Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/rdft/scalar/r2cb/hb2_25.c @ 95:89f5e221ed7b
Add FFTW3
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Wed, 20 Mar 2013 15:35:50 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
94:d278df1123f9 | 95:89f5e221ed7b |
---|---|
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 07:41:30 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 -twiddle-log3 -precompute-twiddles -n 25 -dif -name hb2_25 -include hb.h */ | |
29 | |
30 /* | |
31 * This function contains 440 FP additions, 434 FP multiplications, | |
32 * (or, 84 additions, 78 multiplications, 356 fused multiply/add), | |
33 * 234 stack variables, 47 constants, and 100 memory accesses | |
34 */ | |
35 #include "hb.h" | |
36 | |
37 static void hb2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DK(KP833417178, +0.833417178328688677408962550243238843138996060); | |
40 DK(KP921177326, +0.921177326965143320250447435415066029359282231); | |
41 DK(KP541454447, +0.541454447536312777046285590082819509052033189); | |
42 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
43 DK(KP242145790, +0.242145790282157779872542093866183953459003101); | |
44 DK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
45 DK(KP803003575, +0.803003575438660414833440593570376004635464850); | |
46 DK(KP554608978, +0.554608978404018097464974850792216217022558774); | |
47 DK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
48 DK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
49 DK(KP248028675, +0.248028675328619457762448260696444630363259177); | |
50 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
51 DK(KP831864738, +0.831864738706457140726048799369896829771167132); | |
52 DK(KP871714437, +0.871714437527667770979999223229522602943903653); | |
53 DK(KP851038619, +0.851038619207379630836264138867114231259902550); | |
54 DK(KP943557151, +0.943557151597354104399655195398983005179443399); | |
55 DK(KP726211448, +0.726211448929902658173535992263577167607493062); | |
56 DK(KP525970792, +0.525970792408939708442463226536226366643874659); | |
57 DK(KP912018591, +0.912018591466481957908415381764119056233607330); | |
58 DK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
59 DK(KP994076283, +0.994076283785401014123185814696322018529298887); | |
60 DK(KP614372930, +0.614372930789563808870829930444362096004872855); | |
61 DK(KP621716863, +0.621716863012209892444754556304102309693593202); | |
62 DK(KP860541664, +0.860541664367944677098261680920518816412804187); | |
63 DK(KP949179823, +0.949179823508441261575555465843363271711583843); | |
64 DK(KP557913902, +0.557913902031834264187699648465567037992437152); | |
65 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
66 DK(KP249506682, +0.249506682107067890488084201715862638334226305); | |
67 DK(KP772036680, +0.772036680810363904029489473607579825330539880); | |
68 DK(KP906616052, +0.906616052148196230441134447086066874408359177); | |
69 DK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
70 DK(KP560319534, +0.560319534973832390111614715371676131169633784); | |
71 DK(KP681693190, +0.681693190061530575150324149145440022633095390); | |
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) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(50, rs)) { | |
89 E TN, TQ, T4e, T2y, T4i, T3U, T4u, T4o, T4G, T4C, T2F, T41, T3Q, T4q, T3a; | |
90 E T3F, T4a, T4w, T46, T44; | |
91 { | |
92 E TT, TO, TR, T23, T2d, T2x, TP, TV, T2p, T85, T4d, T25, TX; | |
93 TN = W[0]; | |
94 TT = W[4]; | |
95 TO = W[2]; | |
96 TR = W[3]; | |
97 T23 = W[6]; | |
98 T2d = TN * TT; | |
99 T2x = TO * TT; | |
100 TP = TN * TO; | |
101 TV = TN * TR; | |
102 T2p = TT * T23; | |
103 T85 = TN * T23; | |
104 T4d = TO * T23; | |
105 T25 = W[7]; | |
106 TQ = W[1]; | |
107 TX = W[5]; | |
108 { | |
109 E T86, T4n, TW, T4l, TS, T71, T2q, T4z, T2e, T8a, T2u, T76, T2k, T4B, T6E; | |
110 E T6U, T6Y, T5T, T8i, T1I, T2a, T26, TY, T8d, T8s, T8o, T5C, T5w, T7g, T7c; | |
111 E T5M, T5I, T9, T40, T1R, T3X, T6H, T7F, T5W, T7n, T4N, T68, T1S, T1k, T1T; | |
112 E T1D, T1Y, T1Z, T10, TM, T7K, T7A, T6p, T6w, T4X, T56, T3K, T2U, T7x, T7J; | |
113 E T6v, T6i, T50, T57, T3L, T39, T4Q, T59, T3O, T3E, T67, T7t, T7H, T6y, T63; | |
114 E T4T, T5a, T3N, T3p, T66, T7o; | |
115 { | |
116 E T2A, T2z, T6G, T2E, T5V, T6F; | |
117 { | |
118 E T1, T1J, T3Y, T3Z, T8, T2C, T1M, T1P, T2D, T4h, T89, T2t, T3W, T1Q, T3V; | |
119 T1 = cr[0]; | |
120 T4e = FMA(TR, T25, T4d); | |
121 T4h = TO * T25; | |
122 T89 = TN * T25; | |
123 T2t = TT * T25; | |
124 T86 = FMA(TQ, T25, T85); | |
125 T4n = FNMS(TQ, TO, TV); | |
126 TW = FMA(TQ, TO, TV); | |
127 T4l = FMA(TQ, TR, TP); | |
128 TS = FNMS(TQ, TR, TP); | |
129 T71 = FNMS(TR, TX, T2x); | |
130 T2y = FMA(TR, TX, T2x); | |
131 T2q = FMA(TX, T25, T2p); | |
132 T4z = FMA(TQ, TX, T2d); | |
133 T2e = FNMS(TQ, TX, T2d); | |
134 { | |
135 E T3T, T2j, T4t, T6T; | |
136 T3T = TO * TX; | |
137 T2j = TN * TX; | |
138 T4i = FNMS(TR, T23, T4h); | |
139 T8a = FNMS(TQ, T23, T89); | |
140 T2u = FNMS(TX, T23, T2t); | |
141 T4t = T4l * TX; | |
142 T6T = T4l * T23; | |
143 { | |
144 E T6X, T4m, T1H, T29; | |
145 T6X = T4l * T25; | |
146 T4m = T4l * TT; | |
147 T1H = TS * TX; | |
148 T29 = TS * T25; | |
149 { | |
150 E T24, TU, T4F, T4A; | |
151 T24 = TS * T23; | |
152 TU = TS * TT; | |
153 T4F = T4z * T25; | |
154 T4A = T4z * T23; | |
155 { | |
156 E T8r, T8n, T5B, T5v; | |
157 T8r = T2y * T25; | |
158 T8n = T2y * T23; | |
159 T5B = T2e * T25; | |
160 T5v = T2e * T23; | |
161 T3U = FNMS(TR, TT, T3T); | |
162 T76 = FMA(TR, TT, T3T); | |
163 T2k = FMA(TQ, TT, T2j); | |
164 T4B = FNMS(TQ, TT, T2j); | |
165 T4u = FMA(T4n, TT, T4t); | |
166 T6E = FNMS(T4n, TT, T4t); | |
167 T6U = FMA(T4n, T25, T6T); | |
168 T6Y = FNMS(T4n, T23, T6X); | |
169 T5T = FMA(T4n, TX, T4m); | |
170 T4o = FNMS(T4n, TX, T4m); | |
171 T8i = FMA(TW, TT, T1H); | |
172 T1I = FNMS(TW, TT, T1H); | |
173 T2a = FNMS(TW, T23, T29); | |
174 T26 = FMA(TW, T25, T24); | |
175 TY = FMA(TW, TX, TU); | |
176 T8d = FNMS(TW, TX, TU); | |
177 T8s = FNMS(T3U, T23, T8r); | |
178 T8o = FMA(T3U, T25, T8n); | |
179 T5C = FNMS(T2k, T23, T5B); | |
180 T5w = FMA(T2k, T25, T5v); | |
181 T4G = FNMS(T4B, T23, T4F); | |
182 T4C = FMA(T4B, T25, T4A); | |
183 { | |
184 E T7f, T7b, T5L, T5H; | |
185 T7f = T5T * T25; | |
186 T7b = T5T * T23; | |
187 T5L = TY * T25; | |
188 T5H = TY * T23; | |
189 T7g = FNMS(T6E, T23, T7f); | |
190 T7c = FMA(T6E, T25, T7b); | |
191 T5M = FNMS(T1I, T23, T5L); | |
192 T5I = FMA(T1I, T25, T5H); | |
193 T1J = ci[WS(rs, 24)]; | |
194 } | |
195 } | |
196 } | |
197 } | |
198 } | |
199 { | |
200 E T2, T3, T5, T6; | |
201 T2 = cr[WS(rs, 5)]; | |
202 T3 = ci[WS(rs, 4)]; | |
203 T5 = cr[WS(rs, 10)]; | |
204 T6 = ci[WS(rs, 9)]; | |
205 { | |
206 E T1K, T4, T7, T1L, T1N, T1O; | |
207 T1K = ci[WS(rs, 19)]; | |
208 T3Y = T2 - T3; | |
209 T4 = T2 + T3; | |
210 T3Z = T5 - T6; | |
211 T7 = T5 + T6; | |
212 T1L = cr[WS(rs, 20)]; | |
213 T1N = ci[WS(rs, 14)]; | |
214 T1O = cr[WS(rs, 15)]; | |
215 T8 = T4 + T7; | |
216 T2A = T4 - T7; | |
217 T2C = T1K + T1L; | |
218 T1M = T1K - T1L; | |
219 T1P = T1N - T1O; | |
220 T2D = T1N + T1O; | |
221 } | |
222 } | |
223 T2z = FNMS(KP250000000, T8, T1); | |
224 T9 = T1 + T8; | |
225 T3W = T1M - T1P; | |
226 T1Q = T1M + T1P; | |
227 T40 = FMA(KP618033988, T3Z, T3Y); | |
228 T6G = FNMS(KP618033988, T3Y, T3Z); | |
229 T2E = FMA(KP618033988, T2D, T2C); | |
230 T5V = FNMS(KP618033988, T2C, T2D); | |
231 T1R = T1J + T1Q; | |
232 T3V = FNMS(KP250000000, T1Q, T1J); | |
233 T6F = FNMS(KP559016994, T3W, T3V); | |
234 T3X = FMA(KP559016994, T3W, T3V); | |
235 } | |
236 { | |
237 E T2S, T6n, T2H, T2G, Ti, T5Y, T3C, T3r, TK, T3q, T30, T6d, T33, Tr, T32; | |
238 E T3v, T61, T3y, T1C, T3x, T2L, T6k, T2O, T1a, T2N, T6g, T37, T2W, Tt, T1j; | |
239 E T2V, Tx, T3g, T3j, Tw, T3l, T1t, T3i, Ty; | |
240 { | |
241 E T1u, T1v, T1A, T3u, T1w; | |
242 { | |
243 E TC, TI, T3B, TD, TE; | |
244 { | |
245 E Ta, Te, Tf, Tb, Tc, T5U, T2B, T2R, Tg; | |
246 Ta = cr[WS(rs, 1)]; | |
247 T5U = FNMS(KP559016994, T2A, T2z); | |
248 T2B = FMA(KP559016994, T2A, T2z); | |
249 T6H = FNMS(KP951056516, T6G, T6F); | |
250 T7F = FMA(KP951056516, T6G, T6F); | |
251 Te = cr[WS(rs, 11)]; | |
252 T5W = FMA(KP951056516, T5V, T5U); | |
253 T7n = FNMS(KP951056516, T5V, T5U); | |
254 T4N = FMA(KP951056516, T2E, T2B); | |
255 T2F = FNMS(KP951056516, T2E, T2B); | |
256 Tf = ci[WS(rs, 8)]; | |
257 Tb = cr[WS(rs, 6)]; | |
258 Tc = ci[WS(rs, 3)]; | |
259 TC = cr[WS(rs, 3)]; | |
260 T2R = Tf - Te; | |
261 Tg = Te + Tf; | |
262 { | |
263 E T2Q, Td, Th, TG, TH; | |
264 T2Q = Tb - Tc; | |
265 Td = Tb + Tc; | |
266 TG = ci[WS(rs, 11)]; | |
267 TH = ci[WS(rs, 6)]; | |
268 T2S = FNMS(KP618033988, T2R, T2Q); | |
269 T6n = FMA(KP618033988, T2Q, T2R); | |
270 Th = Td + Tg; | |
271 T2H = Td - Tg; | |
272 TI = TG + TH; | |
273 T3B = TG - TH; | |
274 T2G = FNMS(KP250000000, Th, Ta); | |
275 Ti = Ta + Th; | |
276 TD = cr[WS(rs, 8)]; | |
277 TE = ci[WS(rs, 1)]; | |
278 } | |
279 } | |
280 { | |
281 E Tj, Tk, Tp, T2Z, TJ, Tl; | |
282 Tj = cr[WS(rs, 4)]; | |
283 { | |
284 E Tn, To, T3A, TF; | |
285 Tn = ci[WS(rs, 10)]; | |
286 To = ci[WS(rs, 5)]; | |
287 T3A = TD - TE; | |
288 TF = TD + TE; | |
289 Tk = cr[WS(rs, 9)]; | |
290 Tp = Tn + To; | |
291 T2Z = To - Tn; | |
292 T5Y = FNMS(KP618033988, T3A, T3B); | |
293 T3C = FMA(KP618033988, T3B, T3A); | |
294 T3r = TI - TF; | |
295 TJ = TF + TI; | |
296 Tl = ci[0]; | |
297 } | |
298 T1u = ci[WS(rs, 21)]; | |
299 TK = TC + TJ; | |
300 T3q = FNMS(KP250000000, TJ, TC); | |
301 { | |
302 E T1y, Tm, T2Y, T1z, Tq; | |
303 T1y = cr[WS(rs, 13)]; | |
304 Tm = Tk + Tl; | |
305 T2Y = Tl - Tk; | |
306 T1z = cr[WS(rs, 18)]; | |
307 T1v = ci[WS(rs, 16)]; | |
308 T30 = FMA(KP618033988, T2Z, T2Y); | |
309 T6d = FNMS(KP618033988, T2Y, T2Z); | |
310 T33 = Tm - Tp; | |
311 Tq = Tm + Tp; | |
312 T1A = T1y + T1z; | |
313 T3u = T1z - T1y; | |
314 Tr = Tj + Tq; | |
315 T32 = FMS(KP250000000, Tq, Tj); | |
316 T1w = cr[WS(rs, 23)]; | |
317 } | |
318 } | |
319 } | |
320 { | |
321 E T1b, T1c, T1h, T36, T1d; | |
322 { | |
323 E T12, T13, T18, T2K, T1B, T14; | |
324 T12 = ci[WS(rs, 23)]; | |
325 { | |
326 E T16, T17, T3t, T1x; | |
327 T16 = ci[WS(rs, 13)]; | |
328 T17 = cr[WS(rs, 16)]; | |
329 T3t = T1v + T1w; | |
330 T1x = T1v - T1w; | |
331 T13 = ci[WS(rs, 18)]; | |
332 T18 = T16 - T17; | |
333 T2K = T16 + T17; | |
334 T3v = FMA(KP618033988, T3u, T3t); | |
335 T61 = FNMS(KP618033988, T3t, T3u); | |
336 T3y = T1x + T1A; | |
337 T1B = T1x - T1A; | |
338 T14 = cr[WS(rs, 21)]; | |
339 } | |
340 T1b = ci[WS(rs, 20)]; | |
341 T1C = T1u + T1B; | |
342 T3x = FMS(KP250000000, T1B, T1u); | |
343 { | |
344 E T1f, T15, T2J, T1g, T19; | |
345 T1f = cr[WS(rs, 14)]; | |
346 T15 = T13 - T14; | |
347 T2J = T13 + T14; | |
348 T1g = cr[WS(rs, 19)]; | |
349 T1c = ci[WS(rs, 15)]; | |
350 T2L = FMA(KP618033988, T2K, T2J); | |
351 T6k = FNMS(KP618033988, T2J, T2K); | |
352 T2O = T15 - T18; | |
353 T19 = T15 + T18; | |
354 T1h = T1f + T1g; | |
355 T36 = T1g - T1f; | |
356 T1a = T12 + T19; | |
357 T2N = FNMS(KP250000000, T19, T12); | |
358 T1d = cr[WS(rs, 24)]; | |
359 } | |
360 } | |
361 { | |
362 E T1l, T1p, T1o, T3e, T1i, T1q; | |
363 T1l = ci[WS(rs, 22)]; | |
364 { | |
365 E T1m, T1n, T35, T1e; | |
366 T1m = ci[WS(rs, 17)]; | |
367 T1n = cr[WS(rs, 22)]; | |
368 T35 = T1c + T1d; | |
369 T1e = T1c - T1d; | |
370 T1p = ci[WS(rs, 12)]; | |
371 T1o = T1m - T1n; | |
372 T3e = T1m + T1n; | |
373 T6g = FNMS(KP618033988, T35, T36); | |
374 T37 = FMA(KP618033988, T36, T35); | |
375 T2W = T1e + T1h; | |
376 T1i = T1e - T1h; | |
377 T1q = cr[WS(rs, 17)]; | |
378 } | |
379 Tt = cr[WS(rs, 2)]; | |
380 T1j = T1b + T1i; | |
381 T2V = FMS(KP250000000, T1i, T1b); | |
382 { | |
383 E Tu, T1r, T3f, Tv, T1s; | |
384 Tu = cr[WS(rs, 7)]; | |
385 T1r = T1p - T1q; | |
386 T3f = T1p + T1q; | |
387 Tv = ci[WS(rs, 2)]; | |
388 Tx = cr[WS(rs, 12)]; | |
389 T3g = FMA(KP618033988, T3f, T3e); | |
390 T68 = FNMS(KP618033988, T3e, T3f); | |
391 T3j = T1o - T1r; | |
392 T1s = T1o + T1r; | |
393 Tw = Tu + Tv; | |
394 T3l = Tu - Tv; | |
395 T1t = T1l + T1s; | |
396 T3i = FMS(KP250000000, T1s, T1l); | |
397 Ty = ci[WS(rs, 7)]; | |
398 } | |
399 } | |
400 } | |
401 } | |
402 { | |
403 E T3n, T65, T3c, T3b, T2P, T2M, T4W; | |
404 { | |
405 E TA, T3m, Tz, TB, Ts; | |
406 T3m = Ty - Tx; | |
407 Tz = Tx + Ty; | |
408 T1S = T1a + T1j; | |
409 T1k = T1a - T1j; | |
410 T3n = FNMS(KP618033988, T3m, T3l); | |
411 T65 = FMA(KP618033988, T3l, T3m); | |
412 TA = Tw + Tz; | |
413 T3c = Tz - Tw; | |
414 T3b = FNMS(KP250000000, TA, Tt); | |
415 TB = Tt + TA; | |
416 T1T = T1t + T1C; | |
417 T1D = T1t - T1C; | |
418 T1Y = Ti - Tr; | |
419 Ts = Ti + Tr; | |
420 { | |
421 E T2I, T6j, T6m, TL; | |
422 T2I = FMA(KP559016994, T2H, T2G); | |
423 T6j = FNMS(KP559016994, T2H, T2G); | |
424 T6m = FNMS(KP559016994, T2O, T2N); | |
425 T2P = FMA(KP559016994, T2O, T2N); | |
426 TL = TB + TK; | |
427 T1Z = TB - TK; | |
428 { | |
429 E T6l, T7y, T6o, T7z; | |
430 T6l = FMA(KP951056516, T6k, T6j); | |
431 T7y = FNMS(KP951056516, T6k, T6j); | |
432 T6o = FMA(KP951056516, T6n, T6m); | |
433 T7z = FNMS(KP951056516, T6n, T6m); | |
434 T10 = Ts - TL; | |
435 TM = Ts + TL; | |
436 T2M = FNMS(KP951056516, T2L, T2I); | |
437 T4W = FMA(KP951056516, T2L, T2I); | |
438 T7K = FMA(KP939062505, T7y, T7z); | |
439 T7A = FNMS(KP939062505, T7z, T7y); | |
440 T6p = FNMS(KP549754652, T6o, T6l); | |
441 T6w = FMA(KP549754652, T6l, T6o); | |
442 } | |
443 } | |
444 } | |
445 { | |
446 E T34, T31, T4Y, T60, T3s, T3z, T5X; | |
447 { | |
448 E T2X, T6c, T6f, T4V, T2T; | |
449 T2X = FNMS(KP559016994, T2W, T2V); | |
450 T6c = FMA(KP559016994, T2W, T2V); | |
451 T6f = FMA(KP559016994, T33, T32); | |
452 T34 = FNMS(KP559016994, T33, T32); | |
453 T4V = FNMS(KP951056516, T2S, T2P); | |
454 T2T = FMA(KP951056516, T2S, T2P); | |
455 { | |
456 E T7w, T6e, T7v, T6h; | |
457 T7w = FMA(KP951056516, T6d, T6c); | |
458 T6e = FNMS(KP951056516, T6d, T6c); | |
459 T7v = FMA(KP951056516, T6g, T6f); | |
460 T6h = FNMS(KP951056516, T6g, T6f); | |
461 T4X = FNMS(KP634619297, T4W, T4V); | |
462 T56 = FMA(KP634619297, T4V, T4W); | |
463 T3K = FMA(KP256756360, T2M, T2T); | |
464 T2U = FNMS(KP256756360, T2T, T2M); | |
465 T7x = FMA(KP126329378, T7w, T7v); | |
466 T7J = FNMS(KP126329378, T7v, T7w); | |
467 T6v = FNMS(KP470564281, T6e, T6h); | |
468 T6i = FMA(KP470564281, T6h, T6e); | |
469 T31 = FMA(KP951056516, T30, T2X); | |
470 T4Y = FNMS(KP951056516, T30, T2X); | |
471 } | |
472 T60 = FMA(KP559016994, T3r, T3q); | |
473 T3s = FNMS(KP559016994, T3r, T3q); | |
474 T3z = FNMS(KP559016994, T3y, T3x); | |
475 T5X = FMA(KP559016994, T3y, T3x); | |
476 } | |
477 { | |
478 E T5Z, T7r, T4Z, T38; | |
479 T4Z = FNMS(KP951056516, T37, T34); | |
480 T38 = FMA(KP951056516, T37, T34); | |
481 { | |
482 E T4O, T3w, T4P, T3D; | |
483 T4O = FMA(KP951056516, T3v, T3s); | |
484 T3w = FNMS(KP951056516, T3v, T3s); | |
485 T4P = FMA(KP951056516, T3C, T3z); | |
486 T3D = FNMS(KP951056516, T3C, T3z); | |
487 T50 = FNMS(KP827271945, T4Z, T4Y); | |
488 T57 = FMA(KP827271945, T4Y, T4Z); | |
489 T3L = FMA(KP634619297, T31, T38); | |
490 T39 = FNMS(KP634619297, T38, T31); | |
491 T4Q = FNMS(KP126329378, T4P, T4O); | |
492 T59 = FMA(KP126329378, T4O, T4P); | |
493 T3O = FNMS(KP939062505, T3w, T3D); | |
494 T3E = FMA(KP939062505, T3D, T3w); | |
495 T5Z = FMA(KP951056516, T5Y, T5X); | |
496 T7r = FNMS(KP951056516, T5Y, T5X); | |
497 } | |
498 { | |
499 E T3d, T3k, T64, T7s, T62; | |
500 T67 = FMA(KP559016994, T3c, T3b); | |
501 T3d = FNMS(KP559016994, T3c, T3b); | |
502 T3k = FNMS(KP559016994, T3j, T3i); | |
503 T64 = FMA(KP559016994, T3j, T3i); | |
504 T7s = FNMS(KP951056516, T61, T60); | |
505 T62 = FMA(KP951056516, T61, T60); | |
506 { | |
507 E T4S, T3h, T4R, T3o; | |
508 T4S = FMA(KP951056516, T3g, T3d); | |
509 T3h = FNMS(KP951056516, T3g, T3d); | |
510 T4R = FMA(KP951056516, T3n, T3k); | |
511 T3o = FNMS(KP951056516, T3n, T3k); | |
512 T7t = FNMS(KP827271945, T7s, T7r); | |
513 T7H = FMA(KP827271945, T7r, T7s); | |
514 T6y = FNMS(KP062914667, T5Z, T62); | |
515 T63 = FMA(KP062914667, T62, T5Z); | |
516 T4T = FNMS(KP470564281, T4S, T4R); | |
517 T5a = FMA(KP470564281, T4R, T4S); | |
518 T3N = FNMS(KP549754652, T3h, T3o); | |
519 T3p = FMA(KP549754652, T3o, T3h); | |
520 T66 = FNMS(KP951056516, T65, T64); | |
521 T7o = FMA(KP951056516, T65, T64); | |
522 } | |
523 } | |
524 } | |
525 } | |
526 } | |
527 } | |
528 } | |
529 { | |
530 E T7q, T7G, T6J, T6I, T6q, T6b, T6B, T73, T6Q, T78, T6z, T6a; | |
531 cr[0] = T9 + TM; | |
532 { | |
533 E T1U, T2l, T1X, T2g, T1E, TZ, T2m, T20, T2v, T2n; | |
534 { | |
535 E T1W, T7p, T69, T1V; | |
536 T1W = T1S - T1T; | |
537 T1U = T1S + T1T; | |
538 T7p = FNMS(KP951056516, T68, T67); | |
539 T69 = FMA(KP951056516, T68, T67); | |
540 T1V = FNMS(KP250000000, T1U, T1R); | |
541 T7q = FMA(KP062914667, T7p, T7o); | |
542 T7G = FNMS(KP062914667, T7o, T7p); | |
543 T6z = FNMS(KP634619297, T66, T69); | |
544 T6a = FMA(KP634619297, T69, T66); | |
545 T2l = FNMS(KP559016994, T1W, T1V); | |
546 T1X = FMA(KP559016994, T1W, T1V); | |
547 T2g = FNMS(KP618033988, T1k, T1D); | |
548 T1E = FMA(KP618033988, T1D, T1k); | |
549 TZ = FNMS(KP250000000, TM, T9); | |
550 T2m = FNMS(KP618033988, T1Y, T1Z); | |
551 T20 = FMA(KP618033988, T1Z, T1Y); | |
552 } | |
553 ci[0] = T1R + T1U; | |
554 T2v = FMA(KP951056516, T2m, T2l); | |
555 T2n = FNMS(KP951056516, T2m, T2l); | |
556 { | |
557 E T2b, T21, T2f, T11; | |
558 T2b = FNMS(KP951056516, T20, T1X); | |
559 T21 = FMA(KP951056516, T20, T1X); | |
560 T2f = FNMS(KP559016994, T10, TZ); | |
561 T11 = FMA(KP559016994, T10, TZ); | |
562 { | |
563 E T2h, T2r, T27, T1F; | |
564 T2h = FMA(KP951056516, T2g, T2f); | |
565 T2r = FNMS(KP951056516, T2g, T2f); | |
566 T27 = FMA(KP951056516, T1E, T11); | |
567 T1F = FNMS(KP951056516, T1E, T11); | |
568 { | |
569 E T2o, T2i, T2w, T2s; | |
570 T2o = T2k * T2h; | |
571 T2i = T2e * T2h; | |
572 T2w = T2u * T2r; | |
573 T2s = T2q * T2r; | |
574 { | |
575 E T2c, T28, T22, T1G; | |
576 T2c = T2a * T27; | |
577 T28 = T26 * T27; | |
578 T22 = T1I * T1F; | |
579 T1G = TY * T1F; | |
580 ci[WS(rs, 15)] = FMA(T2q, T2v, T2w); | |
581 cr[WS(rs, 15)] = FNMS(T2u, T2v, T2s); | |
582 ci[WS(rs, 20)] = FMA(T26, T2b, T2c); | |
583 cr[WS(rs, 20)] = FNMS(T2a, T2b, T28); | |
584 ci[WS(rs, 5)] = FMA(TY, T21, T22); | |
585 cr[WS(rs, 5)] = FNMS(T1I, T21, T1G); | |
586 cr[WS(rs, 10)] = FNMS(T2k, T2n, T2i); | |
587 ci[WS(rs, 10)] = FMA(T2e, T2n, T2o); | |
588 } | |
589 } | |
590 } | |
591 } | |
592 } | |
593 { | |
594 E T6x, T6A, T6O, T6P; | |
595 T6x = FMA(KP968479752, T6w, T6v); | |
596 T6J = FNMS(KP968479752, T6w, T6v); | |
597 T6I = FMA(KP845997307, T6z, T6y); | |
598 T6A = FNMS(KP845997307, T6z, T6y); | |
599 T6O = FNMS(KP968479752, T6p, T6i); | |
600 T6q = FMA(KP968479752, T6p, T6i); | |
601 T6b = FMA(KP845997307, T6a, T63); | |
602 T6P = FNMS(KP845997307, T6a, T63); | |
603 T6B = FNMS(KP681693190, T6A, T6x); | |
604 T73 = FMA(KP560319534, T6x, T6A); | |
605 T6Q = FMA(KP681693190, T6P, T6O); | |
606 T78 = FNMS(KP560319534, T6O, T6P); | |
607 } | |
608 { | |
609 E T7U, T8f, T7B, T7u, T82, T8k, T7Y, T7M; | |
610 { | |
611 E T7L, T7I, T80, T81; | |
612 { | |
613 E T7S, T6r, T6t, T6K, T6M, T7T, T6s, T7j; | |
614 T7S = FNMS(KP734762448, T7K, T7J); | |
615 T7L = FMA(KP734762448, T7K, T7J); | |
616 T6r = FMA(KP906616052, T6q, T6b); | |
617 T6t = FNMS(KP906616052, T6q, T6b); | |
618 T6K = FNMS(KP906616052, T6J, T6I); | |
619 T6M = FMA(KP906616052, T6J, T6I); | |
620 T7I = FMA(KP772036680, T7H, T7G); | |
621 T7T = FNMS(KP772036680, T7H, T7G); | |
622 T6s = FNMS(KP249506682, T6r, T5W); | |
623 T7j = FMA(KP998026728, T6r, T5W); | |
624 { | |
625 E T6L, T7l, T72, T6u; | |
626 T6L = FNMS(KP249506682, T6K, T6H); | |
627 T7l = FMA(KP998026728, T6K, T6H); | |
628 T72 = FMA(KP557913902, T6t, T6s); | |
629 T6u = FNMS(KP557913902, T6t, T6s); | |
630 { | |
631 E T7k, T6N, T77, T7m; | |
632 T7k = T4l * T7j; | |
633 T6N = FNMS(KP557913902, T6M, T6L); | |
634 T77 = FMA(KP557913902, T6M, T6L); | |
635 T7m = T4l * T7l; | |
636 { | |
637 E T74, T7d, T6V, T6C; | |
638 T74 = FNMS(KP949179823, T73, T72); | |
639 T7d = FMA(KP949179823, T73, T72); | |
640 T6V = FMA(KP860541664, T6B, T6u); | |
641 T6C = FNMS(KP860541664, T6B, T6u); | |
642 cr[WS(rs, 2)] = FNMS(T4n, T7l, T7k); | |
643 { | |
644 E T7h, T79, T6R, T6Z; | |
645 T7h = FNMS(KP949179823, T78, T77); | |
646 T79 = FMA(KP949179823, T78, T77); | |
647 T6R = FNMS(KP860541664, T6Q, T6N); | |
648 T6Z = FMA(KP860541664, T6Q, T6N); | |
649 ci[WS(rs, 2)] = FMA(T4n, T7j, T7m); | |
650 { | |
651 E T75, T7e, T6W, T6D; | |
652 T75 = T71 * T74; | |
653 T7e = T7c * T7d; | |
654 T6W = T6U * T6V; | |
655 T6D = T5T * T6C; | |
656 { | |
657 E T7a, T7i, T70, T6S; | |
658 T7a = T71 * T79; | |
659 T7i = T7c * T7h; | |
660 T70 = T6U * T6Z; | |
661 T6S = T5T * T6R; | |
662 cr[WS(rs, 12)] = FNMS(T76, T79, T75); | |
663 cr[WS(rs, 17)] = FNMS(T7g, T7h, T7e); | |
664 cr[WS(rs, 22)] = FNMS(T6Y, T6Z, T6W); | |
665 cr[WS(rs, 7)] = FNMS(T6E, T6R, T6D); | |
666 ci[WS(rs, 12)] = FMA(T76, T74, T7a); | |
667 ci[WS(rs, 17)] = FMA(T7g, T7d, T7i); | |
668 ci[WS(rs, 22)] = FMA(T6Y, T6V, T70); | |
669 ci[WS(rs, 7)] = FMA(T6E, T6C, T6S); | |
670 T7U = FNMS(KP621716863, T7T, T7S); | |
671 T8f = FMA(KP614372930, T7S, T7T); | |
672 } | |
673 } | |
674 } | |
675 } | |
676 } | |
677 } | |
678 } | |
679 T80 = FNMS(KP734762448, T7A, T7x); | |
680 T7B = FMA(KP734762448, T7A, T7x); | |
681 T7u = FMA(KP772036680, T7t, T7q); | |
682 T81 = FNMS(KP772036680, T7t, T7q); | |
683 T82 = FNMS(KP621716863, T81, T80); | |
684 T8k = FMA(KP614372930, T80, T81); | |
685 T7Y = FNMS(KP994076283, T7L, T7I); | |
686 T7M = FMA(KP994076283, T7L, T7I); | |
687 } | |
688 { | |
689 E T5y, T5c, T51, T4U, T5f, T5E, T5o, T5i, T5k; | |
690 { | |
691 E T5h, T5g, T5m, T5n, T58, T5b; | |
692 T5h = FMA(KP912575812, T57, T56); | |
693 T58 = FNMS(KP912575812, T57, T56); | |
694 T5b = FNMS(KP912018591, T5a, T59); | |
695 T5g = FMA(KP912018591, T5a, T59); | |
696 { | |
697 E T7X, T7N, T7C, T7Q; | |
698 T7X = FNMS(KP249506682, T7M, T7F); | |
699 T7N = FMA(KP998026728, T7M, T7F); | |
700 T7C = FMA(KP994076283, T7B, T7u); | |
701 T7Q = FNMS(KP994076283, T7B, T7u); | |
702 T5y = FMA(KP525970792, T58, T5b); | |
703 T5c = FNMS(KP726211448, T5b, T58); | |
704 { | |
705 E T7Z, T8j, T7P, T7D; | |
706 T7Z = FNMS(KP557913902, T7Y, T7X); | |
707 T8j = FMA(KP557913902, T7Y, T7X); | |
708 T7P = FNMS(KP249506682, T7C, T7n); | |
709 T7D = FMA(KP998026728, T7C, T7n); | |
710 { | |
711 E T8b, T83, T8t, T8l; | |
712 T8b = FMA(KP943557151, T82, T7Z); | |
713 T83 = FNMS(KP943557151, T82, T7Z); | |
714 T8t = FMA(KP949179823, T8k, T8j); | |
715 T8l = FNMS(KP949179823, T8k, T8j); | |
716 { | |
717 E T8e, T7R, T7O, T7E; | |
718 T8e = FMA(KP557913902, T7Q, T7P); | |
719 T7R = FNMS(KP557913902, T7Q, T7P); | |
720 T7O = TR * T7D; | |
721 T7E = TO * T7D; | |
722 { | |
723 E T8g, T8p, T7V, T87; | |
724 T8g = FMA(KP949179823, T8f, T8e); | |
725 T8p = FNMS(KP949179823, T8f, T8e); | |
726 T7V = FMA(KP943557151, T7U, T7R); | |
727 T87 = FNMS(KP943557151, T7U, T7R); | |
728 ci[WS(rs, 3)] = FMA(TO, T7N, T7O); | |
729 cr[WS(rs, 3)] = FNMS(TR, T7N, T7E); | |
730 { | |
731 E T8m, T8h, T8u, T8q; | |
732 T8m = T8i * T8g; | |
733 T8h = T8d * T8g; | |
734 T8u = T8s * T8p; | |
735 T8q = T8o * T8p; | |
736 { | |
737 E T84, T7W, T8c, T88; | |
738 T84 = T4B * T7V; | |
739 T7W = T4z * T7V; | |
740 T8c = T8a * T87; | |
741 T88 = T86 * T87; | |
742 ci[WS(rs, 13)] = FMA(T8d, T8l, T8m); | |
743 cr[WS(rs, 13)] = FNMS(T8i, T8l, T8h); | |
744 ci[WS(rs, 18)] = FMA(T8o, T8t, T8u); | |
745 cr[WS(rs, 18)] = FNMS(T8s, T8t, T8q); | |
746 ci[WS(rs, 8)] = FMA(T4z, T83, T84); | |
747 cr[WS(rs, 8)] = FNMS(T4B, T83, T7W); | |
748 ci[WS(rs, 23)] = FMA(T86, T8b, T8c); | |
749 cr[WS(rs, 23)] = FNMS(T8a, T8b, T88); | |
750 } | |
751 } | |
752 } | |
753 } | |
754 } | |
755 } | |
756 } | |
757 T51 = FMA(KP912575812, T50, T4X); | |
758 T5m = FNMS(KP912575812, T50, T4X); | |
759 T5n = FMA(KP912018591, T4T, T4Q); | |
760 T4U = FNMS(KP912018591, T4T, T4Q); | |
761 T41 = FMA(KP951056516, T40, T3X); | |
762 T5f = FNMS(KP951056516, T40, T3X); | |
763 T5E = FMA(KP525970792, T5m, T5n); | |
764 T5o = FNMS(KP726211448, T5n, T5m); | |
765 T5i = FMA(KP851038619, T5h, T5g); | |
766 T5k = FNMS(KP851038619, T5h, T5g); | |
767 } | |
768 { | |
769 E T42, T43, T48, T49, T3M, T3P; | |
770 T3M = FMA(KP871714437, T3L, T3K); | |
771 T42 = FNMS(KP871714437, T3L, T3K); | |
772 T43 = FMA(KP831864738, T3O, T3N); | |
773 T3P = FNMS(KP831864738, T3O, T3N); | |
774 { | |
775 E T5R, T5j, T54, T52; | |
776 T5R = FMA(KP992114701, T5i, T5f); | |
777 T5j = FNMS(KP248028675, T5i, T5f); | |
778 T54 = FNMS(KP851038619, T51, T4U); | |
779 T52 = FMA(KP851038619, T51, T4U); | |
780 T3Q = FNMS(KP559154169, T3P, T3M); | |
781 T4q = FMA(KP683113946, T3M, T3P); | |
782 { | |
783 E T5D, T5l, T5P, T53; | |
784 T5D = FMA(KP554608978, T5k, T5j); | |
785 T5l = FNMS(KP554608978, T5k, T5j); | |
786 T5P = FNMS(KP992114701, T52, T4N); | |
787 T53 = FMA(KP248028675, T52, T4N); | |
788 { | |
789 E T5p, T5t, T5F, T5N; | |
790 T5p = FNMS(KP803003575, T5o, T5l); | |
791 T5t = FMA(KP803003575, T5o, T5l); | |
792 T5F = FNMS(KP943557151, T5E, T5D); | |
793 T5N = FMA(KP943557151, T5E, T5D); | |
794 { | |
795 E T55, T5x, T5S, T5Q; | |
796 T55 = FMA(KP554608978, T54, T53); | |
797 T5x = FNMS(KP554608978, T54, T53); | |
798 T5S = TW * T5P; | |
799 T5Q = TS * T5P; | |
800 { | |
801 E T5J, T5z, T5r, T5d; | |
802 T5J = FMA(KP943557151, T5y, T5x); | |
803 T5z = FNMS(KP943557151, T5y, T5x); | |
804 T5r = FMA(KP803003575, T5c, T55); | |
805 T5d = FNMS(KP803003575, T5c, T55); | |
806 ci[WS(rs, 4)] = FMA(TS, T5R, T5S); | |
807 cr[WS(rs, 4)] = FNMS(TW, T5R, T5Q); | |
808 { | |
809 E T5G, T5A, T5O, T5K; | |
810 T5G = T5C * T5z; | |
811 T5A = T5w * T5z; | |
812 T5O = T5M * T5J; | |
813 T5K = T5I * T5J; | |
814 { | |
815 E T5q, T5e, T5u, T5s; | |
816 T5q = TX * T5d; | |
817 T5e = TT * T5d; | |
818 T5u = T25 * T5r; | |
819 T5s = T23 * T5r; | |
820 ci[WS(rs, 14)] = FMA(T5w, T5F, T5G); | |
821 cr[WS(rs, 14)] = FNMS(T5C, T5F, T5A); | |
822 ci[WS(rs, 19)] = FMA(T5I, T5N, T5O); | |
823 cr[WS(rs, 19)] = FNMS(T5M, T5N, T5K); | |
824 ci[WS(rs, 9)] = FMA(TT, T5p, T5q); | |
825 cr[WS(rs, 9)] = FNMS(TX, T5p, T5e); | |
826 ci[WS(rs, 24)] = FMA(T23, T5t, T5u); | |
827 cr[WS(rs, 24)] = FNMS(T25, T5t, T5s); | |
828 } | |
829 } | |
830 } | |
831 } | |
832 } | |
833 } | |
834 } | |
835 T48 = FNMS(KP871714437, T39, T2U); | |
836 T3a = FMA(KP871714437, T39, T2U); | |
837 T3F = FMA(KP831864738, T3E, T3p); | |
838 T49 = FNMS(KP831864738, T3E, T3p); | |
839 T4a = FMA(KP559154169, T49, T48); | |
840 T4w = FNMS(KP683113946, T48, T49); | |
841 T46 = FMA(KP904730450, T43, T42); | |
842 T44 = FNMS(KP904730450, T43, T42); | |
843 } | |
844 } | |
845 } | |
846 } | |
847 } | |
848 } | |
849 { | |
850 E T45, T4L, T3G, T3I; | |
851 T45 = FNMS(KP242145790, T44, T41); | |
852 T4L = FMA(KP968583161, T44, T41); | |
853 T3G = FMA(KP904730450, T3F, T3a); | |
854 T3I = FNMS(KP904730450, T3F, T3a); | |
855 { | |
856 E T4v, T47, T4J, T3H; | |
857 T4v = FNMS(KP541454447, T46, T45); | |
858 T47 = FMA(KP541454447, T46, T45); | |
859 T4J = FMA(KP968583161, T3G, T2F); | |
860 T3H = FNMS(KP242145790, T3G, T2F); | |
861 { | |
862 E T4b, T4j, T4x, T4H; | |
863 T4b = FMA(KP921177326, T4a, T47); | |
864 T4j = FNMS(KP921177326, T4a, T47); | |
865 T4x = FNMS(KP833417178, T4w, T4v); | |
866 T4H = FMA(KP833417178, T4w, T4v); | |
867 { | |
868 E T3J, T4p, T4M, T4K; | |
869 T3J = FMA(KP541454447, T3I, T3H); | |
870 T4p = FNMS(KP541454447, T3I, T3H); | |
871 T4M = TQ * T4J; | |
872 T4K = TN * T4J; | |
873 { | |
874 E T4D, T4r, T4f, T3R; | |
875 T4D = FMA(KP833417178, T4q, T4p); | |
876 T4r = FNMS(KP833417178, T4q, T4p); | |
877 T4f = FMA(KP921177326, T3Q, T3J); | |
878 T3R = FNMS(KP921177326, T3Q, T3J); | |
879 ci[WS(rs, 1)] = FMA(TN, T4L, T4M); | |
880 cr[WS(rs, 1)] = FNMS(TQ, T4L, T4K); | |
881 { | |
882 E T4y, T4s, T4I, T4E; | |
883 T4y = T4u * T4r; | |
884 T4s = T4o * T4r; | |
885 T4I = T4G * T4D; | |
886 T4E = T4C * T4D; | |
887 { | |
888 E T4c, T3S, T4k, T4g; | |
889 T4c = T3U * T3R; | |
890 T3S = T2y * T3R; | |
891 T4k = T4i * T4f; | |
892 T4g = T4e * T4f; | |
893 ci[WS(rs, 11)] = FMA(T4o, T4x, T4y); | |
894 cr[WS(rs, 11)] = FNMS(T4u, T4x, T4s); | |
895 ci[WS(rs, 16)] = FMA(T4C, T4H, T4I); | |
896 cr[WS(rs, 16)] = FNMS(T4G, T4H, T4E); | |
897 ci[WS(rs, 6)] = FMA(T2y, T4b, T4c); | |
898 cr[WS(rs, 6)] = FNMS(T3U, T4b, T3S); | |
899 ci[WS(rs, 21)] = FMA(T4e, T4j, T4k); | |
900 cr[WS(rs, 21)] = FNMS(T4i, T4j, T4g); | |
901 } | |
902 } | |
903 } | |
904 } | |
905 } | |
906 } | |
907 } | |
908 } | |
909 } | |
910 } | |
911 | |
912 static const tw_instr twinstr[] = { | |
913 {TW_CEXP, 1, 1}, | |
914 {TW_CEXP, 1, 3}, | |
915 {TW_CEXP, 1, 9}, | |
916 {TW_CEXP, 1, 24}, | |
917 {TW_NEXT, 1, 0} | |
918 }; | |
919 | |
920 static const hc2hc_desc desc = { 25, "hb2_25", twinstr, &GENUS, {84, 78, 356, 0} }; | |
921 | |
922 void X(codelet_hb2_25) (planner *p) { | |
923 X(khc2hc_register) (p, hb2_25, &desc); | |
924 } | |
925 #else /* HAVE_FMA */ | |
926 | |
927 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 25 -dif -name hb2_25 -include hb.h */ | |
928 | |
929 /* | |
930 * This function contains 440 FP additions, 340 FP multiplications, | |
931 * (or, 280 additions, 180 multiplications, 160 fused multiply/add), | |
932 * 155 stack variables, 20 constants, and 100 memory accesses | |
933 */ | |
934 #include "hb.h" | |
935 | |
936 static void hb2_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms) | |
937 { | |
938 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
939 DK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
940 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
941 DK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
942 DK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
943 DK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
944 DK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
945 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
946 DK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
947 DK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
948 DK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
949 DK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
950 DK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
951 DK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
952 DK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
953 DK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
954 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
955 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
956 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
957 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
958 { | |
959 INT m; | |
960 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)) { | |
961 E TN, TQ, TO, TR, TT, TY, T2t, T2r, TZ, TU, T4f, T4l, T2d, T4v, T5m; | |
962 E T2j, T5l, T4X, T2v, T11, T3R, T1L, T5d, T6x, T5h, T6t, T25, T26, T27, T29; | |
963 E T6D, T7v, T49, T7l, T7p, T7t, T2p, T2n, T4b, T4p, T5n, T6B, T5b, T5p, T6p; | |
964 E T6r, T59, T4r; | |
965 { | |
966 E T2c, T4j, T2h, T4e, T2b, T4k, T2i, T4d; | |
967 { | |
968 E TP, TX, TS, TW; | |
969 TN = W[0]; | |
970 TQ = W[1]; | |
971 TO = W[2]; | |
972 TR = W[3]; | |
973 TP = TN * TO; | |
974 TX = TQ * TO; | |
975 TS = TQ * TR; | |
976 TW = TN * TR; | |
977 TT = TP - TS; | |
978 TY = TW + TX; | |
979 T2t = TW - TX; | |
980 T2r = TP + TS; | |
981 TZ = W[5]; | |
982 T2c = TQ * TZ; | |
983 T4j = TO * TZ; | |
984 T2h = TN * TZ; | |
985 T4e = TR * TZ; | |
986 TU = W[4]; | |
987 T2b = TN * TU; | |
988 T4k = TR * TU; | |
989 T2i = TQ * TU; | |
990 T4d = TO * TU; | |
991 } | |
992 T4f = T4d - T4e; | |
993 T4l = T4j + T4k; | |
994 { | |
995 E T2s, T2u, TV, T10, T3P, T3Q, T1J, T1K; | |
996 T2d = T2b - T2c; | |
997 T4v = T2b + T2c; | |
998 T5m = T4j - T4k; | |
999 T2j = T2h + T2i; | |
1000 T5l = T4d + T4e; | |
1001 T4X = T2h - T2i; | |
1002 T2s = T2r * TU; | |
1003 T2u = T2t * TZ; | |
1004 T2v = T2s + T2u; | |
1005 TV = TT * TU; | |
1006 T10 = TY * TZ; | |
1007 T11 = TV + T10; | |
1008 T3P = T2r * TZ; | |
1009 T3Q = T2t * TU; | |
1010 T3R = T3P - T3Q; | |
1011 T1J = TT * TZ; | |
1012 T1K = TY * TU; | |
1013 T1L = T1J - T1K; | |
1014 T5d = TV - T10; | |
1015 T6x = T3P + T3Q; | |
1016 T5h = T1J + T1K; | |
1017 T6t = T2s - T2u; | |
1018 T25 = W[6]; | |
1019 T26 = W[7]; | |
1020 T27 = FMA(TT, T25, TY * T26); | |
1021 T29 = FNMS(TY, T25, TT * T26); | |
1022 T6D = FNMS(T4X, T25, T4v * T26); | |
1023 T7v = FNMS(T1L, T25, T11 * T26); | |
1024 T49 = FMA(T2r, T25, T2t * T26); | |
1025 T7l = FMA(T2d, T25, T2j * T26); | |
1026 T7p = FNMS(T2j, T25, T2d * T26); | |
1027 T7t = FMA(T11, T25, T1L * T26); | |
1028 T2p = FNMS(TZ, T25, TU * T26); | |
1029 T2n = FMA(TU, T25, TZ * T26); | |
1030 T4b = FNMS(T2t, T25, T2r * T26); | |
1031 T4p = FMA(T2v, T25, T3R * T26); | |
1032 T5n = FMA(T5l, T25, T5m * T26); | |
1033 T6B = FMA(T4v, T25, T4X * T26); | |
1034 T5b = FNMS(TQ, T25, TN * T26); | |
1035 T5p = FNMS(T5m, T25, T5l * T26); | |
1036 T6p = FMA(TO, T25, TR * T26); | |
1037 T6r = FNMS(TR, T25, TO * T26); | |
1038 T59 = FMA(TN, T25, TQ * T26); | |
1039 T4r = FNMS(T3R, T25, T2v * T26); | |
1040 } | |
1041 } | |
1042 { | |
1043 E T9, T6i, T40, T3z, T5Y, Ti, Tr, Ts, T1d, T1m, T1P, T2K, T4P, T3H, T4y; | |
1044 E T5G, T71, T65, T6N, T5z, T70, T64, T6K, T2Z, T4Q, T3I, T4B, T20, T5Z, T3C; | |
1045 E T43, T6j, TB, TK, TL, T1w, T1F, T1Q, T3f, T4S, T3K, T4F, T5V, T74, T68; | |
1046 E T6U, T5O, T73, T67, T6R, T3u, T4T, T3L, T4I; | |
1047 { | |
1048 E T1, T4, T7, T8, T3Z, T3Y, T3x, T3y; | |
1049 T1 = cr[0]; | |
1050 { | |
1051 E T2, T3, T5, T6; | |
1052 T2 = cr[WS(rs, 5)]; | |
1053 T3 = ci[WS(rs, 4)]; | |
1054 T4 = T2 + T3; | |
1055 T5 = cr[WS(rs, 10)]; | |
1056 T6 = ci[WS(rs, 9)]; | |
1057 T7 = T5 + T6; | |
1058 T8 = T4 + T7; | |
1059 T3Z = T5 - T6; | |
1060 T3Y = T2 - T3; | |
1061 } | |
1062 T9 = T1 + T8; | |
1063 T6i = FMA(KP951056516, T3Y, KP587785252 * T3Z); | |
1064 T40 = FNMS(KP951056516, T3Z, KP587785252 * T3Y); | |
1065 T3x = FNMS(KP250000000, T8, T1); | |
1066 T3y = KP559016994 * (T4 - T7); | |
1067 T3z = T3x - T3y; | |
1068 T5Y = T3y + T3x; | |
1069 } | |
1070 { | |
1071 E Ta, T2x, T5w, T2F, Th, T2w, T1e, T2P, T5B, T2X, T1l, T2O, Tj, T2N, T5D; | |
1072 E T2T, Tq, T2S, T15, T2B, T5u, T2H, T1c, T2G; | |
1073 { | |
1074 E Tg, T2E, Td, T2D; | |
1075 Ta = cr[WS(rs, 1)]; | |
1076 { | |
1077 E Te, Tf, Tb, Tc; | |
1078 Te = cr[WS(rs, 11)]; | |
1079 Tf = ci[WS(rs, 8)]; | |
1080 Tg = Te + Tf; | |
1081 T2E = Te - Tf; | |
1082 Tb = cr[WS(rs, 6)]; | |
1083 Tc = ci[WS(rs, 3)]; | |
1084 Td = Tb + Tc; | |
1085 T2D = Tb - Tc; | |
1086 } | |
1087 T2x = KP559016994 * (Td - Tg); | |
1088 T5w = FMA(KP951056516, T2D, KP587785252 * T2E); | |
1089 T2F = FNMS(KP951056516, T2E, KP587785252 * T2D); | |
1090 Th = Td + Tg; | |
1091 T2w = FNMS(KP250000000, Th, Ta); | |
1092 } | |
1093 { | |
1094 E T1k, T2W, T1h, T2V; | |
1095 T1e = ci[WS(rs, 20)]; | |
1096 { | |
1097 E T1i, T1j, T1f, T1g; | |
1098 T1i = cr[WS(rs, 14)]; | |
1099 T1j = cr[WS(rs, 19)]; | |
1100 T1k = T1i + T1j; | |
1101 T2W = T1j - T1i; | |
1102 T1f = ci[WS(rs, 15)]; | |
1103 T1g = cr[WS(rs, 24)]; | |
1104 T1h = T1f - T1g; | |
1105 T2V = T1f + T1g; | |
1106 } | |
1107 T2P = KP559016994 * (T1h + T1k); | |
1108 T5B = FMA(KP951056516, T2V, KP587785252 * T2W); | |
1109 T2X = FNMS(KP951056516, T2W, KP587785252 * T2V); | |
1110 T1l = T1h - T1k; | |
1111 T2O = FNMS(KP250000000, T1l, T1e); | |
1112 } | |
1113 { | |
1114 E Tp, T2M, Tm, T2L; | |
1115 Tj = cr[WS(rs, 4)]; | |
1116 { | |
1117 E Tn, To, Tk, Tl; | |
1118 Tn = ci[WS(rs, 10)]; | |
1119 To = ci[WS(rs, 5)]; | |
1120 Tp = Tn + To; | |
1121 T2M = Tn - To; | |
1122 Tk = cr[WS(rs, 9)]; | |
1123 Tl = ci[0]; | |
1124 Tm = Tk + Tl; | |
1125 T2L = Tk - Tl; | |
1126 } | |
1127 T2N = FNMS(KP951056516, T2M, KP587785252 * T2L); | |
1128 T5D = FMA(KP951056516, T2L, KP587785252 * T2M); | |
1129 T2T = KP559016994 * (Tm - Tp); | |
1130 Tq = Tm + Tp; | |
1131 T2S = FNMS(KP250000000, Tq, Tj); | |
1132 } | |
1133 { | |
1134 E T1b, T2A, T18, T2z; | |
1135 T15 = ci[WS(rs, 23)]; | |
1136 { | |
1137 E T19, T1a, T16, T17; | |
1138 T19 = ci[WS(rs, 13)]; | |
1139 T1a = cr[WS(rs, 16)]; | |
1140 T1b = T19 - T1a; | |
1141 T2A = T19 + T1a; | |
1142 T16 = ci[WS(rs, 18)]; | |
1143 T17 = cr[WS(rs, 21)]; | |
1144 T18 = T16 - T17; | |
1145 T2z = T16 + T17; | |
1146 } | |
1147 T2B = FNMS(KP951056516, T2A, KP587785252 * T2z); | |
1148 T5u = FMA(KP951056516, T2z, KP587785252 * T2A); | |
1149 T2H = KP559016994 * (T18 - T1b); | |
1150 T1c = T18 + T1b; | |
1151 T2G = FNMS(KP250000000, T1c, T15); | |
1152 } | |
1153 Ti = Ta + Th; | |
1154 Tr = Tj + Tq; | |
1155 Ts = Ti + Tr; | |
1156 T1d = T15 + T1c; | |
1157 T1m = T1e + T1l; | |
1158 T1P = T1d + T1m; | |
1159 { | |
1160 E T2C, T4w, T2J, T4x, T2y, T2I; | |
1161 T2y = T2w - T2x; | |
1162 T2C = T2y - T2B; | |
1163 T4w = T2y + T2B; | |
1164 T2I = T2G - T2H; | |
1165 T2J = T2F + T2I; | |
1166 T4x = T2I - T2F; | |
1167 T2K = FNMS(KP481753674, T2J, KP876306680 * T2C); | |
1168 T4P = FMA(KP728968627, T4x, KP684547105 * T4w); | |
1169 T3H = FMA(KP876306680, T2J, KP481753674 * T2C); | |
1170 T4y = FNMS(KP684547105, T4x, KP728968627 * T4w); | |
1171 } | |
1172 { | |
1173 E T5C, T6M, T5F, T6L, T5A, T5E; | |
1174 T5A = T2T + T2S; | |
1175 T5C = T5A - T5B; | |
1176 T6M = T5A + T5B; | |
1177 T5E = T2O + T2P; | |
1178 T5F = T5D + T5E; | |
1179 T6L = T5E - T5D; | |
1180 T5G = FNMS(KP844327925, T5F, KP535826794 * T5C); | |
1181 T71 = FMA(KP637423989, T6L, KP770513242 * T6M); | |
1182 T65 = FMA(KP535826794, T5F, KP844327925 * T5C); | |
1183 T6N = FNMS(KP637423989, T6M, KP770513242 * T6L); | |
1184 } | |
1185 { | |
1186 E T5v, T6I, T5y, T6J, T5t, T5x; | |
1187 T5t = T2x + T2w; | |
1188 T5v = T5t - T5u; | |
1189 T6I = T5t + T5u; | |
1190 T5x = T2H + T2G; | |
1191 T5y = T5w + T5x; | |
1192 T6J = T5x - T5w; | |
1193 T5z = FNMS(KP248689887, T5y, KP968583161 * T5v); | |
1194 T70 = FMA(KP535826794, T6J, KP844327925 * T6I); | |
1195 T64 = FMA(KP968583161, T5y, KP248689887 * T5v); | |
1196 T6K = FNMS(KP844327925, T6J, KP535826794 * T6I); | |
1197 } | |
1198 { | |
1199 E T2R, T4z, T2Y, T4A, T2Q, T2U; | |
1200 T2Q = T2O - T2P; | |
1201 T2R = T2N + T2Q; | |
1202 T4z = T2Q - T2N; | |
1203 T2U = T2S - T2T; | |
1204 T2Y = T2U - T2X; | |
1205 T4A = T2U + T2X; | |
1206 T2Z = FMA(KP904827052, T2R, KP425779291 * T2Y); | |
1207 T4Q = FNMS(KP992114701, T4z, KP125333233 * T4A); | |
1208 T3I = FNMS(KP425779291, T2R, KP904827052 * T2Y); | |
1209 T4B = FMA(KP125333233, T4z, KP992114701 * T4A); | |
1210 } | |
1211 } | |
1212 { | |
1213 E T1S, T1V, T1Y, T1Z, T3B, T3A, T41, T42; | |
1214 T1S = ci[WS(rs, 24)]; | |
1215 { | |
1216 E T1T, T1U, T1W, T1X; | |
1217 T1T = ci[WS(rs, 19)]; | |
1218 T1U = cr[WS(rs, 20)]; | |
1219 T1V = T1T - T1U; | |
1220 T1W = ci[WS(rs, 14)]; | |
1221 T1X = cr[WS(rs, 15)]; | |
1222 T1Y = T1W - T1X; | |
1223 T1Z = T1V + T1Y; | |
1224 T3B = T1W + T1X; | |
1225 T3A = T1T + T1U; | |
1226 } | |
1227 T20 = T1S + T1Z; | |
1228 T5Z = FMA(KP951056516, T3A, KP587785252 * T3B); | |
1229 T3C = FNMS(KP951056516, T3B, KP587785252 * T3A); | |
1230 T41 = FNMS(KP250000000, T1Z, T1S); | |
1231 T42 = KP559016994 * (T1V - T1Y); | |
1232 T43 = T41 - T42; | |
1233 T6j = T42 + T41; | |
1234 } | |
1235 { | |
1236 E Tt, T32, T5L, T3a, TA, T31, T1o, T36, T5J, T3c, T1v, T3b, TC, T3h, T5S; | |
1237 E T3p, TJ, T3g, T1x, T3l, T5Q, T3r, T1E, T3q; | |
1238 { | |
1239 E Tw, T38, Tz, T39; | |
1240 Tt = cr[WS(rs, 2)]; | |
1241 { | |
1242 E Tu, Tv, Tx, Ty; | |
1243 Tu = cr[WS(rs, 7)]; | |
1244 Tv = ci[WS(rs, 2)]; | |
1245 Tw = Tu + Tv; | |
1246 T38 = Tu - Tv; | |
1247 Tx = cr[WS(rs, 12)]; | |
1248 Ty = ci[WS(rs, 7)]; | |
1249 Tz = Tx + Ty; | |
1250 T39 = Tx - Ty; | |
1251 } | |
1252 T32 = KP559016994 * (Tw - Tz); | |
1253 T5L = FMA(KP951056516, T38, KP587785252 * T39); | |
1254 T3a = FNMS(KP951056516, T39, KP587785252 * T38); | |
1255 TA = Tw + Tz; | |
1256 T31 = FNMS(KP250000000, TA, Tt); | |
1257 } | |
1258 { | |
1259 E T1r, T34, T1u, T35; | |
1260 T1o = ci[WS(rs, 22)]; | |
1261 { | |
1262 E T1p, T1q, T1s, T1t; | |
1263 T1p = ci[WS(rs, 17)]; | |
1264 T1q = cr[WS(rs, 22)]; | |
1265 T1r = T1p - T1q; | |
1266 T34 = T1p + T1q; | |
1267 T1s = ci[WS(rs, 12)]; | |
1268 T1t = cr[WS(rs, 17)]; | |
1269 T1u = T1s - T1t; | |
1270 T35 = T1s + T1t; | |
1271 } | |
1272 T36 = FNMS(KP951056516, T35, KP587785252 * T34); | |
1273 T5J = FMA(KP951056516, T34, KP587785252 * T35); | |
1274 T3c = KP559016994 * (T1r - T1u); | |
1275 T1v = T1r + T1u; | |
1276 T3b = FNMS(KP250000000, T1v, T1o); | |
1277 } | |
1278 { | |
1279 E TI, T3o, TF, T3n; | |
1280 TC = cr[WS(rs, 3)]; | |
1281 { | |
1282 E TG, TH, TD, TE; | |
1283 TG = ci[WS(rs, 11)]; | |
1284 TH = ci[WS(rs, 6)]; | |
1285 TI = TG + TH; | |
1286 T3o = TG - TH; | |
1287 TD = cr[WS(rs, 8)]; | |
1288 TE = ci[WS(rs, 1)]; | |
1289 TF = TD + TE; | |
1290 T3n = TD - TE; | |
1291 } | |
1292 T3h = KP559016994 * (TF - TI); | |
1293 T5S = FMA(KP951056516, T3n, KP587785252 * T3o); | |
1294 T3p = FNMS(KP951056516, T3o, KP587785252 * T3n); | |
1295 TJ = TF + TI; | |
1296 T3g = FNMS(KP250000000, TJ, TC); | |
1297 } | |
1298 { | |
1299 E T1D, T3k, T1A, T3j; | |
1300 T1x = ci[WS(rs, 21)]; | |
1301 { | |
1302 E T1B, T1C, T1y, T1z; | |
1303 T1B = cr[WS(rs, 13)]; | |
1304 T1C = cr[WS(rs, 18)]; | |
1305 T1D = T1B + T1C; | |
1306 T3k = T1C - T1B; | |
1307 T1y = ci[WS(rs, 16)]; | |
1308 T1z = cr[WS(rs, 23)]; | |
1309 T1A = T1y - T1z; | |
1310 T3j = T1y + T1z; | |
1311 } | |
1312 T3l = FNMS(KP951056516, T3k, KP587785252 * T3j); | |
1313 T5Q = FMA(KP951056516, T3j, KP587785252 * T3k); | |
1314 T3r = KP559016994 * (T1A + T1D); | |
1315 T1E = T1A - T1D; | |
1316 T3q = FNMS(KP250000000, T1E, T1x); | |
1317 } | |
1318 TB = Tt + TA; | |
1319 TK = TC + TJ; | |
1320 TL = TB + TK; | |
1321 T1w = T1o + T1v; | |
1322 T1F = T1x + T1E; | |
1323 T1Q = T1w + T1F; | |
1324 { | |
1325 E T37, T4D, T3e, T4E, T33, T3d; | |
1326 T33 = T31 - T32; | |
1327 T37 = T33 - T36; | |
1328 T4D = T33 + T36; | |
1329 T3d = T3b - T3c; | |
1330 T3e = T3a + T3d; | |
1331 T4E = T3d - T3a; | |
1332 T3f = FNMS(KP844327925, T3e, KP535826794 * T37); | |
1333 T4S = FMA(KP062790519, T4E, KP998026728 * T4D); | |
1334 T3K = FMA(KP535826794, T3e, KP844327925 * T37); | |
1335 T4F = FNMS(KP998026728, T4E, KP062790519 * T4D); | |
1336 } | |
1337 { | |
1338 E T5R, T6T, T5U, T6S, T5P, T5T; | |
1339 T5P = T3h + T3g; | |
1340 T5R = T5P - T5Q; | |
1341 T6T = T5P + T5Q; | |
1342 T5T = T3q + T3r; | |
1343 T5U = T5S + T5T; | |
1344 T6S = T5T - T5S; | |
1345 T5V = FNMS(KP684547105, T5U, KP728968627 * T5R); | |
1346 T74 = FNMS(KP992114701, T6S, KP125333233 * T6T); | |
1347 T68 = FMA(KP728968627, T5U, KP684547105 * T5R); | |
1348 T6U = FMA(KP125333233, T6S, KP992114701 * T6T); | |
1349 } | |
1350 { | |
1351 E T5K, T6Q, T5N, T6P, T5I, T5M; | |
1352 T5I = T32 + T31; | |
1353 T5K = T5I - T5J; | |
1354 T6Q = T5I + T5J; | |
1355 T5M = T3c + T3b; | |
1356 T5N = T5L + T5M; | |
1357 T6P = T5M - T5L; | |
1358 T5O = FNMS(KP481753674, T5N, KP876306680 * T5K); | |
1359 T73 = FNMS(KP425779291, T6P, KP904827052 * T6Q); | |
1360 T67 = FMA(KP876306680, T5N, KP481753674 * T5K); | |
1361 T6R = FMA(KP904827052, T6P, KP425779291 * T6Q); | |
1362 } | |
1363 { | |
1364 E T3m, T4H, T3t, T4G, T3i, T3s; | |
1365 T3i = T3g - T3h; | |
1366 T3m = T3i - T3l; | |
1367 T4H = T3i + T3l; | |
1368 T3s = T3q - T3r; | |
1369 T3t = T3p + T3s; | |
1370 T4G = T3s - T3p; | |
1371 T3u = FNMS(KP998026728, T3t, KP062790519 * T3m); | |
1372 T4T = FNMS(KP637423989, T4G, KP770513242 * T4H); | |
1373 T3L = FMA(KP062790519, T3t, KP998026728 * T3m); | |
1374 T4I = FMA(KP770513242, T4G, KP637423989 * T4H); | |
1375 } | |
1376 } | |
1377 { | |
1378 E TM, T14, T2e, T21, T23, T2l, T1H, T2f, T1O, T2k; | |
1379 { | |
1380 E T12, T13, T1R, T22; | |
1381 T12 = KP559016994 * (Ts - TL); | |
1382 TM = Ts + TL; | |
1383 T13 = FNMS(KP250000000, TM, T9); | |
1384 T14 = T12 + T13; | |
1385 T2e = T13 - T12; | |
1386 T1R = KP559016994 * (T1P - T1Q); | |
1387 T21 = T1P + T1Q; | |
1388 T22 = FNMS(KP250000000, T21, T20); | |
1389 T23 = T1R + T22; | |
1390 T2l = T22 - T1R; | |
1391 } | |
1392 { | |
1393 E T1n, T1G, T1M, T1N; | |
1394 T1n = T1d - T1m; | |
1395 T1G = T1w - T1F; | |
1396 T1H = FMA(KP951056516, T1n, KP587785252 * T1G); | |
1397 T2f = FNMS(KP951056516, T1G, KP587785252 * T1n); | |
1398 T1M = Ti - Tr; | |
1399 T1N = TB - TK; | |
1400 T1O = FMA(KP951056516, T1M, KP587785252 * T1N); | |
1401 T2k = FNMS(KP951056516, T1N, KP587785252 * T1M); | |
1402 } | |
1403 { | |
1404 E T1I, T24, T2o, T2q; | |
1405 cr[0] = T9 + TM; | |
1406 ci[0] = T20 + T21; | |
1407 T1I = T14 - T1H; | |
1408 T24 = T1O + T23; | |
1409 cr[WS(rs, 5)] = FNMS(T1L, T24, T11 * T1I); | |
1410 ci[WS(rs, 5)] = FMA(T1L, T1I, T11 * T24); | |
1411 T2o = T2e + T2f; | |
1412 T2q = T2l - T2k; | |
1413 cr[WS(rs, 15)] = FNMS(T2p, T2q, T2n * T2o); | |
1414 ci[WS(rs, 15)] = FMA(T2p, T2o, T2n * T2q); | |
1415 { | |
1416 E T2g, T2m, T28, T2a; | |
1417 T2g = T2e - T2f; | |
1418 T2m = T2k + T2l; | |
1419 cr[WS(rs, 10)] = FNMS(T2j, T2m, T2d * T2g); | |
1420 ci[WS(rs, 10)] = FMA(T2j, T2g, T2d * T2m); | |
1421 T28 = T14 + T1H; | |
1422 T2a = T23 - T1O; | |
1423 cr[WS(rs, 20)] = FNMS(T29, T2a, T27 * T28); | |
1424 ci[WS(rs, 20)] = FMA(T29, T28, T27 * T2a); | |
1425 } | |
1426 } | |
1427 } | |
1428 { | |
1429 E T76, T7n, T7a, T7q, T6H, T6W, T6X, T6Y, T7e, T7f, T7d, T7g, T7x, T7y; | |
1430 { | |
1431 E T72, T75, T78, T79; | |
1432 T72 = T70 + T71; | |
1433 T75 = T73 - T74; | |
1434 T76 = FMA(KP951056516, T72, KP587785252 * T75); | |
1435 T7n = FNMS(KP951056516, T75, KP587785252 * T72); | |
1436 T78 = T6K - T6N; | |
1437 T79 = T6U - T6R; | |
1438 T7a = FMA(KP951056516, T78, KP587785252 * T79); | |
1439 T7q = FNMS(KP951056516, T79, KP587785252 * T78); | |
1440 } | |
1441 { | |
1442 E T6O, T6V, T7b, T7c; | |
1443 T6H = T5Y + T5Z; | |
1444 T6O = T6K + T6N; | |
1445 T6V = T6R + T6U; | |
1446 T6W = T6O - T6V; | |
1447 T6X = FNMS(KP250000000, T6W, T6H); | |
1448 T6Y = KP559016994 * (T6O + T6V); | |
1449 T7e = T6j - T6i; | |
1450 T7b = T70 - T71; | |
1451 T7c = T73 + T74; | |
1452 T7f = T7b + T7c; | |
1453 T7d = KP559016994 * (T7b - T7c); | |
1454 T7g = FNMS(KP250000000, T7f, T7e); | |
1455 } | |
1456 T7x = T6H + T6W; | |
1457 T7y = T7e + T7f; | |
1458 cr[WS(rs, 4)] = FNMS(TY, T7y, TT * T7x); | |
1459 ci[WS(rs, 4)] = FMA(TY, T7x, TT * T7y); | |
1460 { | |
1461 E T7o, T7u, T7s, T7w, T7m, T7r; | |
1462 T7m = T6X - T6Y; | |
1463 T7o = T7m - T7n; | |
1464 T7u = T7m + T7n; | |
1465 T7r = T7g - T7d; | |
1466 T7s = T7q + T7r; | |
1467 T7w = T7r - T7q; | |
1468 cr[WS(rs, 14)] = FNMS(T7p, T7s, T7l * T7o); | |
1469 ci[WS(rs, 14)] = FMA(T7p, T7o, T7l * T7s); | |
1470 cr[WS(rs, 19)] = FNMS(T7v, T7w, T7t * T7u); | |
1471 ci[WS(rs, 19)] = FMA(T7v, T7u, T7t * T7w); | |
1472 } | |
1473 { | |
1474 E T77, T7j, T7i, T7k, T6Z, T7h; | |
1475 T6Z = T6X + T6Y; | |
1476 T77 = T6Z - T76; | |
1477 T7j = T6Z + T76; | |
1478 T7h = T7d + T7g; | |
1479 T7i = T7a + T7h; | |
1480 T7k = T7h - T7a; | |
1481 cr[WS(rs, 9)] = FNMS(TZ, T7i, TU * T77); | |
1482 ci[WS(rs, 9)] = FMA(TZ, T77, TU * T7i); | |
1483 cr[WS(rs, 24)] = FNMS(T26, T7k, T25 * T7j); | |
1484 ci[WS(rs, 24)] = FMA(T26, T7j, T25 * T7k); | |
1485 } | |
1486 } | |
1487 { | |
1488 E T3N, T4h, T3U, T4m, T3D, T3E, T3w, T3F, T44, T45, T3X, T46, T4t, T4u; | |
1489 { | |
1490 E T3J, T3M, T3S, T3T; | |
1491 T3J = T3H - T3I; | |
1492 T3M = T3K - T3L; | |
1493 T3N = FMA(KP951056516, T3J, KP587785252 * T3M); | |
1494 T4h = FNMS(KP951056516, T3M, KP587785252 * T3J); | |
1495 T3S = T2K + T2Z; | |
1496 T3T = T3f - T3u; | |
1497 T3U = FMA(KP951056516, T3S, KP587785252 * T3T); | |
1498 T4m = FNMS(KP951056516, T3T, KP587785252 * T3S); | |
1499 } | |
1500 { | |
1501 E T30, T3v, T3V, T3W; | |
1502 T3D = T3z - T3C; | |
1503 T30 = T2K - T2Z; | |
1504 T3v = T3f + T3u; | |
1505 T3E = T30 + T3v; | |
1506 T3w = KP559016994 * (T30 - T3v); | |
1507 T3F = FNMS(KP250000000, T3E, T3D); | |
1508 T44 = T40 + T43; | |
1509 T3V = T3H + T3I; | |
1510 T3W = T3K + T3L; | |
1511 T45 = T3V + T3W; | |
1512 T3X = KP559016994 * (T3V - T3W); | |
1513 T46 = FNMS(KP250000000, T45, T44); | |
1514 } | |
1515 T4t = T3D + T3E; | |
1516 T4u = T44 + T45; | |
1517 cr[WS(rs, 2)] = FNMS(T2t, T4u, T2r * T4t); | |
1518 ci[WS(rs, 2)] = FMA(T2t, T4t, T2r * T4u); | |
1519 { | |
1520 E T4i, T4q, T4o, T4s, T4g, T4n; | |
1521 T4g = T3F - T3w; | |
1522 T4i = T4g - T4h; | |
1523 T4q = T4g + T4h; | |
1524 T4n = T46 - T3X; | |
1525 T4o = T4m + T4n; | |
1526 T4s = T4n - T4m; | |
1527 cr[WS(rs, 12)] = FNMS(T4l, T4o, T4f * T4i); | |
1528 ci[WS(rs, 12)] = FMA(T4l, T4i, T4f * T4o); | |
1529 cr[WS(rs, 17)] = FNMS(T4r, T4s, T4p * T4q); | |
1530 ci[WS(rs, 17)] = FMA(T4r, T4q, T4p * T4s); | |
1531 } | |
1532 { | |
1533 E T3O, T4a, T48, T4c, T3G, T47; | |
1534 T3G = T3w + T3F; | |
1535 T3O = T3G - T3N; | |
1536 T4a = T3G + T3N; | |
1537 T47 = T3X + T46; | |
1538 T48 = T3U + T47; | |
1539 T4c = T47 - T3U; | |
1540 cr[WS(rs, 7)] = FNMS(T3R, T48, T2v * T3O); | |
1541 ci[WS(rs, 7)] = FMA(T3R, T3O, T2v * T48); | |
1542 cr[WS(rs, 22)] = FNMS(T4b, T4c, T49 * T4a); | |
1543 ci[WS(rs, 22)] = FMA(T4b, T4a, T49 * T4c); | |
1544 } | |
1545 } | |
1546 { | |
1547 E T4V, T5f, T50, T5i, T4L, T4M, T4K, T4N, T54, T55, T53, T56, T5r, T5s; | |
1548 { | |
1549 E T4R, T4U, T4Y, T4Z; | |
1550 T4R = T4P - T4Q; | |
1551 T4U = T4S - T4T; | |
1552 T4V = FMA(KP951056516, T4R, KP587785252 * T4U); | |
1553 T5f = FNMS(KP951056516, T4U, KP587785252 * T4R); | |
1554 T4Y = T4y + T4B; | |
1555 T4Z = T4F + T4I; | |
1556 T50 = FMA(KP951056516, T4Y, KP587785252 * T4Z); | |
1557 T5i = FNMS(KP951056516, T4Z, KP587785252 * T4Y); | |
1558 } | |
1559 { | |
1560 E T4C, T4J, T51, T52; | |
1561 T4L = T3z + T3C; | |
1562 T4C = T4y - T4B; | |
1563 T4J = T4F - T4I; | |
1564 T4M = T4C + T4J; | |
1565 T4K = KP559016994 * (T4C - T4J); | |
1566 T4N = FNMS(KP250000000, T4M, T4L); | |
1567 T54 = T43 - T40; | |
1568 T51 = T4P + T4Q; | |
1569 T52 = T4S + T4T; | |
1570 T55 = T51 + T52; | |
1571 T53 = KP559016994 * (T51 - T52); | |
1572 T56 = FNMS(KP250000000, T55, T54); | |
1573 } | |
1574 T5r = T4L + T4M; | |
1575 T5s = T54 + T55; | |
1576 cr[WS(rs, 3)] = FNMS(TR, T5s, TO * T5r); | |
1577 ci[WS(rs, 3)] = FMA(TR, T5r, TO * T5s); | |
1578 { | |
1579 E T5g, T5o, T5k, T5q, T5e, T5j; | |
1580 T5e = T4N - T4K; | |
1581 T5g = T5e - T5f; | |
1582 T5o = T5e + T5f; | |
1583 T5j = T56 - T53; | |
1584 T5k = T5i + T5j; | |
1585 T5q = T5j - T5i; | |
1586 cr[WS(rs, 13)] = FNMS(T5h, T5k, T5d * T5g); | |
1587 ci[WS(rs, 13)] = FMA(T5h, T5g, T5d * T5k); | |
1588 cr[WS(rs, 18)] = FNMS(T5p, T5q, T5n * T5o); | |
1589 ci[WS(rs, 18)] = FMA(T5p, T5o, T5n * T5q); | |
1590 } | |
1591 { | |
1592 E T4W, T5a, T58, T5c, T4O, T57; | |
1593 T4O = T4K + T4N; | |
1594 T4W = T4O - T4V; | |
1595 T5a = T4O + T4V; | |
1596 T57 = T53 + T56; | |
1597 T58 = T50 + T57; | |
1598 T5c = T57 - T50; | |
1599 cr[WS(rs, 8)] = FNMS(T4X, T58, T4v * T4W); | |
1600 ci[WS(rs, 8)] = FMA(T4X, T4W, T4v * T58); | |
1601 cr[WS(rs, 23)] = FNMS(T5b, T5c, T59 * T5a); | |
1602 ci[WS(rs, 23)] = FMA(T5b, T5a, T59 * T5c); | |
1603 } | |
1604 } | |
1605 { | |
1606 E T6a, T6v, T6e, T6y, T60, T61, T5X, T62, T6k, T6l, T6h, T6m, T6F, T6G; | |
1607 { | |
1608 E T66, T69, T6c, T6d; | |
1609 T66 = T64 - T65; | |
1610 T69 = T67 - T68; | |
1611 T6a = FMA(KP951056516, T66, KP587785252 * T69); | |
1612 T6v = FNMS(KP951056516, T69, KP587785252 * T66); | |
1613 T6c = T5z - T5G; | |
1614 T6d = T5O - T5V; | |
1615 T6e = FMA(KP951056516, T6c, KP587785252 * T6d); | |
1616 T6y = FNMS(KP951056516, T6d, KP587785252 * T6c); | |
1617 } | |
1618 { | |
1619 E T5H, T5W, T6f, T6g; | |
1620 T60 = T5Y - T5Z; | |
1621 T5H = T5z + T5G; | |
1622 T5W = T5O + T5V; | |
1623 T61 = T5H + T5W; | |
1624 T5X = KP559016994 * (T5H - T5W); | |
1625 T62 = FNMS(KP250000000, T61, T60); | |
1626 T6k = T6i + T6j; | |
1627 T6f = T64 + T65; | |
1628 T6g = T67 + T68; | |
1629 T6l = T6f + T6g; | |
1630 T6h = KP559016994 * (T6f - T6g); | |
1631 T6m = FNMS(KP250000000, T6l, T6k); | |
1632 } | |
1633 T6F = T60 + T61; | |
1634 T6G = T6k + T6l; | |
1635 cr[WS(rs, 1)] = FNMS(TQ, T6G, TN * T6F); | |
1636 ci[WS(rs, 1)] = FMA(TQ, T6F, TN * T6G); | |
1637 { | |
1638 E T6w, T6C, T6A, T6E, T6u, T6z; | |
1639 T6u = T62 - T5X; | |
1640 T6w = T6u - T6v; | |
1641 T6C = T6u + T6v; | |
1642 T6z = T6m - T6h; | |
1643 T6A = T6y + T6z; | |
1644 T6E = T6z - T6y; | |
1645 cr[WS(rs, 11)] = FNMS(T6x, T6A, T6t * T6w); | |
1646 ci[WS(rs, 11)] = FMA(T6x, T6w, T6t * T6A); | |
1647 cr[WS(rs, 16)] = FNMS(T6D, T6E, T6B * T6C); | |
1648 ci[WS(rs, 16)] = FMA(T6D, T6C, T6B * T6E); | |
1649 } | |
1650 { | |
1651 E T6b, T6q, T6o, T6s, T63, T6n; | |
1652 T63 = T5X + T62; | |
1653 T6b = T63 - T6a; | |
1654 T6q = T63 + T6a; | |
1655 T6n = T6h + T6m; | |
1656 T6o = T6e + T6n; | |
1657 T6s = T6n - T6e; | |
1658 cr[WS(rs, 6)] = FNMS(T5m, T6o, T5l * T6b); | |
1659 ci[WS(rs, 6)] = FMA(T5m, T6b, T5l * T6o); | |
1660 cr[WS(rs, 21)] = FNMS(T6r, T6s, T6p * T6q); | |
1661 ci[WS(rs, 21)] = FMA(T6r, T6q, T6p * T6s); | |
1662 } | |
1663 } | |
1664 } | |
1665 } | |
1666 } | |
1667 } | |
1668 | |
1669 static const tw_instr twinstr[] = { | |
1670 {TW_CEXP, 1, 1}, | |
1671 {TW_CEXP, 1, 3}, | |
1672 {TW_CEXP, 1, 9}, | |
1673 {TW_CEXP, 1, 24}, | |
1674 {TW_NEXT, 1, 0} | |
1675 }; | |
1676 | |
1677 static const hc2hc_desc desc = { 25, "hb2_25", twinstr, &GENUS, {280, 180, 160, 0} }; | |
1678 | |
1679 void X(codelet_hb2_25) (planner *p) { | |
1680 X(khc2hc_register) (p, hb2_25, &desc); | |
1681 } | |
1682 #endif /* HAVE_FMA */ |