Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/scalar/codelets/t1_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 07:35:54 EST 2012 */ | |
23 | |
24 #include "codelet-dft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 25 -name t1_25 -include t.h */ | |
29 | |
30 /* | |
31 * This function contains 400 FP additions, 364 FP multiplications, | |
32 * (or, 84 additions, 48 multiplications, 316 fused multiply/add), | |
33 * 181 stack variables, 47 constants, and 100 memory accesses | |
34 */ | |
35 #include "t.h" | |
36 | |
37 static void t1_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DK(KP860541664, +0.860541664367944677098261680920518816412804187); | |
40 DK(KP681693190, +0.681693190061530575150324149145440022633095390); | |
41 DK(KP560319534, +0.560319534973832390111614715371676131169633784); | |
42 DK(KP949179823, +0.949179823508441261575555465843363271711583843); | |
43 DK(KP557913902, +0.557913902031834264187699648465567037992437152); | |
44 DK(KP249506682, +0.249506682107067890488084201715862638334226305); | |
45 DK(KP906616052, +0.906616052148196230441134447086066874408359177); | |
46 DK(KP968479752, +0.968479752739016373193524836781420152702090879); | |
47 DK(KP621716863, +0.621716863012209892444754556304102309693593202); | |
48 DK(KP614372930, +0.614372930789563808870829930444362096004872855); | |
49 DK(KP845997307, +0.845997307939530944175097360758058292389769300); | |
50 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
51 DK(KP994076283, +0.994076283785401014123185814696322018529298887); | |
52 DK(KP734762448, +0.734762448793050413546343770063151342619912334); | |
53 DK(KP772036680, +0.772036680810363904029489473607579825330539880); | |
54 DK(KP062914667, +0.062914667253649757225485955897349402364686947); | |
55 DK(KP803003575, +0.803003575438660414833440593570376004635464850); | |
56 DK(KP943557151, +0.943557151597354104399655195398983005179443399); | |
57 DK(KP554608978, +0.554608978404018097464974850792216217022558774); | |
58 DK(KP248028675, +0.248028675328619457762448260696444630363259177); | |
59 DK(KP726211448, +0.726211448929902658173535992263577167607493062); | |
60 DK(KP525970792, +0.525970792408939708442463226536226366643874659); | |
61 DK(KP921177326, +0.921177326965143320250447435415066029359282231); | |
62 DK(KP833417178, +0.833417178328688677408962550243238843138996060); | |
63 DK(KP541454447, +0.541454447536312777046285590082819509052033189); | |
64 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
65 DK(KP242145790, +0.242145790282157779872542093866183953459003101); | |
66 DK(KP851038619, +0.851038619207379630836264138867114231259902550); | |
67 DK(KP912575812, +0.912575812670962425556968549836277086778922727); | |
68 DK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
69 DK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
70 DK(KP912018591, +0.912018591466481957908415381764119056233607330); | |
71 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
72 DK(KP470564281, +0.470564281212251493087595091036643380879947982); | |
73 DK(KP827271945, +0.827271945972475634034355757144307982555673741); | |
74 DK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
75 DK(KP126329378, +0.126329378446108174786050455341811215027378105); | |
76 DK(KP831864738, +0.831864738706457140726048799369896829771167132); | |
77 DK(KP549754652, +0.549754652192770074288023275540779861653779767); | |
78 DK(KP871714437, +0.871714437527667770979999223229522602943903653); | |
79 DK(KP634619297, +0.634619297544148100711287640319130485732531031); | |
80 DK(KP939062505, +0.939062505817492352556001843133229685779824606); | |
81 DK(KP256756360, +0.256756360367726783319498520922669048172391148); | |
82 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
83 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
84 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
85 DK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
86 { | |
87 INT m; | |
88 for (m = mb, W = W + (mb * 48); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 48, MAKE_VOLATILE_STRIDE(50, rs)) { | |
89 E T7I, T6Q, T6O, T7O, T7M, T7H, T6P, T6H, T7J, T7N; | |
90 { | |
91 E T78, T5G, T3Y, T3M, T7C, T7c, T77, T6Y, Tt, T3L, T5T, T4P, T5Q, T4W, T3G; | |
92 E T2G, T5P, T4T, T5S, T4M, T65, T45, T68, T4c, T2Z, T11, T67, T49, T64, T42; | |
93 E T5Y, T4r, T61, T4k, T3d, T1z, T60, T4h, T5X, T4o, T3g, T1G, T3q, T4z, T4G; | |
94 E T26, T3i, T1M, T3k, T1S; | |
95 { | |
96 E T3u, T2e, T3E, T4O, T4V, T2E, T3w, T2k, T3y, T2q; | |
97 { | |
98 E T1, T6X, T3P, T7, T3W, Tq, T9, Tc, Tb, T3U, Tk, T3Q, Ta; | |
99 { | |
100 E T3, T6, T2, T5; | |
101 T1 = ri[0]; | |
102 T6X = ii[0]; | |
103 T3 = ri[WS(rs, 5)]; | |
104 T6 = ii[WS(rs, 5)]; | |
105 T2 = W[8]; | |
106 T5 = W[9]; | |
107 { | |
108 E Tm, Tp, To, T3V, Tn, T3O, T4, Tl; | |
109 Tm = ri[WS(rs, 15)]; | |
110 Tp = ii[WS(rs, 15)]; | |
111 T3O = T2 * T6; | |
112 T4 = T2 * T3; | |
113 Tl = W[28]; | |
114 To = W[29]; | |
115 T3P = FNMS(T5, T3, T3O); | |
116 T7 = FMA(T5, T6, T4); | |
117 T3V = Tl * Tp; | |
118 Tn = Tl * Tm; | |
119 { | |
120 E Tg, Tj, Tf, Ti, T3T, Th, T8; | |
121 Tg = ri[WS(rs, 10)]; | |
122 Tj = ii[WS(rs, 10)]; | |
123 T3W = FNMS(To, Tm, T3V); | |
124 Tq = FMA(To, Tp, Tn); | |
125 Tf = W[18]; | |
126 Ti = W[19]; | |
127 T9 = ri[WS(rs, 20)]; | |
128 Tc = ii[WS(rs, 20)]; | |
129 T3T = Tf * Tj; | |
130 Th = Tf * Tg; | |
131 T8 = W[38]; | |
132 Tb = W[39]; | |
133 T3U = FNMS(Ti, Tg, T3T); | |
134 Tk = FMA(Ti, Tj, Th); | |
135 T3Q = T8 * Tc; | |
136 Ta = T8 * T9; | |
137 } | |
138 } | |
139 } | |
140 { | |
141 E T6V, T3X, T7b, Tr, T3R, Td; | |
142 T6V = T3U + T3W; | |
143 T3X = T3U - T3W; | |
144 T7b = Tk - Tq; | |
145 Tr = Tk + Tq; | |
146 T3R = FNMS(Tb, T9, T3Q); | |
147 Td = FMA(Tb, Tc, Ta); | |
148 { | |
149 E T3S, T7a, Te, T6W, T6U, Ts; | |
150 T3S = T3P - T3R; | |
151 T6U = T3P + T3R; | |
152 T7a = T7 - Td; | |
153 Te = T7 + Td; | |
154 T78 = T6U - T6V; | |
155 T6W = T6U + T6V; | |
156 T5G = FNMS(KP618033988, T3S, T3X); | |
157 T3Y = FMA(KP618033988, T3X, T3S); | |
158 T3M = Te - Tr; | |
159 Ts = Te + Tr; | |
160 T7C = FNMS(KP618033988, T7a, T7b); | |
161 T7c = FMA(KP618033988, T7b, T7a); | |
162 T77 = FNMS(KP250000000, T6W, T6X); | |
163 T6Y = T6W + T6X; | |
164 Tt = T1 + Ts; | |
165 T3L = FNMS(KP250000000, Ts, T1); | |
166 } | |
167 } | |
168 } | |
169 { | |
170 E T2g, T2j, T2m, T3v, T2h, T2p, T2l, T2i, T2o, T3x, T2n; | |
171 { | |
172 E T2a, T2d, T29, T2c; | |
173 T2a = ri[WS(rs, 3)]; | |
174 T2d = ii[WS(rs, 3)]; | |
175 T29 = W[4]; | |
176 T2c = W[5]; | |
177 { | |
178 E T2t, T2w, T2z, T3A, T2u, T2C, T2y, T2v, T2B, T3t, T2b, T2s, T2f; | |
179 T2t = ri[WS(rs, 13)]; | |
180 T2w = ii[WS(rs, 13)]; | |
181 T3t = T29 * T2d; | |
182 T2b = T29 * T2a; | |
183 T2s = W[24]; | |
184 T2z = ri[WS(rs, 18)]; | |
185 T3u = FNMS(T2c, T2a, T3t); | |
186 T2e = FMA(T2c, T2d, T2b); | |
187 T3A = T2s * T2w; | |
188 T2u = T2s * T2t; | |
189 T2C = ii[WS(rs, 18)]; | |
190 T2y = W[34]; | |
191 T2v = W[25]; | |
192 T2B = W[35]; | |
193 { | |
194 E T3B, T2x, T3D, T2D, T3C, T2A; | |
195 T2g = ri[WS(rs, 8)]; | |
196 T3C = T2y * T2C; | |
197 T2A = T2y * T2z; | |
198 T3B = FNMS(T2v, T2t, T3A); | |
199 T2x = FMA(T2v, T2w, T2u); | |
200 T3D = FNMS(T2B, T2z, T3C); | |
201 T2D = FMA(T2B, T2C, T2A); | |
202 T2j = ii[WS(rs, 8)]; | |
203 T2f = W[14]; | |
204 T3E = T3B + T3D; | |
205 T4O = T3D - T3B; | |
206 T4V = T2x - T2D; | |
207 T2E = T2x + T2D; | |
208 } | |
209 T2m = ri[WS(rs, 23)]; | |
210 T3v = T2f * T2j; | |
211 T2h = T2f * T2g; | |
212 T2p = ii[WS(rs, 23)]; | |
213 T2l = W[44]; | |
214 T2i = W[15]; | |
215 T2o = W[45]; | |
216 } | |
217 } | |
218 T3x = T2l * T2p; | |
219 T2n = T2l * T2m; | |
220 T3w = FNMS(T2i, T2g, T3v); | |
221 T2k = FMA(T2i, T2j, T2h); | |
222 T3y = FNMS(T2o, T2m, T3x); | |
223 T2q = FMA(T2o, T2p, T2n); | |
224 } | |
225 { | |
226 E T2N, Tz, T2X, T44, T4b, TZ, T2P, TF, T2R, TL; | |
227 { | |
228 E TB, TE, TH, T2O, TC, TK, TG, TD, TJ, T2Q, TI; | |
229 { | |
230 E Tv, Ty, Tu, Tx; | |
231 { | |
232 E T4S, T4L, T4R, T4K, T4N, T3z; | |
233 Tv = ri[WS(rs, 1)]; | |
234 T4N = T3y - T3w; | |
235 T3z = T3w + T3y; | |
236 { | |
237 E T4U, T2r, T3F, T2F; | |
238 T4U = T2k - T2q; | |
239 T2r = T2k + T2q; | |
240 T5T = FNMS(KP618033988, T4N, T4O); | |
241 T4P = FMA(KP618033988, T4O, T4N); | |
242 T3F = T3z + T3E; | |
243 T4S = T3E - T3z; | |
244 T5Q = FNMS(KP618033988, T4U, T4V); | |
245 T4W = FMA(KP618033988, T4V, T4U); | |
246 T2F = T2r + T2E; | |
247 T4L = T2E - T2r; | |
248 T3G = T3u + T3F; | |
249 T4R = FNMS(KP250000000, T3F, T3u); | |
250 T2G = T2e + T2F; | |
251 T4K = FNMS(KP250000000, T2F, T2e); | |
252 Ty = ii[WS(rs, 1)]; | |
253 } | |
254 T5P = FMA(KP559016994, T4S, T4R); | |
255 T4T = FNMS(KP559016994, T4S, T4R); | |
256 T5S = FMA(KP559016994, T4L, T4K); | |
257 T4M = FNMS(KP559016994, T4L, T4K); | |
258 Tu = W[0]; | |
259 } | |
260 Tx = W[1]; | |
261 { | |
262 E TO, TR, TU, T2T, TP, TX, TT, TQ, TW, T2M, Tw, TN, TA; | |
263 TO = ri[WS(rs, 11)]; | |
264 TR = ii[WS(rs, 11)]; | |
265 T2M = Tu * Ty; | |
266 Tw = Tu * Tv; | |
267 TN = W[20]; | |
268 TU = ri[WS(rs, 16)]; | |
269 T2N = FNMS(Tx, Tv, T2M); | |
270 Tz = FMA(Tx, Ty, Tw); | |
271 T2T = TN * TR; | |
272 TP = TN * TO; | |
273 TX = ii[WS(rs, 16)]; | |
274 TT = W[30]; | |
275 TQ = W[21]; | |
276 TW = W[31]; | |
277 { | |
278 E T2U, TS, T2W, TY, T2V, TV; | |
279 TB = ri[WS(rs, 6)]; | |
280 T2V = TT * TX; | |
281 TV = TT * TU; | |
282 T2U = FNMS(TQ, TO, T2T); | |
283 TS = FMA(TQ, TR, TP); | |
284 T2W = FNMS(TW, TU, T2V); | |
285 TY = FMA(TW, TX, TV); | |
286 TE = ii[WS(rs, 6)]; | |
287 TA = W[10]; | |
288 T2X = T2U + T2W; | |
289 T44 = T2W - T2U; | |
290 T4b = TY - TS; | |
291 TZ = TS + TY; | |
292 } | |
293 TH = ri[WS(rs, 21)]; | |
294 T2O = TA * TE; | |
295 TC = TA * TB; | |
296 TK = ii[WS(rs, 21)]; | |
297 TG = W[40]; | |
298 TD = W[11]; | |
299 TJ = W[41]; | |
300 } | |
301 } | |
302 T2Q = TG * TK; | |
303 TI = TG * TH; | |
304 T2P = FNMS(TD, TB, T2O); | |
305 TF = FMA(TD, TE, TC); | |
306 T2R = FNMS(TJ, TH, T2Q); | |
307 TL = FMA(TJ, TK, TI); | |
308 } | |
309 { | |
310 E T31, T17, T3b, T4q, T4j, T1x, T33, T1d, T35, T1j; | |
311 { | |
312 E T19, T1c, T1f, T32, T1a, T1i, T1e, T1b, T1h, T34, T1g; | |
313 { | |
314 E T13, T16, T12, T15; | |
315 { | |
316 E T48, T41, T47, T40, T43, T2S; | |
317 T13 = ri[WS(rs, 4)]; | |
318 T43 = T2P - T2R; | |
319 T2S = T2P + T2R; | |
320 { | |
321 E T4a, TM, T2Y, T10; | |
322 T4a = TL - TF; | |
323 TM = TF + TL; | |
324 T65 = FMA(KP618033988, T43, T44); | |
325 T45 = FNMS(KP618033988, T44, T43); | |
326 T2Y = T2S + T2X; | |
327 T48 = T2S - T2X; | |
328 T68 = FNMS(KP618033988, T4a, T4b); | |
329 T4c = FMA(KP618033988, T4b, T4a); | |
330 T10 = TM + TZ; | |
331 T41 = TM - TZ; | |
332 T2Z = T2N + T2Y; | |
333 T47 = FNMS(KP250000000, T2Y, T2N); | |
334 T11 = Tz + T10; | |
335 T40 = FNMS(KP250000000, T10, Tz); | |
336 T16 = ii[WS(rs, 4)]; | |
337 } | |
338 T67 = FNMS(KP559016994, T48, T47); | |
339 T49 = FMA(KP559016994, T48, T47); | |
340 T64 = FNMS(KP559016994, T41, T40); | |
341 T42 = FMA(KP559016994, T41, T40); | |
342 T12 = W[6]; | |
343 } | |
344 T15 = W[7]; | |
345 { | |
346 E T1m, T1p, T1s, T37, T1n, T1v, T1r, T1o, T1u, T30, T14, T1l, T18; | |
347 T1m = ri[WS(rs, 14)]; | |
348 T1p = ii[WS(rs, 14)]; | |
349 T30 = T12 * T16; | |
350 T14 = T12 * T13; | |
351 T1l = W[26]; | |
352 T1s = ri[WS(rs, 19)]; | |
353 T31 = FNMS(T15, T13, T30); | |
354 T17 = FMA(T15, T16, T14); | |
355 T37 = T1l * T1p; | |
356 T1n = T1l * T1m; | |
357 T1v = ii[WS(rs, 19)]; | |
358 T1r = W[36]; | |
359 T1o = W[27]; | |
360 T1u = W[37]; | |
361 { | |
362 E T38, T1q, T3a, T1w, T39, T1t; | |
363 T19 = ri[WS(rs, 9)]; | |
364 T39 = T1r * T1v; | |
365 T1t = T1r * T1s; | |
366 T38 = FNMS(T1o, T1m, T37); | |
367 T1q = FMA(T1o, T1p, T1n); | |
368 T3a = FNMS(T1u, T1s, T39); | |
369 T1w = FMA(T1u, T1v, T1t); | |
370 T1c = ii[WS(rs, 9)]; | |
371 T18 = W[16]; | |
372 T3b = T38 + T3a; | |
373 T4q = T3a - T38; | |
374 T4j = T1w - T1q; | |
375 T1x = T1q + T1w; | |
376 } | |
377 T1f = ri[WS(rs, 24)]; | |
378 T32 = T18 * T1c; | |
379 T1a = T18 * T19; | |
380 T1i = ii[WS(rs, 24)]; | |
381 T1e = W[46]; | |
382 T1b = W[17]; | |
383 T1h = W[47]; | |
384 } | |
385 } | |
386 T34 = T1e * T1i; | |
387 T1g = T1e * T1f; | |
388 T33 = FNMS(T1b, T19, T32); | |
389 T1d = FMA(T1b, T1c, T1a); | |
390 T35 = FNMS(T1h, T1f, T34); | |
391 T1j = FMA(T1h, T1i, T1g); | |
392 } | |
393 { | |
394 E T1I, T1L, T1O, T3h, T1J, T1R, T1N, T1K, T1Q, T3j, T1P; | |
395 { | |
396 E T1C, T1F, T1B, T1E; | |
397 { | |
398 E T4g, T4n, T4f, T4m, T4p, T36; | |
399 T1C = ri[WS(rs, 2)]; | |
400 T4p = T35 - T33; | |
401 T36 = T33 + T35; | |
402 { | |
403 E T4i, T1k, T3c, T1y; | |
404 T4i = T1j - T1d; | |
405 T1k = T1d + T1j; | |
406 T5Y = FNMS(KP618033988, T4p, T4q); | |
407 T4r = FMA(KP618033988, T4q, T4p); | |
408 T3c = T36 + T3b; | |
409 T4g = T3b - T36; | |
410 T61 = FNMS(KP618033988, T4i, T4j); | |
411 T4k = FMA(KP618033988, T4j, T4i); | |
412 T1y = T1k + T1x; | |
413 T4n = T1k - T1x; | |
414 T3d = T31 + T3c; | |
415 T4f = FNMS(KP250000000, T3c, T31); | |
416 T1z = T17 + T1y; | |
417 T4m = FNMS(KP250000000, T1y, T17); | |
418 T1F = ii[WS(rs, 2)]; | |
419 } | |
420 T60 = FMA(KP559016994, T4g, T4f); | |
421 T4h = FNMS(KP559016994, T4g, T4f); | |
422 T5X = FNMS(KP559016994, T4n, T4m); | |
423 T4o = FMA(KP559016994, T4n, T4m); | |
424 T1B = W[2]; | |
425 } | |
426 T1E = W[3]; | |
427 { | |
428 E T1V, T1Y, T21, T3m, T1W, T24, T20, T1X, T23, T3f, T1D, T1U, T1H; | |
429 T1V = ri[WS(rs, 12)]; | |
430 T1Y = ii[WS(rs, 12)]; | |
431 T3f = T1B * T1F; | |
432 T1D = T1B * T1C; | |
433 T1U = W[22]; | |
434 T21 = ri[WS(rs, 17)]; | |
435 T3g = FNMS(T1E, T1C, T3f); | |
436 T1G = FMA(T1E, T1F, T1D); | |
437 T3m = T1U * T1Y; | |
438 T1W = T1U * T1V; | |
439 T24 = ii[WS(rs, 17)]; | |
440 T20 = W[32]; | |
441 T1X = W[23]; | |
442 T23 = W[33]; | |
443 { | |
444 E T3n, T1Z, T3p, T25, T3o, T22; | |
445 T1I = ri[WS(rs, 7)]; | |
446 T3o = T20 * T24; | |
447 T22 = T20 * T21; | |
448 T3n = FNMS(T1X, T1V, T3m); | |
449 T1Z = FMA(T1X, T1Y, T1W); | |
450 T3p = FNMS(T23, T21, T3o); | |
451 T25 = FMA(T23, T24, T22); | |
452 T1L = ii[WS(rs, 7)]; | |
453 T1H = W[12]; | |
454 T3q = T3n + T3p; | |
455 T4z = T3n - T3p; | |
456 T4G = T25 - T1Z; | |
457 T26 = T1Z + T25; | |
458 } | |
459 T1O = ri[WS(rs, 22)]; | |
460 T3h = T1H * T1L; | |
461 T1J = T1H * T1I; | |
462 T1R = ii[WS(rs, 22)]; | |
463 T1N = W[42]; | |
464 T1K = W[13]; | |
465 T1Q = W[43]; | |
466 } | |
467 } | |
468 T3j = T1N * T1R; | |
469 T1P = T1N * T1O; | |
470 T3i = FNMS(T1K, T1I, T3h); | |
471 T1M = FMA(T1K, T1L, T1J); | |
472 T3k = FNMS(T1Q, T1O, T3j); | |
473 T1S = FMA(T1Q, T1R, T1P); | |
474 } | |
475 } | |
476 } | |
477 } | |
478 { | |
479 E T6R, T5M, T4A, T5J, T4H, T6S, T5I, T4E, T5L, T4x, T3K, T3I, T2K, T74, T76; | |
480 E T2J; | |
481 { | |
482 E T1A, T72, T73, T2H, T28, T2I; | |
483 { | |
484 E T3e, T4D, T4w, T4C, T4v, T3H, T4y, T3l; | |
485 T6R = T2Z + T3d; | |
486 T3e = T2Z - T3d; | |
487 T4y = T3k - T3i; | |
488 T3l = T3i + T3k; | |
489 { | |
490 E T4F, T1T, T3r, T27, T3s; | |
491 T4F = T1S - T1M; | |
492 T1T = T1M + T1S; | |
493 T5M = FMA(KP618033988, T4y, T4z); | |
494 T4A = FNMS(KP618033988, T4z, T4y); | |
495 T3r = T3l + T3q; | |
496 T4D = T3q - T3l; | |
497 T5J = FNMS(KP618033988, T4F, T4G); | |
498 T4H = FMA(KP618033988, T4G, T4F); | |
499 T27 = T1T + T26; | |
500 T4w = T26 - T1T; | |
501 T3s = T3g + T3r; | |
502 T4C = FNMS(KP250000000, T3r, T3g); | |
503 T28 = T1G + T27; | |
504 T4v = FNMS(KP250000000, T27, T1G); | |
505 T3H = T3s - T3G; | |
506 T6S = T3s + T3G; | |
507 } | |
508 T5I = FMA(KP559016994, T4D, T4C); | |
509 T4E = FNMS(KP559016994, T4D, T4C); | |
510 T5L = FMA(KP559016994, T4w, T4v); | |
511 T4x = FNMS(KP559016994, T4w, T4v); | |
512 T3K = FNMS(KP618033988, T3e, T3H); | |
513 T3I = FMA(KP618033988, T3H, T3e); | |
514 } | |
515 T1A = T11 + T1z; | |
516 T72 = T11 - T1z; | |
517 T73 = T28 - T2G; | |
518 T2H = T28 + T2G; | |
519 T2I = T1A + T2H; | |
520 T2K = T1A - T2H; | |
521 T74 = FMA(KP618033988, T73, T72); | |
522 T76 = FNMS(KP618033988, T72, T73); | |
523 ri[0] = Tt + T2I; | |
524 T2J = FNMS(KP250000000, T2I, Tt); | |
525 } | |
526 { | |
527 E T5F, T7B, T7u, T5E, T5C, T7A, T7y, T7t, T5D, T5v; | |
528 { | |
529 E T3Z, T5d, T7p, T7d, T5m, T5l, T56, T7k, T59, T7l, T5z, T5g, T7g, T7i, T52; | |
530 E T50, T5x, T5q, T5A, T5j, T70, T6Z, T3N; | |
531 T5F = FNMS(KP559016994, T3M, T3L); | |
532 T3N = FMA(KP559016994, T3M, T3L); | |
533 { | |
534 E T79, T3J, T2L, T6T; | |
535 T79 = FMA(KP559016994, T78, T77); | |
536 T7B = FNMS(KP559016994, T78, T77); | |
537 T3J = FNMS(KP559016994, T2K, T2J); | |
538 T2L = FMA(KP559016994, T2K, T2J); | |
539 T6T = T6R + T6S; | |
540 T70 = T6R - T6S; | |
541 T3Z = FMA(KP951056516, T3Y, T3N); | |
542 T5d = FNMS(KP951056516, T3Y, T3N); | |
543 ri[WS(rs, 5)] = FMA(KP951056516, T3I, T2L); | |
544 ri[WS(rs, 20)] = FNMS(KP951056516, T3I, T2L); | |
545 ri[WS(rs, 15)] = FMA(KP951056516, T3K, T3J); | |
546 ri[WS(rs, 10)] = FNMS(KP951056516, T3K, T3J); | |
547 ii[0] = T6T + T6Y; | |
548 T6Z = FNMS(KP250000000, T6T, T6Y); | |
549 T7p = FMA(KP951056516, T7c, T79); | |
550 T7d = FNMS(KP951056516, T7c, T79); | |
551 } | |
552 { | |
553 E T5e, T54, T4e, T5f, T5o, T5p, T5i, T4B, T58, T4Y, T55, T4t, T4I, T5h; | |
554 { | |
555 E T4Q, T4X, T4l, T4s; | |
556 { | |
557 E T46, T71, T75, T4d; | |
558 T5m = FNMS(KP951056516, T45, T42); | |
559 T46 = FMA(KP951056516, T45, T42); | |
560 T71 = FMA(KP559016994, T70, T6Z); | |
561 T75 = FNMS(KP559016994, T70, T6Z); | |
562 T4d = FMA(KP951056516, T4c, T49); | |
563 T5l = FNMS(KP951056516, T4c, T49); | |
564 T5e = FMA(KP951056516, T4P, T4M); | |
565 T4Q = FNMS(KP951056516, T4P, T4M); | |
566 ii[WS(rs, 20)] = FMA(KP951056516, T74, T71); | |
567 ii[WS(rs, 5)] = FNMS(KP951056516, T74, T71); | |
568 ii[WS(rs, 15)] = FNMS(KP951056516, T76, T75); | |
569 ii[WS(rs, 10)] = FMA(KP951056516, T76, T75); | |
570 T54 = FNMS(KP256756360, T46, T4d); | |
571 T4e = FMA(KP256756360, T4d, T46); | |
572 T4X = FNMS(KP951056516, T4W, T4T); | |
573 T5f = FMA(KP951056516, T4W, T4T); | |
574 } | |
575 T5o = FNMS(KP951056516, T4k, T4h); | |
576 T4l = FMA(KP951056516, T4k, T4h); | |
577 T4s = FNMS(KP951056516, T4r, T4o); | |
578 T5p = FMA(KP951056516, T4r, T4o); | |
579 T5i = FMA(KP951056516, T4A, T4x); | |
580 T4B = FNMS(KP951056516, T4A, T4x); | |
581 T58 = FNMS(KP939062505, T4Q, T4X); | |
582 T4Y = FMA(KP939062505, T4X, T4Q); | |
583 T55 = FNMS(KP634619297, T4l, T4s); | |
584 T4t = FMA(KP634619297, T4s, T4l); | |
585 T4I = FMA(KP951056516, T4H, T4E); | |
586 T5h = FNMS(KP951056516, T4H, T4E); | |
587 } | |
588 { | |
589 E T7e, T4u, T57, T4J, T7f, T4Z; | |
590 T7e = FNMS(KP871714437, T55, T54); | |
591 T56 = FMA(KP871714437, T55, T54); | |
592 T4u = FMA(KP871714437, T4t, T4e); | |
593 T7k = FNMS(KP871714437, T4t, T4e); | |
594 T57 = FNMS(KP549754652, T4B, T4I); | |
595 T4J = FMA(KP549754652, T4I, T4B); | |
596 T7f = FMA(KP831864738, T58, T57); | |
597 T59 = FNMS(KP831864738, T58, T57); | |
598 T4Z = FMA(KP831864738, T4Y, T4J); | |
599 T7l = FNMS(KP831864738, T4Y, T4J); | |
600 T5z = FMA(KP126329378, T5e, T5f); | |
601 T5g = FNMS(KP126329378, T5f, T5e); | |
602 T7g = FMA(KP904730450, T7f, T7e); | |
603 T7i = FNMS(KP904730450, T7f, T7e); | |
604 T52 = FNMS(KP904730450, T4Z, T4u); | |
605 T50 = FMA(KP904730450, T4Z, T4u); | |
606 } | |
607 T5x = FNMS(KP827271945, T5o, T5p); | |
608 T5q = FMA(KP827271945, T5p, T5o); | |
609 T5A = FMA(KP470564281, T5h, T5i); | |
610 T5j = FNMS(KP470564281, T5i, T5h); | |
611 } | |
612 { | |
613 E T7q, T5B, T5k, T7x, T5w, T5n; | |
614 ri[WS(rs, 1)] = FMA(KP968583161, T50, T3Z); | |
615 T7q = FMA(KP912018591, T5A, T5z); | |
616 T5B = FNMS(KP912018591, T5A, T5z); | |
617 T5k = FNMS(KP912018591, T5j, T5g); | |
618 T7x = FMA(KP912018591, T5j, T5g); | |
619 T5w = FNMS(KP634619297, T5l, T5m); | |
620 T5n = FMA(KP634619297, T5m, T5l); | |
621 ii[WS(rs, 1)] = FMA(KP968583161, T7g, T7d); | |
622 { | |
623 E T5y, T7w, T7s, T5s, T5u, T7o, T7m, T7n, T7j, T5t; | |
624 { | |
625 E T5c, T5a, T51, T7r, T5r, T53, T5b, T7h; | |
626 T5c = FNMS(KP683113946, T56, T59); | |
627 T5a = FMA(KP559154169, T59, T56); | |
628 T7r = FNMS(KP912575812, T5x, T5w); | |
629 T5y = FMA(KP912575812, T5x, T5w); | |
630 T5r = FNMS(KP912575812, T5q, T5n); | |
631 T7w = FMA(KP912575812, T5q, T5n); | |
632 T7s = FMA(KP851038619, T7r, T7q); | |
633 T7u = FNMS(KP851038619, T7r, T7q); | |
634 T5s = FNMS(KP851038619, T5r, T5k); | |
635 T5u = FMA(KP851038619, T5r, T5k); | |
636 T51 = FNMS(KP242145790, T50, T3Z); | |
637 ii[WS(rs, 4)] = FNMS(KP992114701, T7s, T7p); | |
638 ri[WS(rs, 4)] = FNMS(KP992114701, T5s, T5d); | |
639 T7o = FNMS(KP683113946, T7k, T7l); | |
640 T7m = FMA(KP559154169, T7l, T7k); | |
641 T53 = FMA(KP541454447, T52, T51); | |
642 T5b = FNMS(KP541454447, T52, T51); | |
643 T7h = FNMS(KP242145790, T7g, T7d); | |
644 ri[WS(rs, 11)] = FNMS(KP833417178, T5c, T5b); | |
645 ri[WS(rs, 16)] = FMA(KP833417178, T5c, T5b); | |
646 ri[WS(rs, 21)] = FNMS(KP921177326, T5a, T53); | |
647 ri[WS(rs, 6)] = FMA(KP921177326, T5a, T53); | |
648 T7n = FNMS(KP541454447, T7i, T7h); | |
649 T7j = FMA(KP541454447, T7i, T7h); | |
650 } | |
651 T5E = FMA(KP525970792, T5y, T5B); | |
652 T5C = FNMS(KP726211448, T5B, T5y); | |
653 ii[WS(rs, 21)] = FMA(KP921177326, T7m, T7j); | |
654 ii[WS(rs, 6)] = FNMS(KP921177326, T7m, T7j); | |
655 ii[WS(rs, 11)] = FMA(KP833417178, T7o, T7n); | |
656 ii[WS(rs, 16)] = FNMS(KP833417178, T7o, T7n); | |
657 T5t = FMA(KP248028675, T5s, T5d); | |
658 T7A = FNMS(KP525970792, T7w, T7x); | |
659 T7y = FMA(KP726211448, T7x, T7w); | |
660 T7t = FMA(KP248028675, T7s, T7p); | |
661 T5D = FNMS(KP554608978, T5u, T5t); | |
662 T5v = FMA(KP554608978, T5u, T5t); | |
663 } | |
664 } | |
665 } | |
666 { | |
667 E T5H, T6p, T7P, T7D, T6y, T6x, T6l, T7X, T6i, T7W, T6L, T6s, T7S, T7U, T6e; | |
668 E T6c, T6J, T6C, T6M, T6v, T7z, T7v; | |
669 ri[WS(rs, 14)] = FNMS(KP943557151, T5E, T5D); | |
670 ri[WS(rs, 19)] = FMA(KP943557151, T5E, T5D); | |
671 ri[WS(rs, 24)] = FMA(KP803003575, T5C, T5v); | |
672 ri[WS(rs, 9)] = FNMS(KP803003575, T5C, T5v); | |
673 T7z = FNMS(KP554608978, T7u, T7t); | |
674 T7v = FMA(KP554608978, T7u, T7t); | |
675 T5H = FMA(KP951056516, T5G, T5F); | |
676 T6p = FNMS(KP951056516, T5G, T5F); | |
677 ii[WS(rs, 14)] = FMA(KP943557151, T7A, T7z); | |
678 ii[WS(rs, 19)] = FNMS(KP943557151, T7A, T7z); | |
679 ii[WS(rs, 24)] = FMA(KP803003575, T7y, T7v); | |
680 ii[WS(rs, 9)] = FNMS(KP803003575, T7y, T7v); | |
681 { | |
682 E T6t, T6u, T6A, T6j, T5O, T6B, T6q, T6r, T5Z, T6h, T6a, T6k, T5V, T62; | |
683 { | |
684 E T66, T69, T5K, T5N, T5R, T5U; | |
685 T6t = FNMS(KP951056516, T5J, T5I); | |
686 T5K = FMA(KP951056516, T5J, T5I); | |
687 T5N = FMA(KP951056516, T5M, T5L); | |
688 T6u = FNMS(KP951056516, T5M, T5L); | |
689 T6A = FMA(KP951056516, T65, T64); | |
690 T66 = FNMS(KP951056516, T65, T64); | |
691 T7P = FNMS(KP951056516, T7C, T7B); | |
692 T7D = FMA(KP951056516, T7C, T7B); | |
693 T6j = FNMS(KP062914667, T5K, T5N); | |
694 T5O = FMA(KP062914667, T5N, T5K); | |
695 T69 = FMA(KP951056516, T68, T67); | |
696 T6B = FNMS(KP951056516, T68, T67); | |
697 T6q = FMA(KP951056516, T5Q, T5P); | |
698 T5R = FNMS(KP951056516, T5Q, T5P); | |
699 T5U = FNMS(KP951056516, T5T, T5S); | |
700 T6r = FMA(KP951056516, T5T, T5S); | |
701 T6y = FMA(KP951056516, T5Y, T5X); | |
702 T5Z = FNMS(KP951056516, T5Y, T5X); | |
703 T6h = FNMS(KP939062505, T66, T69); | |
704 T6a = FMA(KP939062505, T69, T66); | |
705 T6k = FMA(KP827271945, T5R, T5U); | |
706 T5V = FNMS(KP827271945, T5U, T5R); | |
707 T62 = FMA(KP951056516, T61, T60); | |
708 T6x = FNMS(KP951056516, T61, T60); | |
709 } | |
710 { | |
711 E T7Q, T5W, T6g, T63, T7R, T6b; | |
712 T7Q = FMA(KP772036680, T6k, T6j); | |
713 T6l = FNMS(KP772036680, T6k, T6j); | |
714 T5W = FMA(KP772036680, T5V, T5O); | |
715 T7X = FNMS(KP772036680, T5V, T5O); | |
716 T6g = FMA(KP126329378, T5Z, T62); | |
717 T63 = FNMS(KP126329378, T62, T5Z); | |
718 T7R = FNMS(KP734762448, T6h, T6g); | |
719 T6i = FMA(KP734762448, T6h, T6g); | |
720 T6b = FNMS(KP734762448, T6a, T63); | |
721 T7W = FMA(KP734762448, T6a, T63); | |
722 T6L = FNMS(KP062914667, T6q, T6r); | |
723 T6s = FMA(KP062914667, T6r, T6q); | |
724 T7S = FMA(KP994076283, T7R, T7Q); | |
725 T7U = FNMS(KP994076283, T7R, T7Q); | |
726 T6e = FMA(KP994076283, T6b, T5W); | |
727 T6c = FNMS(KP994076283, T6b, T5W); | |
728 } | |
729 T6J = FNMS(KP549754652, T6A, T6B); | |
730 T6C = FMA(KP549754652, T6B, T6A); | |
731 T6M = FNMS(KP634619297, T6t, T6u); | |
732 T6v = FMA(KP634619297, T6u, T6t); | |
733 } | |
734 { | |
735 E T7E, T6N, T6w, T7L, T6I, T6z; | |
736 ri[WS(rs, 3)] = FMA(KP998026728, T6c, T5H); | |
737 T7E = FMA(KP845997307, T6M, T6L); | |
738 T6N = FNMS(KP845997307, T6M, T6L); | |
739 T6w = FMA(KP845997307, T6v, T6s); | |
740 T7L = FNMS(KP845997307, T6v, T6s); | |
741 T6I = FMA(KP470564281, T6x, T6y); | |
742 T6z = FNMS(KP470564281, T6y, T6x); | |
743 ii[WS(rs, 3)] = FNMS(KP998026728, T7S, T7P); | |
744 { | |
745 E T6K, T7K, T7G, T6E, T6G, T80, T7Y, T7Z, T7V, T6F; | |
746 { | |
747 E T6o, T6m, T6d, T7F, T6D, T6f, T6n, T7T; | |
748 T6o = FMA(KP614372930, T6i, T6l); | |
749 T6m = FNMS(KP621716863, T6l, T6i); | |
750 T7F = FNMS(KP968479752, T6J, T6I); | |
751 T6K = FMA(KP968479752, T6J, T6I); | |
752 T6D = FMA(KP968479752, T6C, T6z); | |
753 T7K = FNMS(KP968479752, T6C, T6z); | |
754 T7G = FMA(KP906616052, T7F, T7E); | |
755 T7I = FNMS(KP906616052, T7F, T7E); | |
756 T6E = FMA(KP906616052, T6D, T6w); | |
757 T6G = FNMS(KP906616052, T6D, T6w); | |
758 T6d = FNMS(KP249506682, T6c, T5H); | |
759 ii[WS(rs, 2)] = FNMS(KP998026728, T7G, T7D); | |
760 ri[WS(rs, 2)] = FMA(KP998026728, T6E, T6p); | |
761 T80 = FNMS(KP614372930, T7W, T7X); | |
762 T7Y = FMA(KP621716863, T7X, T7W); | |
763 T6f = FNMS(KP557913902, T6e, T6d); | |
764 T6n = FMA(KP557913902, T6e, T6d); | |
765 T7T = FMA(KP249506682, T7S, T7P); | |
766 ri[WS(rs, 18)] = FNMS(KP949179823, T6o, T6n); | |
767 ri[WS(rs, 13)] = FMA(KP949179823, T6o, T6n); | |
768 ri[WS(rs, 8)] = FMA(KP943557151, T6m, T6f); | |
769 ri[WS(rs, 23)] = FNMS(KP943557151, T6m, T6f); | |
770 T7Z = FNMS(KP557913902, T7U, T7T); | |
771 T7V = FMA(KP557913902, T7U, T7T); | |
772 } | |
773 T6Q = FNMS(KP560319534, T6K, T6N); | |
774 T6O = FMA(KP681693190, T6N, T6K); | |
775 ii[WS(rs, 23)] = FMA(KP943557151, T7Y, T7V); | |
776 ii[WS(rs, 8)] = FNMS(KP943557151, T7Y, T7V); | |
777 ii[WS(rs, 13)] = FMA(KP949179823, T80, T7Z); | |
778 ii[WS(rs, 18)] = FNMS(KP949179823, T80, T7Z); | |
779 T6F = FNMS(KP249506682, T6E, T6p); | |
780 T7O = FNMS(KP560319534, T7K, T7L); | |
781 T7M = FMA(KP681693190, T7L, T7K); | |
782 T7H = FMA(KP249506682, T7G, T7D); | |
783 T6P = FMA(KP557913902, T6G, T6F); | |
784 T6H = FNMS(KP557913902, T6G, T6F); | |
785 } | |
786 } | |
787 } | |
788 } | |
789 } | |
790 } | |
791 ri[WS(rs, 12)] = FNMS(KP949179823, T6Q, T6P); | |
792 ri[WS(rs, 17)] = FMA(KP949179823, T6Q, T6P); | |
793 ri[WS(rs, 7)] = FMA(KP860541664, T6O, T6H); | |
794 ri[WS(rs, 22)] = FNMS(KP860541664, T6O, T6H); | |
795 T7J = FMA(KP557913902, T7I, T7H); | |
796 T7N = FNMS(KP557913902, T7I, T7H); | |
797 ii[WS(rs, 12)] = FNMS(KP949179823, T7O, T7N); | |
798 ii[WS(rs, 17)] = FMA(KP949179823, T7O, T7N); | |
799 ii[WS(rs, 22)] = FNMS(KP860541664, T7M, T7J); | |
800 ii[WS(rs, 7)] = FMA(KP860541664, T7M, T7J); | |
801 } | |
802 } | |
803 } | |
804 | |
805 static const tw_instr twinstr[] = { | |
806 {TW_FULL, 0, 25}, | |
807 {TW_NEXT, 1, 0} | |
808 }; | |
809 | |
810 static const ct_desc desc = { 25, "t1_25", twinstr, &GENUS, {84, 48, 316, 0}, 0, 0, 0 }; | |
811 | |
812 void X(codelet_t1_25) (planner *p) { | |
813 X(kdft_dit_register) (p, t1_25, &desc); | |
814 } | |
815 #else /* HAVE_FMA */ | |
816 | |
817 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 25 -name t1_25 -include t.h */ | |
818 | |
819 /* | |
820 * This function contains 400 FP additions, 280 FP multiplications, | |
821 * (or, 260 additions, 140 multiplications, 140 fused multiply/add), | |
822 * 101 stack variables, 20 constants, and 100 memory accesses | |
823 */ | |
824 #include "t.h" | |
825 | |
826 static void t1_25(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms) | |
827 { | |
828 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
829 DK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
830 DK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
831 DK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
832 DK(KP992114701, +0.992114701314477831049793042785778521453036709); | |
833 DK(KP125333233, +0.125333233564304245373118759816508793942918247); | |
834 DK(KP637423989, +0.637423989748689710176712811676016195434917298); | |
835 DK(KP770513242, +0.770513242775789230803009636396177847271667672); | |
836 DK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
837 DK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
838 DK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
839 DK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
840 DK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
841 DK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
842 DK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
843 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
844 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
845 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
846 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
847 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
848 { | |
849 INT m; | |
850 for (m = mb, W = W + (mb * 48); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 48, MAKE_VOLATILE_STRIDE(50, rs)) { | |
851 E T1, T6b, T2l, T6o, To, T2m, T6a, T6p, T6t, T6S, T2u, T4I, T2i, T60, T3O; | |
852 E T5D, T4r, T58, T3Z, T5C, T4q, T5b, TS, T5W, T2G, T5s, T4g, T4M, T2R, T5t; | |
853 E T4h, T4P, T1l, T5X, T33, T5w, T4j, T4W, T3e, T5v, T4k, T4T, T1P, T5Z, T3r; | |
854 E T5z, T4o, T51, T3C, T5A, T4n, T54; | |
855 { | |
856 E T6, T2o, Tb, T2p, Tc, T68, Th, T2r, Tm, T2s, Tn, T69; | |
857 T1 = ri[0]; | |
858 T6b = ii[0]; | |
859 { | |
860 E T3, T5, T2, T4; | |
861 T3 = ri[WS(rs, 5)]; | |
862 T5 = ii[WS(rs, 5)]; | |
863 T2 = W[8]; | |
864 T4 = W[9]; | |
865 T6 = FMA(T2, T3, T4 * T5); | |
866 T2o = FNMS(T4, T3, T2 * T5); | |
867 } | |
868 { | |
869 E T8, Ta, T7, T9; | |
870 T8 = ri[WS(rs, 20)]; | |
871 Ta = ii[WS(rs, 20)]; | |
872 T7 = W[38]; | |
873 T9 = W[39]; | |
874 Tb = FMA(T7, T8, T9 * Ta); | |
875 T2p = FNMS(T9, T8, T7 * Ta); | |
876 } | |
877 Tc = T6 + Tb; | |
878 T68 = T2o + T2p; | |
879 { | |
880 E Te, Tg, Td, Tf; | |
881 Te = ri[WS(rs, 10)]; | |
882 Tg = ii[WS(rs, 10)]; | |
883 Td = W[18]; | |
884 Tf = W[19]; | |
885 Th = FMA(Td, Te, Tf * Tg); | |
886 T2r = FNMS(Tf, Te, Td * Tg); | |
887 } | |
888 { | |
889 E Tj, Tl, Ti, Tk; | |
890 Tj = ri[WS(rs, 15)]; | |
891 Tl = ii[WS(rs, 15)]; | |
892 Ti = W[28]; | |
893 Tk = W[29]; | |
894 Tm = FMA(Ti, Tj, Tk * Tl); | |
895 T2s = FNMS(Tk, Tj, Ti * Tl); | |
896 } | |
897 Tn = Th + Tm; | |
898 T69 = T2r + T2s; | |
899 T2l = KP559016994 * (Tc - Tn); | |
900 T6o = KP559016994 * (T68 - T69); | |
901 To = Tc + Tn; | |
902 T2m = FNMS(KP250000000, To, T1); | |
903 T6a = T68 + T69; | |
904 T6p = FNMS(KP250000000, T6a, T6b); | |
905 { | |
906 E T6r, T6s, T2q, T2t; | |
907 T6r = T6 - Tb; | |
908 T6s = Th - Tm; | |
909 T6t = FMA(KP951056516, T6r, KP587785252 * T6s); | |
910 T6S = FNMS(KP587785252, T6r, KP951056516 * T6s); | |
911 T2q = T2o - T2p; | |
912 T2t = T2r - T2s; | |
913 T2u = FMA(KP951056516, T2q, KP587785252 * T2t); | |
914 T4I = FNMS(KP587785252, T2q, KP951056516 * T2t); | |
915 } | |
916 } | |
917 { | |
918 E T1U, T3S, T3J, T3M, T3X, T3W, T3P, T3Q, T3T, T25, T2g, T2h; | |
919 { | |
920 E T1R, T1T, T1Q, T1S; | |
921 T1R = ri[WS(rs, 3)]; | |
922 T1T = ii[WS(rs, 3)]; | |
923 T1Q = W[4]; | |
924 T1S = W[5]; | |
925 T1U = FMA(T1Q, T1R, T1S * T1T); | |
926 T3S = FNMS(T1S, T1R, T1Q * T1T); | |
927 } | |
928 { | |
929 E T1Z, T3H, T2f, T3L, T24, T3I, T2a, T3K; | |
930 { | |
931 E T1W, T1Y, T1V, T1X; | |
932 T1W = ri[WS(rs, 8)]; | |
933 T1Y = ii[WS(rs, 8)]; | |
934 T1V = W[14]; | |
935 T1X = W[15]; | |
936 T1Z = FMA(T1V, T1W, T1X * T1Y); | |
937 T3H = FNMS(T1X, T1W, T1V * T1Y); | |
938 } | |
939 { | |
940 E T2c, T2e, T2b, T2d; | |
941 T2c = ri[WS(rs, 18)]; | |
942 T2e = ii[WS(rs, 18)]; | |
943 T2b = W[34]; | |
944 T2d = W[35]; | |
945 T2f = FMA(T2b, T2c, T2d * T2e); | |
946 T3L = FNMS(T2d, T2c, T2b * T2e); | |
947 } | |
948 { | |
949 E T21, T23, T20, T22; | |
950 T21 = ri[WS(rs, 23)]; | |
951 T23 = ii[WS(rs, 23)]; | |
952 T20 = W[44]; | |
953 T22 = W[45]; | |
954 T24 = FMA(T20, T21, T22 * T23); | |
955 T3I = FNMS(T22, T21, T20 * T23); | |
956 } | |
957 { | |
958 E T27, T29, T26, T28; | |
959 T27 = ri[WS(rs, 13)]; | |
960 T29 = ii[WS(rs, 13)]; | |
961 T26 = W[24]; | |
962 T28 = W[25]; | |
963 T2a = FMA(T26, T27, T28 * T29); | |
964 T3K = FNMS(T28, T27, T26 * T29); | |
965 } | |
966 T3J = T3H - T3I; | |
967 T3M = T3K - T3L; | |
968 T3X = T2a - T2f; | |
969 T3W = T1Z - T24; | |
970 T3P = T3H + T3I; | |
971 T3Q = T3K + T3L; | |
972 T3T = T3P + T3Q; | |
973 T25 = T1Z + T24; | |
974 T2g = T2a + T2f; | |
975 T2h = T25 + T2g; | |
976 } | |
977 T2i = T1U + T2h; | |
978 T60 = T3S + T3T; | |
979 { | |
980 E T3N, T57, T3G, T56, T3E, T3F; | |
981 T3N = FMA(KP951056516, T3J, KP587785252 * T3M); | |
982 T57 = FNMS(KP587785252, T3J, KP951056516 * T3M); | |
983 T3E = KP559016994 * (T25 - T2g); | |
984 T3F = FNMS(KP250000000, T2h, T1U); | |
985 T3G = T3E + T3F; | |
986 T56 = T3F - T3E; | |
987 T3O = T3G + T3N; | |
988 T5D = T56 + T57; | |
989 T4r = T3G - T3N; | |
990 T58 = T56 - T57; | |
991 } | |
992 { | |
993 E T3Y, T59, T3V, T5a, T3R, T3U; | |
994 T3Y = FMA(KP951056516, T3W, KP587785252 * T3X); | |
995 T59 = FNMS(KP587785252, T3W, KP951056516 * T3X); | |
996 T3R = KP559016994 * (T3P - T3Q); | |
997 T3U = FNMS(KP250000000, T3T, T3S); | |
998 T3V = T3R + T3U; | |
999 T5a = T3U - T3R; | |
1000 T3Z = T3V - T3Y; | |
1001 T5C = T5a - T59; | |
1002 T4q = T3Y + T3V; | |
1003 T5b = T59 + T5a; | |
1004 } | |
1005 } | |
1006 { | |
1007 E Tu, T2K, T2B, T2E, T2P, T2O, T2H, T2I, T2L, TF, TQ, TR; | |
1008 { | |
1009 E Tr, Tt, Tq, Ts; | |
1010 Tr = ri[WS(rs, 1)]; | |
1011 Tt = ii[WS(rs, 1)]; | |
1012 Tq = W[0]; | |
1013 Ts = W[1]; | |
1014 Tu = FMA(Tq, Tr, Ts * Tt); | |
1015 T2K = FNMS(Ts, Tr, Tq * Tt); | |
1016 } | |
1017 { | |
1018 E Tz, T2z, TP, T2D, TE, T2A, TK, T2C; | |
1019 { | |
1020 E Tw, Ty, Tv, Tx; | |
1021 Tw = ri[WS(rs, 6)]; | |
1022 Ty = ii[WS(rs, 6)]; | |
1023 Tv = W[10]; | |
1024 Tx = W[11]; | |
1025 Tz = FMA(Tv, Tw, Tx * Ty); | |
1026 T2z = FNMS(Tx, Tw, Tv * Ty); | |
1027 } | |
1028 { | |
1029 E TM, TO, TL, TN; | |
1030 TM = ri[WS(rs, 16)]; | |
1031 TO = ii[WS(rs, 16)]; | |
1032 TL = W[30]; | |
1033 TN = W[31]; | |
1034 TP = FMA(TL, TM, TN * TO); | |
1035 T2D = FNMS(TN, TM, TL * TO); | |
1036 } | |
1037 { | |
1038 E TB, TD, TA, TC; | |
1039 TB = ri[WS(rs, 21)]; | |
1040 TD = ii[WS(rs, 21)]; | |
1041 TA = W[40]; | |
1042 TC = W[41]; | |
1043 TE = FMA(TA, TB, TC * TD); | |
1044 T2A = FNMS(TC, TB, TA * TD); | |
1045 } | |
1046 { | |
1047 E TH, TJ, TG, TI; | |
1048 TH = ri[WS(rs, 11)]; | |
1049 TJ = ii[WS(rs, 11)]; | |
1050 TG = W[20]; | |
1051 TI = W[21]; | |
1052 TK = FMA(TG, TH, TI * TJ); | |
1053 T2C = FNMS(TI, TH, TG * TJ); | |
1054 } | |
1055 T2B = T2z - T2A; | |
1056 T2E = T2C - T2D; | |
1057 T2P = TK - TP; | |
1058 T2O = Tz - TE; | |
1059 T2H = T2z + T2A; | |
1060 T2I = T2C + T2D; | |
1061 T2L = T2H + T2I; | |
1062 TF = Tz + TE; | |
1063 TQ = TK + TP; | |
1064 TR = TF + TQ; | |
1065 } | |
1066 TS = Tu + TR; | |
1067 T5W = T2K + T2L; | |
1068 { | |
1069 E T2F, T4L, T2y, T4K, T2w, T2x; | |
1070 T2F = FMA(KP951056516, T2B, KP587785252 * T2E); | |
1071 T4L = FNMS(KP587785252, T2B, KP951056516 * T2E); | |
1072 T2w = KP559016994 * (TF - TQ); | |
1073 T2x = FNMS(KP250000000, TR, Tu); | |
1074 T2y = T2w + T2x; | |
1075 T4K = T2x - T2w; | |
1076 T2G = T2y + T2F; | |
1077 T5s = T4K + T4L; | |
1078 T4g = T2y - T2F; | |
1079 T4M = T4K - T4L; | |
1080 } | |
1081 { | |
1082 E T2Q, T4N, T2N, T4O, T2J, T2M; | |
1083 T2Q = FMA(KP951056516, T2O, KP587785252 * T2P); | |
1084 T4N = FNMS(KP587785252, T2O, KP951056516 * T2P); | |
1085 T2J = KP559016994 * (T2H - T2I); | |
1086 T2M = FNMS(KP250000000, T2L, T2K); | |
1087 T2N = T2J + T2M; | |
1088 T4O = T2M - T2J; | |
1089 T2R = T2N - T2Q; | |
1090 T5t = T4O - T4N; | |
1091 T4h = T2Q + T2N; | |
1092 T4P = T4N + T4O; | |
1093 } | |
1094 } | |
1095 { | |
1096 E TX, T37, T2Y, T31, T3c, T3b, T34, T35, T38, T18, T1j, T1k; | |
1097 { | |
1098 E TU, TW, TT, TV; | |
1099 TU = ri[WS(rs, 4)]; | |
1100 TW = ii[WS(rs, 4)]; | |
1101 TT = W[6]; | |
1102 TV = W[7]; | |
1103 TX = FMA(TT, TU, TV * TW); | |
1104 T37 = FNMS(TV, TU, TT * TW); | |
1105 } | |
1106 { | |
1107 E T12, T2W, T1i, T30, T17, T2X, T1d, T2Z; | |
1108 { | |
1109 E TZ, T11, TY, T10; | |
1110 TZ = ri[WS(rs, 9)]; | |
1111 T11 = ii[WS(rs, 9)]; | |
1112 TY = W[16]; | |
1113 T10 = W[17]; | |
1114 T12 = FMA(TY, TZ, T10 * T11); | |
1115 T2W = FNMS(T10, TZ, TY * T11); | |
1116 } | |
1117 { | |
1118 E T1f, T1h, T1e, T1g; | |
1119 T1f = ri[WS(rs, 19)]; | |
1120 T1h = ii[WS(rs, 19)]; | |
1121 T1e = W[36]; | |
1122 T1g = W[37]; | |
1123 T1i = FMA(T1e, T1f, T1g * T1h); | |
1124 T30 = FNMS(T1g, T1f, T1e * T1h); | |
1125 } | |
1126 { | |
1127 E T14, T16, T13, T15; | |
1128 T14 = ri[WS(rs, 24)]; | |
1129 T16 = ii[WS(rs, 24)]; | |
1130 T13 = W[46]; | |
1131 T15 = W[47]; | |
1132 T17 = FMA(T13, T14, T15 * T16); | |
1133 T2X = FNMS(T15, T14, T13 * T16); | |
1134 } | |
1135 { | |
1136 E T1a, T1c, T19, T1b; | |
1137 T1a = ri[WS(rs, 14)]; | |
1138 T1c = ii[WS(rs, 14)]; | |
1139 T19 = W[26]; | |
1140 T1b = W[27]; | |
1141 T1d = FMA(T19, T1a, T1b * T1c); | |
1142 T2Z = FNMS(T1b, T1a, T19 * T1c); | |
1143 } | |
1144 T2Y = T2W - T2X; | |
1145 T31 = T2Z - T30; | |
1146 T3c = T1d - T1i; | |
1147 T3b = T12 - T17; | |
1148 T34 = T2W + T2X; | |
1149 T35 = T2Z + T30; | |
1150 T38 = T34 + T35; | |
1151 T18 = T12 + T17; | |
1152 T1j = T1d + T1i; | |
1153 T1k = T18 + T1j; | |
1154 } | |
1155 T1l = TX + T1k; | |
1156 T5X = T37 + T38; | |
1157 { | |
1158 E T32, T4V, T2V, T4U, T2T, T2U; | |
1159 T32 = FMA(KP951056516, T2Y, KP587785252 * T31); | |
1160 T4V = FNMS(KP587785252, T2Y, KP951056516 * T31); | |
1161 T2T = KP559016994 * (T18 - T1j); | |
1162 T2U = FNMS(KP250000000, T1k, TX); | |
1163 T2V = T2T + T2U; | |
1164 T4U = T2U - T2T; | |
1165 T33 = T2V + T32; | |
1166 T5w = T4U + T4V; | |
1167 T4j = T2V - T32; | |
1168 T4W = T4U - T4V; | |
1169 } | |
1170 { | |
1171 E T3d, T4R, T3a, T4S, T36, T39; | |
1172 T3d = FMA(KP951056516, T3b, KP587785252 * T3c); | |
1173 T4R = FNMS(KP587785252, T3b, KP951056516 * T3c); | |
1174 T36 = KP559016994 * (T34 - T35); | |
1175 T39 = FNMS(KP250000000, T38, T37); | |
1176 T3a = T36 + T39; | |
1177 T4S = T39 - T36; | |
1178 T3e = T3a - T3d; | |
1179 T5v = T4S - T4R; | |
1180 T4k = T3d + T3a; | |
1181 T4T = T4R + T4S; | |
1182 } | |
1183 } | |
1184 { | |
1185 E T1r, T3v, T3m, T3p, T3A, T3z, T3s, T3t, T3w, T1C, T1N, T1O; | |
1186 { | |
1187 E T1o, T1q, T1n, T1p; | |
1188 T1o = ri[WS(rs, 2)]; | |
1189 T1q = ii[WS(rs, 2)]; | |
1190 T1n = W[2]; | |
1191 T1p = W[3]; | |
1192 T1r = FMA(T1n, T1o, T1p * T1q); | |
1193 T3v = FNMS(T1p, T1o, T1n * T1q); | |
1194 } | |
1195 { | |
1196 E T1w, T3k, T1M, T3o, T1B, T3l, T1H, T3n; | |
1197 { | |
1198 E T1t, T1v, T1s, T1u; | |
1199 T1t = ri[WS(rs, 7)]; | |
1200 T1v = ii[WS(rs, 7)]; | |
1201 T1s = W[12]; | |
1202 T1u = W[13]; | |
1203 T1w = FMA(T1s, T1t, T1u * T1v); | |
1204 T3k = FNMS(T1u, T1t, T1s * T1v); | |
1205 } | |
1206 { | |
1207 E T1J, T1L, T1I, T1K; | |
1208 T1J = ri[WS(rs, 17)]; | |
1209 T1L = ii[WS(rs, 17)]; | |
1210 T1I = W[32]; | |
1211 T1K = W[33]; | |
1212 T1M = FMA(T1I, T1J, T1K * T1L); | |
1213 T3o = FNMS(T1K, T1J, T1I * T1L); | |
1214 } | |
1215 { | |
1216 E T1y, T1A, T1x, T1z; | |
1217 T1y = ri[WS(rs, 22)]; | |
1218 T1A = ii[WS(rs, 22)]; | |
1219 T1x = W[42]; | |
1220 T1z = W[43]; | |
1221 T1B = FMA(T1x, T1y, T1z * T1A); | |
1222 T3l = FNMS(T1z, T1y, T1x * T1A); | |
1223 } | |
1224 { | |
1225 E T1E, T1G, T1D, T1F; | |
1226 T1E = ri[WS(rs, 12)]; | |
1227 T1G = ii[WS(rs, 12)]; | |
1228 T1D = W[22]; | |
1229 T1F = W[23]; | |
1230 T1H = FMA(T1D, T1E, T1F * T1G); | |
1231 T3n = FNMS(T1F, T1E, T1D * T1G); | |
1232 } | |
1233 T3m = T3k - T3l; | |
1234 T3p = T3n - T3o; | |
1235 T3A = T1H - T1M; | |
1236 T3z = T1w - T1B; | |
1237 T3s = T3k + T3l; | |
1238 T3t = T3n + T3o; | |
1239 T3w = T3s + T3t; | |
1240 T1C = T1w + T1B; | |
1241 T1N = T1H + T1M; | |
1242 T1O = T1C + T1N; | |
1243 } | |
1244 T1P = T1r + T1O; | |
1245 T5Z = T3v + T3w; | |
1246 { | |
1247 E T3q, T50, T3j, T4Z, T3h, T3i; | |
1248 T3q = FMA(KP951056516, T3m, KP587785252 * T3p); | |
1249 T50 = FNMS(KP587785252, T3m, KP951056516 * T3p); | |
1250 T3h = KP559016994 * (T1C - T1N); | |
1251 T3i = FNMS(KP250000000, T1O, T1r); | |
1252 T3j = T3h + T3i; | |
1253 T4Z = T3i - T3h; | |
1254 T3r = T3j + T3q; | |
1255 T5z = T4Z + T50; | |
1256 T4o = T3j - T3q; | |
1257 T51 = T4Z - T50; | |
1258 } | |
1259 { | |
1260 E T3B, T52, T3y, T53, T3u, T3x; | |
1261 T3B = FMA(KP951056516, T3z, KP587785252 * T3A); | |
1262 T52 = FNMS(KP587785252, T3z, KP951056516 * T3A); | |
1263 T3u = KP559016994 * (T3s - T3t); | |
1264 T3x = FNMS(KP250000000, T3w, T3v); | |
1265 T3y = T3u + T3x; | |
1266 T53 = T3x - T3u; | |
1267 T3C = T3y - T3B; | |
1268 T5A = T53 - T52; | |
1269 T4n = T3B + T3y; | |
1270 T54 = T52 + T53; | |
1271 } | |
1272 } | |
1273 { | |
1274 E T62, T64, Tp, T2k, T5T, T5U, T63, T5V; | |
1275 { | |
1276 E T5Y, T61, T1m, T2j; | |
1277 T5Y = T5W - T5X; | |
1278 T61 = T5Z - T60; | |
1279 T62 = FMA(KP951056516, T5Y, KP587785252 * T61); | |
1280 T64 = FNMS(KP587785252, T5Y, KP951056516 * T61); | |
1281 Tp = T1 + To; | |
1282 T1m = TS + T1l; | |
1283 T2j = T1P + T2i; | |
1284 T2k = T1m + T2j; | |
1285 T5T = KP559016994 * (T1m - T2j); | |
1286 T5U = FNMS(KP250000000, T2k, Tp); | |
1287 } | |
1288 ri[0] = Tp + T2k; | |
1289 T63 = T5U - T5T; | |
1290 ri[WS(rs, 10)] = T63 - T64; | |
1291 ri[WS(rs, 15)] = T63 + T64; | |
1292 T5V = T5T + T5U; | |
1293 ri[WS(rs, 20)] = T5V - T62; | |
1294 ri[WS(rs, 5)] = T5V + T62; | |
1295 } | |
1296 { | |
1297 E T6i, T6j, T6c, T67, T6d, T6e, T6k, T6f; | |
1298 { | |
1299 E T6g, T6h, T65, T66; | |
1300 T6g = TS - T1l; | |
1301 T6h = T1P - T2i; | |
1302 T6i = FMA(KP951056516, T6g, KP587785252 * T6h); | |
1303 T6j = FNMS(KP587785252, T6g, KP951056516 * T6h); | |
1304 T6c = T6a + T6b; | |
1305 T65 = T5W + T5X; | |
1306 T66 = T5Z + T60; | |
1307 T67 = T65 + T66; | |
1308 T6d = KP559016994 * (T65 - T66); | |
1309 T6e = FNMS(KP250000000, T67, T6c); | |
1310 } | |
1311 ii[0] = T67 + T6c; | |
1312 T6k = T6e - T6d; | |
1313 ii[WS(rs, 10)] = T6j + T6k; | |
1314 ii[WS(rs, 15)] = T6k - T6j; | |
1315 T6f = T6d + T6e; | |
1316 ii[WS(rs, 5)] = T6f - T6i; | |
1317 ii[WS(rs, 20)] = T6i + T6f; | |
1318 } | |
1319 { | |
1320 E T2v, T4f, T6u, T6G, T42, T6z, T43, T6y, T4A, T6H, T4D, T6F, T4u, T6L, T4v; | |
1321 E T6K, T48, T6v, T4b, T6n, T2n, T6q; | |
1322 T2n = T2l + T2m; | |
1323 T2v = T2n + T2u; | |
1324 T4f = T2n - T2u; | |
1325 T6q = T6o + T6p; | |
1326 T6u = T6q - T6t; | |
1327 T6G = T6t + T6q; | |
1328 { | |
1329 E T2S, T3f, T3g, T3D, T40, T41; | |
1330 T2S = FMA(KP968583161, T2G, KP248689887 * T2R); | |
1331 T3f = FMA(KP535826794, T33, KP844327925 * T3e); | |
1332 T3g = T2S + T3f; | |
1333 T3D = FMA(KP876306680, T3r, KP481753674 * T3C); | |
1334 T40 = FMA(KP728968627, T3O, KP684547105 * T3Z); | |
1335 T41 = T3D + T40; | |
1336 T42 = T3g + T41; | |
1337 T6z = T3D - T40; | |
1338 T43 = KP559016994 * (T3g - T41); | |
1339 T6y = T2S - T3f; | |
1340 } | |
1341 { | |
1342 E T4y, T4z, T6D, T4B, T4C, T6E; | |
1343 T4y = FNMS(KP844327925, T4g, KP535826794 * T4h); | |
1344 T4z = FNMS(KP637423989, T4k, KP770513242 * T4j); | |
1345 T6D = T4y + T4z; | |
1346 T4B = FMA(KP125333233, T4r, KP992114701 * T4q); | |
1347 T4C = FMA(KP904827052, T4o, KP425779291 * T4n); | |
1348 T6E = T4C + T4B; | |
1349 T4A = T4y - T4z; | |
1350 T6H = KP559016994 * (T6D + T6E); | |
1351 T4D = T4B - T4C; | |
1352 T6F = T6D - T6E; | |
1353 } | |
1354 { | |
1355 E T4i, T4l, T4m, T4p, T4s, T4t; | |
1356 T4i = FMA(KP535826794, T4g, KP844327925 * T4h); | |
1357 T4l = FMA(KP637423989, T4j, KP770513242 * T4k); | |
1358 T4m = T4i - T4l; | |
1359 T4p = FNMS(KP425779291, T4o, KP904827052 * T4n); | |
1360 T4s = FNMS(KP992114701, T4r, KP125333233 * T4q); | |
1361 T4t = T4p + T4s; | |
1362 T4u = T4m + T4t; | |
1363 T6L = T4p - T4s; | |
1364 T4v = KP559016994 * (T4m - T4t); | |
1365 T6K = T4i + T4l; | |
1366 } | |
1367 { | |
1368 E T46, T47, T6l, T49, T4a, T6m; | |
1369 T46 = FNMS(KP248689887, T2G, KP968583161 * T2R); | |
1370 T47 = FNMS(KP844327925, T33, KP535826794 * T3e); | |
1371 T6l = T46 + T47; | |
1372 T49 = FNMS(KP481753674, T3r, KP876306680 * T3C); | |
1373 T4a = FNMS(KP684547105, T3O, KP728968627 * T3Z); | |
1374 T6m = T49 + T4a; | |
1375 T48 = T46 - T47; | |
1376 T6v = KP559016994 * (T6l - T6m); | |
1377 T4b = T49 - T4a; | |
1378 T6n = T6l + T6m; | |
1379 } | |
1380 ri[WS(rs, 1)] = T2v + T42; | |
1381 ii[WS(rs, 1)] = T6n + T6u; | |
1382 ri[WS(rs, 4)] = T4f + T4u; | |
1383 ii[WS(rs, 4)] = T6F + T6G; | |
1384 { | |
1385 E T4c, T4e, T45, T4d, T44; | |
1386 T4c = FMA(KP951056516, T48, KP587785252 * T4b); | |
1387 T4e = FNMS(KP587785252, T48, KP951056516 * T4b); | |
1388 T44 = FNMS(KP250000000, T42, T2v); | |
1389 T45 = T43 + T44; | |
1390 T4d = T44 - T43; | |
1391 ri[WS(rs, 21)] = T45 - T4c; | |
1392 ri[WS(rs, 16)] = T4d + T4e; | |
1393 ri[WS(rs, 6)] = T45 + T4c; | |
1394 ri[WS(rs, 11)] = T4d - T4e; | |
1395 } | |
1396 { | |
1397 E T6A, T6B, T6x, T6C, T6w; | |
1398 T6A = FMA(KP951056516, T6y, KP587785252 * T6z); | |
1399 T6B = FNMS(KP587785252, T6y, KP951056516 * T6z); | |
1400 T6w = FNMS(KP250000000, T6n, T6u); | |
1401 T6x = T6v + T6w; | |
1402 T6C = T6w - T6v; | |
1403 ii[WS(rs, 6)] = T6x - T6A; | |
1404 ii[WS(rs, 16)] = T6C - T6B; | |
1405 ii[WS(rs, 21)] = T6A + T6x; | |
1406 ii[WS(rs, 11)] = T6B + T6C; | |
1407 } | |
1408 { | |
1409 E T4E, T4G, T4x, T4F, T4w; | |
1410 T4E = FMA(KP951056516, T4A, KP587785252 * T4D); | |
1411 T4G = FNMS(KP587785252, T4A, KP951056516 * T4D); | |
1412 T4w = FNMS(KP250000000, T4u, T4f); | |
1413 T4x = T4v + T4w; | |
1414 T4F = T4w - T4v; | |
1415 ri[WS(rs, 24)] = T4x - T4E; | |
1416 ri[WS(rs, 19)] = T4F + T4G; | |
1417 ri[WS(rs, 9)] = T4x + T4E; | |
1418 ri[WS(rs, 14)] = T4F - T4G; | |
1419 } | |
1420 { | |
1421 E T6M, T6N, T6J, T6O, T6I; | |
1422 T6M = FMA(KP951056516, T6K, KP587785252 * T6L); | |
1423 T6N = FNMS(KP587785252, T6K, KP951056516 * T6L); | |
1424 T6I = FNMS(KP250000000, T6F, T6G); | |
1425 T6J = T6H + T6I; | |
1426 T6O = T6I - T6H; | |
1427 ii[WS(rs, 9)] = T6J - T6M; | |
1428 ii[WS(rs, 19)] = T6O - T6N; | |
1429 ii[WS(rs, 24)] = T6M + T6J; | |
1430 ii[WS(rs, 14)] = T6N + T6O; | |
1431 } | |
1432 } | |
1433 { | |
1434 E T4J, T5r, T6U, T76, T5e, T6Z, T5f, T6Y, T5M, T77, T5P, T75, T5G, T7b, T5H; | |
1435 E T7a, T5k, T6V, T5n, T6R, T4H, T6T; | |
1436 T4H = T2m - T2l; | |
1437 T4J = T4H - T4I; | |
1438 T5r = T4H + T4I; | |
1439 T6T = T6p - T6o; | |
1440 T6U = T6S + T6T; | |
1441 T76 = T6T - T6S; | |
1442 { | |
1443 E T4Q, T4X, T4Y, T55, T5c, T5d; | |
1444 T4Q = FMA(KP876306680, T4M, KP481753674 * T4P); | |
1445 T4X = FNMS(KP425779291, T4W, KP904827052 * T4T); | |
1446 T4Y = T4Q + T4X; | |
1447 T55 = FMA(KP535826794, T51, KP844327925 * T54); | |
1448 T5c = FMA(KP062790519, T58, KP998026728 * T5b); | |
1449 T5d = T55 + T5c; | |
1450 T5e = T4Y + T5d; | |
1451 T6Z = T55 - T5c; | |
1452 T5f = KP559016994 * (T4Y - T5d); | |
1453 T6Y = T4Q - T4X; | |
1454 } | |
1455 { | |
1456 E T5K, T5L, T73, T5N, T5O, T74; | |
1457 T5K = FNMS(KP684547105, T5s, KP728968627 * T5t); | |
1458 T5L = FMA(KP125333233, T5w, KP992114701 * T5v); | |
1459 T73 = T5K - T5L; | |
1460 T5N = FNMS(KP998026728, T5z, KP062790519 * T5A); | |
1461 T5O = FMA(KP770513242, T5D, KP637423989 * T5C); | |
1462 T74 = T5N - T5O; | |
1463 T5M = T5K + T5L; | |
1464 T77 = KP559016994 * (T73 - T74); | |
1465 T5P = T5N + T5O; | |
1466 T75 = T73 + T74; | |
1467 } | |
1468 { | |
1469 E T5u, T5x, T5y, T5B, T5E, T5F; | |
1470 T5u = FMA(KP728968627, T5s, KP684547105 * T5t); | |
1471 T5x = FNMS(KP992114701, T5w, KP125333233 * T5v); | |
1472 T5y = T5u + T5x; | |
1473 T5B = FMA(KP062790519, T5z, KP998026728 * T5A); | |
1474 T5E = FNMS(KP637423989, T5D, KP770513242 * T5C); | |
1475 T5F = T5B + T5E; | |
1476 T5G = T5y + T5F; | |
1477 T7b = T5B - T5E; | |
1478 T5H = KP559016994 * (T5y - T5F); | |
1479 T7a = T5u - T5x; | |
1480 } | |
1481 { | |
1482 E T5i, T5j, T6P, T5l, T5m, T6Q; | |
1483 T5i = FNMS(KP481753674, T4M, KP876306680 * T4P); | |
1484 T5j = FMA(KP904827052, T4W, KP425779291 * T4T); | |
1485 T6P = T5i - T5j; | |
1486 T5l = FNMS(KP844327925, T51, KP535826794 * T54); | |
1487 T5m = FNMS(KP998026728, T58, KP062790519 * T5b); | |
1488 T6Q = T5l + T5m; | |
1489 T5k = T5i + T5j; | |
1490 T6V = KP559016994 * (T6P - T6Q); | |
1491 T5n = T5l - T5m; | |
1492 T6R = T6P + T6Q; | |
1493 } | |
1494 ri[WS(rs, 2)] = T4J + T5e; | |
1495 ii[WS(rs, 2)] = T6R + T6U; | |
1496 ri[WS(rs, 3)] = T5r + T5G; | |
1497 ii[WS(rs, 3)] = T75 + T76; | |
1498 { | |
1499 E T5o, T5q, T5h, T5p, T5g; | |
1500 T5o = FMA(KP951056516, T5k, KP587785252 * T5n); | |
1501 T5q = FNMS(KP587785252, T5k, KP951056516 * T5n); | |
1502 T5g = FNMS(KP250000000, T5e, T4J); | |
1503 T5h = T5f + T5g; | |
1504 T5p = T5g - T5f; | |
1505 ri[WS(rs, 22)] = T5h - T5o; | |
1506 ri[WS(rs, 17)] = T5p + T5q; | |
1507 ri[WS(rs, 7)] = T5h + T5o; | |
1508 ri[WS(rs, 12)] = T5p - T5q; | |
1509 } | |
1510 { | |
1511 E T70, T71, T6X, T72, T6W; | |
1512 T70 = FMA(KP951056516, T6Y, KP587785252 * T6Z); | |
1513 T71 = FNMS(KP587785252, T6Y, KP951056516 * T6Z); | |
1514 T6W = FNMS(KP250000000, T6R, T6U); | |
1515 T6X = T6V + T6W; | |
1516 T72 = T6W - T6V; | |
1517 ii[WS(rs, 7)] = T6X - T70; | |
1518 ii[WS(rs, 17)] = T72 - T71; | |
1519 ii[WS(rs, 22)] = T70 + T6X; | |
1520 ii[WS(rs, 12)] = T71 + T72; | |
1521 } | |
1522 { | |
1523 E T5Q, T5S, T5J, T5R, T5I; | |
1524 T5Q = FMA(KP951056516, T5M, KP587785252 * T5P); | |
1525 T5S = FNMS(KP587785252, T5M, KP951056516 * T5P); | |
1526 T5I = FNMS(KP250000000, T5G, T5r); | |
1527 T5J = T5H + T5I; | |
1528 T5R = T5I - T5H; | |
1529 ri[WS(rs, 23)] = T5J - T5Q; | |
1530 ri[WS(rs, 18)] = T5R + T5S; | |
1531 ri[WS(rs, 8)] = T5J + T5Q; | |
1532 ri[WS(rs, 13)] = T5R - T5S; | |
1533 } | |
1534 { | |
1535 E T7c, T7d, T79, T7e, T78; | |
1536 T7c = FMA(KP951056516, T7a, KP587785252 * T7b); | |
1537 T7d = FNMS(KP587785252, T7a, KP951056516 * T7b); | |
1538 T78 = FNMS(KP250000000, T75, T76); | |
1539 T79 = T77 + T78; | |
1540 T7e = T78 - T77; | |
1541 ii[WS(rs, 8)] = T79 - T7c; | |
1542 ii[WS(rs, 18)] = T7e - T7d; | |
1543 ii[WS(rs, 23)] = T7c + T79; | |
1544 ii[WS(rs, 13)] = T7d + T7e; | |
1545 } | |
1546 } | |
1547 } | |
1548 } | |
1549 } | |
1550 | |
1551 static const tw_instr twinstr[] = { | |
1552 {TW_FULL, 0, 25}, | |
1553 {TW_NEXT, 1, 0} | |
1554 }; | |
1555 | |
1556 static const ct_desc desc = { 25, "t1_25", twinstr, &GENUS, {260, 140, 140, 0}, 0, 0, 0 }; | |
1557 | |
1558 void X(codelet_t1_25) (planner *p) { | |
1559 X(kdft_dit_register) (p, t1_25, &desc); | |
1560 } | |
1561 #endif /* HAVE_FMA */ |