T4h, T4e, T3m, T3I, T3l; | |
783 T4g = W[1]; | |
784 T3K = FMA(KP904730450, T3H, T3G); | |
785 T3I = FNMS(KP904730450, T3H, T3G); | |
786 T3n = FNMS(KP904730450, T3k, T2P); | |
787 T3l = FMA(KP904730450, T3k, T2P); | |
788 T4d = W[0]; | |
789 T3J = FNMS(KP242145790, T3I, T3F); | |
790 T4h = FMA(KP968583161, T3I, T3F); | |
791 T4e = FMA(KP968583161, T3l, T2k); | |
792 T3m = FNMS(KP242145790, T3l, T2k); | |
793 { | |
794 E T3L, T3o, T45, T40, T3X, T42, T41, T46; | |
795 { | |
796 E T47, T4b, T4a, T49, T4c, T48; | |
797 T47 = W[30]; | |
798 { | |
799 E T43, T3Y, T4f, T4i; | |
800 T43 = FNMS(KP541454447, T3K, T3J); | |
801 T3L = FMA(KP541454447, T3K, T3J); | |
802 T3o = FMA(KP541454447, T3n, T3m); | |
803 T3Y = FNMS(KP541454447, T3n, T3m); | |
804 T4f = T4d * T4e; | |
805 T4i = T4g * T4e; | |
806 T45 = FNMS(KP833417178, T44, T43); | |
807 T4b = FMA(KP833417178, T44, T43); | |
808 T40 = FNMS(KP833417178, T3Z, T3Y); | |
809 T48 = FMA(KP833417178, T3Z, T3Y); | |
810 cr[WS(rs, 1)] = FNMS(T4g, T4h, T4f); | |
811 ci[WS(rs, 1)] = FMA(T4d, T4h, T4i); | |
812 } | |
813 T4a = W[31]; | |
814 T49 = T47 * T48; | |
815 T3X = W[20]; | |
816 T4c = T4a * T48; | |
817 T42 = W[21]; | |
818 cr[WS(rs, 16)] = FNMS(T4a, T4b, T49); | |
819 T41 = T3X * T40; | |
820 ci[WS(rs, 16)] = FMA(T47, T4b, T4c); | |
821 } | |
822 T46 = T42 * T40; | |
823 { | |
824 E T3V, T3R, T3U, T3T, T3W, T3S; | |
825 cr[WS(rs, 11)] = FNMS(T42, T45, T41); | |
826 T3S = FMA(KP921177326, T3v, T3o); | |
827 T3w = FNMS(KP921177326, T3v, T3o); | |
828 ci[WS(rs, 11)] = FMA(T3X, T45, T46); | |
829 T3V = FNMS(KP921177326, T3O, T3L); | |
830 T3P = FMA(KP921177326, T3O, T3L); | |
831 T3R = W[40]; | |
832 T3U = W[41]; | |
833 T2d = W[10]; | |
834 T3T = T3R * T3S; | |
835 T3W = T3U * T3S; | |
836 T3y = W[11]; | |
837 T3x = T2d * T3w; | |
838 cr[WS(rs, 21)] = FNMS(T3U, T3V, T3T); | |
839 ci[WS(rs, 21)] = FMA(T3R, T3V, T3W); | |
840 } | |
841 } | |
842 } | |
843 } | |
844 } | |
845 } | |
846 } | |
847 } | |
848 cr[WS(rs, 6)] = FNMS(T3y, T3P, T3x); | |
849 T3Q = T3y * T3w; | |
850 ci[WS(rs, 6)] = FMA(T2d, T3P, T3Q); | |
851 } | |
852 } | |
853 } | |
854 | |
855 static const tw_instr twinstr[] = { | |
856 {TW_FULL, 1, 25}, | |
857 {TW_NEXT, 1, 0} | |
858 }; | |
859 | |
860 static const hc2hc_desc desc = { 25, "hb_25", twinstr, &GENUS, {84, 48, 316, 0} }; | |
861 | |
862 void X(codelet_hb_25) (planner *p) { | |
863 X(khc2hc_register) (p, hb_25, &desc); | |
864 } | |
865 #else /* HAVE_FMA */ | |
866 | |
867 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -dif -name hb_25 -include hb.h */ | |
868 | |
869 /* | |
870 * This function contains 400 FP additions, 280 FP multiplications, | |
871 * (or, 260 additions, 140 multiplications, 140 fused multiply/add), | |
872 * 107 stack variables, 20 constants, and 100 memory accesses | |
873 */ | |
874 #include "hb.h" | |
875 | |
876 static void hb_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms) | |
877 { | |
878 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
879 DK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
880 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
881 DK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
882 DK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
883 DK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
884 DK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
885 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
886 DK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
887 DK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
888 DK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
889 DK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
890 DK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
891 DK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
892 DK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
893 DK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
894 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
895 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
896 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
897 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
898 { | |
899 INT m; | |
900 for (m = mb, W = W + ((mb - 1) * 48); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 48, MAKE_VOLATILE_STRIDE(50, rs)) { | |
901 E T9, T5Q, T3y, T39, T5v, Ti, Tr, Ts, TZ, T18, T1z, T2k, T4l, T3h, T44; | |
902 E T5d, T6C, T5C, T6o, T56, T6B, T5B, T6l, T2z, T4m, T3i, T47, T1K, T5w, T3c; | |
903 E T3B, T5R, TB, TK, TL, T1i, T1r, T1A, T2P, T4o, T3k, T4b, T5s, T6F, T5F; | |
904 E T6v, T5l, T6E, T5E, T6s, T34, T4p, T3l, T4e; | |
905 { | |
906 E T1, T4, T7, T8, T3x, T3w, T37, T38; | |
907 T1 = cr[0]; | |
908 { | |
909 E T2, T3, T5, T6; | |
910 T2 = cr[WS(rs, 5)]; | |
911 T3 = ci[WS(rs, 4)]; | |
912 T4 = T2 + T3; | |
913 T5 = cr[WS(rs, 10)]; | |
914 T6 = ci[WS(rs, 9)]; | |
915 T7 = T5 + T6; | |
916 T8 = T4 + T7; | |
917 T3x = T5 - T6; | |
918 T3w = T2 - T3; | |
919 } | |
920 T9 = T1 + T8; | |
921 T5Q = FMA(KP951056516, T3w, KP587785252 * T3x); | |
922 T3y = FNMS(KP951056516, T3x, KP587785252 * T3w); | |
923 T37 = FNMS(KP250000000, T8, T1); | |
924 T38 = KP559016994 * (T4 - T7); | |
925 T39 = T37 - T38; | |
926 T5v = T38 + T37; | |
927 } | |
928 { | |
929 E Ta, T27, T53, T2f, Th, T26, T10, T2p, T58, T2x, T17, T2o, Tj, T2n, T5a; | |
930 E T2t, Tq, T2s, TR, T2b, T51, T2h, TY, T2g; | |
931 { | |
932 E Tg, T2e, Td, T2d; | |
933 Ta = cr[WS(rs, 1)]; | |
934 { | |
935 E Te, Tf, Tb, Tc; | |
936 Te = cr[WS(rs, 11)]; | |
937 Tf = ci[WS(rs, 8)]; | |
938 Tg = Te + Tf; | |
939 T2e = Te - Tf; | |
940 Tb = cr[WS(rs, 6)]; | |
941 Tc = ci[WS(rs, 3)]; | |
942 Td = Tb + Tc; | |
943 T2d = Tb - Tc; | |
944 } | |
945 T27 = KP559016994 * (Td - Tg); | |
946 T53 = FMA(KP951056516, T2d, KP587785252 * T2e); | |
947 T2f = FNMS(KP951056516, T2e, KP587785252 * T2d); | |
948 Th = Td + Tg; | |
949 T26 = FNMS(KP250000000, Th, Ta); | |
950 } | |
951 { | |
952 E T16, T2w, T13, T2v; | |
953 T10 = ci[WS(rs, 20)]; | |
954 { | |
955 E T14, T15, T11, T12; | |
956 T14 = cr[WS(rs, 14)]; | |
957 T15 = cr[WS(rs, 19)]; | |
958 T16 = T14 + T15; | |
959 T2w = T15 - T14; | |
960 T11 = ci[WS(rs, 15)]; | |
961 T12 = cr[WS(rs, 24)]; | |
962 T13 = T11 - T12; | |
963 T2v = T11 + T12; | |
964 } | |
965 T2p = KP559016994 * (T13 + T16); | |
966 T58 = FMA(KP951056516, T2v, KP587785252 * T2w); | |
967 T2x = FNMS(KP951056516, T2w, KP587785252 * T2v); | |
968 T17 = T13 - T16; | |
969 T2o = FNMS(KP250000000, T17, T10); | |
970 } | |
971 { | |
972 E Tp, T2m, Tm, T2l; | |
973 Tj = cr[WS(rs, 4)]; | |
974 { | |
975 E Tn, To, Tk, Tl; | |
976 Tn = ci[WS(rs, 10)]; | |
977 To = ci[WS(rs, 5)]; | |
978 Tp = Tn + To; | |
979 T2m = Tn - To; | |
980 Tk = cr[WS(rs, 9)]; | |
981 Tl = ci[0]; | |
982 Tm = Tk + Tl; | |
983 T2l = Tk - Tl; | |
984 } | |
985 T2n = FNMS(KP951056516, T2m, KP587785252 * T2l); | |
986 T5a = FMA(KP951056516, T2l, KP587785252 * T2m); | |
987 T2t = KP559016994 * (Tm - Tp); | |
988 Tq = Tm + Tp; | |
989 T2s = FNMS(KP250000000, Tq, Tj); | |
990 } | |
991 { | |
992 E TX, T2a, TU, T29; | |
993 TR = ci[WS(rs, 23)]; | |
994 { | |
995 E TV, TW, TS, TT; | |
996 TV = ci[WS(rs, 13)]; | |
997 TW = cr[WS(rs, 16)]; | |
998 TX = TV - TW; | |
999 T2a = TV + TW; | |
1000 TS = ci[WS(rs, 18)]; | |
1001 TT = cr[WS(rs, 21)]; | |
1002 TU = TS - TT; | |
1003 T29 = TS + TT; | |
1004 } | |
1005 T2b = FNMS(KP951056516, T2a, KP587785252 * T29); | |
1006 T51 = FMA(KP951056516, T29, KP587785252 * T2a); | |
1007 T2h = KP559016994 * (TU - TX); | |
1008 TY = TU + TX; | |
1009 T2g = FNMS(KP250000000, TY, TR); | |
1010 } | |
1011 Ti = Ta + Th; | |
1012 Tr = Tj + Tq; | |
1013 Ts = Ti + Tr; | |
1014 TZ = TR + TY; | |
1015 T18 = T10 + T17; | |
1016 T1z = TZ + T18; | |
1017 { | |
1018 E T2c, T42, T2j, T43, T28, T2i; | |
1019 T28 = T26 - T27; | |
1020 T2c = T28 - T2b; | |
1021 T42 = T28 + T2b; | |
1022 T2i = T2g - T2h; | |
1023 T2j = T2f + T2i; | |
1024 T43 = T2i - T2f; | |
1025 T2k = FNMS(KP481753674, T2j, KP876306680 * T2c); | |
1026 T4l = FMA(KP728968627, T43, KP684547105 * T42); | |
1027 T3h = FMA(KP876306680, T2j, KP481753674 * T2c); | |
1028 T44 = FNMS(KP684547105, T43, KP728968627 * T42); | |
1029 } | |
1030 { | |
1031 E T59, T6n, T5c, T6m, T57, T5b; | |
1032 T57 = T2t + T2s; | |
1033 T59 = T57 - T58; | |
1034 T6n = T57 + T58; | |
1035 T5b = T2o + T2p; | |
1036 T5c = T5a + T5b; | |
1037 T6m = T5b - T5a; | |
1038 T5d = FNMS(KP844327925, T5c, KP535826794 * T59); | |
1039 T6C = FMA(KP637423989, T6m, KP770513242 * T6n); | |
1040 T5C = FMA(KP535826794, T5c, KP844327925 * T59); | |
1041 T6o = FNMS(KP637423989, T6n, KP770513242 * T6m); | |
1042 } | |
1043 { | |
1044 E T52, T6j, T55, T6k, T50, T54; | |
1045 T50 = T27 + T26; | |
1046 T52 = T50 - T51; | |
1047 T6j = T50 + T51; | |
1048 T54 = T2h + T2g; | |
1049 T55 = T53 + T54; | |
1050 T6k = T54 - T53; | |
1051 T56 = FNMS(KP248689887, T55, KP968583161 * T52); | |
1052 T6B = FMA(KP535826794, T6k, KP844327925 * T6j); | |
1053 T5B = FMA(KP968583161, T55, KP248689887 * T52); | |
1054 T6l = FNMS(KP844327925, T6k, KP535826794 * T6j); | |
1055 } | |
1056 { | |
1057 E T2r, T45, T2y, T46, T2q, T2u; | |
1058 T2q = T2o - T2p; | |
1059 T2r = T2n + T2q; | |
1060 T45 = T2q - T2n; | |
1061 T2u = T2s - T2t; | |
1062 T2y = T2u - T2x; | |
1063 T46 = T2u + T2x; | |
1064 T2z = FMA(KP904827052, T2r, KP425779291 * T2y); | |
1065 T4m = FNMS(KP992114701, T45, KP125333233 * T46); | |
1066 T3i = FNMS(KP425779291, T2r, KP904827052 * T2y); | |
1067 T47 = FMA(KP125333233, T45, KP992114701 * T46); | |
1068 } | |
1069 } | |
1070 { | |
1071 E T1C, T1F, T1I, T1J, T3b, T3a, T3z, T3A; | |
1072 T1C = ci[WS(rs, 24)]; | |
1073 { | |
1074 E T1D, T1E, T1G, T1H; | |
1075 T1D = ci[WS(rs, 19)]; | |
1076 T1E = cr[WS(rs, 20)]; | |
1077 T1F = T1D - T1E; | |
1078 T1G = ci[WS(rs, 14)]; | |
1079 T1H = cr[WS(rs, 15)]; | |
1080 T1I = T1G - T1H; | |
1081 T1J = T1F + T1I; | |
1082 T3b = T1G + T1H; | |
1083 T3a = T1D + T1E; | |
1084 } | |
1085 T1K = T1C + T1J; | |
1086 T5w = FMA(KP951056516, T3a, KP587785252 * T3b); | |
1087 T3c = FNMS(KP951056516, T3b, KP587785252 * T3a); | |
1088 T3z = FNMS(KP250000000, T1J, T1C); | |
1089 T3A = KP559016994 * (T1F - T1I); | |
1090 T3B = T3z - T3A; | |
1091 T5R = T3A + T3z; | |
1092 } | |
1093 { | |
1094 E Tt, T2C, T5i, T2K, TA, T2B, T1a, T2G, T5g, T2M, T1h, T2L, TC, T2R, T5p; | |
1095 E T2Z, TJ, T2Q, T1j, T2V, T5n, T31, T1q, T30; | |
1096 { | |
1097 E Tw, T2I, Tz, T2J; | |
1098 Tt = cr[WS(rs, 2)]; | |
1099 { | |
1100 E Tu, Tv, Tx, Ty; | |
1101 Tu = cr[WS(rs, 7)]; | |
1102 Tv = ci[WS(rs, 2)]; | |
1103 Tw = Tu + Tv; | |
1104 T2I = Tu - Tv; | |
1105 Tx = cr[WS(rs, 12)]; | |
1106 Ty = ci[WS(rs, 7)]; | |
1107 Tz = Tx + Ty; | |
1108 T2J = Tx - Ty; | |
1109 } | |
1110 T2C = KP559016994 * (Tw - Tz); | |
1111 T5i = FMA(KP951056516, T2I, KP587785252 * T2J); | |
1112 T2K = FNMS(KP951056516, T2J, KP587785252 * T2I); | |
1113 TA = Tw + Tz; | |
1114 T2B = FNMS(KP250000000, TA, Tt); | |
1115 } | |
1116 { | |
1117 E T1d, T2E, T1g, T2F; | |
1118 T1a = ci[WS(rs, 22)]; | |
1119 { | |
1120 E T1b, T1c, T1e, T1f; | |
1121 T1b = ci[WS(rs, 17)]; | |
1122 T1c = cr[WS(rs, 22)]; | |
1123 T1d = T1b - T1c; | |
1124 T2E = T1b + T1c; | |
1125 T1e = ci[WS(rs, 12)]; | |
1126 T1f = cr[WS(rs, 17)]; | |
1127 T1g = T1e - T1f; | |
1128 T2F = T1e + T1f; | |
1129 } | |
1130 T2G = FNMS(KP951056516, T2F, KP587785252 * T2E); | |
1131 T5g = FMA(KP951056516, T2E, KP587785252 * T2F); | |
1132 T2M = KP559016994 * (T1d - T1g); | |
1133 T1h = T1d + T1g; | |
1134 T2L = FNMS(KP250000000, T1h, T1a); | |
1135 } | |
1136 { | |
1137 E TI, T2Y, TF, T2X; | |
1138 TC = cr[WS(rs, 3)]; | |
1139 { | |
1140 E TG, TH, TD, TE; | |
1141 TG = ci[WS(rs, 11)]; | |
1142 TH = ci[WS(rs, 6)]; | |
1143 TI = TG + TH; | |
1144 T2Y = TG - TH; | |
1145 TD = cr[WS(rs, 8)]; | |
1146 TE = ci[WS(rs, 1)]; | |
1147 TF = TD + TE; | |
1148 T2X = TD - TE; | |
1149 } | |
1150 T2R = KP559016994 * (TF - TI); | |
1151 T5p = FMA(KP951056516, T2X, KP587785252 * T2Y); | |
1152 T2Z = FNMS(KP951056516, T2Y, KP587785252 * T2X); | |
1153 TJ = TF + TI; | |
1154 T2Q = FNMS(KP250000000, TJ, TC); | |
1155 } | |
1156 { | |
1157 E T1p, T2U, T1m, T2T; | |
1158 T1j = ci[WS(rs, 21)]; | |
1159 { | |
1160 E T1n, T1o, T1k, T1l; | |
1161 T1n = cr[WS(rs, 13)]; | |
1162 T1o = cr[WS(rs, 18)]; | |
1163 T1p = T1n + T1o; | |
1164 T2U = T1o - T1n; | |
1165 T1k = ci[WS(rs, 16)]; | |
1166 T1l = cr[WS(rs, 23)]; | |
1167 T1m = T1k - T1l; | |
1168 T2T = T1k + T1l; | |
1169 } | |
1170 T2V = FNMS(KP951056516, T2U, KP587785252 * T2T); | |
1171 T5n = FMA(KP951056516, T2T, KP587785252 * T2U); | |
1172 T31 = KP559016994 * (T1m + T1p); | |
1173 T1q = T1m - T1p; | |
1174 T30 = FNMS(KP250000000, T1q, T1j); | |
1175 } | |
1176 TB = Tt + TA; | |
1177 TK = TC + TJ; | |
1178 TL = TB + TK; | |
1179 T1i = T1a + T1h; | |
1180 T1r = T1j + T1q; | |
1181 T1A = T1i + T1r; | |
1182 { | |
1183 E T2H, T49, T2O, T4a, T2D, T2N; | |
1184 T2D = T2B - T2C; | |
1185 T2H = T2D - T2G; | |
1186 T49 = T2D + T2G; | |
1187 T2N = T2L - T2M; | |
1188 T2O = T2K + T2N; | |
1189 T4a = T2N - T2K; | |
1190 T2P = FNMS(KP844327925, T2O, KP535826794 * T2H); | |
1191 T4o = FMA(KP062790519, T4a, KP998026728 * T49); | |
1192 T3k = FMA(KP535826794, T2O, KP844327925 * T2H); | |
1193 T4b = FNMS(KP998026728, T4a, KP062790519 * T49); | |
1194 } | |
1195 { | |
1196 E T5o, T6u, T5r, T6t, T5m, T5q; | |
1197 T5m = T2R + T2Q; | |
1198 T5o = T5m - T5n; | |
1199 T6u = T5m + T5n; | |
1200 T5q = T30 + T31; | |
1201 T5r = T5p + T5q; | |
1202 T6t = T5q - T5p; | |
1203 T5s = FNMS(KP684547105, T5r, KP728968627 * T5o); | |
1204 T6F = FNMS(KP992114701, T6t, KP125333233 * T6u); | |
1205 T5F = FMA(KP728968627, T5r, KP684547105 * T5o); | |
1206 T6v = FMA(KP125333233, T6t, KP992114701 * T6u); | |
1207 } | |
1208 { | |
1209 E T5h, T6r, T5k, T6q, T5f, T5j; | |
1210 T5f = T2C + T2B; | |
1211 T5h = T5f - T5g; | |
1212 T6r = T5f + T5g; | |
1213 T5j = T2M + T2L; | |
1214 T5k = T5i + T5j; | |
1215 T6q = T5j - T5i; | |
1216 T5l = FNMS(KP481753674, T5k, KP876306680 * T5h); | |
1217 T6E = FNMS(KP425779291, T6q, KP904827052 * T6r); | |
1218 T5E = FMA(KP876306680, T5k, KP481753674 * T5h); | |
1219 T6s = FMA(KP904827052, T6q, KP425779291 * T6r); | |
1220 } | |
1221 { | |
1222 E T2W, T4d, T33, T4c, T2S, T32; | |
1223 T2S = T2Q - T2R; | |
1224 T2W = T2S - T2V; | |
1225 T4d = T2S + T2V; | |
1226 T32 = T30 - T31; | |
1227 T33 = T2Z + T32; | |
1228 T4c = T32 - T2Z; | |
1229 T34 = FNMS(KP998026728, T33, KP062790519 * T2W); | |
1230 T4p = FNMS(KP637423989, T4c, KP770513242 * T4d); | |
1231 T3l = FMA(KP062790519, T33, KP998026728 * T2W); | |
1232 T4e = FMA(KP770513242, T4c, KP637423989 * T4d); | |
1233 } | |
1234 } | |
1235 { | |
1236 E TM, TQ, T1U, T1L, T1N, T1Z, T1t, T1V, T1y, T1Y; | |
1237 { | |
1238 E TO, TP, T1B, T1M; | |
1239 TO = KP559016994 * (Ts - TL); | |
1240 TM = Ts + TL; | |
1241 TP = FNMS(KP250000000, TM, T9); | |
1242 TQ = TO + TP; | |
1243 T1U = TP - TO; | |
1244 T1B = KP559016994 * (T1z - T1A); | |
1245 T1L = T1z + T1A; | |
1246 T1M = FNMS(KP250000000, T1L, T1K); | |
1247 T1N = T1B + T1M; | |
1248 T1Z = T1M - T1B; | |
1249 } | |
1250 { | |
1251 E T19, T1s, T1w, T1x; | |
1252 T19 = TZ - T18; | |
1253 T1s = T1i - T1r; | |
1254 T1t = FMA(KP951056516, T19, KP587785252 * T1s); | |
1255 T1V = FNMS(KP951056516, T1s, KP587785252 * T19); | |
1256 T1w = Ti - Tr; | |
1257 T1x = TB - TK; | |
1258 T1y = FMA(KP951056516, T1w, KP587785252 * T1x); | |
1259 T1Y = FNMS(KP951056516, T1x, KP587785252 * T1w); | |
1260 } | |
1261 cr[0] = T9 + TM; | |
1262 ci[0] = T1K + T1L; | |
1263 { | |
1264 E T1u, T1O, TN, T1v; | |
1265 T1u = TQ - T1t; | |
1266 T1O = T1y + T1N; | |
1267 TN = W[8]; | |
1268 T1v = W[9]; | |
1269 cr[WS(rs, 5)] = FNMS(T1v, T1O, TN * T1u); | |
1270 ci[WS(rs, 5)] = FMA(T1v, T1u, TN * T1O); | |
1271 } | |
1272 { | |
1273 E T22, T24, T21, T23; | |
1274 T22 = T1U + T1V; | |
1275 T24 = T1Z - T1Y; | |
1276 T21 = W[28]; | |
1277 T23 = W[29]; | |
1278 cr[WS(rs, 15)] = FNMS(T23, T24, T21 * T22); | |
1279 ci[WS(rs, 15)] = FMA(T23, T22, T21 * T24); | |
1280 } | |
1281 { | |
1282 E T1W, T20, T1T, T1X; | |
1283 T1W = T1U - T1V; | |
1284 T20 = T1Y + T1Z; | |
1285 T1T = W[18]; | |
1286 T1X = W[19]; | |
1287 cr[WS(rs, 10)] = FNMS(T1X, T20, T1T * T1W); | |
1288 ci[WS(rs, 10)] = FMA(T1X, T1W, T1T * T20); | |
1289 } | |
1290 { | |
1291 E T1Q, T1S, T1P, T1R; | |
1292 T1Q = TQ + T1t; | |
1293 T1S = T1N - T1y; | |
1294 T1P = W[38]; | |
1295 T1R = W[39]; | |
1296 cr[WS(rs, 20)] = FNMS(T1R, T1S, T1P * T1Q); | |
1297 ci[WS(rs, 20)] = FMA(T1R, T1Q, T1P * T1S); | |
1298 } | |
1299 } | |
1300 { | |
1301 E T6H, T71, T6M, T74, T6i, T6x, T6y, T6z, T6Q, T6R, T6P, T6S; | |
1302 { | |
1303 E T6D, T6G, T6K, T6L; | |
1304 T6D = T6B + T6C; | |
1305 T6G = T6E - T6F; | |
1306 T6H = FMA(KP951056516, T6D, KP587785252 * T6G); | |
1307 T71 = FNMS(KP951056516, T6G, KP587785252 * T6D); | |
1308 T6K = T6l - T6o; | |
1309 T6L = T6v - T6s; | |
1310 T6M = FMA(KP951056516, T6K, KP587785252 * T6L); | |
1311 T74 = FNMS(KP951056516, T6L, KP587785252 * T6K); | |
1312 } | |
1313 { | |
1314 E T6p, T6w, T6N, T6O; | |
1315 T6i = T5v + T5w; | |
1316 T6p = T6l + T6o; | |
1317 T6w = T6s + T6v; | |
1318 T6x = T6p - T6w; | |
1319 T6y = FNMS(KP250000000, T6x, T6i); | |
1320 T6z = KP559016994 * (T6p + T6w); | |
1321 T6Q = T5R - T5Q; | |
1322 T6N = T6B - T6C; | |
1323 T6O = T6E + T6F; | |
1324 T6R = T6N + T6O; | |
1325 T6P = KP559016994 * (T6N - T6O); | |
1326 T6S = FNMS(KP250000000, T6R, T6Q); | |
1327 } | |
1328 { | |
1329 E T7c, T7e, T7b, T7d; | |
1330 T7c = T6i + T6x; | |
1331 T7e = T6Q + T6R; | |
1332 T7b = W[6]; | |
1333 T7d = W[7]; | |
1334 cr[WS(rs, 4)] = FNMS(T7d, T7e, T7b * T7c); | |
1335 ci[WS(rs, 4)] = FMA(T7d, T7c, T7b * T7e); | |
1336 } | |
1337 { | |
1338 E T72, T78, T76, T7a, T70, T75; | |
1339 T70 = T6y - T6z; | |
1340 T72 = T70 - T71; | |
1341 T78 = T70 + T71; | |
1342 T75 = T6S - T6P; | |
1343 T76 = T74 + T75; | |
1344 T7a = T75 - T74; | |
1345 { | |
1346 E T6Z, T73, T77, T79; | |
1347 T6Z = W[26]; | |
1348 T73 = W[27]; | |
1349 cr[WS(rs, 14)] = FNMS(T73, T76, T6Z * T72); | |
1350 ci[WS(rs, 14)] = FMA(T73, T72, T6Z * T76); | |
1351 T77 = W[36]; | |
1352 T79 = W[37]; | |
1353 cr[WS(rs, 19)] = FNMS(T79, T7a, T77 * T78); | |
1354 ci[WS(rs, 19)] = FMA(T79, T78, T77 * T7a); | |
1355 } | |
1356 } | |
1357 { | |
1358 E T6I, T6W, T6U, T6Y, T6A, T6T; | |
1359 T6A = T6y + T6z; | |
1360 T6I = T6A - T6H; | |
1361 T6W = T6A + T6H; | |
1362 T6T = T6P + T6S; | |
1363 T6U = T6M + T6T; | |
1364 T6Y = T6T - T6M; | |
1365 { | |
1366 E T6h, T6J, T6V, T6X; | |
1367 T6h = W[16]; | |
1368 T6J = W[17]; | |
1369 cr[WS(rs, 9)] = FNMS(T6J, T6U, T6h * T6I); | |
1370 ci[WS(rs, 9)] = FMA(T6J, T6I, T6h * T6U); | |
1371 T6V = W[46]; | |
1372 T6X = W[47]; | |
1373 cr[WS(rs, 24)] = FNMS(T6X, T6Y, T6V * T6W); | |
1374 ci[WS(rs, 24)] = FMA(T6X, T6W, T6V * T6Y); | |
1375 } | |
1376 } | |
1377 } | |
1378 { | |
1379 E T3n, T3N, T3s, T3Q, T3d, T3e, T36, T3f, T3C, T3D, T3v, T3E; | |
1380 { | |
1381 E T3j, T3m, T3q, T3r; | |
1382 T3j = T3h - T3i; | |
1383 T3m = T3k - T3l; | |
1384 T3n = FMA(KP951056516, T3j, KP587785252 * T3m); | |
1385 T3N = FNMS(KP951056516, T3m, KP587785252 * T3j); | |
1386 T3q = T2k + T2z; | |
1387 T3r = T2P - T34; | |
1388 T3s = FMA(KP951056516, T3q, KP587785252 * T3r); | |
1389 T3Q = FNMS(KP951056516, T3r, KP587785252 * T3q); | |
1390 } | |
1391 { | |
1392 E T2A, T35, T3t, T3u; | |
1393 T3d = T39 - T3c; | |
1394 T2A = T2k - T2z; | |
1395 T35 = T2P + T34; | |
1396 T3e = T2A + T35; | |
1397 T36 = KP559016994 * (T2A - T35); | |
1398 T3f = FNMS(KP250000000, T3e, T3d); | |
1399 T3C = T3y + T3B; | |
1400 T3t = T3h + T3i; | |
1401 T3u = T3k + T3l; | |
1402 T3D = T3t + T3u; | |
1403 T3v = KP559016994 * (T3t - T3u); | |
1404 T3E = FNMS(KP250000000, T3D, T3C); | |
1405 } | |
1406 { | |
1407 E T3Y, T40, T3X, T3Z; | |
1408 T3Y = T3d + T3e; | |
1409 T40 = T3C + T3D; | |
1410 T3X = W[2]; | |
1411 T3Z = W[3]; | |
1412 cr[WS(rs, 2)] = FNMS(T3Z, T40, T3X * T3Y); | |
1413 ci[WS(rs, 2)] = FMA(T3Z, T3Y, T3X * T40); | |
1414 } | |
1415 { | |
1416 E T3O, T3U, T3S, T3W, T3M, T3R; | |
1417 T3M = T3f - T36; | |
1418 T3O = T3M - T3N; | |
1419 T3U = T3M + T3N; | |
1420 T3R = T3E - T3v; | |
1421 T3S = T3Q + T3R; | |
1422 T3W = T3R - T3Q; | |
1423 { | |
1424 E T3L, T3P, T3T, T3V; | |
1425 T3L = W[22]; | |
1426 T3P = W[23]; | |
1427 cr[WS(rs, 12)] = FNMS(T3P, T3S, T3L * T3O); | |
1428 ci[WS(rs, 12)] = FMA(T3P, T3O, T3L * T3S); | |
1429 T3T = W[32]; | |
1430 T3V = W[33]; | |
1431 cr[WS(rs, 17)] = FNMS(T3V, T3W, T3T * T3U); | |
1432 ci[WS(rs, 17)] = FMA(T3V, T3U, T3T * T3W); | |
1433 } | |
1434 } | |
1435 { | |
1436 E T3o, T3I, T3G, T3K, T3g, T3F; | |
1437 T3g = T36 + T3f; | |
1438 T3o = T3g - T3n; | |
1439 T3I = T3g + T3n; | |
1440 T3F = T3v + T3E; | |
1441 T3G = T3s + T3F; | |
1442 T3K = T3F - T3s; | |
1443 { | |
1444 E T25, T3p, T3H, T3J; | |
1445 T25 = W[12]; | |
1446 T3p = W[13]; | |
1447 cr[WS(rs, 7)] = FNMS(T3p, T3G, T25 * T3o); | |
1448 ci[WS(rs, 7)] = FMA(T3p, T3o, T25 * T3G); | |
1449 T3H = W[42]; | |
1450 T3J = W[43]; | |
1451 cr[WS(rs, 22)] = FNMS(T3J, T3K, T3H * T3I); | |
1452 ci[WS(rs, 22)] = FMA(T3J, T3I, T3H * T3K); | |
1453 } | |
1454 } | |
1455 } | |
1456 { | |
1457 E T4r, T4L, T4w, T4O, T4h, T4i, T4g, T4j, T4A, T4B, T4z, T4C; | |
1458 { | |
1459 E T4n, T4q, T4u, T4v; | |
1460 T4n = T4l - T4m; | |
1461 T4q = T4o - T4p; | |
1462 T4r = FMA(KP951056516, T4n, KP587785252 * T4q); | |
1463 T4L = FNMS(KP951056516, T4q, KP587785252 * T4n); | |
1464 T4u = T44 + T47; | |
1465 T4v = T4b + T4e; | |
1466 T4w = FMA(KP951056516, T4u, KP587785252 * T4v); | |
1467 T4O = FNMS(KP951056516, T4v, KP587785252 * T4u); | |
1468 } | |
1469 { | |
1470 E T48, T4f, T4x, T4y; | |
1471 T4h = T39 + T3c; | |
1472 T48 = T44 - T47; | |
1473 T4f = T4b - T4e; | |
1474 T4i = T48 + T4f; | |
1475 T4g = KP559016994 * (T48 - T4f); | |
1476 T4j = FNMS(KP250000000, T4i, T4h); | |
1477 T4A = T3B - T3y; | |
1478 T4x = T4l + T4m; | |
1479 T4y = T4o + T4p; | |
1480 T4B = T4x + T4y; | |
1481 T4z = KP559016994 * (T4x - T4y); | |
1482 T4C = FNMS(KP250000000, T4B, T4A); | |
1483 } | |
1484 { | |
1485 E T4W, T4Y, T4V, T4X; | |
1486 T4W = T4h + T4i; | |
1487 T4Y = T4A + T4B; | |
1488 T4V = W[4]; | |
1489 T4X = W[5]; | |
1490 cr[WS(rs, 3)] = FNMS(T4X, T4Y, T4V * T4W); | |
1491 ci[WS(rs, 3)] = FMA(T4X, T4W, T4V * T4Y); | |
1492 } | |
1493 { | |
1494 E T4M, T4S, T4Q, T4U, T4K, T4P; | |
1495 T4K = T4j - T4g; | |
1496 T4M = T4K - T4L; | |
1497 T4S = T4K + T4L; | |
1498 T4P = T4C - T4z; | |
1499 T4Q = T4O + T4P; | |
1500 T4U = T4P - T4O; | |
1501 { | |
1502 E T4J, T4N, T4R, T4T; | |
1503 T4J = W[24]; | |
1504 T4N = W[25]; | |
1505 cr[WS(rs, 13)] = FNMS(T4N, T4Q, T4J * T4M); | |
1506 ci[WS(rs, 13)] = FMA(T4N, T4M, T4J * T4Q); | |
1507 T4R = W[34]; | |
1508 T4T = W[35]; | |
1509 cr[WS(rs, 18)] = FNMS(T4T, T4U, T4R * T4S); | |
1510 ci[WS(rs, 18)] = FMA(T4T, T4S, T4R * T4U); | |
1511 } | |
1512 } | |
1513 { | |
1514 E T4s, T4G, T4E, T4I, T4k, T4D; | |
1515 T4k = T4g + T4j; | |
1516 T4s = T4k - T4r; | |
1517 T4G = T4k + T4r; | |
1518 T4D = T4z + T4C; | |
1519 T4E = T4w + T4D; | |
1520 T4I = T4D - T4w; | |
1521 { | |
1522 E T41, T4t, T4F, T4H; | |
1523 T41 = W[14]; | |
1524 T4t = W[15]; | |
1525 cr[WS(rs, 8)] = FNMS(T4t, T4E, T41 * T4s); | |
1526 ci[WS(rs, 8)] = FMA(T4t, T4s, T41 * T4E); | |
1527 T4F = W[44]; | |
1528 T4H = W[45]; | |
1529 cr[WS(rs, 23)] = FNMS(T4H, T4I, T4F * T4G); | |
1530 ci[WS(rs, 23)] = FMA(T4H, T4G, T4F * T4I); | |
1531 } | |
1532 } | |
1533 } | |
1534 { | |
1535 E T5H, T63, T5M, T66, T5x, T5y, T5u, T5z, T5S, T5T, T5P, T5U; | |
1536 { | |
1537 E T5D, T5G, T5K, T5L; | |
1538 T5D = T5B - T5C; | |
1539 T5G = T5E - T5F; | |
1540 T5H = FMA(KP951056516, T5D, KP587785252 * T5G); | |
1541 T63 = FNMS(KP951056516, T5G, KP587785252 * T5D); | |
1542 T5K = T56 - T5d; | |
1543 T5L = T5l - T5s; | |
1544 T5M = FMA(KP951056516, T5K, KP587785252 * T5L); | |
1545 T66 = FNMS(KP951056516, T5L, KP587785252 * T5K); | |
1546 } | |
1547 { | |
1548 E T5e, T5t, T5N, T5O; | |
1549 T5x = T5v - T5w; | |
1550 T5e = T56 + T5d; | |
1551 T5t = T5l + T5s; | |
1552 T5y = T5e + T5t; | |
1553 T5u = KP559016994 * (T5e - T5t); | |
1554 T5z = FNMS(KP250000000, T5y, T5x); | |
1555 T5S = T5Q + T5R; | |
1556 T5N = T5B + T5C; | |
1557 T5O = T5E + T5F; | |
1558 T5T = T5N + T5O; | |
1559 T5P = KP559016994 * (T5N - T5O); | |
1560 T5U = FNMS(KP250000000, T5T, T5S); | |
1561 } | |
1562 { | |
1563 E T6e, T6g, T6d, T6f; | |
1564 T6e = T5x + T5y; | |
1565 T6g = T5S + T5T; | |
1566 T6d = W[0]; | |
1567 T6f = W[1]; | |
1568 cr[WS(rs, 1)] = FNMS(T6f, T6g, T6d * T6e); | |
1569 ci[WS(rs, 1)] = FMA(T6f, T6e, T6d * T6g); | |
1570 } | |
1571 { | |
1572 E T64, T6a, T68, T6c, T62, T67; | |
1573 T62 = T5z - T5u; | |
1574 T64 = T62 - T63; | |
1575 T6a = T62 + T63; | |
1576 T67 = T5U - T5P; | |
1577 T68 = T66 + T67; | |
1578 T6c = T67 - T66; | |
1579 { | |
1580 E T61, T65, T69, T6b; | |
1581 T61 = W[20]; | |
1582 T65 = W[21]; | |
1583 cr[WS(rs, 11)] = FNMS(T65, T68, T61 * T64); | |
1584 ci[WS(rs, 11)] = FMA(T65, T64, T61 * T68); | |
1585 T69 = W[30]; | |
1586 T6b = W[31]; | |
1587 cr[WS(rs, 16)] = FNMS(T6b, T6c, T69 * T6a); | |
1588 ci[WS(rs, 16)] = FMA(T6b, T6a, T69 * T6c); | |
1589 } | |
1590 } | |
1591 { | |
1592 E T5I, T5Y, T5W, T60, T5A, T5V; | |
1593 T5A = T5u + T5z; | |
1594 T5I = T5A - T5H; | |
1595 T5Y = T5A + T5H; | |
1596 T5V = T5P + T5U; | |
1597 T5W = T5M + T5V; | |
1598 T60 = T5V - T5M; | |
1599 { | |
1600 E T4Z, T5J, T5X, T5Z; | |
1601 T4Z = W[10]; | |
1602 T5J = W[11]; | |
1603 cr[WS(rs, 6)] = FNMS(T5J, T5W, T4Z * T5I); | |
1604 ci[WS(rs, 6)] = FMA(T5J, T5I, T4Z * T5W); | |
1605 T5X = W[40]; | |
1606 T5Z = W[41]; | |
1607 cr[WS(rs, 21)] = FNMS(T5Z, T60, T5X * T5Y); | |
1608 ci[WS(rs, 21)] = FMA(T5Z, T5Y, T5X * T60); | |
1609 } | |
1610 } | |
1611 } | |
1612 } | |
1613 } | |
1614 } | |
1615 | |
1616 static const tw_instr twinstr[] = { | |
1617 {TW_FULL, 1, 25}, | |
1618 {TW_NEXT, 1, 0} | |
1619 }; | |
1620 | |
1621 static const hc2hc_desc desc = { 25, "hb_25", twinstr, &GENUS, {260, 140, 140, 0} }; | |
1622 | |
1623 void X(codelet_hb_25) (planner *p) { | |
1624 X(khc2hc_register) (p, hb_25, &desc); | |
1625 } | |
1626 #endif /* HAVE_FMA */ |