Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.5/rdft/scalar/r2cb/r2cb_128.c @ 42:2cd0e3b3e1fd
Current fftw source
author | Chris Cannam |
---|---|
date | Tue, 18 Oct 2016 13:40:26 +0100 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
41:481f5f8c5634 | 42:2cd0e3b3e1fd |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-14 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-14 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 Sat Jul 30 16:49:32 EDT 2016 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include r2cb.h */ | |
29 | |
30 /* | |
31 * This function contains 956 FP additions, 540 FP multiplications, | |
32 * (or, 416 additions, 0 multiplications, 540 fused multiply/add), | |
33 * 242 stack variables, 36 constants, and 256 memory accesses | |
34 */ | |
35 #include "r2cb.h" | |
36 | |
37 static void r2cb_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599); | |
40 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854); | |
41 DK(KP599376933, +0.599376933681923766271389869014404232837890546); | |
42 DK(KP741650546, +0.741650546272035369581266691172079863842265220); | |
43 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067); | |
44 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924); | |
45 DK(KP148335987, +0.148335987538347428753676511486911367000625355); | |
46 DK(KP250486960, +0.250486960191305461595702160124721208578685568); | |
47 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
48 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050); | |
49 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619); | |
50 DK(KP472964775, +0.472964775891319928124438237972992463904131113); | |
51 DK(KP906347169, +0.906347169019147157946142717268914412664134293); | |
52 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229); | |
53 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592); | |
54 DK(KP049126849, +0.049126849769467254105343321271313617079695752); | |
55 DK(KP357805721, +0.357805721314524104672487743774474392487532769); | |
56 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
57 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241); | |
58 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411); | |
59 DK(KP534511135, +0.534511135950791641089685961295362908582039528); | |
60 DK(KP303346683, +0.303346683607342391675883946941299872384187453); | |
61 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624); | |
62 DK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
63 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737); | |
64 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586); | |
65 DK(KP098491403, +0.098491403357164253077197521291327432293052451); | |
66 DK(KP820678790, +0.820678790828660330972281985331011598767386482); | |
67 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462); | |
68 DK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
69 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
70 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252); | |
71 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
72 DK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
73 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875); | |
74 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
75 { | |
76 INT i; | |
77 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) { | |
78 E T9H, T9I, T9X, T9Y; | |
79 { | |
80 E Tdr, T9, Tcl, Ta9, T6b, T2d, T91, T7j, Tg, Tds, Tcm, Tae, T92, T7m, T6c; | |
81 E T2o, Tdu, Tw, Tco, Tap, TeM, Tdx, T6f, T2G, T6e, T2P, T94, T7t, Tcp, Tak; | |
82 E T95, T7q, TdM, T1i, TcL, TbD, Tf0, Te6, T6q, T42, T6B, T5t, T9r, T8j, TcA; | |
83 E TaY, T9g, T7S, TdA, TM, Tcv, TaN, TeP, TdI, T6i, T38, T6l, T3F, T9b, T7J; | |
84 E Tcs, Taw, T98, T7y, T1N, TeW, T6x, T4H, Te8, TdV, T6w, T4Q, T9j, T86, TcO; | |
85 E TcI, T9k, T83, TbI, Tbl, T22, TeV, Te0, Te9, T58, T6u, T6t, T5h, T9m, T8d; | |
86 E TcP, TcF, T9n, T8a, TbJ, Tbw, Te3, T1x, TcB, TbG, Tf1, TdP, T6C, T4p, T6r; | |
87 E T5w, T9h, T8m, TcM, Tb9, T9s, T7Z, TaB, TaG, TdF, T11, Tct, TaQ, TeQ, TdD; | |
88 E T6m, T3v, T7B, T7E, T6j, T3I, T99, T7M; | |
89 { | |
90 E TaU, TaX, T7Q, T7R, Tbk, Tbf; | |
91 { | |
92 E Td, T2e, Tc, Tab, T2m, Te, T2f, T2g; | |
93 { | |
94 E T7h, T27, T2c, T7i; | |
95 { | |
96 E T4, T26, T29, T25, T3, T28, T8, T2a; | |
97 T4 = Cr[WS(csr, 32)]; | |
98 T26 = Ci[WS(csi, 32)]; | |
99 { | |
100 E T1, T2, T6, T7; | |
101 T1 = Cr[0]; | |
102 T2 = Cr[WS(csr, 64)]; | |
103 T6 = Cr[WS(csr, 16)]; | |
104 T7 = Cr[WS(csr, 48)]; | |
105 T29 = Ci[WS(csi, 16)]; | |
106 T25 = T1 - T2; | |
107 T3 = T1 + T2; | |
108 T28 = T6 - T7; | |
109 T8 = T6 + T7; | |
110 T2a = Ci[WS(csi, 48)]; | |
111 } | |
112 { | |
113 E Ta7, T5, Ta8, T2b; | |
114 Ta7 = FNMS(KP2_000000000, T4, T3); | |
115 T5 = FMA(KP2_000000000, T4, T3); | |
116 T7h = FMA(KP2_000000000, T26, T25); | |
117 T27 = FNMS(KP2_000000000, T26, T25); | |
118 Ta8 = T29 - T2a; | |
119 T2b = T29 + T2a; | |
120 Tdr = FNMS(KP2_000000000, T8, T5); | |
121 T9 = FMA(KP2_000000000, T8, T5); | |
122 Tcl = FMA(KP2_000000000, Ta8, Ta7); | |
123 Ta9 = FNMS(KP2_000000000, Ta8, Ta7); | |
124 T2c = T28 - T2b; | |
125 T7i = T28 + T2b; | |
126 } | |
127 } | |
128 { | |
129 E Ta, Tb, T2k, T2l; | |
130 Ta = Cr[WS(csr, 8)]; | |
131 T6b = FNMS(KP1_414213562, T2c, T27); | |
132 T2d = FMA(KP1_414213562, T2c, T27); | |
133 T91 = FMA(KP1_414213562, T7i, T7h); | |
134 T7j = FNMS(KP1_414213562, T7i, T7h); | |
135 Tb = Cr[WS(csr, 56)]; | |
136 T2k = Ci[WS(csi, 8)]; | |
137 T2l = Ci[WS(csi, 56)]; | |
138 Td = Cr[WS(csr, 40)]; | |
139 T2e = Ta - Tb; | |
140 Tc = Ta + Tb; | |
141 Tab = T2k - T2l; | |
142 T2m = T2k + T2l; | |
143 Te = Cr[WS(csr, 24)]; | |
144 T2f = Ci[WS(csi, 40)]; | |
145 T2g = Ci[WS(csi, 24)]; | |
146 } | |
147 } | |
148 { | |
149 E Tag, Taj, T7o, T7p; | |
150 { | |
151 E T2q, Tk, Tam, T2K, T2H, Tn, Tan, T2t, Tu, Tah, T2E, T2N, Tr, T2v, T2y; | |
152 E Tai; | |
153 { | |
154 E Tl, Tm, T2r, T2s; | |
155 { | |
156 E Ti, Tj, T2j, Tf, T2I, T2J; | |
157 Ti = Cr[WS(csr, 4)]; | |
158 T2j = Td - Te; | |
159 Tf = Td + Te; | |
160 { | |
161 E Tac, T2h, T7k, T2n; | |
162 Tac = T2f - T2g; | |
163 T2h = T2f + T2g; | |
164 T7k = T2m - T2j; | |
165 T2n = T2j + T2m; | |
166 { | |
167 E Taa, Tad, T7l, T2i; | |
168 Taa = Tc - Tf; | |
169 Tg = Tc + Tf; | |
170 Tad = Tab - Tac; | |
171 Tds = Tac + Tab; | |
172 T7l = T2e + T2h; | |
173 T2i = T2e - T2h; | |
174 Tcm = Taa + Tad; | |
175 Tae = Taa - Tad; | |
176 T92 = FMA(KP414213562, T7k, T7l); | |
177 T7m = FNMS(KP414213562, T7l, T7k); | |
178 T6c = FMA(KP414213562, T2i, T2n); | |
179 T2o = FNMS(KP414213562, T2n, T2i); | |
180 Tj = Cr[WS(csr, 60)]; | |
181 } | |
182 } | |
183 T2I = Ci[WS(csi, 4)]; | |
184 T2J = Ci[WS(csi, 60)]; | |
185 Tl = Cr[WS(csr, 36)]; | |
186 T2q = Ti - Tj; | |
187 Tk = Ti + Tj; | |
188 Tam = T2I - T2J; | |
189 T2K = T2I + T2J; | |
190 Tm = Cr[WS(csr, 28)]; | |
191 } | |
192 T2r = Ci[WS(csi, 36)]; | |
193 T2s = Ci[WS(csi, 28)]; | |
194 { | |
195 E Ts, Tt, T2B, T2C; | |
196 Ts = Cr[WS(csr, 12)]; | |
197 T2H = Tl - Tm; | |
198 Tn = Tl + Tm; | |
199 Tan = T2r - T2s; | |
200 T2t = T2r + T2s; | |
201 Tt = Cr[WS(csr, 52)]; | |
202 T2B = Ci[WS(csi, 12)]; | |
203 T2C = Ci[WS(csi, 52)]; | |
204 { | |
205 E Tp, T2A, T2D, Tq, T2w, T2x; | |
206 Tp = Cr[WS(csr, 20)]; | |
207 Tu = Ts + Tt; | |
208 T2A = Ts - Tt; | |
209 Tah = T2C - T2B; | |
210 T2D = T2B + T2C; | |
211 Tq = Cr[WS(csr, 44)]; | |
212 T2w = Ci[WS(csi, 20)]; | |
213 T2x = Ci[WS(csi, 44)]; | |
214 T2E = T2A - T2D; | |
215 T2N = T2A + T2D; | |
216 Tr = Tp + Tq; | |
217 T2v = Tp - Tq; | |
218 T2y = T2w + T2x; | |
219 Tai = T2w - T2x; | |
220 } | |
221 } | |
222 } | |
223 { | |
224 E T2M, Tdv, Tdw, T2u, T2F, T7s, T7r, T2L, T2O; | |
225 { | |
226 E To, T2z, Tv, Tal, Tao; | |
227 Tag = Tk - Tn; | |
228 To = Tk + Tn; | |
229 T2M = T2v + T2y; | |
230 T2z = T2v - T2y; | |
231 Tv = Tr + Tu; | |
232 Tal = Tr - Tu; | |
233 Tao = Tam - Tan; | |
234 Tdv = Tan + Tam; | |
235 Tdu = To - Tv; | |
236 Tw = To + Tv; | |
237 Tco = Tao - Tal; | |
238 Tap = Tal + Tao; | |
239 Tdw = Tai + Tah; | |
240 Taj = Tah - Tai; | |
241 T7o = T2q + T2t; | |
242 T2u = T2q - T2t; | |
243 T2F = T2z + T2E; | |
244 T7s = T2E - T2z; | |
245 } | |
246 T7r = T2K - T2H; | |
247 T2L = T2H + T2K; | |
248 TeM = Tdw + Tdv; | |
249 Tdx = Tdv - Tdw; | |
250 T6f = FNMS(KP707106781, T2F, T2u); | |
251 T2G = FMA(KP707106781, T2F, T2u); | |
252 T2O = T2M - T2N; | |
253 T7p = T2M + T2N; | |
254 T6e = FNMS(KP707106781, T2O, T2L); | |
255 T2P = FMA(KP707106781, T2O, T2L); | |
256 T94 = FMA(KP707106781, T7s, T7r); | |
257 T7t = FNMS(KP707106781, T7s, T7r); | |
258 } | |
259 } | |
260 { | |
261 E T3M, T16, TbA, T5o, T5l, T19, TbB, T3P, T1g, TaV, T40, T5r, T1d, T3R, T3U; | |
262 E TaW; | |
263 { | |
264 E T17, T18, T3N, T3O; | |
265 { | |
266 E T14, T15, T5m, T5n; | |
267 T14 = Cr[WS(csr, 1)]; | |
268 Tcp = Tag - Taj; | |
269 Tak = Tag + Taj; | |
270 T95 = FMA(KP707106781, T7p, T7o); | |
271 T7q = FNMS(KP707106781, T7p, T7o); | |
272 T15 = Cr[WS(csr, 63)]; | |
273 T5m = Ci[WS(csi, 1)]; | |
274 T5n = Ci[WS(csi, 63)]; | |
275 T17 = Cr[WS(csr, 33)]; | |
276 T3M = T14 - T15; | |
277 T16 = T14 + T15; | |
278 TbA = T5m - T5n; | |
279 T5o = T5m + T5n; | |
280 T18 = Cr[WS(csr, 31)]; | |
281 } | |
282 T3N = Ci[WS(csi, 33)]; | |
283 T3O = Ci[WS(csi, 31)]; | |
284 { | |
285 E T1e, T1f, T3X, T3Y; | |
286 T1e = Cr[WS(csr, 15)]; | |
287 T5l = T17 - T18; | |
288 T19 = T17 + T18; | |
289 TbB = T3N - T3O; | |
290 T3P = T3N + T3O; | |
291 T1f = Cr[WS(csr, 49)]; | |
292 T3X = Ci[WS(csi, 15)]; | |
293 T3Y = Ci[WS(csi, 49)]; | |
294 { | |
295 E T1b, T3W, T3Z, T1c, T3S, T3T; | |
296 T1b = Cr[WS(csr, 17)]; | |
297 T1g = T1e + T1f; | |
298 T3W = T1e - T1f; | |
299 TaV = T3Y - T3X; | |
300 T3Z = T3X + T3Y; | |
301 T1c = Cr[WS(csr, 47)]; | |
302 T3S = Ci[WS(csi, 17)]; | |
303 T3T = Ci[WS(csi, 47)]; | |
304 T40 = T3W - T3Z; | |
305 T5r = T3W + T3Z; | |
306 T1d = T1b + T1c; | |
307 T3R = T1b - T1c; | |
308 T3U = T3S + T3T; | |
309 TaW = T3S - T3T; | |
310 } | |
311 } | |
312 } | |
313 { | |
314 E T5q, Te4, Te5, T3Q, T41, T8i, T8h, T5p, T5s; | |
315 { | |
316 E T1a, T3V, T1h, Tbz, TbC; | |
317 TaU = T16 - T19; | |
318 T1a = T16 + T19; | |
319 T5q = T3R + T3U; | |
320 T3V = T3R - T3U; | |
321 T1h = T1d + T1g; | |
322 Tbz = T1d - T1g; | |
323 TbC = TbA - TbB; | |
324 Te4 = TbB + TbA; | |
325 TdM = T1a - T1h; | |
326 T1i = T1a + T1h; | |
327 TcL = TbC - Tbz; | |
328 TbD = Tbz + TbC; | |
329 Te5 = TaW + TaV; | |
330 TaX = TaV - TaW; | |
331 T7Q = T3M + T3P; | |
332 T3Q = T3M - T3P; | |
333 T41 = T3V + T40; | |
334 T8i = T40 - T3V; | |
335 } | |
336 T8h = T5o - T5l; | |
337 T5p = T5l + T5o; | |
338 Tf0 = Te5 + Te4; | |
339 Te6 = Te4 - Te5; | |
340 T6q = FNMS(KP707106781, T41, T3Q); | |
341 T42 = FMA(KP707106781, T41, T3Q); | |
342 T5s = T5q - T5r; | |
343 T7R = T5q + T5r; | |
344 T6B = FNMS(KP707106781, T5s, T5p); | |
345 T5t = FMA(KP707106781, T5s, T5p); | |
346 T9r = FMA(KP707106781, T8i, T8h); | |
347 T8j = FNMS(KP707106781, T8i, T8h); | |
348 } | |
349 } | |
350 } | |
351 } | |
352 { | |
353 E Tas, Tav, T7w, T7x; | |
354 { | |
355 E T2S, TA, TaK, T3A, T3x, TD, TaL, T2V, TK, Tat, T36, T3D, TH, T2X, T30; | |
356 E Tau; | |
357 { | |
358 E TB, TC, T2T, T2U; | |
359 { | |
360 E Ty, Tz, T3y, T3z; | |
361 Ty = Cr[WS(csr, 2)]; | |
362 TcA = TaU - TaX; | |
363 TaY = TaU + TaX; | |
364 T9g = FMA(KP707106781, T7R, T7Q); | |
365 T7S = FNMS(KP707106781, T7R, T7Q); | |
366 Tz = Cr[WS(csr, 62)]; | |
367 T3y = Ci[WS(csi, 2)]; | |
368 T3z = Ci[WS(csi, 62)]; | |
369 TB = Cr[WS(csr, 34)]; | |
370 T2S = Ty - Tz; | |
371 TA = Ty + Tz; | |
372 TaK = T3y - T3z; | |
373 T3A = T3y + T3z; | |
374 TC = Cr[WS(csr, 30)]; | |
375 } | |
376 T2T = Ci[WS(csi, 34)]; | |
377 T2U = Ci[WS(csi, 30)]; | |
378 { | |
379 E TI, TJ, T33, T34; | |
380 TI = Cr[WS(csr, 14)]; | |
381 T3x = TB - TC; | |
382 TD = TB + TC; | |
383 TaL = T2T - T2U; | |
384 T2V = T2T + T2U; | |
385 TJ = Cr[WS(csr, 50)]; | |
386 T33 = Ci[WS(csi, 14)]; | |
387 T34 = Ci[WS(csi, 50)]; | |
388 { | |
389 E TF, T32, T35, TG, T2Y, T2Z; | |
390 TF = Cr[WS(csr, 18)]; | |
391 TK = TI + TJ; | |
392 T32 = TI - TJ; | |
393 Tat = T34 - T33; | |
394 T35 = T33 + T34; | |
395 TG = Cr[WS(csr, 46)]; | |
396 T2Y = Ci[WS(csi, 18)]; | |
397 T2Z = Ci[WS(csi, 46)]; | |
398 T36 = T32 - T35; | |
399 T3D = T32 + T35; | |
400 TH = TF + TG; | |
401 T2X = TF - TG; | |
402 T30 = T2Y + T2Z; | |
403 Tau = T2Y - T2Z; | |
404 } | |
405 } | |
406 } | |
407 { | |
408 E T3C, TdG, TdH, T2W, T37, T7I, T7H, T3B, T3E; | |
409 { | |
410 E TE, T31, TL, TaJ, TaM; | |
411 Tas = TA - TD; | |
412 TE = TA + TD; | |
413 T3C = T2X + T30; | |
414 T31 = T2X - T30; | |
415 TL = TH + TK; | |
416 TaJ = TH - TK; | |
417 TaM = TaK - TaL; | |
418 TdG = TaL + TaK; | |
419 TdA = TE - TL; | |
420 TM = TE + TL; | |
421 Tcv = TaM - TaJ; | |
422 TaN = TaJ + TaM; | |
423 TdH = Tau + Tat; | |
424 Tav = Tat - Tau; | |
425 T7w = T2S + T2V; | |
426 T2W = T2S - T2V; | |
427 T37 = T31 + T36; | |
428 T7I = T36 - T31; | |
429 } | |
430 T7H = T3A - T3x; | |
431 T3B = T3x + T3A; | |
432 TeP = TdH + TdG; | |
433 TdI = TdG - TdH; | |
434 T6i = FNMS(KP707106781, T37, T2W); | |
435 T38 = FMA(KP707106781, T37, T2W); | |
436 T3E = T3C - T3D; | |
437 T7x = T3C + T3D; | |
438 T6l = FNMS(KP707106781, T3E, T3B); | |
439 T3F = FMA(KP707106781, T3E, T3B); | |
440 T9b = FMA(KP707106781, T7I, T7H); | |
441 T7J = FNMS(KP707106781, T7I, T7H); | |
442 } | |
443 } | |
444 { | |
445 E T4r, T4I, T1F, Tbb, T4u, T4L, Tbj, TdS, T1I, Tbd, T4N, T4A, T4B, T1L, Tbc; | |
446 E T4E, T1M, Tbg; | |
447 { | |
448 E T1z, T1A, T1C, T1D, Tbi, Tbh; | |
449 T1z = Cr[WS(csr, 5)]; | |
450 Tcs = Tas - Tav; | |
451 Taw = Tas + Tav; | |
452 T98 = FMA(KP707106781, T7x, T7w); | |
453 T7y = FNMS(KP707106781, T7x, T7w); | |
454 T1A = Cr[WS(csr, 59)]; | |
455 T1C = Cr[WS(csr, 37)]; | |
456 T1D = Cr[WS(csr, 27)]; | |
457 { | |
458 E T4s, T1B, T1E, T4t, T4J, T4K; | |
459 T4s = Ci[WS(csi, 37)]; | |
460 T4r = T1z - T1A; | |
461 T1B = T1z + T1A; | |
462 T4I = T1C - T1D; | |
463 T1E = T1C + T1D; | |
464 T4t = Ci[WS(csi, 27)]; | |
465 T4J = Ci[WS(csi, 5)]; | |
466 T4K = Ci[WS(csi, 59)]; | |
467 T1F = T1B + T1E; | |
468 Tbb = T1B - T1E; | |
469 T4u = T4s + T4t; | |
470 Tbi = T4s - T4t; | |
471 Tbh = T4J - T4K; | |
472 T4L = T4J + T4K; | |
473 } | |
474 { | |
475 E T1J, T4w, T4z, T1K, T4C, T4D; | |
476 { | |
477 E T1G, T1H, T4x, T4y; | |
478 T1G = Cr[WS(csr, 21)]; | |
479 Tbj = Tbh - Tbi; | |
480 TdS = Tbi + Tbh; | |
481 T1H = Cr[WS(csr, 43)]; | |
482 T4x = Ci[WS(csi, 21)]; | |
483 T4y = Ci[WS(csi, 43)]; | |
484 T1J = Cr[WS(csr, 11)]; | |
485 T4w = T1G - T1H; | |
486 T1I = T1G + T1H; | |
487 Tbd = T4x - T4y; | |
488 T4z = T4x + T4y; | |
489 T1K = Cr[WS(csr, 53)]; | |
490 T4C = Ci[WS(csi, 11)]; | |
491 T4D = Ci[WS(csi, 53)]; | |
492 } | |
493 T4N = T4w + T4z; | |
494 T4A = T4w - T4z; | |
495 T4B = T1J - T1K; | |
496 T1L = T1J + T1K; | |
497 Tbc = T4D - T4C; | |
498 T4E = T4C + T4D; | |
499 } | |
500 } | |
501 T1M = T1I + T1L; | |
502 Tbg = T1I - T1L; | |
503 { | |
504 E TdT, Tbe, T4F, T4O; | |
505 TdT = Tbd + Tbc; | |
506 Tbe = Tbc - Tbd; | |
507 T4F = T4B - T4E; | |
508 T4O = T4B + T4E; | |
509 { | |
510 E TdR, TdU, T81, T4v, T4G, T85; | |
511 TdR = T1F - T1M; | |
512 T1N = T1F + T1M; | |
513 TeW = TdT + TdS; | |
514 TdU = TdS - TdT; | |
515 T81 = T4r + T4u; | |
516 T4v = T4r - T4u; | |
517 T4G = T4A + T4F; | |
518 T85 = T4F - T4A; | |
519 { | |
520 E T84, T4M, T4P, T82, TcG, TcH; | |
521 T84 = T4L - T4I; | |
522 T4M = T4I + T4L; | |
523 T6x = FNMS(KP707106781, T4G, T4v); | |
524 T4H = FMA(KP707106781, T4G, T4v); | |
525 Te8 = TdR + TdU; | |
526 TdV = TdR - TdU; | |
527 T4P = T4N - T4O; | |
528 T82 = T4N + T4O; | |
529 Tbk = Tbg + Tbj; | |
530 TcG = Tbj - Tbg; | |
531 T6w = FNMS(KP707106781, T4P, T4M); | |
532 T4Q = FMA(KP707106781, T4P, T4M); | |
533 T9j = FMA(KP707106781, T85, T84); | |
534 T86 = FNMS(KP707106781, T85, T84); | |
535 TcH = Tbb - Tbe; | |
536 Tbf = Tbb + Tbe; | |
537 TcO = FMA(KP414213562, TcG, TcH); | |
538 TcI = FNMS(KP414213562, TcH, TcG); | |
539 T9k = FMA(KP707106781, T82, T81); | |
540 T83 = FNMS(KP707106781, T82, T81); | |
541 } | |
542 } | |
543 } | |
544 } | |
545 } | |
546 { | |
547 E T88, T89, Tbv, Tbq; | |
548 { | |
549 E T4S, T59, T4V, Tbm, T1U, T5c, TdX, Tbu, T1X, T53, Tbo, T52, T20, T54, T5e; | |
550 E T51; | |
551 { | |
552 E T1R, T1Q, T1S, T1O, T1P; | |
553 T1O = Cr[WS(csr, 3)]; | |
554 T1P = Cr[WS(csr, 61)]; | |
555 T1R = Cr[WS(csr, 29)]; | |
556 TbI = FMA(KP414213562, Tbf, Tbk); | |
557 Tbl = FNMS(KP414213562, Tbk, Tbf); | |
558 T1Q = T1O + T1P; | |
559 T4S = T1O - T1P; | |
560 T1S = Cr[WS(csr, 35)]; | |
561 { | |
562 E Tbt, Tbs, T4X, T50; | |
563 { | |
564 E T5a, T5b, T4T, T4U, T1T; | |
565 T4T = Ci[WS(csi, 29)]; | |
566 T4U = Ci[WS(csi, 35)]; | |
567 T1T = T1R + T1S; | |
568 T59 = T1R - T1S; | |
569 T5a = Ci[WS(csi, 3)]; | |
570 Tbt = T4T - T4U; | |
571 T4V = T4T + T4U; | |
572 T5b = Ci[WS(csi, 61)]; | |
573 Tbm = T1Q - T1T; | |
574 T1U = T1Q + T1T; | |
575 T5c = T5a + T5b; | |
576 Tbs = T5b - T5a; | |
577 } | |
578 { | |
579 E T4Y, T4Z, T1V, T1W, T1Y, T1Z; | |
580 T1V = Cr[WS(csr, 13)]; | |
581 T1W = Cr[WS(csr, 51)]; | |
582 TdX = Tbt + Tbs; | |
583 Tbu = Tbs - Tbt; | |
584 T4Y = Ci[WS(csi, 13)]; | |
585 T4X = T1V - T1W; | |
586 T1X = T1V + T1W; | |
587 T4Z = Ci[WS(csi, 51)]; | |
588 T1Y = Cr[WS(csr, 19)]; | |
589 T1Z = Cr[WS(csr, 45)]; | |
590 T53 = Ci[WS(csi, 19)]; | |
591 Tbo = T4Y - T4Z; | |
592 T50 = T4Y + T4Z; | |
593 T52 = T1Y - T1Z; | |
594 T20 = T1Y + T1Z; | |
595 T54 = Ci[WS(csi, 45)]; | |
596 } | |
597 T5e = T4X + T50; | |
598 T51 = T4X - T50; | |
599 } | |
600 } | |
601 { | |
602 E T21, Tbr, T55, Tbn; | |
603 T21 = T1X + T20; | |
604 Tbr = T1X - T20; | |
605 T55 = T53 + T54; | |
606 Tbn = T54 - T53; | |
607 { | |
608 E T4W, TdW, Tbp, T5f, TdZ, T57, T8c, TdY, T56; | |
609 T88 = T4S + T4V; | |
610 T4W = T4S - T4V; | |
611 T22 = T1U + T21; | |
612 TdW = T1U - T21; | |
613 TdY = Tbo + Tbn; | |
614 Tbp = Tbn - Tbo; | |
615 T56 = T52 - T55; | |
616 T5f = T52 + T55; | |
617 TeV = TdY + TdX; | |
618 TdZ = TdX - TdY; | |
619 T57 = T51 + T56; | |
620 T8c = T56 - T51; | |
621 { | |
622 E T8b, T5d, T5g, TcD, TcE; | |
623 T8b = T59 + T5c; | |
624 T5d = T59 - T5c; | |
625 T5g = T5e - T5f; | |
626 T89 = T5e + T5f; | |
627 Te0 = TdW + TdZ; | |
628 Te9 = TdZ - TdW; | |
629 T58 = FMA(KP707106781, T57, T4W); | |
630 T6u = FNMS(KP707106781, T57, T4W); | |
631 T6t = FNMS(KP707106781, T5g, T5d); | |
632 T5h = FMA(KP707106781, T5g, T5d); | |
633 Tbv = Tbr + Tbu; | |
634 TcD = Tbu - Tbr; | |
635 TcE = Tbm - Tbp; | |
636 Tbq = Tbm + Tbp; | |
637 T9m = FNMS(KP707106781, T8c, T8b); | |
638 T8d = FMA(KP707106781, T8c, T8b); | |
639 TcP = FNMS(KP414213562, TcD, TcE); | |
640 TcF = FMA(KP414213562, TcE, TcD); | |
641 } | |
642 } | |
643 } | |
644 } | |
645 { | |
646 E Tb3, Tb8, T7V, T7Y; | |
647 { | |
648 E T7T, T4c, TaZ, T1p, TdO, Tb2, T7U, T47, T1t, T4e, T1s, Tb5, T4m, T1u, T4f; | |
649 E T4g; | |
650 { | |
651 E T1m, T43, T1l, Tb0, T4b, T1n, T44, T45; | |
652 { | |
653 E T1j, T1k, T49, T4a; | |
654 T1j = Cr[WS(csr, 9)]; | |
655 T9n = FMA(KP707106781, T89, T88); | |
656 T8a = FNMS(KP707106781, T89, T88); | |
657 TbJ = FNMS(KP414213562, Tbq, Tbv); | |
658 Tbw = FMA(KP414213562, Tbv, Tbq); | |
659 T1k = Cr[WS(csr, 55)]; | |
660 T49 = Ci[WS(csi, 9)]; | |
661 T4a = Ci[WS(csi, 55)]; | |
662 T1m = Cr[WS(csr, 41)]; | |
663 T43 = T1j - T1k; | |
664 T1l = T1j + T1k; | |
665 Tb0 = T49 - T4a; | |
666 T4b = T49 + T4a; | |
667 T1n = Cr[WS(csr, 23)]; | |
668 T44 = Ci[WS(csi, 41)]; | |
669 T45 = Ci[WS(csi, 23)]; | |
670 } | |
671 { | |
672 E T1q, T1r, T4k, T4l; | |
673 T1q = Cr[WS(csr, 7)]; | |
674 { | |
675 E T48, T1o, Tb1, T46; | |
676 T48 = T1m - T1n; | |
677 T1o = T1m + T1n; | |
678 Tb1 = T44 - T45; | |
679 T46 = T44 + T45; | |
680 T7T = T4b - T48; | |
681 T4c = T48 + T4b; | |
682 TaZ = T1l - T1o; | |
683 T1p = T1l + T1o; | |
684 TdO = Tb1 + Tb0; | |
685 Tb2 = Tb0 - Tb1; | |
686 T7U = T43 + T46; | |
687 T47 = T43 - T46; | |
688 T1r = Cr[WS(csr, 57)]; | |
689 } | |
690 T4k = Ci[WS(csi, 7)]; | |
691 T4l = Ci[WS(csi, 57)]; | |
692 T1t = Cr[WS(csr, 25)]; | |
693 T4e = T1q - T1r; | |
694 T1s = T1q + T1r; | |
695 Tb5 = T4l - T4k; | |
696 T4m = T4k + T4l; | |
697 T1u = Cr[WS(csr, 39)]; | |
698 T4f = Ci[WS(csi, 25)]; | |
699 T4g = Ci[WS(csi, 39)]; | |
700 } | |
701 } | |
702 { | |
703 E T7W, TdN, T7X, T5u, T4d, T4o, T5v, T8k, T8l; | |
704 { | |
705 E T4n, T1w, T4i, TbE, TbF, Tb4, Tb7; | |
706 { | |
707 E T4j, T1v, Tb6, T4h; | |
708 T4j = T1t - T1u; | |
709 T1v = T1t + T1u; | |
710 Tb6 = T4f - T4g; | |
711 T4h = T4f + T4g; | |
712 T7W = T4j + T4m; | |
713 T4n = T4j - T4m; | |
714 Tb4 = T1s - T1v; | |
715 T1w = T1s + T1v; | |
716 TdN = Tb6 + Tb5; | |
717 Tb7 = Tb5 - Tb6; | |
718 T7X = T4e + T4h; | |
719 T4i = T4e - T4h; | |
720 } | |
721 Tb3 = TaZ - Tb2; | |
722 TbE = TaZ + Tb2; | |
723 TbF = Tb7 - Tb4; | |
724 Tb8 = Tb4 + Tb7; | |
725 Te3 = T1p - T1w; | |
726 T1x = T1p + T1w; | |
727 TcB = TbE - TbF; | |
728 TbG = TbE + TbF; | |
729 T5u = FMA(KP414213562, T47, T4c); | |
730 T4d = FNMS(KP414213562, T4c, T47); | |
731 T4o = FMA(KP414213562, T4n, T4i); | |
732 T5v = FNMS(KP414213562, T4i, T4n); | |
733 } | |
734 Tf1 = TdO + TdN; | |
735 TdP = TdN - TdO; | |
736 T6C = T4o - T4d; | |
737 T4p = T4d + T4o; | |
738 T7V = FNMS(KP414213562, T7U, T7T); | |
739 T8k = FMA(KP414213562, T7T, T7U); | |
740 T8l = FMA(KP414213562, T7W, T7X); | |
741 T7Y = FNMS(KP414213562, T7X, T7W); | |
742 T6r = T5u - T5v; | |
743 T5w = T5u + T5v; | |
744 T9h = T8k + T8l; | |
745 T8m = T8k - T8l; | |
746 } | |
747 } | |
748 { | |
749 E T7z, T3i, Tax, TT, TdC, TaA, T7A, T3d, TX, T3k, TW, TaD, T3s, TY, T3l; | |
750 E T3m; | |
751 { | |
752 E TQ, T39, TP, Tay, T3h, TR, T3a, T3b; | |
753 { | |
754 E TN, TO, T3f, T3g; | |
755 TN = Cr[WS(csr, 10)]; | |
756 TcM = Tb8 - Tb3; | |
757 Tb9 = Tb3 + Tb8; | |
758 T9s = T7V - T7Y; | |
759 T7Z = T7V + T7Y; | |
760 TO = Cr[WS(csr, 54)]; | |
761 T3f = Ci[WS(csi, 10)]; | |
762 T3g = Ci[WS(csi, 54)]; | |
763 TQ = Cr[WS(csr, 42)]; | |
764 T39 = TN - TO; | |
765 TP = TN + TO; | |
766 Tay = T3f - T3g; | |
767 T3h = T3f + T3g; | |
768 TR = Cr[WS(csr, 22)]; | |
769 T3a = Ci[WS(csi, 42)]; | |
770 T3b = Ci[WS(csi, 22)]; | |
771 } | |
772 { | |
773 E TU, TV, T3q, T3r; | |
774 TU = Cr[WS(csr, 6)]; | |
775 { | |
776 E T3e, TS, Taz, T3c; | |
777 T3e = TQ - TR; | |
778 TS = TQ + TR; | |
779 Taz = T3a - T3b; | |
780 T3c = T3a + T3b; | |
781 T7z = T3h - T3e; | |
782 T3i = T3e + T3h; | |
783 Tax = TP - TS; | |
784 TT = TP + TS; | |
785 TdC = Taz + Tay; | |
786 TaA = Tay - Taz; | |
787 T7A = T39 + T3c; | |
788 T3d = T39 - T3c; | |
789 TV = Cr[WS(csr, 58)]; | |
790 } | |
791 T3q = Ci[WS(csi, 6)]; | |
792 T3r = Ci[WS(csi, 58)]; | |
793 TX = Cr[WS(csr, 26)]; | |
794 T3k = TU - TV; | |
795 TW = TU + TV; | |
796 TaD = T3r - T3q; | |
797 T3s = T3q + T3r; | |
798 TY = Cr[WS(csr, 38)]; | |
799 T3l = Ci[WS(csi, 26)]; | |
800 T3m = Ci[WS(csi, 38)]; | |
801 } | |
802 } | |
803 { | |
804 E T7C, TdB, T7D, T3G, T3j, T3u, T3H, T7K, T7L; | |
805 { | |
806 E T3t, T10, T3o, TaO, TaP, TaC, TaF; | |
807 { | |
808 E T3p, TZ, TaE, T3n; | |
809 T3p = TX - TY; | |
810 TZ = TX + TY; | |
811 TaE = T3l - T3m; | |
812 T3n = T3l + T3m; | |
813 T7C = T3p + T3s; | |
814 T3t = T3p - T3s; | |
815 TaC = TW - TZ; | |
816 T10 = TW + TZ; | |
817 TdB = TaE + TaD; | |
818 TaF = TaD - TaE; | |
819 T7D = T3k + T3n; | |
820 T3o = T3k - T3n; | |
821 } | |
822 TaB = Tax - TaA; | |
823 TaO = Tax + TaA; | |
824 TaP = TaF - TaC; | |
825 TaG = TaC + TaF; | |
826 TdF = TT - T10; | |
827 T11 = TT + T10; | |
828 Tct = TaO - TaP; | |
829 TaQ = TaO + TaP; | |
830 T3G = FMA(KP414213562, T3d, T3i); | |
831 T3j = FNMS(KP414213562, T3i, T3d); | |
832 T3u = FMA(KP414213562, T3t, T3o); | |
833 T3H = FNMS(KP414213562, T3o, T3t); | |
834 } | |
835 TeQ = TdC + TdB; | |
836 TdD = TdB - TdC; | |
837 T6m = T3u - T3j; | |
838 T3v = T3j + T3u; | |
839 T7B = FNMS(KP414213562, T7A, T7z); | |
840 T7K = FMA(KP414213562, T7z, T7A); | |
841 T7L = FMA(KP414213562, T7C, T7D); | |
842 T7E = FNMS(KP414213562, T7D, T7C); | |
843 T6j = T3G - T3H; | |
844 T3I = T3G + T3H; | |
845 T99 = T7K + T7L; | |
846 T7M = T7K - T7L; | |
847 } | |
848 } | |
849 } | |
850 } | |
851 } | |
852 { | |
853 E Tcw, T9c, T7F, Tev, Teu, TeD, Tep, TeG, Tez, TeE, Tes; | |
854 { | |
855 E TbX, TbY, Tc7, TbP, Tar, Tc5, Tc1, Tc0, Tc4, Tba, TbS, TbL, TbQ, TaS, Tbx; | |
856 E Tc8; | |
857 { | |
858 E TeO, TaH, TeR, TeL, TeU, TeZ, Tf2, TeX, Tfh, Tfn, Tfo, Tfm; | |
859 { | |
860 E T12, Tfg, Tfj, Tx, Tff, T24, Tfi, Tfk, Th, T1y, T23; | |
861 TeO = TM - T11; | |
862 T12 = TM + T11; | |
863 Tcw = TaG - TaB; | |
864 TaH = TaB + TaG; | |
865 T9c = T7B - T7E; | |
866 T7F = T7B + T7E; | |
867 Tfg = TeQ + TeP; | |
868 TeR = TeP - TeQ; | |
869 TeL = FNMS(KP2_000000000, Tg, T9); | |
870 Th = FMA(KP2_000000000, Tg, T9); | |
871 T1y = T1i + T1x; | |
872 TeU = T1i - T1x; | |
873 TeZ = T1N - T22; | |
874 T23 = T1N + T22; | |
875 Tfj = Tf1 + Tf0; | |
876 Tf2 = Tf0 - Tf1; | |
877 Tx = FMA(KP2_000000000, Tw, Th); | |
878 Tff = FNMS(KP2_000000000, Tw, Th); | |
879 T24 = T1y + T23; | |
880 Tfi = T1y - T23; | |
881 TeX = TeV - TeW; | |
882 Tfk = TeW + TeV; | |
883 { | |
884 E T13, Tfp, Tfl, Tfq; | |
885 T13 = FMA(KP2_000000000, T12, Tx); | |
886 Tfp = FNMS(KP2_000000000, T12, Tx); | |
887 Tfh = FNMS(KP2_000000000, Tfg, Tff); | |
888 Tfn = FMA(KP2_000000000, Tfg, Tff); | |
889 Tfl = Tfj - Tfk; | |
890 Tfq = Tfk + Tfj; | |
891 R0[0] = FMA(KP2_000000000, T24, T13); | |
892 R0[WS(rs, 32)] = FNMS(KP2_000000000, T24, T13); | |
893 R0[WS(rs, 48)] = FMA(KP2_000000000, Tfq, Tfp); | |
894 R0[WS(rs, 16)] = FNMS(KP2_000000000, Tfq, Tfp); | |
895 Tfo = Tfi + Tfl; | |
896 Tfm = Tfi - Tfl; | |
897 } | |
898 } | |
899 { | |
900 E Tf7, TeN, Tfa, Tf3, Tf8, TeS; | |
901 R0[WS(rs, 8)] = FMA(KP1_414213562, Tfm, Tfh); | |
902 R0[WS(rs, 40)] = FNMS(KP1_414213562, Tfm, Tfh); | |
903 R0[WS(rs, 56)] = FMA(KP1_414213562, Tfo, Tfn); | |
904 R0[WS(rs, 24)] = FNMS(KP1_414213562, Tfo, Tfn); | |
905 Tf7 = FMA(KP2_000000000, TeM, TeL); | |
906 TeN = FNMS(KP2_000000000, TeM, TeL); | |
907 Tfa = Tf2 - TeZ; | |
908 Tf3 = TeZ + Tf2; | |
909 Tf8 = TeO + TeR; | |
910 TeS = TeO - TeR; | |
911 { | |
912 E TbH, TbK, TaI, TaR; | |
913 { | |
914 E Taf, Tf9, Tfd, Tf5, TeT, Tfb, TeY, Taq; | |
915 TbX = FNMS(KP1_414213562, Tae, Ta9); | |
916 Taf = FMA(KP1_414213562, Tae, Ta9); | |
917 Tf9 = FNMS(KP1_414213562, Tf8, Tf7); | |
918 Tfd = FMA(KP1_414213562, Tf8, Tf7); | |
919 Tf5 = FNMS(KP1_414213562, TeS, TeN); | |
920 TeT = FMA(KP1_414213562, TeS, TeN); | |
921 Tfb = TeU - TeX; | |
922 TeY = TeU + TeX; | |
923 Taq = FNMS(KP414213562, Tap, Tak); | |
924 TbY = FMA(KP414213562, Tak, Tap); | |
925 Tc7 = FNMS(KP707106781, TbG, TbD); | |
926 TbH = FMA(KP707106781, TbG, TbD); | |
927 { | |
928 E Tfc, Tfe, Tf6, Tf4; | |
929 Tfc = FNMS(KP414213562, Tfb, Tfa); | |
930 Tfe = FMA(KP414213562, Tfa, Tfb); | |
931 Tf6 = FMA(KP414213562, TeY, Tf3); | |
932 Tf4 = FNMS(KP414213562, Tf3, TeY); | |
933 TbP = FNMS(KP1_847759065, Taq, Taf); | |
934 Tar = FMA(KP1_847759065, Taq, Taf); | |
935 R0[WS(rs, 44)] = FMA(KP1_847759065, Tfc, Tf9); | |
936 R0[WS(rs, 12)] = FNMS(KP1_847759065, Tfc, Tf9); | |
937 R0[WS(rs, 60)] = FMA(KP1_847759065, Tfe, Tfd); | |
938 R0[WS(rs, 28)] = FNMS(KP1_847759065, Tfe, Tfd); | |
939 R0[WS(rs, 52)] = FMA(KP1_847759065, Tf6, Tf5); | |
940 R0[WS(rs, 20)] = FNMS(KP1_847759065, Tf6, Tf5); | |
941 R0[WS(rs, 4)] = FMA(KP1_847759065, Tf4, TeT); | |
942 R0[WS(rs, 36)] = FNMS(KP1_847759065, Tf4, TeT); | |
943 TbK = TbI + TbJ; | |
944 Tc5 = TbI - TbJ; | |
945 } | |
946 } | |
947 Tc1 = FNMS(KP707106781, TaH, Taw); | |
948 TaI = FMA(KP707106781, TaH, Taw); | |
949 TaR = FMA(KP707106781, TaQ, TaN); | |
950 Tc0 = FNMS(KP707106781, TaQ, TaN); | |
951 Tc4 = FNMS(KP707106781, Tb9, TaY); | |
952 Tba = FMA(KP707106781, Tb9, TaY); | |
953 TbS = FNMS(KP923879532, TbK, TbH); | |
954 TbL = FMA(KP923879532, TbK, TbH); | |
955 TbQ = FMA(KP198912367, TaI, TaR); | |
956 TaS = FNMS(KP198912367, TaR, TaI); | |
957 Tbx = Tbl + Tbw; | |
958 Tc8 = Tbw - Tbl; | |
959 } | |
960 } | |
961 } | |
962 { | |
963 E Ten, Teo, Tex, Tef, Tdz, Ter, Teq, TdQ, Tei, Teb, Teg, TdK, Te1, Tey; | |
964 { | |
965 E Te7, Tea, TdE, TdJ; | |
966 { | |
967 E Tdt, TbR, TbV, TbN, TaT, TbT, Tby, Tdy; | |
968 Ten = FMA(KP2_000000000, Tds, Tdr); | |
969 Tdt = FNMS(KP2_000000000, Tds, Tdr); | |
970 TbR = FNMS(KP1_961570560, TbQ, TbP); | |
971 TbV = FMA(KP1_961570560, TbQ, TbP); | |
972 TbN = FNMS(KP1_961570560, TaS, Tar); | |
973 TaT = FMA(KP1_961570560, TaS, Tar); | |
974 TbT = FNMS(KP923879532, Tbx, Tba); | |
975 Tby = FMA(KP923879532, Tbx, Tba); | |
976 Tdy = Tdu - Tdx; | |
977 Teo = Tdu + Tdx; | |
978 Tex = Te6 - Te3; | |
979 Te7 = Te3 + Te6; | |
980 { | |
981 E TbU, TbW, TbO, TbM; | |
982 TbU = FNMS(KP820678790, TbT, TbS); | |
983 TbW = FMA(KP820678790, TbS, TbT); | |
984 TbO = FMA(KP098491403, Tby, TbL); | |
985 TbM = FNMS(KP098491403, TbL, Tby); | |
986 Tef = FNMS(KP1_414213562, Tdy, Tdt); | |
987 Tdz = FMA(KP1_414213562, Tdy, Tdt); | |
988 R0[WS(rs, 41)] = FMA(KP1_546020906, TbU, TbR); | |
989 R0[WS(rs, 9)] = FNMS(KP1_546020906, TbU, TbR); | |
990 R0[WS(rs, 57)] = FMA(KP1_546020906, TbW, TbV); | |
991 R0[WS(rs, 25)] = FNMS(KP1_546020906, TbW, TbV); | |
992 R0[WS(rs, 49)] = FMA(KP1_990369453, TbO, TbN); | |
993 R0[WS(rs, 17)] = FNMS(KP1_990369453, TbO, TbN); | |
994 R0[WS(rs, 1)] = FMA(KP1_990369453, TbM, TaT); | |
995 R0[WS(rs, 33)] = FNMS(KP1_990369453, TbM, TaT); | |
996 Tea = Te8 + Te9; | |
997 Tev = Te8 - Te9; | |
998 } | |
999 } | |
1000 Ter = TdA - TdD; | |
1001 TdE = TdA + TdD; | |
1002 TdJ = TdF + TdI; | |
1003 Teq = TdI - TdF; | |
1004 Teu = TdM - TdP; | |
1005 TdQ = TdM + TdP; | |
1006 Tei = FNMS(KP707106781, Tea, Te7); | |
1007 Teb = FMA(KP707106781, Tea, Te7); | |
1008 Teg = FMA(KP414213562, TdE, TdJ); | |
1009 TdK = FNMS(KP414213562, TdJ, TdE); | |
1010 Te1 = TdV + Te0; | |
1011 Tey = Te0 - TdV; | |
1012 } | |
1013 { | |
1014 E Tcd, TbZ, Tcg, Tc9, Tce, Tc2; | |
1015 { | |
1016 E Teh, Tel, Ted, TdL, Tej, Te2; | |
1017 Teh = FNMS(KP1_847759065, Teg, Tef); | |
1018 Tel = FMA(KP1_847759065, Teg, Tef); | |
1019 Ted = FNMS(KP1_847759065, TdK, Tdz); | |
1020 TdL = FMA(KP1_847759065, TdK, Tdz); | |
1021 Tej = FNMS(KP707106781, Te1, TdQ); | |
1022 Te2 = FMA(KP707106781, Te1, TdQ); | |
1023 { | |
1024 E Tek, Tem, Tee, Tec; | |
1025 Tek = FNMS(KP668178637, Tej, Tei); | |
1026 Tem = FMA(KP668178637, Tei, Tej); | |
1027 Tee = FMA(KP198912367, Te2, Teb); | |
1028 Tec = FNMS(KP198912367, Teb, Te2); | |
1029 Tcd = FMA(KP1_847759065, TbY, TbX); | |
1030 TbZ = FNMS(KP1_847759065, TbY, TbX); | |
1031 R0[WS(rs, 42)] = FMA(KP1_662939224, Tek, Teh); | |
1032 R0[WS(rs, 10)] = FNMS(KP1_662939224, Tek, Teh); | |
1033 R0[WS(rs, 58)] = FMA(KP1_662939224, Tem, Tel); | |
1034 R0[WS(rs, 26)] = FNMS(KP1_662939224, Tem, Tel); | |
1035 R0[WS(rs, 50)] = FMA(KP1_961570560, Tee, Ted); | |
1036 R0[WS(rs, 18)] = FNMS(KP1_961570560, Tee, Ted); | |
1037 R0[WS(rs, 2)] = FMA(KP1_961570560, Tec, TdL); | |
1038 R0[WS(rs, 34)] = FNMS(KP1_961570560, Tec, TdL); | |
1039 } | |
1040 } | |
1041 Tcg = FMA(KP923879532, Tc8, Tc7); | |
1042 Tc9 = FNMS(KP923879532, Tc8, Tc7); | |
1043 Tce = FMA(KP668178637, Tc0, Tc1); | |
1044 Tc2 = FNMS(KP668178637, Tc1, Tc0); | |
1045 { | |
1046 E Tcf, Tcj, Tcb, Tc3, Tch, Tc6; | |
1047 Tcf = FNMS(KP1_662939224, Tce, Tcd); | |
1048 Tcj = FMA(KP1_662939224, Tce, Tcd); | |
1049 Tcb = FMA(KP1_662939224, Tc2, TbZ); | |
1050 Tc3 = FNMS(KP1_662939224, Tc2, TbZ); | |
1051 Tch = FMA(KP923879532, Tc5, Tc4); | |
1052 Tc6 = FNMS(KP923879532, Tc5, Tc4); | |
1053 { | |
1054 E Tci, Tck, Tcc, Tca; | |
1055 Tci = FNMS(KP303346683, Tch, Tcg); | |
1056 Tck = FMA(KP303346683, Tcg, Tch); | |
1057 Tcc = FMA(KP534511135, Tc6, Tc9); | |
1058 Tca = FNMS(KP534511135, Tc9, Tc6); | |
1059 TeD = FMA(KP1_414213562, Teo, Ten); | |
1060 Tep = FNMS(KP1_414213562, Teo, Ten); | |
1061 R0[WS(rs, 45)] = FMA(KP1_913880671, Tci, Tcf); | |
1062 R0[WS(rs, 13)] = FNMS(KP1_913880671, Tci, Tcf); | |
1063 R0[WS(rs, 61)] = FMA(KP1_913880671, Tck, Tcj); | |
1064 R0[WS(rs, 29)] = FNMS(KP1_913880671, Tck, Tcj); | |
1065 R0[WS(rs, 53)] = FMA(KP1_763842528, Tcc, Tcb); | |
1066 R0[WS(rs, 21)] = FNMS(KP1_763842528, Tcc, Tcb); | |
1067 R0[WS(rs, 5)] = FMA(KP1_763842528, Tca, Tc3); | |
1068 R0[WS(rs, 37)] = FNMS(KP1_763842528, Tca, Tc3); | |
1069 } | |
1070 } | |
1071 TeG = FMA(KP707106781, Tey, Tex); | |
1072 Tez = FNMS(KP707106781, Tey, Tex); | |
1073 TeE = FMA(KP414213562, Teq, Ter); | |
1074 Tes = FNMS(KP414213562, Ter, Teq); | |
1075 } | |
1076 } | |
1077 } | |
1078 { | |
1079 E T5L, T5M, T61, T62; | |
1080 { | |
1081 E Td3, Td4, Tdd, TcV, Tcr, Tdb, Td7, Td6, Tda, TcC, TcY, TcR, TcW, Tcy, TcJ; | |
1082 E Tde; | |
1083 { | |
1084 E TcN, TcQ, Tcu, Tcx; | |
1085 { | |
1086 E Tcn, TeF, TeJ, TeB, Tet, TeH, Tew, Tcq; | |
1087 Td3 = FMA(KP1_414213562, Tcm, Tcl); | |
1088 Tcn = FNMS(KP1_414213562, Tcm, Tcl); | |
1089 TeF = FNMS(KP1_847759065, TeE, TeD); | |
1090 TeJ = FMA(KP1_847759065, TeE, TeD); | |
1091 TeB = FMA(KP1_847759065, Tes, Tep); | |
1092 Tet = FNMS(KP1_847759065, Tes, Tep); | |
1093 TeH = FMA(KP707106781, Tev, Teu); | |
1094 Tew = FNMS(KP707106781, Tev, Teu); | |
1095 Tcq = FNMS(KP414213562, Tcp, Tco); | |
1096 Td4 = FMA(KP414213562, Tco, Tcp); | |
1097 Tdd = FMA(KP707106781, TcM, TcL); | |
1098 TcN = FNMS(KP707106781, TcM, TcL); | |
1099 { | |
1100 E TeI, TeK, TeC, TeA; | |
1101 TeI = FNMS(KP198912367, TeH, TeG); | |
1102 TeK = FMA(KP198912367, TeG, TeH); | |
1103 TeC = FMA(KP668178637, Tew, Tez); | |
1104 TeA = FNMS(KP668178637, Tez, Tew); | |
1105 TcV = FMA(KP1_847759065, Tcq, Tcn); | |
1106 Tcr = FNMS(KP1_847759065, Tcq, Tcn); | |
1107 R0[WS(rs, 46)] = FMA(KP1_961570560, TeI, TeF); | |
1108 R0[WS(rs, 14)] = FNMS(KP1_961570560, TeI, TeF); | |
1109 R0[WS(rs, 62)] = FMA(KP1_961570560, TeK, TeJ); | |
1110 R0[WS(rs, 30)] = FNMS(KP1_961570560, TeK, TeJ); | |
1111 R0[WS(rs, 54)] = FMA(KP1_662939224, TeC, TeB); | |
1112 R0[WS(rs, 22)] = FNMS(KP1_662939224, TeC, TeB); | |
1113 R0[WS(rs, 6)] = FMA(KP1_662939224, TeA, Tet); | |
1114 R0[WS(rs, 38)] = FNMS(KP1_662939224, TeA, Tet); | |
1115 TcQ = TcO - TcP; | |
1116 Tdb = TcO + TcP; | |
1117 } | |
1118 } | |
1119 Td7 = FMA(KP707106781, Tct, Tcs); | |
1120 Tcu = FNMS(KP707106781, Tct, Tcs); | |
1121 Tcx = FNMS(KP707106781, Tcw, Tcv); | |
1122 Td6 = FMA(KP707106781, Tcw, Tcv); | |
1123 Tda = FMA(KP707106781, TcB, TcA); | |
1124 TcC = FNMS(KP707106781, TcB, TcA); | |
1125 TcY = FNMS(KP923879532, TcQ, TcN); | |
1126 TcR = FMA(KP923879532, TcQ, TcN); | |
1127 TcW = FMA(KP668178637, Tcu, Tcx); | |
1128 Tcy = FNMS(KP668178637, Tcx, Tcu); | |
1129 TcJ = TcF - TcI; | |
1130 Tde = TcI + TcF; | |
1131 } | |
1132 { | |
1133 E Tdj, Td5, Tdm, Tdf, Tdk, Td8; | |
1134 { | |
1135 E TcX, Td1, TcT, Tcz, TcZ, TcK; | |
1136 TcX = FNMS(KP1_662939224, TcW, TcV); | |
1137 Td1 = FMA(KP1_662939224, TcW, TcV); | |
1138 TcT = FNMS(KP1_662939224, Tcy, Tcr); | |
1139 Tcz = FMA(KP1_662939224, Tcy, Tcr); | |
1140 TcZ = FNMS(KP923879532, TcJ, TcC); | |
1141 TcK = FMA(KP923879532, TcJ, TcC); | |
1142 { | |
1143 E Td0, Td2, TcU, TcS; | |
1144 Td0 = FNMS(KP534511135, TcZ, TcY); | |
1145 Td2 = FMA(KP534511135, TcY, TcZ); | |
1146 TcU = FMA(KP303346683, TcK, TcR); | |
1147 TcS = FNMS(KP303346683, TcR, TcK); | |
1148 Tdj = FMA(KP1_847759065, Td4, Td3); | |
1149 Td5 = FNMS(KP1_847759065, Td4, Td3); | |
1150 R0[WS(rs, 43)] = FMA(KP1_763842528, Td0, TcX); | |
1151 R0[WS(rs, 11)] = FNMS(KP1_763842528, Td0, TcX); | |
1152 R0[WS(rs, 59)] = FMA(KP1_763842528, Td2, Td1); | |
1153 R0[WS(rs, 27)] = FNMS(KP1_763842528, Td2, Td1); | |
1154 R0[WS(rs, 51)] = FMA(KP1_913880671, TcU, TcT); | |
1155 R0[WS(rs, 19)] = FNMS(KP1_913880671, TcU, TcT); | |
1156 R0[WS(rs, 3)] = FMA(KP1_913880671, TcS, Tcz); | |
1157 R0[WS(rs, 35)] = FNMS(KP1_913880671, TcS, Tcz); | |
1158 } | |
1159 } | |
1160 Tdm = FMA(KP923879532, Tde, Tdd); | |
1161 Tdf = FNMS(KP923879532, Tde, Tdd); | |
1162 Tdk = FMA(KP198912367, Td6, Td7); | |
1163 Td8 = FNMS(KP198912367, Td7, Td6); | |
1164 { | |
1165 E T5F, T2R, T5G, T3K, T64, T5S, T5X, T5x, T5U, T4q, T4R, T63, T5P, T5i, T5V; | |
1166 E T5A; | |
1167 { | |
1168 E T5N, T5O, T5R, T3w, T3J, T5Q, T5y, T5z; | |
1169 { | |
1170 E T2p, Tdl, Tdp, Tdh, Td9, Tdn, Tdc, T2Q; | |
1171 T5N = FNMS(KP1_847759065, T2o, T2d); | |
1172 T2p = FMA(KP1_847759065, T2o, T2d); | |
1173 Tdl = FNMS(KP1_961570560, Tdk, Tdj); | |
1174 Tdp = FMA(KP1_961570560, Tdk, Tdj); | |
1175 Tdh = FMA(KP1_961570560, Td8, Td5); | |
1176 Td9 = FNMS(KP1_961570560, Td8, Td5); | |
1177 Tdn = FMA(KP923879532, Tdb, Tda); | |
1178 Tdc = FNMS(KP923879532, Tdb, Tda); | |
1179 T2Q = FNMS(KP198912367, T2P, T2G); | |
1180 T5O = FMA(KP198912367, T2G, T2P); | |
1181 T5R = FNMS(KP923879532, T3v, T38); | |
1182 T3w = FMA(KP923879532, T3v, T38); | |
1183 { | |
1184 E Tdo, Tdq, Tdi, Tdg; | |
1185 Tdo = FNMS(KP098491403, Tdn, Tdm); | |
1186 Tdq = FMA(KP098491403, Tdm, Tdn); | |
1187 Tdi = FMA(KP820678790, Tdc, Tdf); | |
1188 Tdg = FNMS(KP820678790, Tdf, Tdc); | |
1189 T5F = FNMS(KP1_961570560, T2Q, T2p); | |
1190 T2R = FMA(KP1_961570560, T2Q, T2p); | |
1191 R0[WS(rs, 47)] = FMA(KP1_990369453, Tdo, Tdl); | |
1192 R0[WS(rs, 15)] = FNMS(KP1_990369453, Tdo, Tdl); | |
1193 R0[WS(rs, 63)] = FMA(KP1_990369453, Tdq, Tdp); | |
1194 R0[WS(rs, 31)] = FNMS(KP1_990369453, Tdq, Tdp); | |
1195 R0[WS(rs, 55)] = FMA(KP1_546020906, Tdi, Tdh); | |
1196 R0[WS(rs, 23)] = FNMS(KP1_546020906, Tdi, Tdh); | |
1197 R0[WS(rs, 7)] = FMA(KP1_546020906, Tdg, Td9); | |
1198 R0[WS(rs, 39)] = FNMS(KP1_546020906, Tdg, Td9); | |
1199 T3J = FMA(KP923879532, T3I, T3F); | |
1200 T5Q = FNMS(KP923879532, T3I, T3F); | |
1201 } | |
1202 } | |
1203 T5G = FMA(KP098491403, T3w, T3J); | |
1204 T3K = FNMS(KP098491403, T3J, T3w); | |
1205 T64 = FMA(KP820678790, T5Q, T5R); | |
1206 T5S = FNMS(KP820678790, T5R, T5Q); | |
1207 T5X = FNMS(KP923879532, T5w, T5t); | |
1208 T5x = FMA(KP923879532, T5w, T5t); | |
1209 T5U = FNMS(KP923879532, T4p, T42); | |
1210 T4q = FMA(KP923879532, T4p, T42); | |
1211 T4R = FNMS(KP198912367, T4Q, T4H); | |
1212 T5y = FMA(KP198912367, T4H, T4Q); | |
1213 T63 = FMA(KP1_961570560, T5O, T5N); | |
1214 T5P = FNMS(KP1_961570560, T5O, T5N); | |
1215 T5z = FNMS(KP198912367, T58, T5h); | |
1216 T5i = FMA(KP198912367, T5h, T58); | |
1217 T5V = T5y - T5z; | |
1218 T5A = T5y + T5z; | |
1219 } | |
1220 { | |
1221 E T5W, T5I, T5Z, T5J; | |
1222 { | |
1223 E T5D, T3L, T67, T5B, T5Y, T5j, T65, T69, T66, T5k; | |
1224 T5D = FNMS(KP1_990369453, T3K, T2R); | |
1225 T3L = FMA(KP1_990369453, T3K, T2R); | |
1226 T5W = FNMS(KP980785280, T5V, T5U); | |
1227 T67 = FMA(KP980785280, T5V, T5U); | |
1228 T5I = FNMS(KP980785280, T5A, T5x); | |
1229 T5B = FMA(KP980785280, T5A, T5x); | |
1230 T5Y = T5i - T4R; | |
1231 T5j = T4R + T5i; | |
1232 T65 = FNMS(KP1_546020906, T64, T63); | |
1233 T69 = FMA(KP1_546020906, T64, T63); | |
1234 T5Z = FNMS(KP980785280, T5Y, T5X); | |
1235 T66 = FMA(KP980785280, T5Y, T5X); | |
1236 T5J = FNMS(KP980785280, T5j, T4q); | |
1237 T5k = FMA(KP980785280, T5j, T4q); | |
1238 { | |
1239 E T68, T6a, T5E, T5C; | |
1240 T68 = FNMS(KP357805721, T67, T66); | |
1241 T6a = FMA(KP357805721, T66, T67); | |
1242 T5E = FMA(KP049126849, T5k, T5B); | |
1243 T5C = FNMS(KP049126849, T5B, T5k); | |
1244 R1[WS(rs, 60)] = FMA(KP1_883088130, T6a, T69); | |
1245 R1[WS(rs, 28)] = FNMS(KP1_883088130, T6a, T69); | |
1246 R1[WS(rs, 44)] = FMA(KP1_883088130, T68, T65); | |
1247 R1[WS(rs, 12)] = FNMS(KP1_883088130, T68, T65); | |
1248 R1[0] = FMA(KP1_997590912, T5C, T3L); | |
1249 R1[WS(rs, 32)] = FNMS(KP1_997590912, T5C, T3L); | |
1250 R1[WS(rs, 16)] = FNMS(KP1_997590912, T5E, T5D); | |
1251 R1[WS(rs, 48)] = FMA(KP1_997590912, T5E, T5D); | |
1252 } | |
1253 } | |
1254 { | |
1255 E T5H, T5K, T5T, T60; | |
1256 T5L = FMA(KP1_990369453, T5G, T5F); | |
1257 T5H = FNMS(KP1_990369453, T5G, T5F); | |
1258 T5K = FNMS(KP906347169, T5J, T5I); | |
1259 T5M = FMA(KP906347169, T5I, T5J); | |
1260 T61 = FMA(KP1_546020906, T5S, T5P); | |
1261 T5T = FNMS(KP1_546020906, T5S, T5P); | |
1262 T60 = FNMS(KP472964775, T5Z, T5W); | |
1263 T62 = FMA(KP472964775, T5W, T5Z); | |
1264 R1[WS(rs, 40)] = FMA(KP1_481902250, T5K, T5H); | |
1265 R1[WS(rs, 8)] = FNMS(KP1_481902250, T5K, T5H); | |
1266 R1[WS(rs, 4)] = FMA(KP1_807978586, T60, T5T); | |
1267 R1[WS(rs, 36)] = FNMS(KP1_807978586, T60, T5T); | |
1268 } | |
1269 } | |
1270 } | |
1271 } | |
1272 } | |
1273 { | |
1274 E T8B, T8C, T8R, T8S; | |
1275 { | |
1276 E T8v, T7v, T8w, T7O, T8N, T8n, T8U, T8I, T8T, T8F, T8K, T80, T87, T8e, T8L; | |
1277 E T8q; | |
1278 { | |
1279 E T8D, T8E, T8H, T8G, T8o, T8p; | |
1280 { | |
1281 E T7n, T7u, T7G, T7N; | |
1282 T8D = FMA(KP1_847759065, T7m, T7j); | |
1283 T7n = FNMS(KP1_847759065, T7m, T7j); | |
1284 R1[WS(rs, 52)] = FMA(KP1_807978586, T62, T61); | |
1285 R1[WS(rs, 20)] = FNMS(KP1_807978586, T62, T61); | |
1286 R1[WS(rs, 56)] = FMA(KP1_481902250, T5M, T5L); | |
1287 R1[WS(rs, 24)] = FNMS(KP1_481902250, T5M, T5L); | |
1288 T7u = FNMS(KP668178637, T7t, T7q); | |
1289 T8E = FMA(KP668178637, T7q, T7t); | |
1290 T8H = FMA(KP923879532, T7F, T7y); | |
1291 T7G = FNMS(KP923879532, T7F, T7y); | |
1292 T7N = FMA(KP923879532, T7M, T7J); | |
1293 T8G = FNMS(KP923879532, T7M, T7J); | |
1294 T8v = FNMS(KP1_662939224, T7u, T7n); | |
1295 T7v = FMA(KP1_662939224, T7u, T7n); | |
1296 T8w = FMA(KP303346683, T7G, T7N); | |
1297 T7O = FNMS(KP303346683, T7N, T7G); | |
1298 } | |
1299 T8N = FNMS(KP923879532, T8m, T8j); | |
1300 T8n = FMA(KP923879532, T8m, T8j); | |
1301 T8U = FMA(KP534511135, T8G, T8H); | |
1302 T8I = FNMS(KP534511135, T8H, T8G); | |
1303 T8T = FMA(KP1_662939224, T8E, T8D); | |
1304 T8F = FNMS(KP1_662939224, T8E, T8D); | |
1305 T8K = FMA(KP923879532, T7Z, T7S); | |
1306 T80 = FNMS(KP923879532, T7Z, T7S); | |
1307 T87 = FNMS(KP668178637, T86, T83); | |
1308 T8o = FMA(KP668178637, T83, T86); | |
1309 T8p = FMA(KP668178637, T8a, T8d); | |
1310 T8e = FNMS(KP668178637, T8d, T8a); | |
1311 T8L = T8o + T8p; | |
1312 T8q = T8o - T8p; | |
1313 } | |
1314 { | |
1315 E T8M, T8y, T8P, T8z; | |
1316 { | |
1317 E T8t, T7P, T8X, T8r, T8O, T8f, T8V, T8Z, T8W, T8g; | |
1318 T8t = FNMS(KP1_913880671, T7O, T7v); | |
1319 T7P = FMA(KP1_913880671, T7O, T7v); | |
1320 T8M = FNMS(KP831469612, T8L, T8K); | |
1321 T8X = FMA(KP831469612, T8L, T8K); | |
1322 T8y = FNMS(KP831469612, T8q, T8n); | |
1323 T8r = FMA(KP831469612, T8q, T8n); | |
1324 T8O = T8e - T87; | |
1325 T8f = T87 + T8e; | |
1326 T8V = FNMS(KP1_763842528, T8U, T8T); | |
1327 T8Z = FMA(KP1_763842528, T8U, T8T); | |
1328 T8P = FNMS(KP831469612, T8O, T8N); | |
1329 T8W = FMA(KP831469612, T8O, T8N); | |
1330 T8z = FNMS(KP831469612, T8f, T80); | |
1331 T8g = FMA(KP831469612, T8f, T80); | |
1332 { | |
1333 E T8Y, T90, T8u, T8s; | |
1334 T8Y = FNMS(KP250486960, T8X, T8W); | |
1335 T90 = FMA(KP250486960, T8W, T8X); | |
1336 T8u = FMA(KP148335987, T8g, T8r); | |
1337 T8s = FNMS(KP148335987, T8r, T8g); | |
1338 R1[WS(rs, 61)] = FMA(KP1_940062506, T90, T8Z); | |
1339 R1[WS(rs, 29)] = FNMS(KP1_940062506, T90, T8Z); | |
1340 R1[WS(rs, 45)] = FMA(KP1_940062506, T8Y, T8V); | |
1341 R1[WS(rs, 13)] = FNMS(KP1_940062506, T8Y, T8V); | |
1342 R1[WS(rs, 1)] = FMA(KP1_978353019, T8s, T7P); | |
1343 R1[WS(rs, 33)] = FNMS(KP1_978353019, T8s, T7P); | |
1344 R1[WS(rs, 17)] = FNMS(KP1_978353019, T8u, T8t); | |
1345 R1[WS(rs, 49)] = FMA(KP1_978353019, T8u, T8t); | |
1346 } | |
1347 } | |
1348 { | |
1349 E T8x, T8A, T8J, T8Q; | |
1350 T8B = FMA(KP1_913880671, T8w, T8v); | |
1351 T8x = FNMS(KP1_913880671, T8w, T8v); | |
1352 T8A = FNMS(KP741650546, T8z, T8y); | |
1353 T8C = FMA(KP741650546, T8y, T8z); | |
1354 T8R = FMA(KP1_763842528, T8I, T8F); | |
1355 T8J = FNMS(KP1_763842528, T8I, T8F); | |
1356 T8Q = FNMS(KP599376933, T8P, T8M); | |
1357 T8S = FMA(KP599376933, T8M, T8P); | |
1358 R1[WS(rs, 41)] = FMA(KP1_606415062, T8A, T8x); | |
1359 R1[WS(rs, 9)] = FNMS(KP1_606415062, T8A, T8x); | |
1360 R1[WS(rs, 5)] = FMA(KP1_715457220, T8Q, T8J); | |
1361 R1[WS(rs, 37)] = FNMS(KP1_715457220, T8Q, T8J); | |
1362 } | |
1363 } | |
1364 } | |
1365 { | |
1366 E T6R, T6S, T77, T78; | |
1367 { | |
1368 E T6L, T6h, T6M, T6o, T73, T6D, T7a, T6Y, T79, T6V, T70, T6s, T6y, T6v, T71; | |
1369 E T6G; | |
1370 { | |
1371 E T6T, T6U, T6X, T6W, T6E, T6F; | |
1372 { | |
1373 E T6d, T6g, T6k, T6n; | |
1374 T6T = FMA(KP1_847759065, T6c, T6b); | |
1375 T6d = FNMS(KP1_847759065, T6c, T6b); | |
1376 R1[WS(rs, 53)] = FMA(KP1_715457220, T8S, T8R); | |
1377 R1[WS(rs, 21)] = FNMS(KP1_715457220, T8S, T8R); | |
1378 R1[WS(rs, 57)] = FMA(KP1_606415062, T8C, T8B); | |
1379 R1[WS(rs, 25)] = FNMS(KP1_606415062, T8C, T8B); | |
1380 T6g = FNMS(KP668178637, T6f, T6e); | |
1381 T6U = FMA(KP668178637, T6e, T6f); | |
1382 T6X = FMA(KP923879532, T6j, T6i); | |
1383 T6k = FNMS(KP923879532, T6j, T6i); | |
1384 T6n = FNMS(KP923879532, T6m, T6l); | |
1385 T6W = FMA(KP923879532, T6m, T6l); | |
1386 T6L = FMA(KP1_662939224, T6g, T6d); | |
1387 T6h = FNMS(KP1_662939224, T6g, T6d); | |
1388 T6M = FMA(KP534511135, T6k, T6n); | |
1389 T6o = FNMS(KP534511135, T6n, T6k); | |
1390 } | |
1391 T73 = FMA(KP923879532, T6C, T6B); | |
1392 T6D = FNMS(KP923879532, T6C, T6B); | |
1393 T7a = FMA(KP303346683, T6W, T6X); | |
1394 T6Y = FNMS(KP303346683, T6X, T6W); | |
1395 T79 = FMA(KP1_662939224, T6U, T6T); | |
1396 T6V = FNMS(KP1_662939224, T6U, T6T); | |
1397 T70 = FMA(KP923879532, T6r, T6q); | |
1398 T6s = FNMS(KP923879532, T6r, T6q); | |
1399 T6y = FNMS(KP668178637, T6x, T6w); | |
1400 T6E = FMA(KP668178637, T6w, T6x); | |
1401 T6F = FNMS(KP668178637, T6t, T6u); | |
1402 T6v = FMA(KP668178637, T6u, T6t); | |
1403 T71 = T6E + T6F; | |
1404 T6G = T6E - T6F; | |
1405 } | |
1406 { | |
1407 E T72, T6O, T75, T6P; | |
1408 { | |
1409 E T6J, T6p, T7d, T6H, T74, T6z, T7b, T7f, T7c, T6A; | |
1410 T6J = FNMS(KP1_763842528, T6o, T6h); | |
1411 T6p = FMA(KP1_763842528, T6o, T6h); | |
1412 T72 = FNMS(KP831469612, T71, T70); | |
1413 T7d = FMA(KP831469612, T71, T70); | |
1414 T6O = FNMS(KP831469612, T6G, T6D); | |
1415 T6H = FMA(KP831469612, T6G, T6D); | |
1416 T74 = T6y + T6v; | |
1417 T6z = T6v - T6y; | |
1418 T7b = FNMS(KP1_913880671, T7a, T79); | |
1419 T7f = FMA(KP1_913880671, T7a, T79); | |
1420 T75 = FNMS(KP831469612, T74, T73); | |
1421 T7c = FMA(KP831469612, T74, T73); | |
1422 T6P = FNMS(KP831469612, T6z, T6s); | |
1423 T6A = FMA(KP831469612, T6z, T6s); | |
1424 { | |
1425 E T7e, T7g, T6K, T6I; | |
1426 T7e = FNMS(KP148335987, T7d, T7c); | |
1427 T7g = FMA(KP148335987, T7c, T7d); | |
1428 T6K = FMA(KP250486960, T6A, T6H); | |
1429 T6I = FNMS(KP250486960, T6H, T6A); | |
1430 R1[WS(rs, 62)] = FMA(KP1_978353019, T7g, T7f); | |
1431 R1[WS(rs, 30)] = FNMS(KP1_978353019, T7g, T7f); | |
1432 R1[WS(rs, 46)] = FMA(KP1_978353019, T7e, T7b); | |
1433 R1[WS(rs, 14)] = FNMS(KP1_978353019, T7e, T7b); | |
1434 R1[WS(rs, 2)] = FMA(KP1_940062506, T6I, T6p); | |
1435 R1[WS(rs, 34)] = FNMS(KP1_940062506, T6I, T6p); | |
1436 R1[WS(rs, 18)] = FNMS(KP1_940062506, T6K, T6J); | |
1437 R1[WS(rs, 50)] = FMA(KP1_940062506, T6K, T6J); | |
1438 } | |
1439 } | |
1440 { | |
1441 E T6N, T6Q, T6Z, T76; | |
1442 T6R = FMA(KP1_763842528, T6M, T6L); | |
1443 T6N = FNMS(KP1_763842528, T6M, T6L); | |
1444 T6Q = FNMS(KP599376933, T6P, T6O); | |
1445 T6S = FMA(KP599376933, T6O, T6P); | |
1446 T77 = FMA(KP1_913880671, T6Y, T6V); | |
1447 T6Z = FNMS(KP1_913880671, T6Y, T6V); | |
1448 T76 = FNMS(KP741650546, T75, T72); | |
1449 T78 = FMA(KP741650546, T72, T75); | |
1450 R1[WS(rs, 42)] = FMA(KP1_715457220, T6Q, T6N); | |
1451 R1[WS(rs, 10)] = FNMS(KP1_715457220, T6Q, T6N); | |
1452 R1[WS(rs, 6)] = FMA(KP1_606415062, T76, T6Z); | |
1453 R1[WS(rs, 38)] = FNMS(KP1_606415062, T76, T6Z); | |
1454 } | |
1455 } | |
1456 } | |
1457 { | |
1458 E T9B, T97, T9C, T9e, T9T, T9t, Ta0, T9O, T9Z, T9L, T9Q, T9i, T9l, T9o, T9R; | |
1459 E T9w; | |
1460 { | |
1461 E T9J, T9K, T9N, T9M, T9u, T9v; | |
1462 { | |
1463 E T93, T96, T9a, T9d; | |
1464 T9J = FMA(KP1_847759065, T92, T91); | |
1465 T93 = FNMS(KP1_847759065, T92, T91); | |
1466 R1[WS(rs, 54)] = FMA(KP1_606415062, T78, T77); | |
1467 R1[WS(rs, 22)] = FNMS(KP1_606415062, T78, T77); | |
1468 R1[WS(rs, 58)] = FMA(KP1_715457220, T6S, T6R); | |
1469 R1[WS(rs, 26)] = FNMS(KP1_715457220, T6S, T6R); | |
1470 T96 = FNMS(KP198912367, T95, T94); | |
1471 T9K = FMA(KP198912367, T94, T95); | |
1472 T9N = FMA(KP923879532, T99, T98); | |
1473 T9a = FNMS(KP923879532, T99, T98); | |
1474 T9d = FNMS(KP923879532, T9c, T9b); | |
1475 T9M = FMA(KP923879532, T9c, T9b); | |
1476 T9B = FMA(KP1_961570560, T96, T93); | |
1477 T97 = FNMS(KP1_961570560, T96, T93); | |
1478 T9C = FMA(KP820678790, T9a, T9d); | |
1479 T9e = FNMS(KP820678790, T9d, T9a); | |
1480 } | |
1481 T9T = FMA(KP923879532, T9s, T9r); | |
1482 T9t = FNMS(KP923879532, T9s, T9r); | |
1483 Ta0 = FMA(KP098491403, T9M, T9N); | |
1484 T9O = FNMS(KP098491403, T9N, T9M); | |
1485 T9Z = FMA(KP1_961570560, T9K, T9J); | |
1486 T9L = FNMS(KP1_961570560, T9K, T9J); | |
1487 T9Q = FMA(KP923879532, T9h, T9g); | |
1488 T9i = FNMS(KP923879532, T9h, T9g); | |
1489 T9l = FNMS(KP198912367, T9k, T9j); | |
1490 T9u = FMA(KP198912367, T9j, T9k); | |
1491 T9v = FMA(KP198912367, T9m, T9n); | |
1492 T9o = FNMS(KP198912367, T9n, T9m); | |
1493 T9R = T9u + T9v; | |
1494 T9w = T9u - T9v; | |
1495 } | |
1496 { | |
1497 E T9S, T9E, T9V, T9F; | |
1498 { | |
1499 E T9z, T9f, Ta3, T9x, T9U, T9p, Ta1, Ta5, Ta2, T9q; | |
1500 T9z = FNMS(KP1_546020906, T9e, T97); | |
1501 T9f = FMA(KP1_546020906, T9e, T97); | |
1502 T9S = FNMS(KP980785280, T9R, T9Q); | |
1503 Ta3 = FMA(KP980785280, T9R, T9Q); | |
1504 T9E = FNMS(KP980785280, T9w, T9t); | |
1505 T9x = FMA(KP980785280, T9w, T9t); | |
1506 T9U = T9l - T9o; | |
1507 T9p = T9l + T9o; | |
1508 Ta1 = FNMS(KP1_990369453, Ta0, T9Z); | |
1509 Ta5 = FMA(KP1_990369453, Ta0, T9Z); | |
1510 T9V = FNMS(KP980785280, T9U, T9T); | |
1511 Ta2 = FMA(KP980785280, T9U, T9T); | |
1512 T9F = FMA(KP980785280, T9p, T9i); | |
1513 T9q = FNMS(KP980785280, T9p, T9i); | |
1514 { | |
1515 E Ta4, Ta6, T9A, T9y; | |
1516 Ta4 = FNMS(KP049126849, Ta3, Ta2); | |
1517 Ta6 = FMA(KP049126849, Ta2, Ta3); | |
1518 T9A = FMA(KP357805721, T9q, T9x); | |
1519 T9y = FNMS(KP357805721, T9x, T9q); | |
1520 R1[WS(rs, 63)] = FMA(KP1_997590912, Ta6, Ta5); | |
1521 R1[WS(rs, 31)] = FNMS(KP1_997590912, Ta6, Ta5); | |
1522 R1[WS(rs, 47)] = FMA(KP1_997590912, Ta4, Ta1); | |
1523 R1[WS(rs, 15)] = FNMS(KP1_997590912, Ta4, Ta1); | |
1524 R1[WS(rs, 3)] = FMA(KP1_883088130, T9y, T9f); | |
1525 R1[WS(rs, 35)] = FNMS(KP1_883088130, T9y, T9f); | |
1526 R1[WS(rs, 19)] = FNMS(KP1_883088130, T9A, T9z); | |
1527 R1[WS(rs, 51)] = FMA(KP1_883088130, T9A, T9z); | |
1528 } | |
1529 } | |
1530 { | |
1531 E T9D, T9G, T9P, T9W; | |
1532 T9H = FMA(KP1_546020906, T9C, T9B); | |
1533 T9D = FNMS(KP1_546020906, T9C, T9B); | |
1534 T9G = FNMS(KP472964775, T9F, T9E); | |
1535 T9I = FMA(KP472964775, T9E, T9F); | |
1536 T9X = FMA(KP1_990369453, T9O, T9L); | |
1537 T9P = FNMS(KP1_990369453, T9O, T9L); | |
1538 T9W = FNMS(KP906347169, T9V, T9S); | |
1539 T9Y = FMA(KP906347169, T9S, T9V); | |
1540 R1[WS(rs, 43)] = FMA(KP1_807978586, T9G, T9D); | |
1541 R1[WS(rs, 11)] = FNMS(KP1_807978586, T9G, T9D); | |
1542 R1[WS(rs, 7)] = FMA(KP1_481902250, T9W, T9P); | |
1543 R1[WS(rs, 39)] = FNMS(KP1_481902250, T9W, T9P); | |
1544 } | |
1545 } | |
1546 } | |
1547 } | |
1548 } | |
1549 } | |
1550 } | |
1551 } | |
1552 R1[WS(rs, 55)] = FMA(KP1_481902250, T9Y, T9X); | |
1553 R1[WS(rs, 23)] = FNMS(KP1_481902250, T9Y, T9X); | |
1554 R1[WS(rs, 59)] = FMA(KP1_807978586, T9I, T9H); | |
1555 R1[WS(rs, 27)] = FNMS(KP1_807978586, T9I, T9H); | |
1556 } | |
1557 } | |
1558 } | |
1559 | |
1560 static const kr2c_desc desc = { 128, "r2cb_128", {416, 0, 540, 0}, &GENUS }; | |
1561 | |
1562 void X(codelet_r2cb_128) (planner *p) { | |
1563 X(kr2c_register) (p, r2cb_128, &desc); | |
1564 } | |
1565 | |
1566 #else /* HAVE_FMA */ | |
1567 | |
1568 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include r2cb.h */ | |
1569 | |
1570 /* | |
1571 * This function contains 956 FP additions, 342 FP multiplications, | |
1572 * (or, 812 additions, 198 multiplications, 144 fused multiply/add), | |
1573 * 198 stack variables, 39 constants, and 256 memory accesses | |
1574 */ | |
1575 #include "r2cb.h" | |
1576 | |
1577 static void r2cb_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
1578 { | |
1579 DK(KP1_028205488, +1.028205488386443453187387677937631545216098241); | |
1580 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599); | |
1581 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854); | |
1582 DK(KP1_191398608, +1.191398608984866686934073057659939779023852677); | |
1583 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924); | |
1584 DK(KP485960359, +0.485960359806527779896548324154942236641981567); | |
1585 DK(KP293460948, +0.293460948910723503317700259293435639412430633); | |
1586 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067); | |
1587 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
1588 DK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
1589 DK(KP855110186, +0.855110186860564188641933713777597068609157259); | |
1590 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050); | |
1591 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619); | |
1592 DK(KP1_343117909, +1.343117909694036801250753700854843606457501264); | |
1593 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592); | |
1594 DK(KP673779706, +0.673779706784440101378506425238295140955533559); | |
1595 DK(KP098135348, +0.098135348654836028509909953885365316629490726); | |
1596 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229); | |
1597 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
1598 DK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
1599 DK(KP580569354, +0.580569354508924735272384751634790549382952557); | |
1600 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411); | |
1601 DK(KP942793473, +0.942793473651995297112775251810508755314920638); | |
1602 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241); | |
1603 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382); | |
1604 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624); | |
1605 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190); | |
1606 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586); | |
1607 DK(KP196034280, +0.196034280659121203988391127777283691722273346); | |
1608 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737); | |
1609 DK(KP390180644, +0.390180644032256535696569736954044481855383236); | |
1610 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462); | |
1611 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
1612 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
1613 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
1614 DK(KP765366864, +0.765366864730179543456919968060797733522689125); | |
1615 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252); | |
1616 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875); | |
1617 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
1618 { | |
1619 INT i; | |
1620 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) { | |
1621 E Ta, T6q, T2a, T5k, T8x, Tbx, TcF, Ten, Th, T6r, T2j, T5l, T8E, Tby, TcI; | |
1622 E Teo, Tx, T6t, TcM, Teq, TcP, Ter, T2t, T5n, T2C, T5o, T8Q, TbA, T8X, TbB; | |
1623 E T6w, T7L, T1j, T6L, Tde, TeC, TdL, TeR, T3v, T5z, T4I, T5O, T9O, TbM, TaV; | |
1624 E Tc1, T78, T7Z, TN, T6z, TcU, Teu, Td8, Tey, T2N, T5r, T3j, T5v, T9a, TbE; | |
1625 E T9A, TbI, T6H, T7O, T1O, T7V, T48, T4u, Tds, TeG, T5E, T5K, Taf, TbP, Tdp; | |
1626 E TeF, T6U, T72, Tam, TbQ, T23, T7U, T4r, T4v, Tdz, TeJ, T5H, T5L, Tay, TbS; | |
1627 E Tdw, TeI, T6Z, T73, TaF, TbT, T1y, T75, Tdl, TeQ, TdI, TeD, T3O, T5N, T4z; | |
1628 E T5A, Ta3, Tc0, TaO, TbN, T6O, T80, T12, T6E, Td1, Tex, Td5, Tev, T36, T5u; | |
1629 E T3a, T5s, T9p, TbH, T9t, TbF, T6C, T7P; | |
1630 { | |
1631 E T5, T8s, T3, T8q, T9, T8u, T29, T8v, T6, T26; | |
1632 { | |
1633 E T4, T8r, T1, T2; | |
1634 T4 = Cr[WS(csr, 32)]; | |
1635 T5 = KP2_000000000 * T4; | |
1636 T8r = Ci[WS(csi, 32)]; | |
1637 T8s = KP2_000000000 * T8r; | |
1638 T1 = Cr[0]; | |
1639 T2 = Cr[WS(csr, 64)]; | |
1640 T3 = T1 + T2; | |
1641 T8q = T1 - T2; | |
1642 { | |
1643 E T7, T8, T27, T28; | |
1644 T7 = Cr[WS(csr, 16)]; | |
1645 T8 = Cr[WS(csr, 48)]; | |
1646 T9 = KP2_000000000 * (T7 + T8); | |
1647 T8u = T7 - T8; | |
1648 T27 = Ci[WS(csi, 16)]; | |
1649 T28 = Ci[WS(csi, 48)]; | |
1650 T29 = KP2_000000000 * (T27 - T28); | |
1651 T8v = T27 + T28; | |
1652 } | |
1653 } | |
1654 T6 = T3 + T5; | |
1655 Ta = T6 + T9; | |
1656 T6q = T6 - T9; | |
1657 T26 = T3 - T5; | |
1658 T2a = T26 - T29; | |
1659 T5k = T26 + T29; | |
1660 { | |
1661 E T8t, T8w, TcD, TcE; | |
1662 T8t = T8q - T8s; | |
1663 T8w = KP1_414213562 * (T8u - T8v); | |
1664 T8x = T8t + T8w; | |
1665 Tbx = T8t - T8w; | |
1666 TcD = T8q + T8s; | |
1667 TcE = KP1_414213562 * (T8u + T8v); | |
1668 TcF = TcD - TcE; | |
1669 Ten = TcD + TcE; | |
1670 } | |
1671 } | |
1672 { | |
1673 E Td, T8y, T2e, T8C, Tg, T8B, T2h, T8z, T2b, T2i; | |
1674 { | |
1675 E Tb, Tc, T2c, T2d; | |
1676 Tb = Cr[WS(csr, 8)]; | |
1677 Tc = Cr[WS(csr, 56)]; | |
1678 Td = Tb + Tc; | |
1679 T8y = Tb - Tc; | |
1680 T2c = Ci[WS(csi, 8)]; | |
1681 T2d = Ci[WS(csi, 56)]; | |
1682 T2e = T2c - T2d; | |
1683 T8C = T2c + T2d; | |
1684 } | |
1685 { | |
1686 E Te, Tf, T2f, T2g; | |
1687 Te = Cr[WS(csr, 40)]; | |
1688 Tf = Cr[WS(csr, 24)]; | |
1689 Tg = Te + Tf; | |
1690 T8B = Te - Tf; | |
1691 T2f = Ci[WS(csi, 40)]; | |
1692 T2g = Ci[WS(csi, 24)]; | |
1693 T2h = T2f - T2g; | |
1694 T8z = T2f + T2g; | |
1695 } | |
1696 Th = KP2_000000000 * (Td + Tg); | |
1697 T6r = KP2_000000000 * (T2h + T2e); | |
1698 T2b = Td - Tg; | |
1699 T2i = T2e - T2h; | |
1700 T2j = KP1_414213562 * (T2b - T2i); | |
1701 T5l = KP1_414213562 * (T2b + T2i); | |
1702 { | |
1703 E T8A, T8D, TcG, TcH; | |
1704 T8A = T8y - T8z; | |
1705 T8D = T8B + T8C; | |
1706 T8E = FNMS(KP765366864, T8D, KP1_847759065 * T8A); | |
1707 Tby = FMA(KP765366864, T8A, KP1_847759065 * T8D); | |
1708 TcG = T8y + T8z; | |
1709 TcH = T8C - T8B; | |
1710 TcI = FNMS(KP1_847759065, TcH, KP765366864 * TcG); | |
1711 Teo = FMA(KP1_847759065, TcG, KP765366864 * TcH); | |
1712 } | |
1713 } | |
1714 { | |
1715 E Tl, T8G, T2x, T8V, To, T8U, T2A, T8H, Tv, T8S, T2o, T8O, Ts, T8R, T2r; | |
1716 E T8L; | |
1717 { | |
1718 E Tj, Tk, T2y, T2z; | |
1719 Tj = Cr[WS(csr, 4)]; | |
1720 Tk = Cr[WS(csr, 60)]; | |
1721 Tl = Tj + Tk; | |
1722 T8G = Tj - Tk; | |
1723 { | |
1724 E T2v, T2w, Tm, Tn; | |
1725 T2v = Ci[WS(csi, 4)]; | |
1726 T2w = Ci[WS(csi, 60)]; | |
1727 T2x = T2v - T2w; | |
1728 T8V = T2v + T2w; | |
1729 Tm = Cr[WS(csr, 36)]; | |
1730 Tn = Cr[WS(csr, 28)]; | |
1731 To = Tm + Tn; | |
1732 T8U = Tm - Tn; | |
1733 } | |
1734 T2y = Ci[WS(csi, 36)]; | |
1735 T2z = Ci[WS(csi, 28)]; | |
1736 T2A = T2y - T2z; | |
1737 T8H = T2y + T2z; | |
1738 { | |
1739 E Tt, Tu, T8M, T2m, T2n, T8N; | |
1740 Tt = Cr[WS(csr, 12)]; | |
1741 Tu = Cr[WS(csr, 52)]; | |
1742 T8M = Tt - Tu; | |
1743 T2m = Ci[WS(csi, 52)]; | |
1744 T2n = Ci[WS(csi, 12)]; | |
1745 T8N = T2n + T2m; | |
1746 Tv = Tt + Tu; | |
1747 T8S = T8M + T8N; | |
1748 T2o = T2m - T2n; | |
1749 T8O = T8M - T8N; | |
1750 } | |
1751 { | |
1752 E Tq, Tr, T8J, T2p, T2q, T8K; | |
1753 Tq = Cr[WS(csr, 20)]; | |
1754 Tr = Cr[WS(csr, 44)]; | |
1755 T8J = Tq - Tr; | |
1756 T2p = Ci[WS(csi, 20)]; | |
1757 T2q = Ci[WS(csi, 44)]; | |
1758 T8K = T2p + T2q; | |
1759 Ts = Tq + Tr; | |
1760 T8R = T8J + T8K; | |
1761 T2r = T2p - T2q; | |
1762 T8L = T8J - T8K; | |
1763 } | |
1764 } | |
1765 { | |
1766 E Tp, Tw, TcK, TcL; | |
1767 Tp = Tl + To; | |
1768 Tw = Ts + Tv; | |
1769 Tx = KP2_000000000 * (Tp + Tw); | |
1770 T6t = Tp - Tw; | |
1771 TcK = T8G + T8H; | |
1772 TcL = KP707106781 * (T8R + T8S); | |
1773 TcM = TcK - TcL; | |
1774 Teq = TcK + TcL; | |
1775 } | |
1776 { | |
1777 E TcN, TcO, T2l, T2s; | |
1778 TcN = KP707106781 * (T8L - T8O); | |
1779 TcO = T8V - T8U; | |
1780 TcP = TcN + TcO; | |
1781 Ter = TcO - TcN; | |
1782 T2l = Tl - To; | |
1783 T2s = T2o - T2r; | |
1784 T2t = T2l + T2s; | |
1785 T5n = T2l - T2s; | |
1786 } | |
1787 { | |
1788 E T2u, T2B, T8I, T8P; | |
1789 T2u = Ts - Tv; | |
1790 T2B = T2x - T2A; | |
1791 T2C = T2u + T2B; | |
1792 T5o = T2B - T2u; | |
1793 T8I = T8G - T8H; | |
1794 T8P = KP707106781 * (T8L + T8O); | |
1795 T8Q = T8I + T8P; | |
1796 TbA = T8I - T8P; | |
1797 } | |
1798 { | |
1799 E T8T, T8W, T6u, T6v; | |
1800 T8T = KP707106781 * (T8R - T8S); | |
1801 T8W = T8U + T8V; | |
1802 T8X = T8T + T8W; | |
1803 TbB = T8W - T8T; | |
1804 T6u = T2A + T2x; | |
1805 T6v = T2r + T2o; | |
1806 T6w = T6u - T6v; | |
1807 T7L = KP2_000000000 * (T6v + T6u); | |
1808 } | |
1809 } | |
1810 { | |
1811 E T17, T9E, T4D, TaT, T1a, TaS, T4G, T9F, T1h, TaQ, T3q, T9M, T1e, TaP, T3t; | |
1812 E T9J; | |
1813 { | |
1814 E T15, T16, T4E, T4F; | |
1815 T15 = Cr[WS(csr, 1)]; | |
1816 T16 = Cr[WS(csr, 63)]; | |
1817 T17 = T15 + T16; | |
1818 T9E = T15 - T16; | |
1819 { | |
1820 E T4B, T4C, T18, T19; | |
1821 T4B = Ci[WS(csi, 1)]; | |
1822 T4C = Ci[WS(csi, 63)]; | |
1823 T4D = T4B - T4C; | |
1824 TaT = T4B + T4C; | |
1825 T18 = Cr[WS(csr, 33)]; | |
1826 T19 = Cr[WS(csr, 31)]; | |
1827 T1a = T18 + T19; | |
1828 TaS = T18 - T19; | |
1829 } | |
1830 T4E = Ci[WS(csi, 33)]; | |
1831 T4F = Ci[WS(csi, 31)]; | |
1832 T4G = T4E - T4F; | |
1833 T9F = T4E + T4F; | |
1834 { | |
1835 E T1f, T1g, T9K, T3o, T3p, T9L; | |
1836 T1f = Cr[WS(csr, 15)]; | |
1837 T1g = Cr[WS(csr, 49)]; | |
1838 T9K = T1f - T1g; | |
1839 T3o = Ci[WS(csi, 49)]; | |
1840 T3p = Ci[WS(csi, 15)]; | |
1841 T9L = T3p + T3o; | |
1842 T1h = T1f + T1g; | |
1843 TaQ = T9K + T9L; | |
1844 T3q = T3o - T3p; | |
1845 T9M = T9K - T9L; | |
1846 } | |
1847 { | |
1848 E T1c, T1d, T9H, T3r, T3s, T9I; | |
1849 T1c = Cr[WS(csr, 17)]; | |
1850 T1d = Cr[WS(csr, 47)]; | |
1851 T9H = T1c - T1d; | |
1852 T3r = Ci[WS(csi, 17)]; | |
1853 T3s = Ci[WS(csi, 47)]; | |
1854 T9I = T3r + T3s; | |
1855 T1e = T1c + T1d; | |
1856 TaP = T9H + T9I; | |
1857 T3t = T3r - T3s; | |
1858 T9J = T9H - T9I; | |
1859 } | |
1860 } | |
1861 { | |
1862 E T1b, T1i, Tdc, Tdd; | |
1863 T1b = T17 + T1a; | |
1864 T1i = T1e + T1h; | |
1865 T1j = T1b + T1i; | |
1866 T6L = T1b - T1i; | |
1867 Tdc = T9E + T9F; | |
1868 Tdd = KP707106781 * (TaP + TaQ); | |
1869 Tde = Tdc - Tdd; | |
1870 TeC = Tdc + Tdd; | |
1871 } | |
1872 { | |
1873 E TdJ, TdK, T3n, T3u; | |
1874 TdJ = KP707106781 * (T9J - T9M); | |
1875 TdK = TaT - TaS; | |
1876 TdL = TdJ + TdK; | |
1877 TeR = TdK - TdJ; | |
1878 T3n = T17 - T1a; | |
1879 T3u = T3q - T3t; | |
1880 T3v = T3n + T3u; | |
1881 T5z = T3n - T3u; | |
1882 } | |
1883 { | |
1884 E T4A, T4H, T9G, T9N; | |
1885 T4A = T1e - T1h; | |
1886 T4H = T4D - T4G; | |
1887 T4I = T4A + T4H; | |
1888 T5O = T4H - T4A; | |
1889 T9G = T9E - T9F; | |
1890 T9N = KP707106781 * (T9J + T9M); | |
1891 T9O = T9G + T9N; | |
1892 TbM = T9G - T9N; | |
1893 } | |
1894 { | |
1895 E TaR, TaU, T76, T77; | |
1896 TaR = KP707106781 * (TaP - TaQ); | |
1897 TaU = TaS + TaT; | |
1898 TaV = TaR + TaU; | |
1899 Tc1 = TaU - TaR; | |
1900 T76 = T4G + T4D; | |
1901 T77 = T3t + T3q; | |
1902 T78 = T76 - T77; | |
1903 T7Z = T77 + T76; | |
1904 } | |
1905 } | |
1906 { | |
1907 E TB, T90, T3e, T9y, TE, T9x, T3h, T91, TL, T9v, T2I, T98, TI, T9u, T2L; | |
1908 E T95; | |
1909 { | |
1910 E Tz, TA, T3f, T3g; | |
1911 Tz = Cr[WS(csr, 2)]; | |
1912 TA = Cr[WS(csr, 62)]; | |
1913 TB = Tz + TA; | |
1914 T90 = Tz - TA; | |
1915 { | |
1916 E T3c, T3d, TC, TD; | |
1917 T3c = Ci[WS(csi, 2)]; | |
1918 T3d = Ci[WS(csi, 62)]; | |
1919 T3e = T3c - T3d; | |
1920 T9y = T3c + T3d; | |
1921 TC = Cr[WS(csr, 34)]; | |
1922 TD = Cr[WS(csr, 30)]; | |
1923 TE = TC + TD; | |
1924 T9x = TC - TD; | |
1925 } | |
1926 T3f = Ci[WS(csi, 34)]; | |
1927 T3g = Ci[WS(csi, 30)]; | |
1928 T3h = T3f - T3g; | |
1929 T91 = T3f + T3g; | |
1930 { | |
1931 E TJ, TK, T96, T2G, T2H, T97; | |
1932 TJ = Cr[WS(csr, 14)]; | |
1933 TK = Cr[WS(csr, 50)]; | |
1934 T96 = TJ - TK; | |
1935 T2G = Ci[WS(csi, 50)]; | |
1936 T2H = Ci[WS(csi, 14)]; | |
1937 T97 = T2H + T2G; | |
1938 TL = TJ + TK; | |
1939 T9v = T96 + T97; | |
1940 T2I = T2G - T2H; | |
1941 T98 = T96 - T97; | |
1942 } | |
1943 { | |
1944 E TG, TH, T93, T2J, T2K, T94; | |
1945 TG = Cr[WS(csr, 18)]; | |
1946 TH = Cr[WS(csr, 46)]; | |
1947 T93 = TG - TH; | |
1948 T2J = Ci[WS(csi, 18)]; | |
1949 T2K = Ci[WS(csi, 46)]; | |
1950 T94 = T2J + T2K; | |
1951 TI = TG + TH; | |
1952 T9u = T93 + T94; | |
1953 T2L = T2J - T2K; | |
1954 T95 = T93 - T94; | |
1955 } | |
1956 } | |
1957 { | |
1958 E TF, TM, TcS, TcT; | |
1959 TF = TB + TE; | |
1960 TM = TI + TL; | |
1961 TN = TF + TM; | |
1962 T6z = TF - TM; | |
1963 TcS = T90 + T91; | |
1964 TcT = KP707106781 * (T9u + T9v); | |
1965 TcU = TcS - TcT; | |
1966 Teu = TcS + TcT; | |
1967 } | |
1968 { | |
1969 E Td6, Td7, T2F, T2M; | |
1970 Td6 = KP707106781 * (T95 - T98); | |
1971 Td7 = T9y - T9x; | |
1972 Td8 = Td6 + Td7; | |
1973 Tey = Td7 - Td6; | |
1974 T2F = TB - TE; | |
1975 T2M = T2I - T2L; | |
1976 T2N = T2F + T2M; | |
1977 T5r = T2F - T2M; | |
1978 } | |
1979 { | |
1980 E T3b, T3i, T92, T99; | |
1981 T3b = TI - TL; | |
1982 T3i = T3e - T3h; | |
1983 T3j = T3b + T3i; | |
1984 T5v = T3i - T3b; | |
1985 T92 = T90 - T91; | |
1986 T99 = KP707106781 * (T95 + T98); | |
1987 T9a = T92 + T99; | |
1988 TbE = T92 - T99; | |
1989 } | |
1990 { | |
1991 E T9w, T9z, T6F, T6G; | |
1992 T9w = KP707106781 * (T9u - T9v); | |
1993 T9z = T9x + T9y; | |
1994 T9A = T9w + T9z; | |
1995 TbI = T9z - T9w; | |
1996 T6F = T3h + T3e; | |
1997 T6G = T2L + T2I; | |
1998 T6H = T6F - T6G; | |
1999 T7O = T6G + T6F; | |
2000 } | |
2001 } | |
2002 { | |
2003 E T1G, Taj, T3Q, Ta5, T46, Tak, T6R, Ta6, T1N, Tag, Tah, T3X, T3Z, Taa, Tad; | |
2004 E T6S, Tdn, Tdo; | |
2005 { | |
2006 E T1A, T1B, T1C, T1D, T1E, T1F; | |
2007 T1A = Cr[WS(csr, 5)]; | |
2008 T1B = Cr[WS(csr, 59)]; | |
2009 T1C = T1A + T1B; | |
2010 T1D = Cr[WS(csr, 37)]; | |
2011 T1E = Cr[WS(csr, 27)]; | |
2012 T1F = T1D + T1E; | |
2013 T1G = T1C + T1F; | |
2014 Taj = T1D - T1E; | |
2015 T3Q = T1C - T1F; | |
2016 Ta5 = T1A - T1B; | |
2017 } | |
2018 { | |
2019 E T40, T41, T42, T43, T44, T45; | |
2020 T40 = Ci[WS(csi, 5)]; | |
2021 T41 = Ci[WS(csi, 59)]; | |
2022 T42 = T40 - T41; | |
2023 T43 = Ci[WS(csi, 37)]; | |
2024 T44 = Ci[WS(csi, 27)]; | |
2025 T45 = T43 - T44; | |
2026 T46 = T42 - T45; | |
2027 Tak = T40 + T41; | |
2028 T6R = T45 + T42; | |
2029 Ta6 = T43 + T44; | |
2030 } | |
2031 { | |
2032 E T1J, Ta8, T3W, Ta9, T1M, Tab, T3T, Tac; | |
2033 { | |
2034 E T1H, T1I, T3U, T3V; | |
2035 T1H = Cr[WS(csr, 21)]; | |
2036 T1I = Cr[WS(csr, 43)]; | |
2037 T1J = T1H + T1I; | |
2038 Ta8 = T1H - T1I; | |
2039 T3U = Ci[WS(csi, 21)]; | |
2040 T3V = Ci[WS(csi, 43)]; | |
2041 T3W = T3U - T3V; | |
2042 Ta9 = T3U + T3V; | |
2043 } | |
2044 { | |
2045 E T1K, T1L, T3R, T3S; | |
2046 T1K = Cr[WS(csr, 11)]; | |
2047 T1L = Cr[WS(csr, 53)]; | |
2048 T1M = T1K + T1L; | |
2049 Tab = T1K - T1L; | |
2050 T3R = Ci[WS(csi, 53)]; | |
2051 T3S = Ci[WS(csi, 11)]; | |
2052 T3T = T3R - T3S; | |
2053 Tac = T3S + T3R; | |
2054 } | |
2055 T1N = T1J + T1M; | |
2056 Tag = Ta8 + Ta9; | |
2057 Tah = Tab + Tac; | |
2058 T3X = T3T - T3W; | |
2059 T3Z = T1J - T1M; | |
2060 Taa = Ta8 - Ta9; | |
2061 Tad = Tab - Tac; | |
2062 T6S = T3W + T3T; | |
2063 } | |
2064 T1O = T1G + T1N; | |
2065 T7V = T6S + T6R; | |
2066 { | |
2067 E T3Y, T47, Tdq, Tdr; | |
2068 T3Y = T3Q + T3X; | |
2069 T47 = T3Z + T46; | |
2070 T48 = FNMS(KP382683432, T47, KP923879532 * T3Y); | |
2071 T4u = FMA(KP382683432, T3Y, KP923879532 * T47); | |
2072 Tdq = KP707106781 * (Taa - Tad); | |
2073 Tdr = Tak - Taj; | |
2074 Tds = Tdq + Tdr; | |
2075 TeG = Tdr - Tdq; | |
2076 } | |
2077 { | |
2078 E T5C, T5D, Ta7, Tae; | |
2079 T5C = T3Q - T3X; | |
2080 T5D = T46 - T3Z; | |
2081 T5E = FNMS(KP923879532, T5D, KP382683432 * T5C); | |
2082 T5K = FMA(KP923879532, T5C, KP382683432 * T5D); | |
2083 Ta7 = Ta5 - Ta6; | |
2084 Tae = KP707106781 * (Taa + Tad); | |
2085 Taf = Ta7 + Tae; | |
2086 TbP = Ta7 - Tae; | |
2087 } | |
2088 Tdn = Ta5 + Ta6; | |
2089 Tdo = KP707106781 * (Tag + Tah); | |
2090 Tdp = Tdn - Tdo; | |
2091 TeF = Tdn + Tdo; | |
2092 { | |
2093 E T6Q, T6T, Tai, Tal; | |
2094 T6Q = T1G - T1N; | |
2095 T6T = T6R - T6S; | |
2096 T6U = T6Q - T6T; | |
2097 T72 = T6Q + T6T; | |
2098 Tai = KP707106781 * (Tag - Tah); | |
2099 Tal = Taj + Tak; | |
2100 Tam = Tai + Tal; | |
2101 TbQ = Tal - Tai; | |
2102 } | |
2103 } | |
2104 { | |
2105 E T1V, TaC, T49, Tao, T4p, TaD, T6W, Tap, T22, Taz, TaA, T4g, T4i, Tat, Taw; | |
2106 E T6X, Tdu, Tdv; | |
2107 { | |
2108 E T1P, T1Q, T1R, T1S, T1T, T1U; | |
2109 T1P = Cr[WS(csr, 3)]; | |
2110 T1Q = Cr[WS(csr, 61)]; | |
2111 T1R = T1P + T1Q; | |
2112 T1S = Cr[WS(csr, 29)]; | |
2113 T1T = Cr[WS(csr, 35)]; | |
2114 T1U = T1S + T1T; | |
2115 T1V = T1R + T1U; | |
2116 TaC = T1S - T1T; | |
2117 T49 = T1R - T1U; | |
2118 Tao = T1P - T1Q; | |
2119 } | |
2120 { | |
2121 E T4j, T4k, T4l, T4m, T4n, T4o; | |
2122 T4j = Ci[WS(csi, 61)]; | |
2123 T4k = Ci[WS(csi, 3)]; | |
2124 T4l = T4j - T4k; | |
2125 T4m = Ci[WS(csi, 29)]; | |
2126 T4n = Ci[WS(csi, 35)]; | |
2127 T4o = T4m - T4n; | |
2128 T4p = T4l - T4o; | |
2129 TaD = T4k + T4j; | |
2130 T6W = T4o + T4l; | |
2131 Tap = T4m + T4n; | |
2132 } | |
2133 { | |
2134 E T1Y, Tar, T4f, Tas, T21, Tau, T4c, Tav; | |
2135 { | |
2136 E T1W, T1X, T4d, T4e; | |
2137 T1W = Cr[WS(csr, 13)]; | |
2138 T1X = Cr[WS(csr, 51)]; | |
2139 T1Y = T1W + T1X; | |
2140 Tar = T1W - T1X; | |
2141 T4d = Ci[WS(csi, 13)]; | |
2142 T4e = Ci[WS(csi, 51)]; | |
2143 T4f = T4d - T4e; | |
2144 Tas = T4d + T4e; | |
2145 } | |
2146 { | |
2147 E T1Z, T20, T4a, T4b; | |
2148 T1Z = Cr[WS(csr, 19)]; | |
2149 T20 = Cr[WS(csr, 45)]; | |
2150 T21 = T1Z + T20; | |
2151 Tau = T1Z - T20; | |
2152 T4a = Ci[WS(csi, 45)]; | |
2153 T4b = Ci[WS(csi, 19)]; | |
2154 T4c = T4a - T4b; | |
2155 Tav = T4b + T4a; | |
2156 } | |
2157 T22 = T1Y + T21; | |
2158 Taz = Tar + Tas; | |
2159 TaA = Tau + Tav; | |
2160 T4g = T4c - T4f; | |
2161 T4i = T1Y - T21; | |
2162 Tat = Tar - Tas; | |
2163 Taw = Tau - Tav; | |
2164 T6X = T4f + T4c; | |
2165 } | |
2166 T23 = T1V + T22; | |
2167 T7U = T6X + T6W; | |
2168 { | |
2169 E T4h, T4q, Tdx, Tdy; | |
2170 T4h = T49 + T4g; | |
2171 T4q = T4i + T4p; | |
2172 T4r = FMA(KP923879532, T4h, KP382683432 * T4q); | |
2173 T4v = FNMS(KP382683432, T4h, KP923879532 * T4q); | |
2174 Tdx = KP707106781 * (Tat - Taw); | |
2175 Tdy = TaC + TaD; | |
2176 Tdz = Tdx - Tdy; | |
2177 TeJ = Tdx + Tdy; | |
2178 } | |
2179 { | |
2180 E T5F, T5G, Taq, Tax; | |
2181 T5F = T49 - T4g; | |
2182 T5G = T4p - T4i; | |
2183 T5H = FMA(KP382683432, T5F, KP923879532 * T5G); | |
2184 T5L = FNMS(KP923879532, T5F, KP382683432 * T5G); | |
2185 Taq = Tao - Tap; | |
2186 Tax = KP707106781 * (Tat + Taw); | |
2187 Tay = Taq + Tax; | |
2188 TbS = Taq - Tax; | |
2189 } | |
2190 Tdu = Tao + Tap; | |
2191 Tdv = KP707106781 * (Taz + TaA); | |
2192 Tdw = Tdu - Tdv; | |
2193 TeI = Tdu + Tdv; | |
2194 { | |
2195 E T6V, T6Y, TaB, TaE; | |
2196 T6V = T1V - T22; | |
2197 T6Y = T6W - T6X; | |
2198 T6Z = T6V + T6Y; | |
2199 T73 = T6Y - T6V; | |
2200 TaB = KP707106781 * (Taz - TaA); | |
2201 TaE = TaC - TaD; | |
2202 TaF = TaB + TaE; | |
2203 TbT = TaE - TaB; | |
2204 } | |
2205 } | |
2206 { | |
2207 E T1m, T3z, T1p, T3C, T3w, T3D, Tdg, Tdf, T9U, T9R, T1t, T3I, T1w, T3L, T3F; | |
2208 E T3M, Tdj, Tdi, Ta1, T9Y; | |
2209 { | |
2210 E T9P, T9T, T9S, T9Q; | |
2211 { | |
2212 E T1k, T1l, T3x, T3y; | |
2213 T1k = Cr[WS(csr, 9)]; | |
2214 T1l = Cr[WS(csr, 55)]; | |
2215 T1m = T1k + T1l; | |
2216 T9P = T1k - T1l; | |
2217 T3x = Ci[WS(csi, 9)]; | |
2218 T3y = Ci[WS(csi, 55)]; | |
2219 T3z = T3x - T3y; | |
2220 T9T = T3x + T3y; | |
2221 } | |
2222 { | |
2223 E T1n, T1o, T3A, T3B; | |
2224 T1n = Cr[WS(csr, 41)]; | |
2225 T1o = Cr[WS(csr, 23)]; | |
2226 T1p = T1n + T1o; | |
2227 T9S = T1n - T1o; | |
2228 T3A = Ci[WS(csi, 41)]; | |
2229 T3B = Ci[WS(csi, 23)]; | |
2230 T3C = T3A - T3B; | |
2231 T9Q = T3A + T3B; | |
2232 } | |
2233 T3w = T1m - T1p; | |
2234 T3D = T3z - T3C; | |
2235 Tdg = T9T - T9S; | |
2236 Tdf = T9P + T9Q; | |
2237 T9U = T9S + T9T; | |
2238 T9R = T9P - T9Q; | |
2239 } | |
2240 { | |
2241 E T9W, Ta0, T9Z, T9X; | |
2242 { | |
2243 E T1r, T1s, T3G, T3H; | |
2244 T1r = Cr[WS(csr, 7)]; | |
2245 T1s = Cr[WS(csr, 57)]; | |
2246 T1t = T1r + T1s; | |
2247 T9W = T1r - T1s; | |
2248 T3G = Ci[WS(csi, 57)]; | |
2249 T3H = Ci[WS(csi, 7)]; | |
2250 T3I = T3G - T3H; | |
2251 Ta0 = T3H + T3G; | |
2252 } | |
2253 { | |
2254 E T1u, T1v, T3J, T3K; | |
2255 T1u = Cr[WS(csr, 25)]; | |
2256 T1v = Cr[WS(csr, 39)]; | |
2257 T1w = T1u + T1v; | |
2258 T9Z = T1u - T1v; | |
2259 T3J = Ci[WS(csi, 25)]; | |
2260 T3K = Ci[WS(csi, 39)]; | |
2261 T3L = T3J - T3K; | |
2262 T9X = T3J + T3K; | |
2263 } | |
2264 T3F = T1t - T1w; | |
2265 T3M = T3I - T3L; | |
2266 Tdj = T9Z + Ta0; | |
2267 Tdi = T9W + T9X; | |
2268 Ta1 = T9Z - Ta0; | |
2269 T9Y = T9W - T9X; | |
2270 } | |
2271 { | |
2272 E T1q, T1x, Tdh, Tdk; | |
2273 T1q = T1m + T1p; | |
2274 T1x = T1t + T1w; | |
2275 T1y = T1q + T1x; | |
2276 T75 = T1q - T1x; | |
2277 Tdh = FNMS(KP923879532, Tdg, KP382683432 * Tdf); | |
2278 Tdk = FNMS(KP923879532, Tdj, KP382683432 * Tdi); | |
2279 Tdl = Tdh + Tdk; | |
2280 TeQ = Tdh - Tdk; | |
2281 } | |
2282 { | |
2283 E TdG, TdH, T3E, T3N; | |
2284 TdG = FMA(KP923879532, Tdf, KP382683432 * Tdg); | |
2285 TdH = FMA(KP923879532, Tdi, KP382683432 * Tdj); | |
2286 TdI = TdG - TdH; | |
2287 TeD = TdG + TdH; | |
2288 T3E = T3w - T3D; | |
2289 T3N = T3F + T3M; | |
2290 T3O = KP707106781 * (T3E + T3N); | |
2291 T5N = KP707106781 * (T3E - T3N); | |
2292 } | |
2293 { | |
2294 E T4x, T4y, T9V, Ta2; | |
2295 T4x = T3w + T3D; | |
2296 T4y = T3M - T3F; | |
2297 T4z = KP707106781 * (T4x + T4y); | |
2298 T5A = KP707106781 * (T4y - T4x); | |
2299 T9V = FNMS(KP382683432, T9U, KP923879532 * T9R); | |
2300 Ta2 = FMA(KP923879532, T9Y, KP382683432 * Ta1); | |
2301 Ta3 = T9V + Ta2; | |
2302 Tc0 = T9V - Ta2; | |
2303 } | |
2304 { | |
2305 E TaM, TaN, T6M, T6N; | |
2306 TaM = FMA(KP382683432, T9R, KP923879532 * T9U); | |
2307 TaN = FNMS(KP382683432, T9Y, KP923879532 * Ta1); | |
2308 TaO = TaM + TaN; | |
2309 TbN = TaN - TaM; | |
2310 T6M = T3L + T3I; | |
2311 T6N = T3C + T3z; | |
2312 T6O = T6M - T6N; | |
2313 T80 = T6N + T6M; | |
2314 } | |
2315 } | |
2316 { | |
2317 E TQ, T2R, TT, T2U, T2O, T2V, TcW, TcV, T9g, T9d, TX, T30, T10, T33, T2X; | |
2318 E T34, TcZ, TcY, T9n, T9k; | |
2319 { | |
2320 E T9b, T9f, T9e, T9c; | |
2321 { | |
2322 E TO, TP, T2P, T2Q; | |
2323 TO = Cr[WS(csr, 10)]; | |
2324 TP = Cr[WS(csr, 54)]; | |
2325 TQ = TO + TP; | |
2326 T9b = TO - TP; | |
2327 T2P = Ci[WS(csi, 10)]; | |
2328 T2Q = Ci[WS(csi, 54)]; | |
2329 T2R = T2P - T2Q; | |
2330 T9f = T2P + T2Q; | |
2331 } | |
2332 { | |
2333 E TR, TS, T2S, T2T; | |
2334 TR = Cr[WS(csr, 42)]; | |
2335 TS = Cr[WS(csr, 22)]; | |
2336 TT = TR + TS; | |
2337 T9e = TR - TS; | |
2338 T2S = Ci[WS(csi, 42)]; | |
2339 T2T = Ci[WS(csi, 22)]; | |
2340 T2U = T2S - T2T; | |
2341 T9c = T2S + T2T; | |
2342 } | |
2343 T2O = TQ - TT; | |
2344 T2V = T2R - T2U; | |
2345 TcW = T9f - T9e; | |
2346 TcV = T9b + T9c; | |
2347 T9g = T9e + T9f; | |
2348 T9d = T9b - T9c; | |
2349 } | |
2350 { | |
2351 E T9i, T9m, T9l, T9j; | |
2352 { | |
2353 E TV, TW, T2Y, T2Z; | |
2354 TV = Cr[WS(csr, 6)]; | |
2355 TW = Cr[WS(csr, 58)]; | |
2356 TX = TV + TW; | |
2357 T9i = TV - TW; | |
2358 T2Y = Ci[WS(csi, 58)]; | |
2359 T2Z = Ci[WS(csi, 6)]; | |
2360 T30 = T2Y - T2Z; | |
2361 T9m = T2Z + T2Y; | |
2362 } | |
2363 { | |
2364 E TY, TZ, T31, T32; | |
2365 TY = Cr[WS(csr, 26)]; | |
2366 TZ = Cr[WS(csr, 38)]; | |
2367 T10 = TY + TZ; | |
2368 T9l = TY - TZ; | |
2369 T31 = Ci[WS(csi, 26)]; | |
2370 T32 = Ci[WS(csi, 38)]; | |
2371 T33 = T31 - T32; | |
2372 T9j = T31 + T32; | |
2373 } | |
2374 T2X = TX - T10; | |
2375 T34 = T30 - T33; | |
2376 TcZ = T9l + T9m; | |
2377 TcY = T9i + T9j; | |
2378 T9n = T9l - T9m; | |
2379 T9k = T9i - T9j; | |
2380 } | |
2381 { | |
2382 E TU, T11, TcX, Td0; | |
2383 TU = TQ + TT; | |
2384 T11 = TX + T10; | |
2385 T12 = TU + T11; | |
2386 T6E = TU - T11; | |
2387 TcX = FNMS(KP923879532, TcW, KP382683432 * TcV); | |
2388 Td0 = FNMS(KP923879532, TcZ, KP382683432 * TcY); | |
2389 Td1 = TcX + Td0; | |
2390 Tex = TcX - Td0; | |
2391 } | |
2392 { | |
2393 E Td3, Td4, T2W, T35; | |
2394 Td3 = FMA(KP923879532, TcV, KP382683432 * TcW); | |
2395 Td4 = FMA(KP923879532, TcY, KP382683432 * TcZ); | |
2396 Td5 = Td3 - Td4; | |
2397 Tev = Td3 + Td4; | |
2398 T2W = T2O - T2V; | |
2399 T35 = T2X + T34; | |
2400 T36 = KP707106781 * (T2W + T35); | |
2401 T5u = KP707106781 * (T2W - T35); | |
2402 } | |
2403 { | |
2404 E T38, T39, T9h, T9o; | |
2405 T38 = T2O + T2V; | |
2406 T39 = T34 - T2X; | |
2407 T3a = KP707106781 * (T38 + T39); | |
2408 T5s = KP707106781 * (T39 - T38); | |
2409 T9h = FNMS(KP382683432, T9g, KP923879532 * T9d); | |
2410 T9o = FMA(KP923879532, T9k, KP382683432 * T9n); | |
2411 T9p = T9h + T9o; | |
2412 TbH = T9h - T9o; | |
2413 } | |
2414 { | |
2415 E T9r, T9s, T6A, T6B; | |
2416 T9r = FMA(KP382683432, T9d, KP923879532 * T9g); | |
2417 T9s = FNMS(KP382683432, T9k, KP923879532 * T9n); | |
2418 T9t = T9r + T9s; | |
2419 TbF = T9s - T9r; | |
2420 T6A = T33 + T30; | |
2421 T6B = T2U + T2R; | |
2422 T6C = T6A - T6B; | |
2423 T7P = T6B + T6A; | |
2424 } | |
2425 } | |
2426 { | |
2427 E T13, T8f, Ty, T8e, T25, T8h, T8k, T8p, Ti, T14, T8o; | |
2428 T13 = KP2_000000000 * (TN + T12); | |
2429 T8f = KP2_000000000 * (T7P + T7O); | |
2430 Ti = Ta + Th; | |
2431 Ty = Ti + Tx; | |
2432 T8e = Ti - Tx; | |
2433 { | |
2434 E T1z, T24, T8i, T8j; | |
2435 T1z = T1j + T1y; | |
2436 T24 = T1O + T23; | |
2437 T25 = KP2_000000000 * (T1z + T24); | |
2438 T8h = T1z - T24; | |
2439 T8i = T80 + T7Z; | |
2440 T8j = T7V + T7U; | |
2441 T8k = T8i - T8j; | |
2442 T8p = KP2_000000000 * (T8j + T8i); | |
2443 } | |
2444 T14 = Ty + T13; | |
2445 R0[WS(rs, 32)] = T14 - T25; | |
2446 R0[0] = T14 + T25; | |
2447 T8o = Ty - T13; | |
2448 R0[WS(rs, 16)] = T8o - T8p; | |
2449 R0[WS(rs, 48)] = T8o + T8p; | |
2450 { | |
2451 E T8g, T8l, T8m, T8n; | |
2452 T8g = T8e - T8f; | |
2453 T8l = KP1_414213562 * (T8h - T8k); | |
2454 R0[WS(rs, 40)] = T8g - T8l; | |
2455 R0[WS(rs, 8)] = T8g + T8l; | |
2456 T8m = T8e + T8f; | |
2457 T8n = KP1_414213562 * (T8h + T8k); | |
2458 R0[WS(rs, 24)] = T8m - T8n; | |
2459 R0[WS(rs, 56)] = T8m + T8n; | |
2460 } | |
2461 } | |
2462 { | |
2463 E T7M, T86, T82, T8a, T7R, T87, T7X, T89, T7K, T7Y, T81; | |
2464 T7K = Ta - Th; | |
2465 T7M = T7K - T7L; | |
2466 T86 = T7K + T7L; | |
2467 T7Y = T1O - T23; | |
2468 T81 = T7Z - T80; | |
2469 T82 = T7Y + T81; | |
2470 T8a = T81 - T7Y; | |
2471 { | |
2472 E T7N, T7Q, T7T, T7W; | |
2473 T7N = TN - T12; | |
2474 T7Q = T7O - T7P; | |
2475 T7R = KP1_414213562 * (T7N - T7Q); | |
2476 T87 = KP1_414213562 * (T7N + T7Q); | |
2477 T7T = T1j - T1y; | |
2478 T7W = T7U - T7V; | |
2479 T7X = T7T + T7W; | |
2480 T89 = T7T - T7W; | |
2481 } | |
2482 { | |
2483 E T7S, T83, T8c, T8d; | |
2484 T7S = T7M + T7R; | |
2485 T83 = FNMS(KP765366864, T82, KP1_847759065 * T7X); | |
2486 R0[WS(rs, 36)] = T7S - T83; | |
2487 R0[WS(rs, 4)] = T7S + T83; | |
2488 T8c = T86 + T87; | |
2489 T8d = FMA(KP1_847759065, T89, KP765366864 * T8a); | |
2490 R0[WS(rs, 28)] = T8c - T8d; | |
2491 R0[WS(rs, 60)] = T8c + T8d; | |
2492 } | |
2493 { | |
2494 E T84, T85, T88, T8b; | |
2495 T84 = T7M - T7R; | |
2496 T85 = FMA(KP765366864, T7X, KP1_847759065 * T82); | |
2497 R0[WS(rs, 20)] = T84 - T85; | |
2498 R0[WS(rs, 52)] = T84 + T85; | |
2499 T88 = T86 - T87; | |
2500 T8b = FNMS(KP1_847759065, T8a, KP765366864 * T89); | |
2501 R0[WS(rs, 44)] = T88 - T8b; | |
2502 R0[WS(rs, 12)] = T88 + T8b; | |
2503 } | |
2504 } | |
2505 { | |
2506 E T2E, T4O, T4K, T4S, T3l, T4P, T4t, T4R; | |
2507 { | |
2508 E T2k, T2D, T4w, T4J; | |
2509 T2k = T2a + T2j; | |
2510 T2D = FNMS(KP765366864, T2C, KP1_847759065 * T2t); | |
2511 T2E = T2k + T2D; | |
2512 T4O = T2k - T2D; | |
2513 T4w = T4u + T4v; | |
2514 T4J = T4z + T4I; | |
2515 T4K = T4w + T4J; | |
2516 T4S = T4J - T4w; | |
2517 } | |
2518 { | |
2519 E T37, T3k, T3P, T4s; | |
2520 T37 = T2N + T36; | |
2521 T3k = T3a + T3j; | |
2522 T3l = FNMS(KP390180644, T3k, KP1_961570560 * T37); | |
2523 T4P = FMA(KP390180644, T37, KP1_961570560 * T3k); | |
2524 T3P = T3v + T3O; | |
2525 T4s = T48 + T4r; | |
2526 T4t = T3P + T4s; | |
2527 T4R = T3P - T4s; | |
2528 } | |
2529 { | |
2530 E T3m, T4L, T4U, T4V; | |
2531 T3m = T2E + T3l; | |
2532 T4L = FNMS(KP196034280, T4K, KP1_990369453 * T4t); | |
2533 R0[WS(rs, 33)] = T3m - T4L; | |
2534 R0[WS(rs, 1)] = T3m + T4L; | |
2535 T4U = T4O + T4P; | |
2536 T4V = FMA(KP1_546020906, T4R, KP1_268786568 * T4S); | |
2537 R0[WS(rs, 25)] = T4U - T4V; | |
2538 R0[WS(rs, 57)] = T4U + T4V; | |
2539 } | |
2540 { | |
2541 E T4M, T4N, T4Q, T4T; | |
2542 T4M = T2E - T3l; | |
2543 T4N = FMA(KP196034280, T4t, KP1_990369453 * T4K); | |
2544 R0[WS(rs, 17)] = T4M - T4N; | |
2545 R0[WS(rs, 49)] = T4M + T4N; | |
2546 T4Q = T4O - T4P; | |
2547 T4T = FNMS(KP1_546020906, T4S, KP1_268786568 * T4R); | |
2548 R0[WS(rs, 41)] = T4Q - T4T; | |
2549 R0[WS(rs, 9)] = T4Q + T4T; | |
2550 } | |
2551 } | |
2552 { | |
2553 E T6y, T7e, T7a, T7i, T6J, T7f, T71, T7h; | |
2554 { | |
2555 E T6s, T6x, T74, T79; | |
2556 T6s = T6q - T6r; | |
2557 T6x = KP1_414213562 * (T6t - T6w); | |
2558 T6y = T6s + T6x; | |
2559 T7e = T6s - T6x; | |
2560 T74 = KP707106781 * (T72 + T73); | |
2561 T79 = T75 + T78; | |
2562 T7a = T74 + T79; | |
2563 T7i = T79 - T74; | |
2564 } | |
2565 { | |
2566 E T6D, T6I, T6P, T70; | |
2567 T6D = T6z + T6C; | |
2568 T6I = T6E + T6H; | |
2569 T6J = FNMS(KP765366864, T6I, KP1_847759065 * T6D); | |
2570 T7f = FMA(KP765366864, T6D, KP1_847759065 * T6I); | |
2571 T6P = T6L + T6O; | |
2572 T70 = KP707106781 * (T6U + T6Z); | |
2573 T71 = T6P + T70; | |
2574 T7h = T6P - T70; | |
2575 } | |
2576 { | |
2577 E T6K, T7b, T7k, T7l; | |
2578 T6K = T6y + T6J; | |
2579 T7b = FNMS(KP390180644, T7a, KP1_961570560 * T71); | |
2580 R0[WS(rs, 34)] = T6K - T7b; | |
2581 R0[WS(rs, 2)] = T6K + T7b; | |
2582 T7k = T7e + T7f; | |
2583 T7l = FMA(KP1_662939224, T7h, KP1_111140466 * T7i); | |
2584 R0[WS(rs, 26)] = T7k - T7l; | |
2585 R0[WS(rs, 58)] = T7k + T7l; | |
2586 } | |
2587 { | |
2588 E T7c, T7d, T7g, T7j; | |
2589 T7c = T6y - T6J; | |
2590 T7d = FMA(KP390180644, T71, KP1_961570560 * T7a); | |
2591 R0[WS(rs, 18)] = T7c - T7d; | |
2592 R0[WS(rs, 50)] = T7c + T7d; | |
2593 T7g = T7e - T7f; | |
2594 T7j = FNMS(KP1_662939224, T7i, KP1_111140466 * T7h); | |
2595 R0[WS(rs, 42)] = T7g - T7j; | |
2596 R0[WS(rs, 10)] = T7g + T7j; | |
2597 } | |
2598 } | |
2599 { | |
2600 E T4Y, T5c, T58, T5g, T51, T5d, T55, T5f; | |
2601 { | |
2602 E T4W, T4X, T56, T57; | |
2603 T4W = T2a - T2j; | |
2604 T4X = FMA(KP765366864, T2t, KP1_847759065 * T2C); | |
2605 T4Y = T4W - T4X; | |
2606 T5c = T4W + T4X; | |
2607 T56 = T48 - T4r; | |
2608 T57 = T4I - T4z; | |
2609 T58 = T56 + T57; | |
2610 T5g = T57 - T56; | |
2611 } | |
2612 { | |
2613 E T4Z, T50, T53, T54; | |
2614 T4Z = T2N - T36; | |
2615 T50 = T3j - T3a; | |
2616 T51 = FNMS(KP1_662939224, T50, KP1_111140466 * T4Z); | |
2617 T5d = FMA(KP1_662939224, T4Z, KP1_111140466 * T50); | |
2618 T53 = T3v - T3O; | |
2619 T54 = T4v - T4u; | |
2620 T55 = T53 + T54; | |
2621 T5f = T53 - T54; | |
2622 } | |
2623 { | |
2624 E T52, T59, T5i, T5j; | |
2625 T52 = T4Y + T51; | |
2626 T59 = FNMS(KP942793473, T58, KP1_763842528 * T55); | |
2627 R0[WS(rs, 37)] = T52 - T59; | |
2628 R0[WS(rs, 5)] = T52 + T59; | |
2629 T5i = T5c + T5d; | |
2630 T5j = FMA(KP1_913880671, T5f, KP580569354 * T5g); | |
2631 R0[WS(rs, 29)] = T5i - T5j; | |
2632 R0[WS(rs, 61)] = T5i + T5j; | |
2633 } | |
2634 { | |
2635 E T5a, T5b, T5e, T5h; | |
2636 T5a = T4Y - T51; | |
2637 T5b = FMA(KP942793473, T55, KP1_763842528 * T58); | |
2638 R0[WS(rs, 21)] = T5a - T5b; | |
2639 R0[WS(rs, 53)] = T5a + T5b; | |
2640 T5e = T5c - T5d; | |
2641 T5h = FNMS(KP1_913880671, T5g, KP580569354 * T5f); | |
2642 R0[WS(rs, 45)] = T5e - T5h; | |
2643 R0[WS(rs, 13)] = T5e + T5h; | |
2644 } | |
2645 } | |
2646 { | |
2647 E T7o, T7C, T7y, T7G, T7r, T7D, T7v, T7F; | |
2648 { | |
2649 E T7m, T7n, T7w, T7x; | |
2650 T7m = T6q + T6r; | |
2651 T7n = KP1_414213562 * (T6t + T6w); | |
2652 T7o = T7m - T7n; | |
2653 T7C = T7m + T7n; | |
2654 T7w = KP707106781 * (T6U - T6Z); | |
2655 T7x = T78 - T75; | |
2656 T7y = T7w + T7x; | |
2657 T7G = T7x - T7w; | |
2658 } | |
2659 { | |
2660 E T7p, T7q, T7t, T7u; | |
2661 T7p = T6z - T6C; | |
2662 T7q = T6H - T6E; | |
2663 T7r = FNMS(KP1_847759065, T7q, KP765366864 * T7p); | |
2664 T7D = FMA(KP1_847759065, T7p, KP765366864 * T7q); | |
2665 T7t = T6L - T6O; | |
2666 T7u = KP707106781 * (T73 - T72); | |
2667 T7v = T7t + T7u; | |
2668 T7F = T7t - T7u; | |
2669 } | |
2670 { | |
2671 E T7s, T7z, T7I, T7J; | |
2672 T7s = T7o + T7r; | |
2673 T7z = FNMS(KP1_111140466, T7y, KP1_662939224 * T7v); | |
2674 R0[WS(rs, 38)] = T7s - T7z; | |
2675 R0[WS(rs, 6)] = T7s + T7z; | |
2676 T7I = T7C + T7D; | |
2677 T7J = FMA(KP1_961570560, T7F, KP390180644 * T7G); | |
2678 R0[WS(rs, 30)] = T7I - T7J; | |
2679 R0[WS(rs, 62)] = T7I + T7J; | |
2680 } | |
2681 { | |
2682 E T7A, T7B, T7E, T7H; | |
2683 T7A = T7o - T7r; | |
2684 T7B = FMA(KP1_111140466, T7v, KP1_662939224 * T7y); | |
2685 R0[WS(rs, 22)] = T7A - T7B; | |
2686 R0[WS(rs, 54)] = T7A + T7B; | |
2687 T7E = T7C - T7D; | |
2688 T7H = FNMS(KP1_961570560, T7G, KP390180644 * T7F); | |
2689 R0[WS(rs, 46)] = T7E - T7H; | |
2690 R0[WS(rs, 14)] = T7E + T7H; | |
2691 } | |
2692 } | |
2693 { | |
2694 E T5q, T5U, T5Q, T5Y, T5x, T5V, T5J, T5X; | |
2695 { | |
2696 E T5m, T5p, T5M, T5P; | |
2697 T5m = T5k - T5l; | |
2698 T5p = FNMS(KP1_847759065, T5o, KP765366864 * T5n); | |
2699 T5q = T5m + T5p; | |
2700 T5U = T5m - T5p; | |
2701 T5M = T5K + T5L; | |
2702 T5P = T5N + T5O; | |
2703 T5Q = T5M + T5P; | |
2704 T5Y = T5P - T5M; | |
2705 } | |
2706 { | |
2707 E T5t, T5w, T5B, T5I; | |
2708 T5t = T5r + T5s; | |
2709 T5w = T5u + T5v; | |
2710 T5x = FNMS(KP1_111140466, T5w, KP1_662939224 * T5t); | |
2711 T5V = FMA(KP1_111140466, T5t, KP1_662939224 * T5w); | |
2712 T5B = T5z + T5A; | |
2713 T5I = T5E + T5H; | |
2714 T5J = T5B + T5I; | |
2715 T5X = T5B - T5I; | |
2716 } | |
2717 { | |
2718 E T5y, T5R, T60, T61; | |
2719 T5y = T5q + T5x; | |
2720 T5R = FNMS(KP580569354, T5Q, KP1_913880671 * T5J); | |
2721 R0[WS(rs, 35)] = T5y - T5R; | |
2722 R0[WS(rs, 3)] = T5y + T5R; | |
2723 T60 = T5U + T5V; | |
2724 T61 = FMA(KP1_763842528, T5X, KP942793473 * T5Y); | |
2725 R0[WS(rs, 27)] = T60 - T61; | |
2726 R0[WS(rs, 59)] = T60 + T61; | |
2727 } | |
2728 { | |
2729 E T5S, T5T, T5W, T5Z; | |
2730 T5S = T5q - T5x; | |
2731 T5T = FMA(KP580569354, T5J, KP1_913880671 * T5Q); | |
2732 R0[WS(rs, 19)] = T5S - T5T; | |
2733 R0[WS(rs, 51)] = T5S + T5T; | |
2734 T5W = T5U - T5V; | |
2735 T5Z = FNMS(KP1_763842528, T5Y, KP942793473 * T5X); | |
2736 R0[WS(rs, 43)] = T5W - T5Z; | |
2737 R0[WS(rs, 11)] = T5W + T5Z; | |
2738 } | |
2739 } | |
2740 { | |
2741 E T64, T6i, T6e, T6m, T67, T6j, T6b, T6l; | |
2742 { | |
2743 E T62, T63, T6c, T6d; | |
2744 T62 = T5k + T5l; | |
2745 T63 = FMA(KP1_847759065, T5n, KP765366864 * T5o); | |
2746 T64 = T62 - T63; | |
2747 T6i = T62 + T63; | |
2748 T6c = T5E - T5H; | |
2749 T6d = T5O - T5N; | |
2750 T6e = T6c + T6d; | |
2751 T6m = T6d - T6c; | |
2752 } | |
2753 { | |
2754 E T65, T66, T69, T6a; | |
2755 T65 = T5r - T5s; | |
2756 T66 = T5v - T5u; | |
2757 T67 = FNMS(KP1_961570560, T66, KP390180644 * T65); | |
2758 T6j = FMA(KP1_961570560, T65, KP390180644 * T66); | |
2759 T69 = T5z - T5A; | |
2760 T6a = T5L - T5K; | |
2761 T6b = T69 + T6a; | |
2762 T6l = T69 - T6a; | |
2763 } | |
2764 { | |
2765 E T68, T6f, T6o, T6p; | |
2766 T68 = T64 + T67; | |
2767 T6f = FNMS(KP1_268786568, T6e, KP1_546020906 * T6b); | |
2768 R0[WS(rs, 39)] = T68 - T6f; | |
2769 R0[WS(rs, 7)] = T68 + T6f; | |
2770 T6o = T6i + T6j; | |
2771 T6p = FMA(KP1_990369453, T6l, KP196034280 * T6m); | |
2772 R0[WS(rs, 31)] = T6o - T6p; | |
2773 R0[WS(rs, 63)] = T6o + T6p; | |
2774 } | |
2775 { | |
2776 E T6g, T6h, T6k, T6n; | |
2777 T6g = T64 - T67; | |
2778 T6h = FMA(KP1_268786568, T6b, KP1_546020906 * T6e); | |
2779 R0[WS(rs, 23)] = T6g - T6h; | |
2780 R0[WS(rs, 55)] = T6g + T6h; | |
2781 T6k = T6i - T6j; | |
2782 T6n = FNMS(KP1_990369453, T6m, KP196034280 * T6l); | |
2783 R0[WS(rs, 47)] = T6k - T6n; | |
2784 R0[WS(rs, 15)] = T6k + T6n; | |
2785 } | |
2786 } | |
2787 { | |
2788 E T8Z, Tb1, T9C, Tb2, Tbe, Tbq, Tbb, Tbp, TaX, Tbs, Tb5, Tbi, TaI, Tbt, Tb4; | |
2789 E Tbl; | |
2790 { | |
2791 E T8F, T8Y, Tb9, Tba; | |
2792 T8F = T8x + T8E; | |
2793 T8Y = FNMS(KP390180644, T8X, KP1_961570560 * T8Q); | |
2794 T8Z = T8F + T8Y; | |
2795 Tb1 = T8F - T8Y; | |
2796 { | |
2797 E T9q, T9B, Tbc, Tbd; | |
2798 T9q = T9a + T9p; | |
2799 T9B = T9t + T9A; | |
2800 T9C = FNMS(KP196034280, T9B, KP1_990369453 * T9q); | |
2801 Tb2 = FMA(KP196034280, T9q, KP1_990369453 * T9B); | |
2802 Tbc = T9a - T9p; | |
2803 Tbd = T9A - T9t; | |
2804 Tbe = FNMS(KP1_546020906, Tbd, KP1_268786568 * Tbc); | |
2805 Tbq = FMA(KP1_546020906, Tbc, KP1_268786568 * Tbd); | |
2806 } | |
2807 Tb9 = T8x - T8E; | |
2808 Tba = FMA(KP390180644, T8Q, KP1_961570560 * T8X); | |
2809 Tbb = Tb9 - Tba; | |
2810 Tbp = Tb9 + Tba; | |
2811 { | |
2812 E TaW, Tbg, TaL, Tbh, TaJ, TaK; | |
2813 TaW = TaO + TaV; | |
2814 Tbg = T9O - Ta3; | |
2815 TaJ = FMA(KP195090322, Taf, KP980785280 * Tam); | |
2816 TaK = FNMS(KP195090322, Tay, KP980785280 * TaF); | |
2817 TaL = TaJ + TaK; | |
2818 Tbh = TaK - TaJ; | |
2819 TaX = TaL + TaW; | |
2820 Tbs = Tbg - Tbh; | |
2821 Tb5 = TaW - TaL; | |
2822 Tbi = Tbg + Tbh; | |
2823 } | |
2824 { | |
2825 E Ta4, Tbk, TaH, Tbj, Tan, TaG; | |
2826 Ta4 = T9O + Ta3; | |
2827 Tbk = TaV - TaO; | |
2828 Tan = FNMS(KP195090322, Tam, KP980785280 * Taf); | |
2829 TaG = FMA(KP980785280, Tay, KP195090322 * TaF); | |
2830 TaH = Tan + TaG; | |
2831 Tbj = Tan - TaG; | |
2832 TaI = Ta4 + TaH; | |
2833 Tbt = Tbk - Tbj; | |
2834 Tb4 = Ta4 - TaH; | |
2835 Tbl = Tbj + Tbk; | |
2836 } | |
2837 } | |
2838 { | |
2839 E T9D, TaY, Tbr, Tbu; | |
2840 T9D = T8Z + T9C; | |
2841 TaY = FNMS(KP098135348, TaX, KP1_997590912 * TaI); | |
2842 R1[WS(rs, 32)] = T9D - TaY; | |
2843 R1[0] = T9D + TaY; | |
2844 Tbr = Tbp - Tbq; | |
2845 Tbu = FNMS(KP1_883088130, Tbt, KP673779706 * Tbs); | |
2846 R1[WS(rs, 44)] = Tbr - Tbu; | |
2847 R1[WS(rs, 12)] = Tbr + Tbu; | |
2848 } | |
2849 { | |
2850 E Tbv, Tbw, TaZ, Tb0; | |
2851 Tbv = Tbp + Tbq; | |
2852 Tbw = FMA(KP1_883088130, Tbs, KP673779706 * Tbt); | |
2853 R1[WS(rs, 28)] = Tbv - Tbw; | |
2854 R1[WS(rs, 60)] = Tbv + Tbw; | |
2855 TaZ = T8Z - T9C; | |
2856 Tb0 = FMA(KP098135348, TaI, KP1_997590912 * TaX); | |
2857 R1[WS(rs, 16)] = TaZ - Tb0; | |
2858 R1[WS(rs, 48)] = TaZ + Tb0; | |
2859 } | |
2860 { | |
2861 E Tb3, Tb6, Tbf, Tbm; | |
2862 Tb3 = Tb1 - Tb2; | |
2863 Tb6 = FNMS(KP1_481902250, Tb5, KP1_343117909 * Tb4); | |
2864 R1[WS(rs, 40)] = Tb3 - Tb6; | |
2865 R1[WS(rs, 8)] = Tb3 + Tb6; | |
2866 Tbf = Tbb + Tbe; | |
2867 Tbm = FNMS(KP855110186, Tbl, KP1_807978586 * Tbi); | |
2868 R1[WS(rs, 36)] = Tbf - Tbm; | |
2869 R1[WS(rs, 4)] = Tbf + Tbm; | |
2870 } | |
2871 { | |
2872 E Tbn, Tbo, Tb7, Tb8; | |
2873 Tbn = Tbb - Tbe; | |
2874 Tbo = FMA(KP855110186, Tbi, KP1_807978586 * Tbl); | |
2875 R1[WS(rs, 20)] = Tbn - Tbo; | |
2876 R1[WS(rs, 52)] = Tbn + Tbo; | |
2877 Tb7 = Tb1 + Tb2; | |
2878 Tb8 = FMA(KP1_481902250, Tb4, KP1_343117909 * Tb5); | |
2879 R1[WS(rs, 24)] = Tb7 - Tb8; | |
2880 R1[WS(rs, 56)] = Tb7 + Tb8; | |
2881 } | |
2882 } | |
2883 { | |
2884 E TcR, TdR, Tda, TdS, Te4, Teg, Te1, Tef, TdN, Tei, TdV, Te8, TdC, Tej, TdU; | |
2885 E Teb; | |
2886 { | |
2887 E TcJ, TcQ, TdZ, Te0; | |
2888 TcJ = TcF + TcI; | |
2889 TcQ = FNMS(KP1_111140466, TcP, KP1_662939224 * TcM); | |
2890 TcR = TcJ + TcQ; | |
2891 TdR = TcJ - TcQ; | |
2892 { | |
2893 E Td2, Td9, Te2, Te3; | |
2894 Td2 = TcU + Td1; | |
2895 Td9 = Td5 + Td8; | |
2896 Tda = FNMS(KP580569354, Td9, KP1_913880671 * Td2); | |
2897 TdS = FMA(KP580569354, Td2, KP1_913880671 * Td9); | |
2898 Te2 = TcU - Td1; | |
2899 Te3 = Td8 - Td5; | |
2900 Te4 = FNMS(KP1_763842528, Te3, KP942793473 * Te2); | |
2901 Teg = FMA(KP1_763842528, Te2, KP942793473 * Te3); | |
2902 } | |
2903 TdZ = TcF - TcI; | |
2904 Te0 = FMA(KP1_111140466, TcM, KP1_662939224 * TcP); | |
2905 Te1 = TdZ - Te0; | |
2906 Tef = TdZ + Te0; | |
2907 { | |
2908 E TdM, Te6, TdF, Te7, TdD, TdE; | |
2909 TdM = TdI + TdL; | |
2910 Te6 = Tde - Tdl; | |
2911 TdD = FMA(KP555570233, Tdp, KP831469612 * Tds); | |
2912 TdE = FNMS(KP555570233, Tdw, KP831469612 * Tdz); | |
2913 TdF = TdD + TdE; | |
2914 Te7 = TdE - TdD; | |
2915 TdN = TdF + TdM; | |
2916 Tei = Te6 - Te7; | |
2917 TdV = TdM - TdF; | |
2918 Te8 = Te6 + Te7; | |
2919 } | |
2920 { | |
2921 E Tdm, Tea, TdB, Te9, Tdt, TdA; | |
2922 Tdm = Tde + Tdl; | |
2923 Tea = TdL - TdI; | |
2924 Tdt = FNMS(KP555570233, Tds, KP831469612 * Tdp); | |
2925 TdA = FMA(KP831469612, Tdw, KP555570233 * Tdz); | |
2926 TdB = Tdt + TdA; | |
2927 Te9 = Tdt - TdA; | |
2928 TdC = Tdm + TdB; | |
2929 Tej = Tea - Te9; | |
2930 TdU = Tdm - TdB; | |
2931 Teb = Te9 + Tea; | |
2932 } | |
2933 } | |
2934 { | |
2935 E Tdb, TdO, Teh, Tek; | |
2936 Tdb = TcR + Tda; | |
2937 TdO = FNMS(KP293460948, TdN, KP1_978353019 * TdC); | |
2938 R1[WS(rs, 33)] = Tdb - TdO; | |
2939 R1[WS(rs, 1)] = Tdb + TdO; | |
2940 Teh = Tef - Teg; | |
2941 Tek = FNMS(KP1_940062506, Tej, KP485960359 * Tei); | |
2942 R1[WS(rs, 45)] = Teh - Tek; | |
2943 R1[WS(rs, 13)] = Teh + Tek; | |
2944 } | |
2945 { | |
2946 E Tel, Tem, TdP, TdQ; | |
2947 Tel = Tef + Teg; | |
2948 Tem = FMA(KP1_940062506, Tei, KP485960359 * Tej); | |
2949 R1[WS(rs, 29)] = Tel - Tem; | |
2950 R1[WS(rs, 61)] = Tel + Tem; | |
2951 TdP = TcR - Tda; | |
2952 TdQ = FMA(KP293460948, TdC, KP1_978353019 * TdN); | |
2953 R1[WS(rs, 17)] = TdP - TdQ; | |
2954 R1[WS(rs, 49)] = TdP + TdQ; | |
2955 } | |
2956 { | |
2957 E TdT, TdW, Te5, Tec; | |
2958 TdT = TdR - TdS; | |
2959 TdW = FNMS(KP1_606415062, TdV, KP1_191398608 * TdU); | |
2960 R1[WS(rs, 41)] = TdT - TdW; | |
2961 R1[WS(rs, 9)] = TdT + TdW; | |
2962 Te5 = Te1 + Te4; | |
2963 Tec = FNMS(KP1_028205488, Teb, KP1_715457220 * Te8); | |
2964 R1[WS(rs, 37)] = Te5 - Tec; | |
2965 R1[WS(rs, 5)] = Te5 + Tec; | |
2966 } | |
2967 { | |
2968 E Ted, Tee, TdX, TdY; | |
2969 Ted = Te1 - Te4; | |
2970 Tee = FMA(KP1_028205488, Te8, KP1_715457220 * Teb); | |
2971 R1[WS(rs, 21)] = Ted - Tee; | |
2972 R1[WS(rs, 53)] = Ted + Tee; | |
2973 TdX = TdR + TdS; | |
2974 TdY = FMA(KP1_606415062, TdU, KP1_191398608 * TdV); | |
2975 R1[WS(rs, 25)] = TdX - TdY; | |
2976 R1[WS(rs, 57)] = TdX + TdY; | |
2977 } | |
2978 } | |
2979 { | |
2980 E TbD, Tc7, TbK, Tc8, Tck, Tcw, Tch, Tcv, Tc3, Tcy, Tcb, Tco, TbW, Tcz, Tca; | |
2981 E Tcr; | |
2982 { | |
2983 E Tbz, TbC, Tcf, Tcg; | |
2984 Tbz = Tbx - Tby; | |
2985 TbC = FNMS(KP1_662939224, TbB, KP1_111140466 * TbA); | |
2986 TbD = Tbz + TbC; | |
2987 Tc7 = Tbz - TbC; | |
2988 { | |
2989 E TbG, TbJ, Tci, Tcj; | |
2990 TbG = TbE + TbF; | |
2991 TbJ = TbH + TbI; | |
2992 TbK = FNMS(KP942793473, TbJ, KP1_763842528 * TbG); | |
2993 Tc8 = FMA(KP942793473, TbG, KP1_763842528 * TbJ); | |
2994 Tci = TbE - TbF; | |
2995 Tcj = TbI - TbH; | |
2996 Tck = FNMS(KP1_913880671, Tcj, KP580569354 * Tci); | |
2997 Tcw = FMA(KP1_913880671, Tci, KP580569354 * Tcj); | |
2998 } | |
2999 Tcf = Tbx + Tby; | |
3000 Tcg = FMA(KP1_662939224, TbA, KP1_111140466 * TbB); | |
3001 Tch = Tcf - Tcg; | |
3002 Tcv = Tcf + Tcg; | |
3003 { | |
3004 E Tc2, Tcm, TbZ, Tcn, TbX, TbY; | |
3005 Tc2 = Tc0 + Tc1; | |
3006 Tcm = TbM - TbN; | |
3007 TbX = FMA(KP831469612, TbP, KP555570233 * TbQ); | |
3008 TbY = FNMS(KP831469612, TbS, KP555570233 * TbT); | |
3009 TbZ = TbX + TbY; | |
3010 Tcn = TbY - TbX; | |
3011 Tc3 = TbZ + Tc2; | |
3012 Tcy = Tcm - Tcn; | |
3013 Tcb = Tc2 - TbZ; | |
3014 Tco = Tcm + Tcn; | |
3015 } | |
3016 { | |
3017 E TbO, Tcq, TbV, Tcp, TbR, TbU; | |
3018 TbO = TbM + TbN; | |
3019 Tcq = Tc1 - Tc0; | |
3020 TbR = FNMS(KP831469612, TbQ, KP555570233 * TbP); | |
3021 TbU = FMA(KP555570233, TbS, KP831469612 * TbT); | |
3022 TbV = TbR + TbU; | |
3023 Tcp = TbR - TbU; | |
3024 TbW = TbO + TbV; | |
3025 Tcz = Tcq - Tcp; | |
3026 Tca = TbO - TbV; | |
3027 Tcr = Tcp + Tcq; | |
3028 } | |
3029 } | |
3030 { | |
3031 E TbL, Tc4, Tcx, TcA; | |
3032 TbL = TbD + TbK; | |
3033 Tc4 = FNMS(KP485960359, Tc3, KP1_940062506 * TbW); | |
3034 R1[WS(rs, 34)] = TbL - Tc4; | |
3035 R1[WS(rs, 2)] = TbL + Tc4; | |
3036 Tcx = Tcv - Tcw; | |
3037 TcA = FNMS(KP1_978353019, Tcz, KP293460948 * Tcy); | |
3038 R1[WS(rs, 46)] = Tcx - TcA; | |
3039 R1[WS(rs, 14)] = Tcx + TcA; | |
3040 } | |
3041 { | |
3042 E TcB, TcC, Tc5, Tc6; | |
3043 TcB = Tcv + Tcw; | |
3044 TcC = FMA(KP1_978353019, Tcy, KP293460948 * Tcz); | |
3045 R1[WS(rs, 30)] = TcB - TcC; | |
3046 R1[WS(rs, 62)] = TcB + TcC; | |
3047 Tc5 = TbD - TbK; | |
3048 Tc6 = FMA(KP485960359, TbW, KP1_940062506 * Tc3); | |
3049 R1[WS(rs, 18)] = Tc5 - Tc6; | |
3050 R1[WS(rs, 50)] = Tc5 + Tc6; | |
3051 } | |
3052 { | |
3053 E Tc9, Tcc, Tcl, Tcs; | |
3054 Tc9 = Tc7 - Tc8; | |
3055 Tcc = FNMS(KP1_715457220, Tcb, KP1_028205488 * Tca); | |
3056 R1[WS(rs, 42)] = Tc9 - Tcc; | |
3057 R1[WS(rs, 10)] = Tc9 + Tcc; | |
3058 Tcl = Tch + Tck; | |
3059 Tcs = FNMS(KP1_191398608, Tcr, KP1_606415062 * Tco); | |
3060 R1[WS(rs, 38)] = Tcl - Tcs; | |
3061 R1[WS(rs, 6)] = Tcl + Tcs; | |
3062 } | |
3063 { | |
3064 E Tct, Tcu, Tcd, Tce; | |
3065 Tct = Tch - Tck; | |
3066 Tcu = FMA(KP1_191398608, Tco, KP1_606415062 * Tcr); | |
3067 R1[WS(rs, 22)] = Tct - Tcu; | |
3068 R1[WS(rs, 54)] = Tct + Tcu; | |
3069 Tcd = Tc7 + Tc8; | |
3070 Tce = FMA(KP1_715457220, Tca, KP1_028205488 * Tcb); | |
3071 R1[WS(rs, 26)] = Tcd - Tce; | |
3072 R1[WS(rs, 58)] = Tcd + Tce; | |
3073 } | |
3074 } | |
3075 { | |
3076 E Tet, TeX, TeA, TeY, Tfa, Tfm, Tf7, Tfl, TeT, Tfo, Tf1, Tfe, TeM, Tfp, Tf0; | |
3077 E Tfh; | |
3078 { | |
3079 E Tep, Tes, Tf5, Tf6; | |
3080 Tep = Ten - Teo; | |
3081 Tes = FNMS(KP1_961570560, Ter, KP390180644 * Teq); | |
3082 Tet = Tep + Tes; | |
3083 TeX = Tep - Tes; | |
3084 { | |
3085 E Tew, Tez, Tf8, Tf9; | |
3086 Tew = Teu - Tev; | |
3087 Tez = Tex + Tey; | |
3088 TeA = FNMS(KP1_268786568, Tez, KP1_546020906 * Tew); | |
3089 TeY = FMA(KP1_268786568, Tew, KP1_546020906 * Tez); | |
3090 Tf8 = Teu + Tev; | |
3091 Tf9 = Tey - Tex; | |
3092 Tfa = FNMS(KP1_990369453, Tf9, KP196034280 * Tf8); | |
3093 Tfm = FMA(KP1_990369453, Tf8, KP196034280 * Tf9); | |
3094 } | |
3095 Tf5 = Ten + Teo; | |
3096 Tf6 = FMA(KP1_961570560, Teq, KP390180644 * Ter); | |
3097 Tf7 = Tf5 - Tf6; | |
3098 Tfl = Tf5 + Tf6; | |
3099 { | |
3100 E TeS, Tfc, TeP, Tfd, TeN, TeO; | |
3101 TeS = TeQ + TeR; | |
3102 Tfc = TeC + TeD; | |
3103 TeN = FMA(KP980785280, TeF, KP195090322 * TeG); | |
3104 TeO = FMA(KP980785280, TeI, KP195090322 * TeJ); | |
3105 TeP = TeN - TeO; | |
3106 Tfd = TeN + TeO; | |
3107 TeT = TeP + TeS; | |
3108 Tfo = Tfc + Tfd; | |
3109 Tf1 = TeS - TeP; | |
3110 Tfe = Tfc - Tfd; | |
3111 } | |
3112 { | |
3113 E TeE, Tfg, TeL, Tff, TeH, TeK; | |
3114 TeE = TeC - TeD; | |
3115 Tfg = TeR - TeQ; | |
3116 TeH = FNMS(KP980785280, TeG, KP195090322 * TeF); | |
3117 TeK = FNMS(KP980785280, TeJ, KP195090322 * TeI); | |
3118 TeL = TeH + TeK; | |
3119 Tff = TeH - TeK; | |
3120 TeM = TeE + TeL; | |
3121 Tfp = Tfg - Tff; | |
3122 Tf0 = TeE - TeL; | |
3123 Tfh = Tff + Tfg; | |
3124 } | |
3125 } | |
3126 { | |
3127 E TeB, TeU, Tfn, Tfq; | |
3128 TeB = Tet + TeA; | |
3129 TeU = FNMS(KP673779706, TeT, KP1_883088130 * TeM); | |
3130 R1[WS(rs, 35)] = TeB - TeU; | |
3131 R1[WS(rs, 3)] = TeB + TeU; | |
3132 Tfn = Tfl - Tfm; | |
3133 Tfq = FNMS(KP1_997590912, Tfp, KP098135348 * Tfo); | |
3134 R1[WS(rs, 47)] = Tfn - Tfq; | |
3135 R1[WS(rs, 15)] = Tfn + Tfq; | |
3136 } | |
3137 { | |
3138 E Tfr, Tfs, TeV, TeW; | |
3139 Tfr = Tfl + Tfm; | |
3140 Tfs = FMA(KP1_997590912, Tfo, KP098135348 * Tfp); | |
3141 R1[WS(rs, 31)] = Tfr - Tfs; | |
3142 R1[WS(rs, 63)] = Tfr + Tfs; | |
3143 TeV = Tet - TeA; | |
3144 TeW = FMA(KP673779706, TeM, KP1_883088130 * TeT); | |
3145 R1[WS(rs, 19)] = TeV - TeW; | |
3146 R1[WS(rs, 51)] = TeV + TeW; | |
3147 } | |
3148 { | |
3149 E TeZ, Tf2, Tfb, Tfi; | |
3150 TeZ = TeX - TeY; | |
3151 Tf2 = FNMS(KP1_807978586, Tf1, KP855110186 * Tf0); | |
3152 R1[WS(rs, 43)] = TeZ - Tf2; | |
3153 R1[WS(rs, 11)] = TeZ + Tf2; | |
3154 Tfb = Tf7 + Tfa; | |
3155 Tfi = FNMS(KP1_343117909, Tfh, KP1_481902250 * Tfe); | |
3156 R1[WS(rs, 39)] = Tfb - Tfi; | |
3157 R1[WS(rs, 7)] = Tfb + Tfi; | |
3158 } | |
3159 { | |
3160 E Tfj, Tfk, Tf3, Tf4; | |
3161 Tfj = Tf7 - Tfa; | |
3162 Tfk = FMA(KP1_343117909, Tfe, KP1_481902250 * Tfh); | |
3163 R1[WS(rs, 23)] = Tfj - Tfk; | |
3164 R1[WS(rs, 55)] = Tfj + Tfk; | |
3165 Tf3 = TeX + TeY; | |
3166 Tf4 = FMA(KP1_807978586, Tf0, KP855110186 * Tf1); | |
3167 R1[WS(rs, 27)] = Tf3 - Tf4; | |
3168 R1[WS(rs, 59)] = Tf3 + Tf4; | |
3169 } | |
3170 } | |
3171 } | |
3172 } | |
3173 } | |
3174 | |
3175 static const kr2c_desc desc = { 128, "r2cb_128", {812, 198, 144, 0}, &GENUS }; | |
3176 | |
3177 void X(codelet_r2cb_128) (planner *p) { | |
3178 X(kr2c_register) (p, r2cb_128, &desc); | |
3179 } | |
3180 | |
3181 #endif /* HAVE_FMA */ |