Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/rdft/scalar/r2cb/r2cb_128.c @ 167:bd3cc4d1df30
Add FFTW 3.3.8 source, and a Linux build
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Tue, 19 Nov 2019 14:52:55 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
166:cbd6d7e562c7 | 167:bd3cc4d1df30 |
---|---|
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 Thu May 24 08:07:29 EDT 2018 */ | |
23 | |
24 #include "rdft/codelet-rdft.h" | |
25 | |
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA) | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include rdft/scalar/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 * 195 stack variables, 36 constants, and 256 memory accesses | |
34 */ | |
35 #include "rdft/scalar/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(KP599376933, +0.599376933681923766271389869014404232837890546); | |
41 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854); | |
42 DK(KP741650546, +0.741650546272035369581266691172079863842265220); | |
43 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924); | |
44 DK(KP250486960, +0.250486960191305461595702160124721208578685568); | |
45 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067); | |
46 DK(KP148335987, +0.148335987538347428753676511486911367000625355); | |
47 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
48 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050); | |
49 DK(KP472964775, +0.472964775891319928124438237972992463904131113); | |
50 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619); | |
51 DK(KP906347169, +0.906347169019147157946142717268914412664134293); | |
52 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592); | |
53 DK(KP357805721, +0.357805721314524104672487743774474392487532769); | |
54 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229); | |
55 DK(KP049126849, +0.049126849769467254105343321271313617079695752); | |
56 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
57 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411); | |
58 DK(KP303346683, +0.303346683607342391675883946941299872384187453); | |
59 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241); | |
60 DK(KP534511135, +0.534511135950791641089685961295362908582039528); | |
61 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624); | |
62 DK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
63 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586); | |
64 DK(KP820678790, +0.820678790828660330972281985331011598767386482); | |
65 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737); | |
66 DK(KP098491403, +0.098491403357164253077197521291327432293052451); | |
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 T9, Tdr, Ta9, Tcl, T2d, T6b, T7j, T91, Tg, Tds, T2o, T6c, T7m, T92, Tae; | |
79 E Tcm, Tw, Tdu, Tap, Tco, Tdx, TeM, T2G, T6f, T2P, T6e, T7t, T94, Tak, Tcp; | |
80 E T7q, T95, T1i, TdM, TbD, TcL, Te6, Tf0, T42, T6q, T5t, T6B, T8j, T9r, TaY; | |
81 E TcA, T7S, T9g, TM, TdA, TaN, Tcv, TdI, TeP, T38, T6i, T3F, T6l, T7J, T9b; | |
82 E Taw, Tcs, T7y, T98, T1N, TeW, T4H, T6x, TdV, Te8, T4Q, T6w, T86, T9j, TcI; | |
83 E TcO, T83, T9k, Tbl, TbI, T22, TeV, T58, T6u, Te0, Te9, T5h, T6t, T8d, T9m; | |
84 E TcF, TcP, T8a, T9n, Tbw, TbJ, T1x, Te3, TbG, TcB, TdP, Tf1, T4p, T6C, T5w; | |
85 E T6r, T8m, T9h, Tb9, TcM, T7Z, T9s, T11, TdF, TaQ, Tct, TdD, TeQ, T3v, T6m; | |
86 E T3I, T6j, T7M, T99, TaH, Tcw, T7F, T9c; | |
87 { | |
88 E T4, T26, T3, T25, T8, T28, T2b, Ta8, T1, T2, T5, Ta7; | |
89 T4 = Cr[WS(csr, 32)]; | |
90 T26 = Ci[WS(csi, 32)]; | |
91 T1 = Cr[0]; | |
92 T2 = Cr[WS(csr, 64)]; | |
93 T3 = T1 + T2; | |
94 T25 = T1 - T2; | |
95 { | |
96 E T6, T7, T29, T2a; | |
97 T6 = Cr[WS(csr, 16)]; | |
98 T7 = Cr[WS(csr, 48)]; | |
99 T8 = T6 + T7; | |
100 T28 = T6 - T7; | |
101 T29 = Ci[WS(csi, 16)]; | |
102 T2a = Ci[WS(csi, 48)]; | |
103 T2b = T29 + T2a; | |
104 Ta8 = T29 - T2a; | |
105 } | |
106 T5 = FMA(KP2_000000000, T4, T3); | |
107 T9 = FMA(KP2_000000000, T8, T5); | |
108 Tdr = FNMS(KP2_000000000, T8, T5); | |
109 Ta7 = FNMS(KP2_000000000, T4, T3); | |
110 Ta9 = FNMS(KP2_000000000, Ta8, Ta7); | |
111 Tcl = FMA(KP2_000000000, Ta8, Ta7); | |
112 { | |
113 E T27, T2c, T7h, T7i; | |
114 T27 = FNMS(KP2_000000000, T26, T25); | |
115 T2c = T28 - T2b; | |
116 T2d = FMA(KP1_414213562, T2c, T27); | |
117 T6b = FNMS(KP1_414213562, T2c, T27); | |
118 T7h = FMA(KP2_000000000, T26, T25); | |
119 T7i = T28 + T2b; | |
120 T7j = FNMS(KP1_414213562, T7i, T7h); | |
121 T91 = FMA(KP1_414213562, T7i, T7h); | |
122 } | |
123 } | |
124 { | |
125 E Tc, T2e, T2m, Tab, Tf, T2j, T2h, Tac, T2i, T2n; | |
126 { | |
127 E Ta, Tb, T2k, T2l; | |
128 Ta = Cr[WS(csr, 8)]; | |
129 Tb = Cr[WS(csr, 56)]; | |
130 Tc = Ta + Tb; | |
131 T2e = Ta - Tb; | |
132 T2k = Ci[WS(csi, 8)]; | |
133 T2l = Ci[WS(csi, 56)]; | |
134 T2m = T2k + T2l; | |
135 Tab = T2k - T2l; | |
136 } | |
137 { | |
138 E Td, Te, T2f, T2g; | |
139 Td = Cr[WS(csr, 40)]; | |
140 Te = Cr[WS(csr, 24)]; | |
141 Tf = Td + Te; | |
142 T2j = Td - Te; | |
143 T2f = Ci[WS(csi, 40)]; | |
144 T2g = Ci[WS(csi, 24)]; | |
145 T2h = T2f + T2g; | |
146 Tac = T2f - T2g; | |
147 } | |
148 Tg = Tc + Tf; | |
149 Tds = Tac + Tab; | |
150 T2i = T2e - T2h; | |
151 T2n = T2j + T2m; | |
152 T2o = FNMS(KP414213562, T2n, T2i); | |
153 T6c = FMA(KP414213562, T2i, T2n); | |
154 { | |
155 E T7k, T7l, Taa, Tad; | |
156 T7k = T2m - T2j; | |
157 T7l = T2e + T2h; | |
158 T7m = FNMS(KP414213562, T7l, T7k); | |
159 T92 = FMA(KP414213562, T7k, T7l); | |
160 Taa = Tc - Tf; | |
161 Tad = Tab - Tac; | |
162 Tae = Taa - Tad; | |
163 Tcm = Taa + Tad; | |
164 } | |
165 } | |
166 { | |
167 E Tk, T2q, T2K, Tam, Tn, T2H, T2t, Tan, Tu, Tah, T2E, T2N, Tr, Tai, T2z; | |
168 E T2M; | |
169 { | |
170 E Ti, Tj, T2r, T2s; | |
171 Ti = Cr[WS(csr, 4)]; | |
172 Tj = Cr[WS(csr, 60)]; | |
173 Tk = Ti + Tj; | |
174 T2q = Ti - Tj; | |
175 { | |
176 E T2I, T2J, Tl, Tm; | |
177 T2I = Ci[WS(csi, 4)]; | |
178 T2J = Ci[WS(csi, 60)]; | |
179 T2K = T2I + T2J; | |
180 Tam = T2I - T2J; | |
181 Tl = Cr[WS(csr, 36)]; | |
182 Tm = Cr[WS(csr, 28)]; | |
183 Tn = Tl + Tm; | |
184 T2H = Tl - Tm; | |
185 } | |
186 T2r = Ci[WS(csi, 36)]; | |
187 T2s = Ci[WS(csi, 28)]; | |
188 T2t = T2r + T2s; | |
189 Tan = T2r - T2s; | |
190 { | |
191 E Ts, Tt, T2A, T2B, T2C, T2D; | |
192 Ts = Cr[WS(csr, 12)]; | |
193 Tt = Cr[WS(csr, 52)]; | |
194 T2A = Ts - Tt; | |
195 T2B = Ci[WS(csi, 12)]; | |
196 T2C = Ci[WS(csi, 52)]; | |
197 T2D = T2B + T2C; | |
198 Tu = Ts + Tt; | |
199 Tah = T2C - T2B; | |
200 T2E = T2A - T2D; | |
201 T2N = T2A + T2D; | |
202 } | |
203 { | |
204 E Tp, Tq, T2v, T2w, T2x, T2y; | |
205 Tp = Cr[WS(csr, 20)]; | |
206 Tq = Cr[WS(csr, 44)]; | |
207 T2v = Tp - Tq; | |
208 T2w = Ci[WS(csi, 20)]; | |
209 T2x = Ci[WS(csi, 44)]; | |
210 T2y = T2w + T2x; | |
211 Tr = Tp + Tq; | |
212 Tai = T2w - T2x; | |
213 T2z = T2v - T2y; | |
214 T2M = T2v + T2y; | |
215 } | |
216 } | |
217 { | |
218 E To, Tv, Tal, Tao; | |
219 To = Tk + Tn; | |
220 Tv = Tr + Tu; | |
221 Tw = To + Tv; | |
222 Tdu = To - Tv; | |
223 Tal = Tr - Tu; | |
224 Tao = Tam - Tan; | |
225 Tap = Tal + Tao; | |
226 Tco = Tao - Tal; | |
227 } | |
228 { | |
229 E Tdv, Tdw, T2u, T2F; | |
230 Tdv = Tan + Tam; | |
231 Tdw = Tai + Tah; | |
232 Tdx = Tdv - Tdw; | |
233 TeM = Tdw + Tdv; | |
234 T2u = T2q - T2t; | |
235 T2F = T2z + T2E; | |
236 T2G = FMA(KP707106781, T2F, T2u); | |
237 T6f = FNMS(KP707106781, T2F, T2u); | |
238 } | |
239 { | |
240 E T2L, T2O, T7r, T7s; | |
241 T2L = T2H + T2K; | |
242 T2O = T2M - T2N; | |
243 T2P = FMA(KP707106781, T2O, T2L); | |
244 T6e = FNMS(KP707106781, T2O, T2L); | |
245 T7r = T2K - T2H; | |
246 T7s = T2E - T2z; | |
247 T7t = FNMS(KP707106781, T7s, T7r); | |
248 T94 = FMA(KP707106781, T7s, T7r); | |
249 } | |
250 { | |
251 E Tag, Taj, T7o, T7p; | |
252 Tag = Tk - Tn; | |
253 Taj = Tah - Tai; | |
254 Tak = Tag + Taj; | |
255 Tcp = Tag - Taj; | |
256 T7o = T2q + T2t; | |
257 T7p = T2M + T2N; | |
258 T7q = FNMS(KP707106781, T7p, T7o); | |
259 T95 = FMA(KP707106781, T7p, T7o); | |
260 } | |
261 } | |
262 { | |
263 E T16, T3M, T5o, TbA, T19, T5l, T3P, TbB, T1g, TaV, T40, T5r, T1d, TaW, T3V; | |
264 E T5q; | |
265 { | |
266 E T14, T15, T3N, T3O; | |
267 T14 = Cr[WS(csr, 1)]; | |
268 T15 = Cr[WS(csr, 63)]; | |
269 T16 = T14 + T15; | |
270 T3M = T14 - T15; | |
271 { | |
272 E T5m, T5n, T17, T18; | |
273 T5m = Ci[WS(csi, 1)]; | |
274 T5n = Ci[WS(csi, 63)]; | |
275 T5o = T5m + T5n; | |
276 TbA = T5m - T5n; | |
277 T17 = Cr[WS(csr, 33)]; | |
278 T18 = Cr[WS(csr, 31)]; | |
279 T19 = T17 + T18; | |
280 T5l = T17 - T18; | |
281 } | |
282 T3N = Ci[WS(csi, 33)]; | |
283 T3O = Ci[WS(csi, 31)]; | |
284 T3P = T3N + T3O; | |
285 TbB = T3N - T3O; | |
286 { | |
287 E T1e, T1f, T3W, T3X, T3Y, T3Z; | |
288 T1e = Cr[WS(csr, 15)]; | |
289 T1f = Cr[WS(csr, 49)]; | |
290 T3W = T1e - T1f; | |
291 T3X = Ci[WS(csi, 15)]; | |
292 T3Y = Ci[WS(csi, 49)]; | |
293 T3Z = T3X + T3Y; | |
294 T1g = T1e + T1f; | |
295 TaV = T3Y - T3X; | |
296 T40 = T3W - T3Z; | |
297 T5r = T3W + T3Z; | |
298 } | |
299 { | |
300 E T1b, T1c, T3R, T3S, T3T, T3U; | |
301 T1b = Cr[WS(csr, 17)]; | |
302 T1c = Cr[WS(csr, 47)]; | |
303 T3R = T1b - T1c; | |
304 T3S = Ci[WS(csi, 17)]; | |
305 T3T = Ci[WS(csi, 47)]; | |
306 T3U = T3S + T3T; | |
307 T1d = T1b + T1c; | |
308 TaW = T3S - T3T; | |
309 T3V = T3R - T3U; | |
310 T5q = T3R + T3U; | |
311 } | |
312 } | |
313 { | |
314 E T1a, T1h, Tbz, TbC; | |
315 T1a = T16 + T19; | |
316 T1h = T1d + T1g; | |
317 T1i = T1a + T1h; | |
318 TdM = T1a - T1h; | |
319 Tbz = T1d - T1g; | |
320 TbC = TbA - TbB; | |
321 TbD = Tbz + TbC; | |
322 TcL = TbC - Tbz; | |
323 } | |
324 { | |
325 E Te4, Te5, T3Q, T41; | |
326 Te4 = TbB + TbA; | |
327 Te5 = TaW + TaV; | |
328 Te6 = Te4 - Te5; | |
329 Tf0 = Te5 + Te4; | |
330 T3Q = T3M - T3P; | |
331 T41 = T3V + T40; | |
332 T42 = FMA(KP707106781, T41, T3Q); | |
333 T6q = FNMS(KP707106781, T41, T3Q); | |
334 } | |
335 { | |
336 E T5p, T5s, T8h, T8i; | |
337 T5p = T5l + T5o; | |
338 T5s = T5q - T5r; | |
339 T5t = FMA(KP707106781, T5s, T5p); | |
340 T6B = FNMS(KP707106781, T5s, T5p); | |
341 T8h = T5o - T5l; | |
342 T8i = T40 - T3V; | |
343 T8j = FNMS(KP707106781, T8i, T8h); | |
344 T9r = FMA(KP707106781, T8i, T8h); | |
345 } | |
346 { | |
347 E TaU, TaX, T7Q, T7R; | |
348 TaU = T16 - T19; | |
349 TaX = TaV - TaW; | |
350 TaY = TaU + TaX; | |
351 TcA = TaU - TaX; | |
352 T7Q = T3M + T3P; | |
353 T7R = T5q + T5r; | |
354 T7S = FNMS(KP707106781, T7R, T7Q); | |
355 T9g = FMA(KP707106781, T7R, T7Q); | |
356 } | |
357 } | |
358 { | |
359 E TA, T2S, T3A, TaK, TD, T3x, T2V, TaL, TK, Tat, T36, T3D, TH, Tau, T31; | |
360 E T3C; | |
361 { | |
362 E Ty, Tz, T2T, T2U; | |
363 Ty = Cr[WS(csr, 2)]; | |
364 Tz = Cr[WS(csr, 62)]; | |
365 TA = Ty + Tz; | |
366 T2S = Ty - Tz; | |
367 { | |
368 E T3y, T3z, TB, TC; | |
369 T3y = Ci[WS(csi, 2)]; | |
370 T3z = Ci[WS(csi, 62)]; | |
371 T3A = T3y + T3z; | |
372 TaK = T3y - T3z; | |
373 TB = Cr[WS(csr, 34)]; | |
374 TC = Cr[WS(csr, 30)]; | |
375 TD = TB + TC; | |
376 T3x = TB - TC; | |
377 } | |
378 T2T = Ci[WS(csi, 34)]; | |
379 T2U = Ci[WS(csi, 30)]; | |
380 T2V = T2T + T2U; | |
381 TaL = T2T - T2U; | |
382 { | |
383 E TI, TJ, T32, T33, T34, T35; | |
384 TI = Cr[WS(csr, 14)]; | |
385 TJ = Cr[WS(csr, 50)]; | |
386 T32 = TI - TJ; | |
387 T33 = Ci[WS(csi, 14)]; | |
388 T34 = Ci[WS(csi, 50)]; | |
389 T35 = T33 + T34; | |
390 TK = TI + TJ; | |
391 Tat = T34 - T33; | |
392 T36 = T32 - T35; | |
393 T3D = T32 + T35; | |
394 } | |
395 { | |
396 E TF, TG, T2X, T2Y, T2Z, T30; | |
397 TF = Cr[WS(csr, 18)]; | |
398 TG = Cr[WS(csr, 46)]; | |
399 T2X = TF - TG; | |
400 T2Y = Ci[WS(csi, 18)]; | |
401 T2Z = Ci[WS(csi, 46)]; | |
402 T30 = T2Y + T2Z; | |
403 TH = TF + TG; | |
404 Tau = T2Y - T2Z; | |
405 T31 = T2X - T30; | |
406 T3C = T2X + T30; | |
407 } | |
408 } | |
409 { | |
410 E TE, TL, TaJ, TaM; | |
411 TE = TA + TD; | |
412 TL = TH + TK; | |
413 TM = TE + TL; | |
414 TdA = TE - TL; | |
415 TaJ = TH - TK; | |
416 TaM = TaK - TaL; | |
417 TaN = TaJ + TaM; | |
418 Tcv = TaM - TaJ; | |
419 } | |
420 { | |
421 E TdG, TdH, T2W, T37; | |
422 TdG = TaL + TaK; | |
423 TdH = Tau + Tat; | |
424 TdI = TdG - TdH; | |
425 TeP = TdH + TdG; | |
426 T2W = T2S - T2V; | |
427 T37 = T31 + T36; | |
428 T38 = FMA(KP707106781, T37, T2W); | |
429 T6i = FNMS(KP707106781, T37, T2W); | |
430 } | |
431 { | |
432 E T3B, T3E, T7H, T7I; | |
433 T3B = T3x + T3A; | |
434 T3E = T3C - T3D; | |
435 T3F = FMA(KP707106781, T3E, T3B); | |
436 T6l = FNMS(KP707106781, T3E, T3B); | |
437 T7H = T3A - T3x; | |
438 T7I = T36 - T31; | |
439 T7J = FNMS(KP707106781, T7I, T7H); | |
440 T9b = FMA(KP707106781, T7I, T7H); | |
441 } | |
442 { | |
443 E Tas, Tav, T7w, T7x; | |
444 Tas = TA - TD; | |
445 Tav = Tat - Tau; | |
446 Taw = Tas + Tav; | |
447 Tcs = Tas - Tav; | |
448 T7w = T2S + T2V; | |
449 T7x = T3C + T3D; | |
450 T7y = FNMS(KP707106781, T7x, T7w); | |
451 T98 = FMA(KP707106781, T7x, T7w); | |
452 } | |
453 } | |
454 { | |
455 E T1F, Tbb, T4r, T4I, T4u, TdS, T4L, Tbj, T1M, Tbg, TdT, T4A, T4F, T4O, Tbe; | |
456 E T4N, TcG, TcH; | |
457 { | |
458 E T1z, T1A, T1B, T1C, T1D, T1E; | |
459 T1z = Cr[WS(csr, 5)]; | |
460 T1A = Cr[WS(csr, 59)]; | |
461 T1B = T1z + T1A; | |
462 T1C = Cr[WS(csr, 37)]; | |
463 T1D = Cr[WS(csr, 27)]; | |
464 T1E = T1C + T1D; | |
465 T1F = T1B + T1E; | |
466 Tbb = T1B - T1E; | |
467 T4r = T1z - T1A; | |
468 T4I = T1C - T1D; | |
469 } | |
470 { | |
471 E T4s, T4t, Tbi, T4J, T4K, Tbh; | |
472 T4s = Ci[WS(csi, 37)]; | |
473 T4t = Ci[WS(csi, 27)]; | |
474 Tbi = T4s - T4t; | |
475 T4J = Ci[WS(csi, 5)]; | |
476 T4K = Ci[WS(csi, 59)]; | |
477 Tbh = T4J - T4K; | |
478 T4u = T4s + T4t; | |
479 TdS = Tbi + Tbh; | |
480 T4L = T4J + T4K; | |
481 Tbj = Tbh - Tbi; | |
482 } | |
483 { | |
484 E T1I, T4w, T4z, Tbd, T1L, T4B, T4E, Tbc; | |
485 { | |
486 E T1G, T1H, T4x, T4y; | |
487 T1G = Cr[WS(csr, 21)]; | |
488 T1H = Cr[WS(csr, 43)]; | |
489 T1I = T1G + T1H; | |
490 T4w = T1G - T1H; | |
491 T4x = Ci[WS(csi, 21)]; | |
492 T4y = Ci[WS(csi, 43)]; | |
493 T4z = T4x + T4y; | |
494 Tbd = T4x - T4y; | |
495 } | |
496 { | |
497 E T1J, T1K, T4C, T4D; | |
498 T1J = Cr[WS(csr, 11)]; | |
499 T1K = Cr[WS(csr, 53)]; | |
500 T1L = T1J + T1K; | |
501 T4B = T1J - T1K; | |
502 T4C = Ci[WS(csi, 11)]; | |
503 T4D = Ci[WS(csi, 53)]; | |
504 T4E = T4C + T4D; | |
505 Tbc = T4D - T4C; | |
506 } | |
507 T1M = T1I + T1L; | |
508 Tbg = T1I - T1L; | |
509 TdT = Tbd + Tbc; | |
510 T4A = T4w - T4z; | |
511 T4F = T4B - T4E; | |
512 T4O = T4B + T4E; | |
513 Tbe = Tbc - Tbd; | |
514 T4N = T4w + T4z; | |
515 } | |
516 T1N = T1F + T1M; | |
517 TeW = TdT + TdS; | |
518 { | |
519 E T4v, T4G, TdR, TdU; | |
520 T4v = T4r - T4u; | |
521 T4G = T4A + T4F; | |
522 T4H = FMA(KP707106781, T4G, T4v); | |
523 T6x = FNMS(KP707106781, T4G, T4v); | |
524 TdR = T1F - T1M; | |
525 TdU = TdS - TdT; | |
526 TdV = TdR - TdU; | |
527 Te8 = TdR + TdU; | |
528 } | |
529 { | |
530 E T4M, T4P, T84, T85; | |
531 T4M = T4I + T4L; | |
532 T4P = T4N - T4O; | |
533 T4Q = FMA(KP707106781, T4P, T4M); | |
534 T6w = FNMS(KP707106781, T4P, T4M); | |
535 T84 = T4L - T4I; | |
536 T85 = T4F - T4A; | |
537 T86 = FNMS(KP707106781, T85, T84); | |
538 T9j = FMA(KP707106781, T85, T84); | |
539 } | |
540 TcG = Tbj - Tbg; | |
541 TcH = Tbb - Tbe; | |
542 TcI = FNMS(KP414213562, TcH, TcG); | |
543 TcO = FMA(KP414213562, TcG, TcH); | |
544 { | |
545 E T81, T82, Tbf, Tbk; | |
546 T81 = T4r + T4u; | |
547 T82 = T4N + T4O; | |
548 T83 = FNMS(KP707106781, T82, T81); | |
549 T9k = FMA(KP707106781, T82, T81); | |
550 Tbf = Tbb + Tbe; | |
551 Tbk = Tbg + Tbj; | |
552 Tbl = FNMS(KP414213562, Tbk, Tbf); | |
553 TbI = FMA(KP414213562, Tbf, Tbk); | |
554 } | |
555 } | |
556 { | |
557 E T1U, Tbm, T4S, T59, T4V, TdX, T5c, Tbu, T21, Tbr, TdY, T51, T56, T5f, Tbp; | |
558 E T5e, TcD, TcE; | |
559 { | |
560 E T1O, T1P, T1Q, T1R, T1S, T1T; | |
561 T1O = Cr[WS(csr, 3)]; | |
562 T1P = Cr[WS(csr, 61)]; | |
563 T1Q = T1O + T1P; | |
564 T1R = Cr[WS(csr, 29)]; | |
565 T1S = Cr[WS(csr, 35)]; | |
566 T1T = T1R + T1S; | |
567 T1U = T1Q + T1T; | |
568 Tbm = T1Q - T1T; | |
569 T4S = T1O - T1P; | |
570 T59 = T1R - T1S; | |
571 } | |
572 { | |
573 E T4T, T4U, Tbt, T5a, T5b, Tbs; | |
574 T4T = Ci[WS(csi, 29)]; | |
575 T4U = Ci[WS(csi, 35)]; | |
576 Tbt = T4T - T4U; | |
577 T5a = Ci[WS(csi, 3)]; | |
578 T5b = Ci[WS(csi, 61)]; | |
579 Tbs = T5b - T5a; | |
580 T4V = T4T + T4U; | |
581 TdX = Tbt + Tbs; | |
582 T5c = T5a + T5b; | |
583 Tbu = Tbs - Tbt; | |
584 } | |
585 { | |
586 E T1X, T4X, T50, Tbo, T20, T52, T55, Tbn; | |
587 { | |
588 E T1V, T1W, T4Y, T4Z; | |
589 T1V = Cr[WS(csr, 13)]; | |
590 T1W = Cr[WS(csr, 51)]; | |
591 T1X = T1V + T1W; | |
592 T4X = T1V - T1W; | |
593 T4Y = Ci[WS(csi, 13)]; | |
594 T4Z = Ci[WS(csi, 51)]; | |
595 T50 = T4Y + T4Z; | |
596 Tbo = T4Y - T4Z; | |
597 } | |
598 { | |
599 E T1Y, T1Z, T53, T54; | |
600 T1Y = Cr[WS(csr, 19)]; | |
601 T1Z = Cr[WS(csr, 45)]; | |
602 T20 = T1Y + T1Z; | |
603 T52 = T1Y - T1Z; | |
604 T53 = Ci[WS(csi, 19)]; | |
605 T54 = Ci[WS(csi, 45)]; | |
606 T55 = T53 + T54; | |
607 Tbn = T54 - T53; | |
608 } | |
609 T21 = T1X + T20; | |
610 Tbr = T1X - T20; | |
611 TdY = Tbo + Tbn; | |
612 T51 = T4X - T50; | |
613 T56 = T52 - T55; | |
614 T5f = T52 + T55; | |
615 Tbp = Tbn - Tbo; | |
616 T5e = T4X + T50; | |
617 } | |
618 T22 = T1U + T21; | |
619 TeV = TdY + TdX; | |
620 { | |
621 E T4W, T57, TdW, TdZ; | |
622 T4W = T4S - T4V; | |
623 T57 = T51 + T56; | |
624 T58 = FMA(KP707106781, T57, T4W); | |
625 T6u = FNMS(KP707106781, T57, T4W); | |
626 TdW = T1U - T21; | |
627 TdZ = TdX - TdY; | |
628 Te0 = TdW + TdZ; | |
629 Te9 = TdZ - TdW; | |
630 } | |
631 { | |
632 E T5d, T5g, T8b, T8c; | |
633 T5d = T59 - T5c; | |
634 T5g = T5e - T5f; | |
635 T5h = FMA(KP707106781, T5g, T5d); | |
636 T6t = FNMS(KP707106781, T5g, T5d); | |
637 T8b = T59 + T5c; | |
638 T8c = T56 - T51; | |
639 T8d = FMA(KP707106781, T8c, T8b); | |
640 T9m = FNMS(KP707106781, T8c, T8b); | |
641 } | |
642 TcD = Tbu - Tbr; | |
643 TcE = Tbm - Tbp; | |
644 TcF = FMA(KP414213562, TcE, TcD); | |
645 TcP = FNMS(KP414213562, TcD, TcE); | |
646 { | |
647 E T88, T89, Tbq, Tbv; | |
648 T88 = T4S + T4V; | |
649 T89 = T5e + T5f; | |
650 T8a = FNMS(KP707106781, T89, T88); | |
651 T9n = FMA(KP707106781, T89, T88); | |
652 Tbq = Tbm + Tbp; | |
653 Tbv = Tbr + Tbu; | |
654 Tbw = FMA(KP414213562, Tbv, Tbq); | |
655 TbJ = FNMS(KP414213562, Tbq, Tbv); | |
656 } | |
657 } | |
658 { | |
659 E T1l, Tb0, T1o, Tb1, T47, T4c, Tb2, TaZ, T7U, T7T, T1s, Tb5, T1v, Tb6, T4i; | |
660 E T4n, Tb7, Tb4, T7X, T7W; | |
661 { | |
662 E T43, T4b, T48, T46; | |
663 { | |
664 E T1j, T1k, T49, T4a; | |
665 T1j = Cr[WS(csr, 9)]; | |
666 T1k = Cr[WS(csr, 55)]; | |
667 T1l = T1j + T1k; | |
668 T43 = T1j - T1k; | |
669 T49 = Ci[WS(csi, 9)]; | |
670 T4a = Ci[WS(csi, 55)]; | |
671 T4b = T49 + T4a; | |
672 Tb0 = T49 - T4a; | |
673 } | |
674 { | |
675 E T1m, T1n, T44, T45; | |
676 T1m = Cr[WS(csr, 41)]; | |
677 T1n = Cr[WS(csr, 23)]; | |
678 T1o = T1m + T1n; | |
679 T48 = T1m - T1n; | |
680 T44 = Ci[WS(csi, 41)]; | |
681 T45 = Ci[WS(csi, 23)]; | |
682 T46 = T44 + T45; | |
683 Tb1 = T44 - T45; | |
684 } | |
685 T47 = T43 - T46; | |
686 T4c = T48 + T4b; | |
687 Tb2 = Tb0 - Tb1; | |
688 TaZ = T1l - T1o; | |
689 T7U = T43 + T46; | |
690 T7T = T4b - T48; | |
691 } | |
692 { | |
693 E T4e, T4m, T4j, T4h; | |
694 { | |
695 E T1q, T1r, T4k, T4l; | |
696 T1q = Cr[WS(csr, 7)]; | |
697 T1r = Cr[WS(csr, 57)]; | |
698 T1s = T1q + T1r; | |
699 T4e = T1q - T1r; | |
700 T4k = Ci[WS(csi, 7)]; | |
701 T4l = Ci[WS(csi, 57)]; | |
702 T4m = T4k + T4l; | |
703 Tb5 = T4l - T4k; | |
704 } | |
705 { | |
706 E T1t, T1u, T4f, T4g; | |
707 T1t = Cr[WS(csr, 25)]; | |
708 T1u = Cr[WS(csr, 39)]; | |
709 T1v = T1t + T1u; | |
710 T4j = T1t - T1u; | |
711 T4f = Ci[WS(csi, 25)]; | |
712 T4g = Ci[WS(csi, 39)]; | |
713 T4h = T4f + T4g; | |
714 Tb6 = T4f - T4g; | |
715 } | |
716 T4i = T4e - T4h; | |
717 T4n = T4j - T4m; | |
718 Tb7 = Tb5 - Tb6; | |
719 Tb4 = T1s - T1v; | |
720 T7X = T4e + T4h; | |
721 T7W = T4j + T4m; | |
722 } | |
723 { | |
724 E T1p, T1w, TbE, TbF; | |
725 T1p = T1l + T1o; | |
726 T1w = T1s + T1v; | |
727 T1x = T1p + T1w; | |
728 Te3 = T1p - T1w; | |
729 TbE = TaZ + Tb2; | |
730 TbF = Tb7 - Tb4; | |
731 TbG = TbE + TbF; | |
732 TcB = TbE - TbF; | |
733 } | |
734 { | |
735 E TdN, TdO, T4d, T4o; | |
736 TdN = Tb6 + Tb5; | |
737 TdO = Tb1 + Tb0; | |
738 TdP = TdN - TdO; | |
739 Tf1 = TdO + TdN; | |
740 T4d = FNMS(KP414213562, T4c, T47); | |
741 T4o = FMA(KP414213562, T4n, T4i); | |
742 T4p = T4d + T4o; | |
743 T6C = T4o - T4d; | |
744 } | |
745 { | |
746 E T5u, T5v, T8k, T8l; | |
747 T5u = FMA(KP414213562, T47, T4c); | |
748 T5v = FNMS(KP414213562, T4i, T4n); | |
749 T5w = T5u + T5v; | |
750 T6r = T5u - T5v; | |
751 T8k = FMA(KP414213562, T7T, T7U); | |
752 T8l = FMA(KP414213562, T7W, T7X); | |
753 T8m = T8k - T8l; | |
754 T9h = T8k + T8l; | |
755 } | |
756 { | |
757 E Tb3, Tb8, T7V, T7Y; | |
758 Tb3 = TaZ - Tb2; | |
759 Tb8 = Tb4 + Tb7; | |
760 Tb9 = Tb3 + Tb8; | |
761 TcM = Tb8 - Tb3; | |
762 T7V = FNMS(KP414213562, T7U, T7T); | |
763 T7Y = FNMS(KP414213562, T7X, T7W); | |
764 T7Z = T7V + T7Y; | |
765 T9s = T7V - T7Y; | |
766 } | |
767 } | |
768 { | |
769 E TP, Tay, TS, Taz, T3d, T3i, TaA, Tax, T7A, T7z, TW, TaD, TZ, TaE, T3o; | |
770 E T3t, TaF, TaC, T7D, T7C; | |
771 { | |
772 E T39, T3h, T3e, T3c; | |
773 { | |
774 E TN, TO, T3f, T3g; | |
775 TN = Cr[WS(csr, 10)]; | |
776 TO = Cr[WS(csr, 54)]; | |
777 TP = TN + TO; | |
778 T39 = TN - TO; | |
779 T3f = Ci[WS(csi, 10)]; | |
780 T3g = Ci[WS(csi, 54)]; | |
781 T3h = T3f + T3g; | |
782 Tay = T3f - T3g; | |
783 } | |
784 { | |
785 E TQ, TR, T3a, T3b; | |
786 TQ = Cr[WS(csr, 42)]; | |
787 TR = Cr[WS(csr, 22)]; | |
788 TS = TQ + TR; | |
789 T3e = TQ - TR; | |
790 T3a = Ci[WS(csi, 42)]; | |
791 T3b = Ci[WS(csi, 22)]; | |
792 T3c = T3a + T3b; | |
793 Taz = T3a - T3b; | |
794 } | |
795 T3d = T39 - T3c; | |
796 T3i = T3e + T3h; | |
797 TaA = Tay - Taz; | |
798 Tax = TP - TS; | |
799 T7A = T39 + T3c; | |
800 T7z = T3h - T3e; | |
801 } | |
802 { | |
803 E T3k, T3s, T3p, T3n; | |
804 { | |
805 E TU, TV, T3q, T3r; | |
806 TU = Cr[WS(csr, 6)]; | |
807 TV = Cr[WS(csr, 58)]; | |
808 TW = TU + TV; | |
809 T3k = TU - TV; | |
810 T3q = Ci[WS(csi, 6)]; | |
811 T3r = Ci[WS(csi, 58)]; | |
812 T3s = T3q + T3r; | |
813 TaD = T3r - T3q; | |
814 } | |
815 { | |
816 E TX, TY, T3l, T3m; | |
817 TX = Cr[WS(csr, 26)]; | |
818 TY = Cr[WS(csr, 38)]; | |
819 TZ = TX + TY; | |
820 T3p = TX - TY; | |
821 T3l = Ci[WS(csi, 26)]; | |
822 T3m = Ci[WS(csi, 38)]; | |
823 T3n = T3l + T3m; | |
824 TaE = T3l - T3m; | |
825 } | |
826 T3o = T3k - T3n; | |
827 T3t = T3p - T3s; | |
828 TaF = TaD - TaE; | |
829 TaC = TW - TZ; | |
830 T7D = T3k + T3n; | |
831 T7C = T3p + T3s; | |
832 } | |
833 { | |
834 E TT, T10, TaO, TaP; | |
835 TT = TP + TS; | |
836 T10 = TW + TZ; | |
837 T11 = TT + T10; | |
838 TdF = TT - T10; | |
839 TaO = Tax + TaA; | |
840 TaP = TaF - TaC; | |
841 TaQ = TaO + TaP; | |
842 Tct = TaO - TaP; | |
843 } | |
844 { | |
845 E TdB, TdC, T3j, T3u; | |
846 TdB = TaE + TaD; | |
847 TdC = Taz + Tay; | |
848 TdD = TdB - TdC; | |
849 TeQ = TdC + TdB; | |
850 T3j = FNMS(KP414213562, T3i, T3d); | |
851 T3u = FMA(KP414213562, T3t, T3o); | |
852 T3v = T3j + T3u; | |
853 T6m = T3u - T3j; | |
854 } | |
855 { | |
856 E T3G, T3H, T7K, T7L; | |
857 T3G = FMA(KP414213562, T3d, T3i); | |
858 T3H = FNMS(KP414213562, T3o, T3t); | |
859 T3I = T3G + T3H; | |
860 T6j = T3G - T3H; | |
861 T7K = FMA(KP414213562, T7z, T7A); | |
862 T7L = FMA(KP414213562, T7C, T7D); | |
863 T7M = T7K - T7L; | |
864 T99 = T7K + T7L; | |
865 } | |
866 { | |
867 E TaB, TaG, T7B, T7E; | |
868 TaB = Tax - TaA; | |
869 TaG = TaC + TaF; | |
870 TaH = TaB + TaG; | |
871 Tcw = TaG - TaB; | |
872 T7B = FNMS(KP414213562, T7A, T7z); | |
873 T7E = FNMS(KP414213562, T7D, T7C); | |
874 T7F = T7B + T7E; | |
875 T9c = T7B - T7E; | |
876 } | |
877 } | |
878 { | |
879 E T12, Tfg, Tx, Tff, T24, Tfi, Tfl, Tfq, Th, T13, Tfp; | |
880 T12 = TM + T11; | |
881 Tfg = TeQ + TeP; | |
882 Th = FMA(KP2_000000000, Tg, T9); | |
883 Tx = FMA(KP2_000000000, Tw, Th); | |
884 Tff = FNMS(KP2_000000000, Tw, Th); | |
885 { | |
886 E T1y, T23, Tfj, Tfk; | |
887 T1y = T1i + T1x; | |
888 T23 = T1N + T22; | |
889 T24 = T1y + T23; | |
890 Tfi = T1y - T23; | |
891 Tfj = Tf1 + Tf0; | |
892 Tfk = TeW + TeV; | |
893 Tfl = Tfj - Tfk; | |
894 Tfq = Tfk + Tfj; | |
895 } | |
896 T13 = FMA(KP2_000000000, T12, Tx); | |
897 R0[WS(rs, 32)] = FNMS(KP2_000000000, T24, T13); | |
898 R0[0] = FMA(KP2_000000000, T24, T13); | |
899 Tfp = FNMS(KP2_000000000, T12, Tx); | |
900 R0[WS(rs, 16)] = FNMS(KP2_000000000, Tfq, Tfp); | |
901 R0[WS(rs, 48)] = FMA(KP2_000000000, Tfq, Tfp); | |
902 { | |
903 E Tfh, Tfm, Tfn, Tfo; | |
904 Tfh = FNMS(KP2_000000000, Tfg, Tff); | |
905 Tfm = Tfi - Tfl; | |
906 R0[WS(rs, 40)] = FNMS(KP1_414213562, Tfm, Tfh); | |
907 R0[WS(rs, 8)] = FMA(KP1_414213562, Tfm, Tfh); | |
908 Tfn = FMA(KP2_000000000, Tfg, Tff); | |
909 Tfo = Tfi + Tfl; | |
910 R0[WS(rs, 24)] = FNMS(KP1_414213562, Tfo, Tfn); | |
911 R0[WS(rs, 56)] = FMA(KP1_414213562, Tfo, Tfn); | |
912 } | |
913 } | |
914 { | |
915 E TeN, Tf7, Tf3, Tfa, TeS, Tf8, TeY, Tfb, TeL, TeZ, Tf2; | |
916 TeL = FNMS(KP2_000000000, Tg, T9); | |
917 TeN = FNMS(KP2_000000000, TeM, TeL); | |
918 Tf7 = FMA(KP2_000000000, TeM, TeL); | |
919 TeZ = T1N - T22; | |
920 Tf2 = Tf0 - Tf1; | |
921 Tf3 = TeZ + Tf2; | |
922 Tfa = Tf2 - TeZ; | |
923 { | |
924 E TeO, TeR, TeU, TeX; | |
925 TeO = TM - T11; | |
926 TeR = TeP - TeQ; | |
927 TeS = TeO - TeR; | |
928 Tf8 = TeO + TeR; | |
929 TeU = T1i - T1x; | |
930 TeX = TeV - TeW; | |
931 TeY = TeU + TeX; | |
932 Tfb = TeU - TeX; | |
933 } | |
934 { | |
935 E TeT, Tf4, Tfd, Tfe; | |
936 TeT = FMA(KP1_414213562, TeS, TeN); | |
937 Tf4 = FNMS(KP414213562, Tf3, TeY); | |
938 R0[WS(rs, 36)] = FNMS(KP1_847759065, Tf4, TeT); | |
939 R0[WS(rs, 4)] = FMA(KP1_847759065, Tf4, TeT); | |
940 Tfd = FMA(KP1_414213562, Tf8, Tf7); | |
941 Tfe = FMA(KP414213562, Tfa, Tfb); | |
942 R0[WS(rs, 28)] = FNMS(KP1_847759065, Tfe, Tfd); | |
943 R0[WS(rs, 60)] = FMA(KP1_847759065, Tfe, Tfd); | |
944 } | |
945 { | |
946 E Tf5, Tf6, Tf9, Tfc; | |
947 Tf5 = FNMS(KP1_414213562, TeS, TeN); | |
948 Tf6 = FMA(KP414213562, TeY, Tf3); | |
949 R0[WS(rs, 20)] = FNMS(KP1_847759065, Tf6, Tf5); | |
950 R0[WS(rs, 52)] = FMA(KP1_847759065, Tf6, Tf5); | |
951 Tf9 = FNMS(KP1_414213562, Tf8, Tf7); | |
952 Tfc = FNMS(KP414213562, Tfb, Tfa); | |
953 R0[WS(rs, 12)] = FNMS(KP1_847759065, Tfc, Tf9); | |
954 R0[WS(rs, 44)] = FMA(KP1_847759065, Tfc, Tf9); | |
955 } | |
956 } | |
957 { | |
958 E Tar, TbP, TbL, TbS, TaS, TbQ, Tby, TbT; | |
959 { | |
960 E Taf, Taq, TbH, TbK; | |
961 Taf = FMA(KP1_414213562, Tae, Ta9); | |
962 Taq = FNMS(KP414213562, Tap, Tak); | |
963 Tar = FMA(KP1_847759065, Taq, Taf); | |
964 TbP = FNMS(KP1_847759065, Taq, Taf); | |
965 TbH = FMA(KP707106781, TbG, TbD); | |
966 TbK = TbI + TbJ; | |
967 TbL = FMA(KP923879532, TbK, TbH); | |
968 TbS = FNMS(KP923879532, TbK, TbH); | |
969 } | |
970 { | |
971 E TaI, TaR, Tba, Tbx; | |
972 TaI = FMA(KP707106781, TaH, Taw); | |
973 TaR = FMA(KP707106781, TaQ, TaN); | |
974 TaS = FNMS(KP198912367, TaR, TaI); | |
975 TbQ = FMA(KP198912367, TaI, TaR); | |
976 Tba = FMA(KP707106781, Tb9, TaY); | |
977 Tbx = Tbl + Tbw; | |
978 Tby = FMA(KP923879532, Tbx, Tba); | |
979 TbT = FNMS(KP923879532, Tbx, Tba); | |
980 } | |
981 { | |
982 E TaT, TbM, TbV, TbW; | |
983 TaT = FMA(KP1_961570560, TaS, Tar); | |
984 TbM = FNMS(KP098491403, TbL, Tby); | |
985 R0[WS(rs, 33)] = FNMS(KP1_990369453, TbM, TaT); | |
986 R0[WS(rs, 1)] = FMA(KP1_990369453, TbM, TaT); | |
987 TbV = FMA(KP1_961570560, TbQ, TbP); | |
988 TbW = FMA(KP820678790, TbS, TbT); | |
989 R0[WS(rs, 25)] = FNMS(KP1_546020906, TbW, TbV); | |
990 R0[WS(rs, 57)] = FMA(KP1_546020906, TbW, TbV); | |
991 } | |
992 { | |
993 E TbN, TbO, TbR, TbU; | |
994 TbN = FNMS(KP1_961570560, TaS, Tar); | |
995 TbO = FMA(KP098491403, Tby, TbL); | |
996 R0[WS(rs, 17)] = FNMS(KP1_990369453, TbO, TbN); | |
997 R0[WS(rs, 49)] = FMA(KP1_990369453, TbO, TbN); | |
998 TbR = FNMS(KP1_961570560, TbQ, TbP); | |
999 TbU = FNMS(KP820678790, TbT, TbS); | |
1000 R0[WS(rs, 9)] = FNMS(KP1_546020906, TbU, TbR); | |
1001 R0[WS(rs, 41)] = FMA(KP1_546020906, TbU, TbR); | |
1002 } | |
1003 } | |
1004 { | |
1005 E Tdz, Tef, Teb, Tei, TdK, Teg, Te2, Tej; | |
1006 { | |
1007 E Tdt, Tdy, Te7, Tea; | |
1008 Tdt = FNMS(KP2_000000000, Tds, Tdr); | |
1009 Tdy = Tdu - Tdx; | |
1010 Tdz = FMA(KP1_414213562, Tdy, Tdt); | |
1011 Tef = FNMS(KP1_414213562, Tdy, Tdt); | |
1012 Te7 = Te3 + Te6; | |
1013 Tea = Te8 + Te9; | |
1014 Teb = FMA(KP707106781, Tea, Te7); | |
1015 Tei = FNMS(KP707106781, Tea, Te7); | |
1016 } | |
1017 { | |
1018 E TdE, TdJ, TdQ, Te1; | |
1019 TdE = TdA + TdD; | |
1020 TdJ = TdF + TdI; | |
1021 TdK = FNMS(KP414213562, TdJ, TdE); | |
1022 Teg = FMA(KP414213562, TdE, TdJ); | |
1023 TdQ = TdM + TdP; | |
1024 Te1 = TdV + Te0; | |
1025 Te2 = FMA(KP707106781, Te1, TdQ); | |
1026 Tej = FNMS(KP707106781, Te1, TdQ); | |
1027 } | |
1028 { | |
1029 E TdL, Tec, Tel, Tem; | |
1030 TdL = FMA(KP1_847759065, TdK, Tdz); | |
1031 Tec = FNMS(KP198912367, Teb, Te2); | |
1032 R0[WS(rs, 34)] = FNMS(KP1_961570560, Tec, TdL); | |
1033 R0[WS(rs, 2)] = FMA(KP1_961570560, Tec, TdL); | |
1034 Tel = FMA(KP1_847759065, Teg, Tef); | |
1035 Tem = FMA(KP668178637, Tei, Tej); | |
1036 R0[WS(rs, 26)] = FNMS(KP1_662939224, Tem, Tel); | |
1037 R0[WS(rs, 58)] = FMA(KP1_662939224, Tem, Tel); | |
1038 } | |
1039 { | |
1040 E Ted, Tee, Teh, Tek; | |
1041 Ted = FNMS(KP1_847759065, TdK, Tdz); | |
1042 Tee = FMA(KP198912367, Te2, Teb); | |
1043 R0[WS(rs, 18)] = FNMS(KP1_961570560, Tee, Ted); | |
1044 R0[WS(rs, 50)] = FMA(KP1_961570560, Tee, Ted); | |
1045 Teh = FNMS(KP1_847759065, Teg, Tef); | |
1046 Tek = FNMS(KP668178637, Tej, Tei); | |
1047 R0[WS(rs, 10)] = FNMS(KP1_662939224, Tek, Teh); | |
1048 R0[WS(rs, 42)] = FMA(KP1_662939224, Tek, Teh); | |
1049 } | |
1050 } | |
1051 { | |
1052 E TbZ, Tcd, Tc9, Tcg, Tc2, Tce, Tc6, Tch; | |
1053 { | |
1054 E TbX, TbY, Tc7, Tc8; | |
1055 TbX = FNMS(KP1_414213562, Tae, Ta9); | |
1056 TbY = FMA(KP414213562, Tak, Tap); | |
1057 TbZ = FNMS(KP1_847759065, TbY, TbX); | |
1058 Tcd = FMA(KP1_847759065, TbY, TbX); | |
1059 Tc7 = FNMS(KP707106781, TbG, TbD); | |
1060 Tc8 = Tbw - Tbl; | |
1061 Tc9 = FNMS(KP923879532, Tc8, Tc7); | |
1062 Tcg = FMA(KP923879532, Tc8, Tc7); | |
1063 } | |
1064 { | |
1065 E Tc0, Tc1, Tc4, Tc5; | |
1066 Tc0 = FNMS(KP707106781, TaQ, TaN); | |
1067 Tc1 = FNMS(KP707106781, TaH, Taw); | |
1068 Tc2 = FNMS(KP668178637, Tc1, Tc0); | |
1069 Tce = FMA(KP668178637, Tc0, Tc1); | |
1070 Tc4 = FNMS(KP707106781, Tb9, TaY); | |
1071 Tc5 = TbI - TbJ; | |
1072 Tc6 = FNMS(KP923879532, Tc5, Tc4); | |
1073 Tch = FMA(KP923879532, Tc5, Tc4); | |
1074 } | |
1075 { | |
1076 E Tc3, Tca, Tcj, Tck; | |
1077 Tc3 = FNMS(KP1_662939224, Tc2, TbZ); | |
1078 Tca = FNMS(KP534511135, Tc9, Tc6); | |
1079 R0[WS(rs, 37)] = FNMS(KP1_763842528, Tca, Tc3); | |
1080 R0[WS(rs, 5)] = FMA(KP1_763842528, Tca, Tc3); | |
1081 Tcj = FMA(KP1_662939224, Tce, Tcd); | |
1082 Tck = FMA(KP303346683, Tcg, Tch); | |
1083 R0[WS(rs, 29)] = FNMS(KP1_913880671, Tck, Tcj); | |
1084 R0[WS(rs, 61)] = FMA(KP1_913880671, Tck, Tcj); | |
1085 } | |
1086 { | |
1087 E Tcb, Tcc, Tcf, Tci; | |
1088 Tcb = FMA(KP1_662939224, Tc2, TbZ); | |
1089 Tcc = FMA(KP534511135, Tc6, Tc9); | |
1090 R0[WS(rs, 21)] = FNMS(KP1_763842528, Tcc, Tcb); | |
1091 R0[WS(rs, 53)] = FMA(KP1_763842528, Tcc, Tcb); | |
1092 Tcf = FNMS(KP1_662939224, Tce, Tcd); | |
1093 Tci = FNMS(KP303346683, Tch, Tcg); | |
1094 R0[WS(rs, 13)] = FNMS(KP1_913880671, Tci, Tcf); | |
1095 R0[WS(rs, 45)] = FMA(KP1_913880671, Tci, Tcf); | |
1096 } | |
1097 } | |
1098 { | |
1099 E Tep, TeD, Tez, TeG, Tes, TeE, Tew, TeH; | |
1100 { | |
1101 E Ten, Teo, Tex, Tey; | |
1102 Ten = FMA(KP2_000000000, Tds, Tdr); | |
1103 Teo = Tdu + Tdx; | |
1104 Tep = FNMS(KP1_414213562, Teo, Ten); | |
1105 TeD = FMA(KP1_414213562, Teo, Ten); | |
1106 Tex = Te6 - Te3; | |
1107 Tey = Te0 - TdV; | |
1108 Tez = FNMS(KP707106781, Tey, Tex); | |
1109 TeG = FMA(KP707106781, Tey, Tex); | |
1110 } | |
1111 { | |
1112 E Teq, Ter, Teu, Tev; | |
1113 Teq = TdI - TdF; | |
1114 Ter = TdA - TdD; | |
1115 Tes = FNMS(KP414213562, Ter, Teq); | |
1116 TeE = FMA(KP414213562, Teq, Ter); | |
1117 Teu = TdM - TdP; | |
1118 Tev = Te8 - Te9; | |
1119 Tew = FNMS(KP707106781, Tev, Teu); | |
1120 TeH = FMA(KP707106781, Tev, Teu); | |
1121 } | |
1122 { | |
1123 E Tet, TeA, TeJ, TeK; | |
1124 Tet = FNMS(KP1_847759065, Tes, Tep); | |
1125 TeA = FNMS(KP668178637, Tez, Tew); | |
1126 R0[WS(rs, 38)] = FNMS(KP1_662939224, TeA, Tet); | |
1127 R0[WS(rs, 6)] = FMA(KP1_662939224, TeA, Tet); | |
1128 TeJ = FMA(KP1_847759065, TeE, TeD); | |
1129 TeK = FMA(KP198912367, TeG, TeH); | |
1130 R0[WS(rs, 30)] = FNMS(KP1_961570560, TeK, TeJ); | |
1131 R0[WS(rs, 62)] = FMA(KP1_961570560, TeK, TeJ); | |
1132 } | |
1133 { | |
1134 E TeB, TeC, TeF, TeI; | |
1135 TeB = FMA(KP1_847759065, Tes, Tep); | |
1136 TeC = FMA(KP668178637, Tew, Tez); | |
1137 R0[WS(rs, 22)] = FNMS(KP1_662939224, TeC, TeB); | |
1138 R0[WS(rs, 54)] = FMA(KP1_662939224, TeC, TeB); | |
1139 TeF = FNMS(KP1_847759065, TeE, TeD); | |
1140 TeI = FNMS(KP198912367, TeH, TeG); | |
1141 R0[WS(rs, 14)] = FNMS(KP1_961570560, TeI, TeF); | |
1142 R0[WS(rs, 46)] = FMA(KP1_961570560, TeI, TeF); | |
1143 } | |
1144 } | |
1145 { | |
1146 E Tcr, TcV, TcR, TcY, Tcy, TcW, TcK, TcZ; | |
1147 { | |
1148 E Tcn, Tcq, TcN, TcQ; | |
1149 Tcn = FNMS(KP1_414213562, Tcm, Tcl); | |
1150 Tcq = FNMS(KP414213562, Tcp, Tco); | |
1151 Tcr = FNMS(KP1_847759065, Tcq, Tcn); | |
1152 TcV = FMA(KP1_847759065, Tcq, Tcn); | |
1153 TcN = FNMS(KP707106781, TcM, TcL); | |
1154 TcQ = TcO - TcP; | |
1155 TcR = FMA(KP923879532, TcQ, TcN); | |
1156 TcY = FNMS(KP923879532, TcQ, TcN); | |
1157 } | |
1158 { | |
1159 E Tcu, Tcx, TcC, TcJ; | |
1160 Tcu = FNMS(KP707106781, Tct, Tcs); | |
1161 Tcx = FNMS(KP707106781, Tcw, Tcv); | |
1162 Tcy = FNMS(KP668178637, Tcx, Tcu); | |
1163 TcW = FMA(KP668178637, Tcu, Tcx); | |
1164 TcC = FNMS(KP707106781, TcB, TcA); | |
1165 TcJ = TcF - TcI; | |
1166 TcK = FMA(KP923879532, TcJ, TcC); | |
1167 TcZ = FNMS(KP923879532, TcJ, TcC); | |
1168 } | |
1169 { | |
1170 E Tcz, TcS, Td1, Td2; | |
1171 Tcz = FMA(KP1_662939224, Tcy, Tcr); | |
1172 TcS = FNMS(KP303346683, TcR, TcK); | |
1173 R0[WS(rs, 35)] = FNMS(KP1_913880671, TcS, Tcz); | |
1174 R0[WS(rs, 3)] = FMA(KP1_913880671, TcS, Tcz); | |
1175 Td1 = FMA(KP1_662939224, TcW, TcV); | |
1176 Td2 = FMA(KP534511135, TcY, TcZ); | |
1177 R0[WS(rs, 27)] = FNMS(KP1_763842528, Td2, Td1); | |
1178 R0[WS(rs, 59)] = FMA(KP1_763842528, Td2, Td1); | |
1179 } | |
1180 { | |
1181 E TcT, TcU, TcX, Td0; | |
1182 TcT = FNMS(KP1_662939224, Tcy, Tcr); | |
1183 TcU = FMA(KP303346683, TcK, TcR); | |
1184 R0[WS(rs, 19)] = FNMS(KP1_913880671, TcU, TcT); | |
1185 R0[WS(rs, 51)] = FMA(KP1_913880671, TcU, TcT); | |
1186 TcX = FNMS(KP1_662939224, TcW, TcV); | |
1187 Td0 = FNMS(KP534511135, TcZ, TcY); | |
1188 R0[WS(rs, 11)] = FNMS(KP1_763842528, Td0, TcX); | |
1189 R0[WS(rs, 43)] = FMA(KP1_763842528, Td0, TcX); | |
1190 } | |
1191 } | |
1192 { | |
1193 E Td5, Tdj, Tdf, Tdm, Td8, Tdk, Tdc, Tdn; | |
1194 { | |
1195 E Td3, Td4, Tdd, Tde; | |
1196 Td3 = FMA(KP1_414213562, Tcm, Tcl); | |
1197 Td4 = FMA(KP414213562, Tco, Tcp); | |
1198 Td5 = FNMS(KP1_847759065, Td4, Td3); | |
1199 Tdj = FMA(KP1_847759065, Td4, Td3); | |
1200 Tdd = FMA(KP707106781, TcM, TcL); | |
1201 Tde = TcI + TcF; | |
1202 Tdf = FNMS(KP923879532, Tde, Tdd); | |
1203 Tdm = FMA(KP923879532, Tde, Tdd); | |
1204 } | |
1205 { | |
1206 E Td6, Td7, Tda, Tdb; | |
1207 Td6 = FMA(KP707106781, Tcw, Tcv); | |
1208 Td7 = FMA(KP707106781, Tct, Tcs); | |
1209 Td8 = FNMS(KP198912367, Td7, Td6); | |
1210 Tdk = FMA(KP198912367, Td6, Td7); | |
1211 Tda = FMA(KP707106781, TcB, TcA); | |
1212 Tdb = TcO + TcP; | |
1213 Tdc = FNMS(KP923879532, Tdb, Tda); | |
1214 Tdn = FMA(KP923879532, Tdb, Tda); | |
1215 } | |
1216 { | |
1217 E Td9, Tdg, Tdp, Tdq; | |
1218 Td9 = FNMS(KP1_961570560, Td8, Td5); | |
1219 Tdg = FNMS(KP820678790, Tdf, Tdc); | |
1220 R0[WS(rs, 39)] = FNMS(KP1_546020906, Tdg, Td9); | |
1221 R0[WS(rs, 7)] = FMA(KP1_546020906, Tdg, Td9); | |
1222 Tdp = FMA(KP1_961570560, Tdk, Tdj); | |
1223 Tdq = FMA(KP098491403, Tdm, Tdn); | |
1224 R0[WS(rs, 31)] = FNMS(KP1_990369453, Tdq, Tdp); | |
1225 R0[WS(rs, 63)] = FMA(KP1_990369453, Tdq, Tdp); | |
1226 } | |
1227 { | |
1228 E Tdh, Tdi, Tdl, Tdo; | |
1229 Tdh = FMA(KP1_961570560, Td8, Td5); | |
1230 Tdi = FMA(KP820678790, Tdc, Tdf); | |
1231 R0[WS(rs, 23)] = FNMS(KP1_546020906, Tdi, Tdh); | |
1232 R0[WS(rs, 55)] = FMA(KP1_546020906, Tdi, Tdh); | |
1233 Tdl = FNMS(KP1_961570560, Tdk, Tdj); | |
1234 Tdo = FNMS(KP098491403, Tdn, Tdm); | |
1235 R0[WS(rs, 15)] = FNMS(KP1_990369453, Tdo, Tdl); | |
1236 R0[WS(rs, 47)] = FMA(KP1_990369453, Tdo, Tdl); | |
1237 } | |
1238 } | |
1239 { | |
1240 E T2R, T5F, T3K, T5G, T5S, T64, T5P, T63, T5B, T67, T5I, T5W, T5k, T66, T5J; | |
1241 E T5Z; | |
1242 { | |
1243 E T2p, T2Q, T5N, T5O; | |
1244 T2p = FMA(KP1_847759065, T2o, T2d); | |
1245 T2Q = FNMS(KP198912367, T2P, T2G); | |
1246 T2R = FMA(KP1_961570560, T2Q, T2p); | |
1247 T5F = FNMS(KP1_961570560, T2Q, T2p); | |
1248 { | |
1249 E T3w, T3J, T5Q, T5R; | |
1250 T3w = FMA(KP923879532, T3v, T38); | |
1251 T3J = FMA(KP923879532, T3I, T3F); | |
1252 T3K = FNMS(KP098491403, T3J, T3w); | |
1253 T5G = FMA(KP098491403, T3w, T3J); | |
1254 T5Q = FNMS(KP923879532, T3I, T3F); | |
1255 T5R = FNMS(KP923879532, T3v, T38); | |
1256 T5S = FNMS(KP820678790, T5R, T5Q); | |
1257 T64 = FMA(KP820678790, T5Q, T5R); | |
1258 } | |
1259 T5N = FNMS(KP1_847759065, T2o, T2d); | |
1260 T5O = FMA(KP198912367, T2G, T2P); | |
1261 T5P = FNMS(KP1_961570560, T5O, T5N); | |
1262 T63 = FMA(KP1_961570560, T5O, T5N); | |
1263 { | |
1264 E T5x, T5U, T5A, T5V, T5y, T5z; | |
1265 T5x = FMA(KP923879532, T5w, T5t); | |
1266 T5U = FNMS(KP923879532, T4p, T42); | |
1267 T5y = FMA(KP198912367, T4H, T4Q); | |
1268 T5z = FNMS(KP198912367, T58, T5h); | |
1269 T5A = T5y + T5z; | |
1270 T5V = T5y - T5z; | |
1271 T5B = FMA(KP980785280, T5A, T5x); | |
1272 T67 = FMA(KP980785280, T5V, T5U); | |
1273 T5I = FNMS(KP980785280, T5A, T5x); | |
1274 T5W = FNMS(KP980785280, T5V, T5U); | |
1275 } | |
1276 { | |
1277 E T4q, T5X, T5j, T5Y, T4R, T5i; | |
1278 T4q = FMA(KP923879532, T4p, T42); | |
1279 T5X = FNMS(KP923879532, T5w, T5t); | |
1280 T4R = FNMS(KP198912367, T4Q, T4H); | |
1281 T5i = FMA(KP198912367, T5h, T58); | |
1282 T5j = T4R + T5i; | |
1283 T5Y = T5i - T4R; | |
1284 T5k = FMA(KP980785280, T5j, T4q); | |
1285 T66 = FMA(KP980785280, T5Y, T5X); | |
1286 T5J = FNMS(KP980785280, T5j, T4q); | |
1287 T5Z = FNMS(KP980785280, T5Y, T5X); | |
1288 } | |
1289 } | |
1290 { | |
1291 E T3L, T5C, T65, T68; | |
1292 T3L = FMA(KP1_990369453, T3K, T2R); | |
1293 T5C = FNMS(KP049126849, T5B, T5k); | |
1294 R1[WS(rs, 32)] = FNMS(KP1_997590912, T5C, T3L); | |
1295 R1[0] = FMA(KP1_997590912, T5C, T3L); | |
1296 T65 = FNMS(KP1_546020906, T64, T63); | |
1297 T68 = FNMS(KP357805721, T67, T66); | |
1298 R1[WS(rs, 12)] = FNMS(KP1_883088130, T68, T65); | |
1299 R1[WS(rs, 44)] = FMA(KP1_883088130, T68, T65); | |
1300 } | |
1301 { | |
1302 E T69, T6a, T5D, T5E; | |
1303 T69 = FMA(KP1_546020906, T64, T63); | |
1304 T6a = FMA(KP357805721, T66, T67); | |
1305 R1[WS(rs, 28)] = FNMS(KP1_883088130, T6a, T69); | |
1306 R1[WS(rs, 60)] = FMA(KP1_883088130, T6a, T69); | |
1307 T5D = FNMS(KP1_990369453, T3K, T2R); | |
1308 T5E = FMA(KP049126849, T5k, T5B); | |
1309 R1[WS(rs, 16)] = FNMS(KP1_997590912, T5E, T5D); | |
1310 R1[WS(rs, 48)] = FMA(KP1_997590912, T5E, T5D); | |
1311 } | |
1312 { | |
1313 E T5H, T5K, T5T, T60; | |
1314 T5H = FNMS(KP1_990369453, T5G, T5F); | |
1315 T5K = FNMS(KP906347169, T5J, T5I); | |
1316 R1[WS(rs, 8)] = FNMS(KP1_481902250, T5K, T5H); | |
1317 R1[WS(rs, 40)] = FMA(KP1_481902250, T5K, T5H); | |
1318 T5T = FNMS(KP1_546020906, T5S, T5P); | |
1319 T60 = FNMS(KP472964775, T5Z, T5W); | |
1320 R1[WS(rs, 36)] = FNMS(KP1_807978586, T60, T5T); | |
1321 R1[WS(rs, 4)] = FMA(KP1_807978586, T60, T5T); | |
1322 } | |
1323 { | |
1324 E T61, T62, T5L, T5M; | |
1325 T61 = FMA(KP1_546020906, T5S, T5P); | |
1326 T62 = FMA(KP472964775, T5W, T5Z); | |
1327 R1[WS(rs, 20)] = FNMS(KP1_807978586, T62, T61); | |
1328 R1[WS(rs, 52)] = FMA(KP1_807978586, T62, T61); | |
1329 T5L = FMA(KP1_990369453, T5G, T5F); | |
1330 T5M = FMA(KP906347169, T5I, T5J); | |
1331 R1[WS(rs, 24)] = FNMS(KP1_481902250, T5M, T5L); | |
1332 R1[WS(rs, 56)] = FMA(KP1_481902250, T5M, T5L); | |
1333 } | |
1334 } | |
1335 { | |
1336 E T7v, T8v, T7O, T8w, T8I, T8U, T8F, T8T, T8r, T8X, T8y, T8M, T8g, T8W, T8z; | |
1337 E T8P; | |
1338 { | |
1339 E T7n, T7u, T8D, T8E; | |
1340 T7n = FNMS(KP1_847759065, T7m, T7j); | |
1341 T7u = FNMS(KP668178637, T7t, T7q); | |
1342 T7v = FMA(KP1_662939224, T7u, T7n); | |
1343 T8v = FNMS(KP1_662939224, T7u, T7n); | |
1344 { | |
1345 E T7G, T7N, T8G, T8H; | |
1346 T7G = FNMS(KP923879532, T7F, T7y); | |
1347 T7N = FMA(KP923879532, T7M, T7J); | |
1348 T7O = FNMS(KP303346683, T7N, T7G); | |
1349 T8w = FMA(KP303346683, T7G, T7N); | |
1350 T8G = FNMS(KP923879532, T7M, T7J); | |
1351 T8H = FMA(KP923879532, T7F, T7y); | |
1352 T8I = FNMS(KP534511135, T8H, T8G); | |
1353 T8U = FMA(KP534511135, T8G, T8H); | |
1354 } | |
1355 T8D = FMA(KP1_847759065, T7m, T7j); | |
1356 T8E = FMA(KP668178637, T7q, T7t); | |
1357 T8F = FNMS(KP1_662939224, T8E, T8D); | |
1358 T8T = FMA(KP1_662939224, T8E, T8D); | |
1359 { | |
1360 E T8n, T8K, T8q, T8L, T8o, T8p; | |
1361 T8n = FMA(KP923879532, T8m, T8j); | |
1362 T8K = FMA(KP923879532, T7Z, T7S); | |
1363 T8o = FMA(KP668178637, T83, T86); | |
1364 T8p = FMA(KP668178637, T8a, T8d); | |
1365 T8q = T8o - T8p; | |
1366 T8L = T8o + T8p; | |
1367 T8r = FMA(KP831469612, T8q, T8n); | |
1368 T8X = FMA(KP831469612, T8L, T8K); | |
1369 T8y = FNMS(KP831469612, T8q, T8n); | |
1370 T8M = FNMS(KP831469612, T8L, T8K); | |
1371 } | |
1372 { | |
1373 E T80, T8N, T8f, T8O, T87, T8e; | |
1374 T80 = FNMS(KP923879532, T7Z, T7S); | |
1375 T8N = FNMS(KP923879532, T8m, T8j); | |
1376 T87 = FNMS(KP668178637, T86, T83); | |
1377 T8e = FNMS(KP668178637, T8d, T8a); | |
1378 T8f = T87 + T8e; | |
1379 T8O = T8e - T87; | |
1380 T8g = FMA(KP831469612, T8f, T80); | |
1381 T8W = FMA(KP831469612, T8O, T8N); | |
1382 T8z = FNMS(KP831469612, T8f, T80); | |
1383 T8P = FNMS(KP831469612, T8O, T8N); | |
1384 } | |
1385 } | |
1386 { | |
1387 E T7P, T8s, T8V, T8Y; | |
1388 T7P = FMA(KP1_913880671, T7O, T7v); | |
1389 T8s = FNMS(KP148335987, T8r, T8g); | |
1390 R1[WS(rs, 33)] = FNMS(KP1_978353019, T8s, T7P); | |
1391 R1[WS(rs, 1)] = FMA(KP1_978353019, T8s, T7P); | |
1392 T8V = FNMS(KP1_763842528, T8U, T8T); | |
1393 T8Y = FNMS(KP250486960, T8X, T8W); | |
1394 R1[WS(rs, 13)] = FNMS(KP1_940062506, T8Y, T8V); | |
1395 R1[WS(rs, 45)] = FMA(KP1_940062506, T8Y, T8V); | |
1396 } | |
1397 { | |
1398 E T8Z, T90, T8t, T8u; | |
1399 T8Z = FMA(KP1_763842528, T8U, T8T); | |
1400 T90 = FMA(KP250486960, T8W, T8X); | |
1401 R1[WS(rs, 29)] = FNMS(KP1_940062506, T90, T8Z); | |
1402 R1[WS(rs, 61)] = FMA(KP1_940062506, T90, T8Z); | |
1403 T8t = FNMS(KP1_913880671, T7O, T7v); | |
1404 T8u = FMA(KP148335987, T8g, T8r); | |
1405 R1[WS(rs, 17)] = FNMS(KP1_978353019, T8u, T8t); | |
1406 R1[WS(rs, 49)] = FMA(KP1_978353019, T8u, T8t); | |
1407 } | |
1408 { | |
1409 E T8x, T8A, T8J, T8Q; | |
1410 T8x = FNMS(KP1_913880671, T8w, T8v); | |
1411 T8A = FNMS(KP741650546, T8z, T8y); | |
1412 R1[WS(rs, 9)] = FNMS(KP1_606415062, T8A, T8x); | |
1413 R1[WS(rs, 41)] = FMA(KP1_606415062, T8A, T8x); | |
1414 T8J = FNMS(KP1_763842528, T8I, T8F); | |
1415 T8Q = FNMS(KP599376933, T8P, T8M); | |
1416 R1[WS(rs, 37)] = FNMS(KP1_715457220, T8Q, T8J); | |
1417 R1[WS(rs, 5)] = FMA(KP1_715457220, T8Q, T8J); | |
1418 } | |
1419 { | |
1420 E T8R, T8S, T8B, T8C; | |
1421 T8R = FMA(KP1_763842528, T8I, T8F); | |
1422 T8S = FMA(KP599376933, T8M, T8P); | |
1423 R1[WS(rs, 21)] = FNMS(KP1_715457220, T8S, T8R); | |
1424 R1[WS(rs, 53)] = FMA(KP1_715457220, T8S, T8R); | |
1425 T8B = FMA(KP1_913880671, T8w, T8v); | |
1426 T8C = FMA(KP741650546, T8y, T8z); | |
1427 R1[WS(rs, 25)] = FNMS(KP1_606415062, T8C, T8B); | |
1428 R1[WS(rs, 57)] = FMA(KP1_606415062, T8C, T8B); | |
1429 } | |
1430 } | |
1431 { | |
1432 E T6h, T6L, T6o, T6M, T6Y, T7a, T6V, T79, T6H, T7d, T6O, T72, T6A, T7c, T6P; | |
1433 E T75; | |
1434 { | |
1435 E T6d, T6g, T6T, T6U; | |
1436 T6d = FNMS(KP1_847759065, T6c, T6b); | |
1437 T6g = FNMS(KP668178637, T6f, T6e); | |
1438 T6h = FNMS(KP1_662939224, T6g, T6d); | |
1439 T6L = FMA(KP1_662939224, T6g, T6d); | |
1440 { | |
1441 E T6k, T6n, T6W, T6X; | |
1442 T6k = FNMS(KP923879532, T6j, T6i); | |
1443 T6n = FNMS(KP923879532, T6m, T6l); | |
1444 T6o = FNMS(KP534511135, T6n, T6k); | |
1445 T6M = FMA(KP534511135, T6k, T6n); | |
1446 T6W = FMA(KP923879532, T6m, T6l); | |
1447 T6X = FMA(KP923879532, T6j, T6i); | |
1448 T6Y = FNMS(KP303346683, T6X, T6W); | |
1449 T7a = FMA(KP303346683, T6W, T6X); | |
1450 } | |
1451 T6T = FMA(KP1_847759065, T6c, T6b); | |
1452 T6U = FMA(KP668178637, T6e, T6f); | |
1453 T6V = FNMS(KP1_662939224, T6U, T6T); | |
1454 T79 = FMA(KP1_662939224, T6U, T6T); | |
1455 { | |
1456 E T6D, T70, T6G, T71, T6E, T6F; | |
1457 T6D = FNMS(KP923879532, T6C, T6B); | |
1458 T70 = FMA(KP923879532, T6r, T6q); | |
1459 T6E = FMA(KP668178637, T6w, T6x); | |
1460 T6F = FNMS(KP668178637, T6t, T6u); | |
1461 T6G = T6E - T6F; | |
1462 T71 = T6E + T6F; | |
1463 T6H = FMA(KP831469612, T6G, T6D); | |
1464 T7d = FMA(KP831469612, T71, T70); | |
1465 T6O = FNMS(KP831469612, T6G, T6D); | |
1466 T72 = FNMS(KP831469612, T71, T70); | |
1467 } | |
1468 { | |
1469 E T6s, T73, T6z, T74, T6v, T6y; | |
1470 T6s = FNMS(KP923879532, T6r, T6q); | |
1471 T73 = FMA(KP923879532, T6C, T6B); | |
1472 T6v = FMA(KP668178637, T6u, T6t); | |
1473 T6y = FNMS(KP668178637, T6x, T6w); | |
1474 T6z = T6v - T6y; | |
1475 T74 = T6y + T6v; | |
1476 T6A = FMA(KP831469612, T6z, T6s); | |
1477 T7c = FMA(KP831469612, T74, T73); | |
1478 T6P = FNMS(KP831469612, T6z, T6s); | |
1479 T75 = FNMS(KP831469612, T74, T73); | |
1480 } | |
1481 } | |
1482 { | |
1483 E T6p, T6I, T7b, T7e; | |
1484 T6p = FMA(KP1_763842528, T6o, T6h); | |
1485 T6I = FNMS(KP250486960, T6H, T6A); | |
1486 R1[WS(rs, 34)] = FNMS(KP1_940062506, T6I, T6p); | |
1487 R1[WS(rs, 2)] = FMA(KP1_940062506, T6I, T6p); | |
1488 T7b = FNMS(KP1_913880671, T7a, T79); | |
1489 T7e = FNMS(KP148335987, T7d, T7c); | |
1490 R1[WS(rs, 14)] = FNMS(KP1_978353019, T7e, T7b); | |
1491 R1[WS(rs, 46)] = FMA(KP1_978353019, T7e, T7b); | |
1492 } | |
1493 { | |
1494 E T7f, T7g, T6J, T6K; | |
1495 T7f = FMA(KP1_913880671, T7a, T79); | |
1496 T7g = FMA(KP148335987, T7c, T7d); | |
1497 R1[WS(rs, 30)] = FNMS(KP1_978353019, T7g, T7f); | |
1498 R1[WS(rs, 62)] = FMA(KP1_978353019, T7g, T7f); | |
1499 T6J = FNMS(KP1_763842528, T6o, T6h); | |
1500 T6K = FMA(KP250486960, T6A, T6H); | |
1501 R1[WS(rs, 18)] = FNMS(KP1_940062506, T6K, T6J); | |
1502 R1[WS(rs, 50)] = FMA(KP1_940062506, T6K, T6J); | |
1503 } | |
1504 { | |
1505 E T6N, T6Q, T6Z, T76; | |
1506 T6N = FNMS(KP1_763842528, T6M, T6L); | |
1507 T6Q = FNMS(KP599376933, T6P, T6O); | |
1508 R1[WS(rs, 10)] = FNMS(KP1_715457220, T6Q, T6N); | |
1509 R1[WS(rs, 42)] = FMA(KP1_715457220, T6Q, T6N); | |
1510 T6Z = FNMS(KP1_913880671, T6Y, T6V); | |
1511 T76 = FNMS(KP741650546, T75, T72); | |
1512 R1[WS(rs, 38)] = FNMS(KP1_606415062, T76, T6Z); | |
1513 R1[WS(rs, 6)] = FMA(KP1_606415062, T76, T6Z); | |
1514 } | |
1515 { | |
1516 E T77, T78, T6R, T6S; | |
1517 T77 = FMA(KP1_913880671, T6Y, T6V); | |
1518 T78 = FMA(KP741650546, T72, T75); | |
1519 R1[WS(rs, 22)] = FNMS(KP1_606415062, T78, T77); | |
1520 R1[WS(rs, 54)] = FMA(KP1_606415062, T78, T77); | |
1521 T6R = FMA(KP1_763842528, T6M, T6L); | |
1522 T6S = FMA(KP599376933, T6O, T6P); | |
1523 R1[WS(rs, 26)] = FNMS(KP1_715457220, T6S, T6R); | |
1524 R1[WS(rs, 58)] = FMA(KP1_715457220, T6S, T6R); | |
1525 } | |
1526 } | |
1527 { | |
1528 E T97, T9B, T9e, T9C, T9O, Ta0, T9L, T9Z, T9x, Ta3, T9E, T9S, T9q, Ta2, T9F; | |
1529 E T9V; | |
1530 { | |
1531 E T93, T96, T9J, T9K; | |
1532 T93 = FNMS(KP1_847759065, T92, T91); | |
1533 T96 = FNMS(KP198912367, T95, T94); | |
1534 T97 = FNMS(KP1_961570560, T96, T93); | |
1535 T9B = FMA(KP1_961570560, T96, T93); | |
1536 { | |
1537 E T9a, T9d, T9M, T9N; | |
1538 T9a = FNMS(KP923879532, T99, T98); | |
1539 T9d = FNMS(KP923879532, T9c, T9b); | |
1540 T9e = FNMS(KP820678790, T9d, T9a); | |
1541 T9C = FMA(KP820678790, T9a, T9d); | |
1542 T9M = FMA(KP923879532, T9c, T9b); | |
1543 T9N = FMA(KP923879532, T99, T98); | |
1544 T9O = FNMS(KP098491403, T9N, T9M); | |
1545 Ta0 = FMA(KP098491403, T9M, T9N); | |
1546 } | |
1547 T9J = FMA(KP1_847759065, T92, T91); | |
1548 T9K = FMA(KP198912367, T94, T95); | |
1549 T9L = FNMS(KP1_961570560, T9K, T9J); | |
1550 T9Z = FMA(KP1_961570560, T9K, T9J); | |
1551 { | |
1552 E T9t, T9Q, T9w, T9R, T9u, T9v; | |
1553 T9t = FNMS(KP923879532, T9s, T9r); | |
1554 T9Q = FMA(KP923879532, T9h, T9g); | |
1555 T9u = FMA(KP198912367, T9j, T9k); | |
1556 T9v = FMA(KP198912367, T9m, T9n); | |
1557 T9w = T9u - T9v; | |
1558 T9R = T9u + T9v; | |
1559 T9x = FMA(KP980785280, T9w, T9t); | |
1560 Ta3 = FMA(KP980785280, T9R, T9Q); | |
1561 T9E = FNMS(KP980785280, T9w, T9t); | |
1562 T9S = FNMS(KP980785280, T9R, T9Q); | |
1563 } | |
1564 { | |
1565 E T9i, T9T, T9p, T9U, T9l, T9o; | |
1566 T9i = FNMS(KP923879532, T9h, T9g); | |
1567 T9T = FMA(KP923879532, T9s, T9r); | |
1568 T9l = FNMS(KP198912367, T9k, T9j); | |
1569 T9o = FNMS(KP198912367, T9n, T9m); | |
1570 T9p = T9l + T9o; | |
1571 T9U = T9l - T9o; | |
1572 T9q = FNMS(KP980785280, T9p, T9i); | |
1573 Ta2 = FMA(KP980785280, T9U, T9T); | |
1574 T9F = FMA(KP980785280, T9p, T9i); | |
1575 T9V = FNMS(KP980785280, T9U, T9T); | |
1576 } | |
1577 } | |
1578 { | |
1579 E T9f, T9y, Ta1, Ta4; | |
1580 T9f = FMA(KP1_546020906, T9e, T97); | |
1581 T9y = FNMS(KP357805721, T9x, T9q); | |
1582 R1[WS(rs, 35)] = FNMS(KP1_883088130, T9y, T9f); | |
1583 R1[WS(rs, 3)] = FMA(KP1_883088130, T9y, T9f); | |
1584 Ta1 = FNMS(KP1_990369453, Ta0, T9Z); | |
1585 Ta4 = FNMS(KP049126849, Ta3, Ta2); | |
1586 R1[WS(rs, 15)] = FNMS(KP1_997590912, Ta4, Ta1); | |
1587 R1[WS(rs, 47)] = FMA(KP1_997590912, Ta4, Ta1); | |
1588 } | |
1589 { | |
1590 E Ta5, Ta6, T9z, T9A; | |
1591 Ta5 = FMA(KP1_990369453, Ta0, T9Z); | |
1592 Ta6 = FMA(KP049126849, Ta2, Ta3); | |
1593 R1[WS(rs, 31)] = FNMS(KP1_997590912, Ta6, Ta5); | |
1594 R1[WS(rs, 63)] = FMA(KP1_997590912, Ta6, Ta5); | |
1595 T9z = FNMS(KP1_546020906, T9e, T97); | |
1596 T9A = FMA(KP357805721, T9q, T9x); | |
1597 R1[WS(rs, 19)] = FNMS(KP1_883088130, T9A, T9z); | |
1598 R1[WS(rs, 51)] = FMA(KP1_883088130, T9A, T9z); | |
1599 } | |
1600 { | |
1601 E T9D, T9G, T9P, T9W; | |
1602 T9D = FNMS(KP1_546020906, T9C, T9B); | |
1603 T9G = FNMS(KP472964775, T9F, T9E); | |
1604 R1[WS(rs, 11)] = FNMS(KP1_807978586, T9G, T9D); | |
1605 R1[WS(rs, 43)] = FMA(KP1_807978586, T9G, T9D); | |
1606 T9P = FNMS(KP1_990369453, T9O, T9L); | |
1607 T9W = FNMS(KP906347169, T9V, T9S); | |
1608 R1[WS(rs, 39)] = FNMS(KP1_481902250, T9W, T9P); | |
1609 R1[WS(rs, 7)] = FMA(KP1_481902250, T9W, T9P); | |
1610 } | |
1611 { | |
1612 E T9X, T9Y, T9H, T9I; | |
1613 T9X = FMA(KP1_990369453, T9O, T9L); | |
1614 T9Y = FMA(KP906347169, T9S, T9V); | |
1615 R1[WS(rs, 23)] = FNMS(KP1_481902250, T9Y, T9X); | |
1616 R1[WS(rs, 55)] = FMA(KP1_481902250, T9Y, T9X); | |
1617 T9H = FMA(KP1_546020906, T9C, T9B); | |
1618 T9I = FMA(KP472964775, T9E, T9F); | |
1619 R1[WS(rs, 27)] = FNMS(KP1_807978586, T9I, T9H); | |
1620 R1[WS(rs, 59)] = FMA(KP1_807978586, T9I, T9H); | |
1621 } | |
1622 } | |
1623 } | |
1624 } | |
1625 } | |
1626 | |
1627 static const kr2c_desc desc = { 128, "r2cb_128", {416, 0, 540, 0}, &GENUS }; | |
1628 | |
1629 void X(codelet_r2cb_128) (planner *p) { | |
1630 X(kr2c_register) (p, r2cb_128, &desc); | |
1631 } | |
1632 | |
1633 #else | |
1634 | |
1635 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include rdft/scalar/r2cb.h */ | |
1636 | |
1637 /* | |
1638 * This function contains 956 FP additions, 342 FP multiplications, | |
1639 * (or, 812 additions, 198 multiplications, 144 fused multiply/add), | |
1640 * 198 stack variables, 39 constants, and 256 memory accesses | |
1641 */ | |
1642 #include "rdft/scalar/r2cb.h" | |
1643 | |
1644 static void r2cb_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
1645 { | |
1646 DK(KP1_028205488, +1.028205488386443453187387677937631545216098241); | |
1647 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599); | |
1648 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854); | |
1649 DK(KP1_191398608, +1.191398608984866686934073057659939779023852677); | |
1650 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924); | |
1651 DK(KP485960359, +0.485960359806527779896548324154942236641981567); | |
1652 DK(KP293460948, +0.293460948910723503317700259293435639412430633); | |
1653 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067); | |
1654 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
1655 DK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
1656 DK(KP855110186, +0.855110186860564188641933713777597068609157259); | |
1657 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050); | |
1658 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619); | |
1659 DK(KP1_343117909, +1.343117909694036801250753700854843606457501264); | |
1660 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592); | |
1661 DK(KP673779706, +0.673779706784440101378506425238295140955533559); | |
1662 DK(KP098135348, +0.098135348654836028509909953885365316629490726); | |
1663 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229); | |
1664 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
1665 DK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
1666 DK(KP580569354, +0.580569354508924735272384751634790549382952557); | |
1667 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411); | |
1668 DK(KP942793473, +0.942793473651995297112775251810508755314920638); | |
1669 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241); | |
1670 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382); | |
1671 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624); | |
1672 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190); | |
1673 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586); | |
1674 DK(KP196034280, +0.196034280659121203988391127777283691722273346); | |
1675 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737); | |
1676 DK(KP390180644, +0.390180644032256535696569736954044481855383236); | |
1677 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462); | |
1678 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
1679 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
1680 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
1681 DK(KP765366864, +0.765366864730179543456919968060797733522689125); | |
1682 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252); | |
1683 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875); | |
1684 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
1685 { | |
1686 INT i; | |
1687 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)) { | |
1688 E Ta, T6q, T2a, T5k, T8x, Tbx, TcF, Ten, Th, T6r, T2j, T5l, T8E, Tby, TcI; | |
1689 E Teo, Tx, T6t, TcM, Teq, TcP, Ter, T2t, T5n, T2C, T5o, T8Q, TbA, T8X, TbB; | |
1690 E T6w, T7L, T1j, T6L, Tde, TeC, TdL, TeR, T3v, T5z, T4I, T5O, T9O, TbM, TaV; | |
1691 E Tc1, T78, T7Z, TN, T6z, TcU, Teu, Td8, Tey, T2N, T5r, T3j, T5v, T9a, TbE; | |
1692 E T9A, TbI, T6H, T7O, T1O, T7V, T48, T4u, Tds, TeG, T5E, T5K, Taf, TbP, Tdp; | |
1693 E TeF, T6U, T72, Tam, TbQ, T23, T7U, T4r, T4v, Tdz, TeJ, T5H, T5L, Tay, TbS; | |
1694 E Tdw, TeI, T6Z, T73, TaF, TbT, T1y, T75, Tdl, TeQ, TdI, TeD, T3O, T5N, T4z; | |
1695 E T5A, Ta3, Tc0, TaO, TbN, T6O, T80, T12, T6E, Td1, Tex, Td5, Tev, T36, T5u; | |
1696 E T3a, T5s, T9p, TbH, T9t, TbF, T6C, T7P; | |
1697 { | |
1698 E T5, T8s, T3, T8q, T9, T8u, T29, T8v, T6, T26; | |
1699 { | |
1700 E T4, T8r, T1, T2; | |
1701 T4 = Cr[WS(csr, 32)]; | |
1702 T5 = KP2_000000000 * T4; | |
1703 T8r = Ci[WS(csi, 32)]; | |
1704 T8s = KP2_000000000 * T8r; | |
1705 T1 = Cr[0]; | |
1706 T2 = Cr[WS(csr, 64)]; | |
1707 T3 = T1 + T2; | |
1708 T8q = T1 - T2; | |
1709 { | |
1710 E T7, T8, T27, T28; | |
1711 T7 = Cr[WS(csr, 16)]; | |
1712 T8 = Cr[WS(csr, 48)]; | |
1713 T9 = KP2_000000000 * (T7 + T8); | |
1714 T8u = T7 - T8; | |
1715 T27 = Ci[WS(csi, 16)]; | |
1716 T28 = Ci[WS(csi, 48)]; | |
1717 T29 = KP2_000000000 * (T27 - T28); | |
1718 T8v = T27 + T28; | |
1719 } | |
1720 } | |
1721 T6 = T3 + T5; | |
1722 Ta = T6 + T9; | |
1723 T6q = T6 - T9; | |
1724 T26 = T3 - T5; | |
1725 T2a = T26 - T29; | |
1726 T5k = T26 + T29; | |
1727 { | |
1728 E T8t, T8w, TcD, TcE; | |
1729 T8t = T8q - T8s; | |
1730 T8w = KP1_414213562 * (T8u - T8v); | |
1731 T8x = T8t + T8w; | |
1732 Tbx = T8t - T8w; | |
1733 TcD = T8q + T8s; | |
1734 TcE = KP1_414213562 * (T8u + T8v); | |
1735 TcF = TcD - TcE; | |
1736 Ten = TcD + TcE; | |
1737 } | |
1738 } | |
1739 { | |
1740 E Td, T8y, T2e, T8C, Tg, T8B, T2h, T8z, T2b, T2i; | |
1741 { | |
1742 E Tb, Tc, T2c, T2d; | |
1743 Tb = Cr[WS(csr, 8)]; | |
1744 Tc = Cr[WS(csr, 56)]; | |
1745 Td = Tb + Tc; | |
1746 T8y = Tb - Tc; | |
1747 T2c = Ci[WS(csi, 8)]; | |
1748 T2d = Ci[WS(csi, 56)]; | |
1749 T2e = T2c - T2d; | |
1750 T8C = T2c + T2d; | |
1751 } | |
1752 { | |
1753 E Te, Tf, T2f, T2g; | |
1754 Te = Cr[WS(csr, 40)]; | |
1755 Tf = Cr[WS(csr, 24)]; | |
1756 Tg = Te + Tf; | |
1757 T8B = Te - Tf; | |
1758 T2f = Ci[WS(csi, 40)]; | |
1759 T2g = Ci[WS(csi, 24)]; | |
1760 T2h = T2f - T2g; | |
1761 T8z = T2f + T2g; | |
1762 } | |
1763 Th = KP2_000000000 * (Td + Tg); | |
1764 T6r = KP2_000000000 * (T2h + T2e); | |
1765 T2b = Td - Tg; | |
1766 T2i = T2e - T2h; | |
1767 T2j = KP1_414213562 * (T2b - T2i); | |
1768 T5l = KP1_414213562 * (T2b + T2i); | |
1769 { | |
1770 E T8A, T8D, TcG, TcH; | |
1771 T8A = T8y - T8z; | |
1772 T8D = T8B + T8C; | |
1773 T8E = FNMS(KP765366864, T8D, KP1_847759065 * T8A); | |
1774 Tby = FMA(KP765366864, T8A, KP1_847759065 * T8D); | |
1775 TcG = T8y + T8z; | |
1776 TcH = T8C - T8B; | |
1777 TcI = FNMS(KP1_847759065, TcH, KP765366864 * TcG); | |
1778 Teo = FMA(KP1_847759065, TcG, KP765366864 * TcH); | |
1779 } | |
1780 } | |
1781 { | |
1782 E Tl, T8G, T2x, T8V, To, T8U, T2A, T8H, Tv, T8S, T2o, T8O, Ts, T8R, T2r; | |
1783 E T8L; | |
1784 { | |
1785 E Tj, Tk, T2y, T2z; | |
1786 Tj = Cr[WS(csr, 4)]; | |
1787 Tk = Cr[WS(csr, 60)]; | |
1788 Tl = Tj + Tk; | |
1789 T8G = Tj - Tk; | |
1790 { | |
1791 E T2v, T2w, Tm, Tn; | |
1792 T2v = Ci[WS(csi, 4)]; | |
1793 T2w = Ci[WS(csi, 60)]; | |
1794 T2x = T2v - T2w; | |
1795 T8V = T2v + T2w; | |
1796 Tm = Cr[WS(csr, 36)]; | |
1797 Tn = Cr[WS(csr, 28)]; | |
1798 To = Tm + Tn; | |
1799 T8U = Tm - Tn; | |
1800 } | |
1801 T2y = Ci[WS(csi, 36)]; | |
1802 T2z = Ci[WS(csi, 28)]; | |
1803 T2A = T2y - T2z; | |
1804 T8H = T2y + T2z; | |
1805 { | |
1806 E Tt, Tu, T8M, T2m, T2n, T8N; | |
1807 Tt = Cr[WS(csr, 12)]; | |
1808 Tu = Cr[WS(csr, 52)]; | |
1809 T8M = Tt - Tu; | |
1810 T2m = Ci[WS(csi, 52)]; | |
1811 T2n = Ci[WS(csi, 12)]; | |
1812 T8N = T2n + T2m; | |
1813 Tv = Tt + Tu; | |
1814 T8S = T8M + T8N; | |
1815 T2o = T2m - T2n; | |
1816 T8O = T8M - T8N; | |
1817 } | |
1818 { | |
1819 E Tq, Tr, T8J, T2p, T2q, T8K; | |
1820 Tq = Cr[WS(csr, 20)]; | |
1821 Tr = Cr[WS(csr, 44)]; | |
1822 T8J = Tq - Tr; | |
1823 T2p = Ci[WS(csi, 20)]; | |
1824 T2q = Ci[WS(csi, 44)]; | |
1825 T8K = T2p + T2q; | |
1826 Ts = Tq + Tr; | |
1827 T8R = T8J + T8K; | |
1828 T2r = T2p - T2q; | |
1829 T8L = T8J - T8K; | |
1830 } | |
1831 } | |
1832 { | |
1833 E Tp, Tw, TcK, TcL; | |
1834 Tp = Tl + To; | |
1835 Tw = Ts + Tv; | |
1836 Tx = KP2_000000000 * (Tp + Tw); | |
1837 T6t = Tp - Tw; | |
1838 TcK = T8G + T8H; | |
1839 TcL = KP707106781 * (T8R + T8S); | |
1840 TcM = TcK - TcL; | |
1841 Teq = TcK + TcL; | |
1842 } | |
1843 { | |
1844 E TcN, TcO, T2l, T2s; | |
1845 TcN = KP707106781 * (T8L - T8O); | |
1846 TcO = T8V - T8U; | |
1847 TcP = TcN + TcO; | |
1848 Ter = TcO - TcN; | |
1849 T2l = Tl - To; | |
1850 T2s = T2o - T2r; | |
1851 T2t = T2l + T2s; | |
1852 T5n = T2l - T2s; | |
1853 } | |
1854 { | |
1855 E T2u, T2B, T8I, T8P; | |
1856 T2u = Ts - Tv; | |
1857 T2B = T2x - T2A; | |
1858 T2C = T2u + T2B; | |
1859 T5o = T2B - T2u; | |
1860 T8I = T8G - T8H; | |
1861 T8P = KP707106781 * (T8L + T8O); | |
1862 T8Q = T8I + T8P; | |
1863 TbA = T8I - T8P; | |
1864 } | |
1865 { | |
1866 E T8T, T8W, T6u, T6v; | |
1867 T8T = KP707106781 * (T8R - T8S); | |
1868 T8W = T8U + T8V; | |
1869 T8X = T8T + T8W; | |
1870 TbB = T8W - T8T; | |
1871 T6u = T2A + T2x; | |
1872 T6v = T2r + T2o; | |
1873 T6w = T6u - T6v; | |
1874 T7L = KP2_000000000 * (T6v + T6u); | |
1875 } | |
1876 } | |
1877 { | |
1878 E T17, T9E, T4D, TaT, T1a, TaS, T4G, T9F, T1h, TaQ, T3q, T9M, T1e, TaP, T3t; | |
1879 E T9J; | |
1880 { | |
1881 E T15, T16, T4E, T4F; | |
1882 T15 = Cr[WS(csr, 1)]; | |
1883 T16 = Cr[WS(csr, 63)]; | |
1884 T17 = T15 + T16; | |
1885 T9E = T15 - T16; | |
1886 { | |
1887 E T4B, T4C, T18, T19; | |
1888 T4B = Ci[WS(csi, 1)]; | |
1889 T4C = Ci[WS(csi, 63)]; | |
1890 T4D = T4B - T4C; | |
1891 TaT = T4B + T4C; | |
1892 T18 = Cr[WS(csr, 33)]; | |
1893 T19 = Cr[WS(csr, 31)]; | |
1894 T1a = T18 + T19; | |
1895 TaS = T18 - T19; | |
1896 } | |
1897 T4E = Ci[WS(csi, 33)]; | |
1898 T4F = Ci[WS(csi, 31)]; | |
1899 T4G = T4E - T4F; | |
1900 T9F = T4E + T4F; | |
1901 { | |
1902 E T1f, T1g, T9K, T3o, T3p, T9L; | |
1903 T1f = Cr[WS(csr, 15)]; | |
1904 T1g = Cr[WS(csr, 49)]; | |
1905 T9K = T1f - T1g; | |
1906 T3o = Ci[WS(csi, 49)]; | |
1907 T3p = Ci[WS(csi, 15)]; | |
1908 T9L = T3p + T3o; | |
1909 T1h = T1f + T1g; | |
1910 TaQ = T9K + T9L; | |
1911 T3q = T3o - T3p; | |
1912 T9M = T9K - T9L; | |
1913 } | |
1914 { | |
1915 E T1c, T1d, T9H, T3r, T3s, T9I; | |
1916 T1c = Cr[WS(csr, 17)]; | |
1917 T1d = Cr[WS(csr, 47)]; | |
1918 T9H = T1c - T1d; | |
1919 T3r = Ci[WS(csi, 17)]; | |
1920 T3s = Ci[WS(csi, 47)]; | |
1921 T9I = T3r + T3s; | |
1922 T1e = T1c + T1d; | |
1923 TaP = T9H + T9I; | |
1924 T3t = T3r - T3s; | |
1925 T9J = T9H - T9I; | |
1926 } | |
1927 } | |
1928 { | |
1929 E T1b, T1i, Tdc, Tdd; | |
1930 T1b = T17 + T1a; | |
1931 T1i = T1e + T1h; | |
1932 T1j = T1b + T1i; | |
1933 T6L = T1b - T1i; | |
1934 Tdc = T9E + T9F; | |
1935 Tdd = KP707106781 * (TaP + TaQ); | |
1936 Tde = Tdc - Tdd; | |
1937 TeC = Tdc + Tdd; | |
1938 } | |
1939 { | |
1940 E TdJ, TdK, T3n, T3u; | |
1941 TdJ = KP707106781 * (T9J - T9M); | |
1942 TdK = TaT - TaS; | |
1943 TdL = TdJ + TdK; | |
1944 TeR = TdK - TdJ; | |
1945 T3n = T17 - T1a; | |
1946 T3u = T3q - T3t; | |
1947 T3v = T3n + T3u; | |
1948 T5z = T3n - T3u; | |
1949 } | |
1950 { | |
1951 E T4A, T4H, T9G, T9N; | |
1952 T4A = T1e - T1h; | |
1953 T4H = T4D - T4G; | |
1954 T4I = T4A + T4H; | |
1955 T5O = T4H - T4A; | |
1956 T9G = T9E - T9F; | |
1957 T9N = KP707106781 * (T9J + T9M); | |
1958 T9O = T9G + T9N; | |
1959 TbM = T9G - T9N; | |
1960 } | |
1961 { | |
1962 E TaR, TaU, T76, T77; | |
1963 TaR = KP707106781 * (TaP - TaQ); | |
1964 TaU = TaS + TaT; | |
1965 TaV = TaR + TaU; | |
1966 Tc1 = TaU - TaR; | |
1967 T76 = T4G + T4D; | |
1968 T77 = T3t + T3q; | |
1969 T78 = T76 - T77; | |
1970 T7Z = T77 + T76; | |
1971 } | |
1972 } | |
1973 { | |
1974 E TB, T90, T3e, T9y, TE, T9x, T3h, T91, TL, T9v, T2I, T98, TI, T9u, T2L; | |
1975 E T95; | |
1976 { | |
1977 E Tz, TA, T3f, T3g; | |
1978 Tz = Cr[WS(csr, 2)]; | |
1979 TA = Cr[WS(csr, 62)]; | |
1980 TB = Tz + TA; | |
1981 T90 = Tz - TA; | |
1982 { | |
1983 E T3c, T3d, TC, TD; | |
1984 T3c = Ci[WS(csi, 2)]; | |
1985 T3d = Ci[WS(csi, 62)]; | |
1986 T3e = T3c - T3d; | |
1987 T9y = T3c + T3d; | |
1988 TC = Cr[WS(csr, 34)]; | |
1989 TD = Cr[WS(csr, 30)]; | |
1990 TE = TC + TD; | |
1991 T9x = TC - TD; | |
1992 } | |
1993 T3f = Ci[WS(csi, 34)]; | |
1994 T3g = Ci[WS(csi, 30)]; | |
1995 T3h = T3f - T3g; | |
1996 T91 = T3f + T3g; | |
1997 { | |
1998 E TJ, TK, T96, T2G, T2H, T97; | |
1999 TJ = Cr[WS(csr, 14)]; | |
2000 TK = Cr[WS(csr, 50)]; | |
2001 T96 = TJ - TK; | |
2002 T2G = Ci[WS(csi, 50)]; | |
2003 T2H = Ci[WS(csi, 14)]; | |
2004 T97 = T2H + T2G; | |
2005 TL = TJ + TK; | |
2006 T9v = T96 + T97; | |
2007 T2I = T2G - T2H; | |
2008 T98 = T96 - T97; | |
2009 } | |
2010 { | |
2011 E TG, TH, T93, T2J, T2K, T94; | |
2012 TG = Cr[WS(csr, 18)]; | |
2013 TH = Cr[WS(csr, 46)]; | |
2014 T93 = TG - TH; | |
2015 T2J = Ci[WS(csi, 18)]; | |
2016 T2K = Ci[WS(csi, 46)]; | |
2017 T94 = T2J + T2K; | |
2018 TI = TG + TH; | |
2019 T9u = T93 + T94; | |
2020 T2L = T2J - T2K; | |
2021 T95 = T93 - T94; | |
2022 } | |
2023 } | |
2024 { | |
2025 E TF, TM, TcS, TcT; | |
2026 TF = TB + TE; | |
2027 TM = TI + TL; | |
2028 TN = TF + TM; | |
2029 T6z = TF - TM; | |
2030 TcS = T90 + T91; | |
2031 TcT = KP707106781 * (T9u + T9v); | |
2032 TcU = TcS - TcT; | |
2033 Teu = TcS + TcT; | |
2034 } | |
2035 { | |
2036 E Td6, Td7, T2F, T2M; | |
2037 Td6 = KP707106781 * (T95 - T98); | |
2038 Td7 = T9y - T9x; | |
2039 Td8 = Td6 + Td7; | |
2040 Tey = Td7 - Td6; | |
2041 T2F = TB - TE; | |
2042 T2M = T2I - T2L; | |
2043 T2N = T2F + T2M; | |
2044 T5r = T2F - T2M; | |
2045 } | |
2046 { | |
2047 E T3b, T3i, T92, T99; | |
2048 T3b = TI - TL; | |
2049 T3i = T3e - T3h; | |
2050 T3j = T3b + T3i; | |
2051 T5v = T3i - T3b; | |
2052 T92 = T90 - T91; | |
2053 T99 = KP707106781 * (T95 + T98); | |
2054 T9a = T92 + T99; | |
2055 TbE = T92 - T99; | |
2056 } | |
2057 { | |
2058 E T9w, T9z, T6F, T6G; | |
2059 T9w = KP707106781 * (T9u - T9v); | |
2060 T9z = T9x + T9y; | |
2061 T9A = T9w + T9z; | |
2062 TbI = T9z - T9w; | |
2063 T6F = T3h + T3e; | |
2064 T6G = T2L + T2I; | |
2065 T6H = T6F - T6G; | |
2066 T7O = T6G + T6F; | |
2067 } | |
2068 } | |
2069 { | |
2070 E T1G, Taj, T3Q, Ta5, T46, Tak, T6R, Ta6, T1N, Tag, Tah, T3X, T3Z, Taa, Tad; | |
2071 E T6S, Tdn, Tdo; | |
2072 { | |
2073 E T1A, T1B, T1C, T1D, T1E, T1F; | |
2074 T1A = Cr[WS(csr, 5)]; | |
2075 T1B = Cr[WS(csr, 59)]; | |
2076 T1C = T1A + T1B; | |
2077 T1D = Cr[WS(csr, 37)]; | |
2078 T1E = Cr[WS(csr, 27)]; | |
2079 T1F = T1D + T1E; | |
2080 T1G = T1C + T1F; | |
2081 Taj = T1D - T1E; | |
2082 T3Q = T1C - T1F; | |
2083 Ta5 = T1A - T1B; | |
2084 } | |
2085 { | |
2086 E T40, T41, T42, T43, T44, T45; | |
2087 T40 = Ci[WS(csi, 5)]; | |
2088 T41 = Ci[WS(csi, 59)]; | |
2089 T42 = T40 - T41; | |
2090 T43 = Ci[WS(csi, 37)]; | |
2091 T44 = Ci[WS(csi, 27)]; | |
2092 T45 = T43 - T44; | |
2093 T46 = T42 - T45; | |
2094 Tak = T40 + T41; | |
2095 T6R = T45 + T42; | |
2096 Ta6 = T43 + T44; | |
2097 } | |
2098 { | |
2099 E T1J, Ta8, T3W, Ta9, T1M, Tab, T3T, Tac; | |
2100 { | |
2101 E T1H, T1I, T3U, T3V; | |
2102 T1H = Cr[WS(csr, 21)]; | |
2103 T1I = Cr[WS(csr, 43)]; | |
2104 T1J = T1H + T1I; | |
2105 Ta8 = T1H - T1I; | |
2106 T3U = Ci[WS(csi, 21)]; | |
2107 T3V = Ci[WS(csi, 43)]; | |
2108 T3W = T3U - T3V; | |
2109 Ta9 = T3U + T3V; | |
2110 } | |
2111 { | |
2112 E T1K, T1L, T3R, T3S; | |
2113 T1K = Cr[WS(csr, 11)]; | |
2114 T1L = Cr[WS(csr, 53)]; | |
2115 T1M = T1K + T1L; | |
2116 Tab = T1K - T1L; | |
2117 T3R = Ci[WS(csi, 53)]; | |
2118 T3S = Ci[WS(csi, 11)]; | |
2119 T3T = T3R - T3S; | |
2120 Tac = T3S + T3R; | |
2121 } | |
2122 T1N = T1J + T1M; | |
2123 Tag = Ta8 + Ta9; | |
2124 Tah = Tab + Tac; | |
2125 T3X = T3T - T3W; | |
2126 T3Z = T1J - T1M; | |
2127 Taa = Ta8 - Ta9; | |
2128 Tad = Tab - Tac; | |
2129 T6S = T3W + T3T; | |
2130 } | |
2131 T1O = T1G + T1N; | |
2132 T7V = T6S + T6R; | |
2133 { | |
2134 E T3Y, T47, Tdq, Tdr; | |
2135 T3Y = T3Q + T3X; | |
2136 T47 = T3Z + T46; | |
2137 T48 = FNMS(KP382683432, T47, KP923879532 * T3Y); | |
2138 T4u = FMA(KP382683432, T3Y, KP923879532 * T47); | |
2139 Tdq = KP707106781 * (Taa - Tad); | |
2140 Tdr = Tak - Taj; | |
2141 Tds = Tdq + Tdr; | |
2142 TeG = Tdr - Tdq; | |
2143 } | |
2144 { | |
2145 E T5C, T5D, Ta7, Tae; | |
2146 T5C = T3Q - T3X; | |
2147 T5D = T46 - T3Z; | |
2148 T5E = FNMS(KP923879532, T5D, KP382683432 * T5C); | |
2149 T5K = FMA(KP923879532, T5C, KP382683432 * T5D); | |
2150 Ta7 = Ta5 - Ta6; | |
2151 Tae = KP707106781 * (Taa + Tad); | |
2152 Taf = Ta7 + Tae; | |
2153 TbP = Ta7 - Tae; | |
2154 } | |
2155 Tdn = Ta5 + Ta6; | |
2156 Tdo = KP707106781 * (Tag + Tah); | |
2157 Tdp = Tdn - Tdo; | |
2158 TeF = Tdn + Tdo; | |
2159 { | |
2160 E T6Q, T6T, Tai, Tal; | |
2161 T6Q = T1G - T1N; | |
2162 T6T = T6R - T6S; | |
2163 T6U = T6Q - T6T; | |
2164 T72 = T6Q + T6T; | |
2165 Tai = KP707106781 * (Tag - Tah); | |
2166 Tal = Taj + Tak; | |
2167 Tam = Tai + Tal; | |
2168 TbQ = Tal - Tai; | |
2169 } | |
2170 } | |
2171 { | |
2172 E T1V, TaC, T49, Tao, T4p, TaD, T6W, Tap, T22, Taz, TaA, T4g, T4i, Tat, Taw; | |
2173 E T6X, Tdu, Tdv; | |
2174 { | |
2175 E T1P, T1Q, T1R, T1S, T1T, T1U; | |
2176 T1P = Cr[WS(csr, 3)]; | |
2177 T1Q = Cr[WS(csr, 61)]; | |
2178 T1R = T1P + T1Q; | |
2179 T1S = Cr[WS(csr, 29)]; | |
2180 T1T = Cr[WS(csr, 35)]; | |
2181 T1U = T1S + T1T; | |
2182 T1V = T1R + T1U; | |
2183 TaC = T1S - T1T; | |
2184 T49 = T1R - T1U; | |
2185 Tao = T1P - T1Q; | |
2186 } | |
2187 { | |
2188 E T4j, T4k, T4l, T4m, T4n, T4o; | |
2189 T4j = Ci[WS(csi, 61)]; | |
2190 T4k = Ci[WS(csi, 3)]; | |
2191 T4l = T4j - T4k; | |
2192 T4m = Ci[WS(csi, 29)]; | |
2193 T4n = Ci[WS(csi, 35)]; | |
2194 T4o = T4m - T4n; | |
2195 T4p = T4l - T4o; | |
2196 TaD = T4k + T4j; | |
2197 T6W = T4o + T4l; | |
2198 Tap = T4m + T4n; | |
2199 } | |
2200 { | |
2201 E T1Y, Tar, T4f, Tas, T21, Tau, T4c, Tav; | |
2202 { | |
2203 E T1W, T1X, T4d, T4e; | |
2204 T1W = Cr[WS(csr, 13)]; | |
2205 T1X = Cr[WS(csr, 51)]; | |
2206 T1Y = T1W + T1X; | |
2207 Tar = T1W - T1X; | |
2208 T4d = Ci[WS(csi, 13)]; | |
2209 T4e = Ci[WS(csi, 51)]; | |
2210 T4f = T4d - T4e; | |
2211 Tas = T4d + T4e; | |
2212 } | |
2213 { | |
2214 E T1Z, T20, T4a, T4b; | |
2215 T1Z = Cr[WS(csr, 19)]; | |
2216 T20 = Cr[WS(csr, 45)]; | |
2217 T21 = T1Z + T20; | |
2218 Tau = T1Z - T20; | |
2219 T4a = Ci[WS(csi, 45)]; | |
2220 T4b = Ci[WS(csi, 19)]; | |
2221 T4c = T4a - T4b; | |
2222 Tav = T4b + T4a; | |
2223 } | |
2224 T22 = T1Y + T21; | |
2225 Taz = Tar + Tas; | |
2226 TaA = Tau + Tav; | |
2227 T4g = T4c - T4f; | |
2228 T4i = T1Y - T21; | |
2229 Tat = Tar - Tas; | |
2230 Taw = Tau - Tav; | |
2231 T6X = T4f + T4c; | |
2232 } | |
2233 T23 = T1V + T22; | |
2234 T7U = T6X + T6W; | |
2235 { | |
2236 E T4h, T4q, Tdx, Tdy; | |
2237 T4h = T49 + T4g; | |
2238 T4q = T4i + T4p; | |
2239 T4r = FMA(KP923879532, T4h, KP382683432 * T4q); | |
2240 T4v = FNMS(KP382683432, T4h, KP923879532 * T4q); | |
2241 Tdx = KP707106781 * (Tat - Taw); | |
2242 Tdy = TaC + TaD; | |
2243 Tdz = Tdx - Tdy; | |
2244 TeJ = Tdx + Tdy; | |
2245 } | |
2246 { | |
2247 E T5F, T5G, Taq, Tax; | |
2248 T5F = T49 - T4g; | |
2249 T5G = T4p - T4i; | |
2250 T5H = FMA(KP382683432, T5F, KP923879532 * T5G); | |
2251 T5L = FNMS(KP923879532, T5F, KP382683432 * T5G); | |
2252 Taq = Tao - Tap; | |
2253 Tax = KP707106781 * (Tat + Taw); | |
2254 Tay = Taq + Tax; | |
2255 TbS = Taq - Tax; | |
2256 } | |
2257 Tdu = Tao + Tap; | |
2258 Tdv = KP707106781 * (Taz + TaA); | |
2259 Tdw = Tdu - Tdv; | |
2260 TeI = Tdu + Tdv; | |
2261 { | |
2262 E T6V, T6Y, TaB, TaE; | |
2263 T6V = T1V - T22; | |
2264 T6Y = T6W - T6X; | |
2265 T6Z = T6V + T6Y; | |
2266 T73 = T6Y - T6V; | |
2267 TaB = KP707106781 * (Taz - TaA); | |
2268 TaE = TaC - TaD; | |
2269 TaF = TaB + TaE; | |
2270 TbT = TaE - TaB; | |
2271 } | |
2272 } | |
2273 { | |
2274 E T1m, T3z, T1p, T3C, T3w, T3D, Tdg, Tdf, T9U, T9R, T1t, T3I, T1w, T3L, T3F; | |
2275 E T3M, Tdj, Tdi, Ta1, T9Y; | |
2276 { | |
2277 E T9P, T9T, T9S, T9Q; | |
2278 { | |
2279 E T1k, T1l, T3x, T3y; | |
2280 T1k = Cr[WS(csr, 9)]; | |
2281 T1l = Cr[WS(csr, 55)]; | |
2282 T1m = T1k + T1l; | |
2283 T9P = T1k - T1l; | |
2284 T3x = Ci[WS(csi, 9)]; | |
2285 T3y = Ci[WS(csi, 55)]; | |
2286 T3z = T3x - T3y; | |
2287 T9T = T3x + T3y; | |
2288 } | |
2289 { | |
2290 E T1n, T1o, T3A, T3B; | |
2291 T1n = Cr[WS(csr, 41)]; | |
2292 T1o = Cr[WS(csr, 23)]; | |
2293 T1p = T1n + T1o; | |
2294 T9S = T1n - T1o; | |
2295 T3A = Ci[WS(csi, 41)]; | |
2296 T3B = Ci[WS(csi, 23)]; | |
2297 T3C = T3A - T3B; | |
2298 T9Q = T3A + T3B; | |
2299 } | |
2300 T3w = T1m - T1p; | |
2301 T3D = T3z - T3C; | |
2302 Tdg = T9T - T9S; | |
2303 Tdf = T9P + T9Q; | |
2304 T9U = T9S + T9T; | |
2305 T9R = T9P - T9Q; | |
2306 } | |
2307 { | |
2308 E T9W, Ta0, T9Z, T9X; | |
2309 { | |
2310 E T1r, T1s, T3G, T3H; | |
2311 T1r = Cr[WS(csr, 7)]; | |
2312 T1s = Cr[WS(csr, 57)]; | |
2313 T1t = T1r + T1s; | |
2314 T9W = T1r - T1s; | |
2315 T3G = Ci[WS(csi, 57)]; | |
2316 T3H = Ci[WS(csi, 7)]; | |
2317 T3I = T3G - T3H; | |
2318 Ta0 = T3H + T3G; | |
2319 } | |
2320 { | |
2321 E T1u, T1v, T3J, T3K; | |
2322 T1u = Cr[WS(csr, 25)]; | |
2323 T1v = Cr[WS(csr, 39)]; | |
2324 T1w = T1u + T1v; | |
2325 T9Z = T1u - T1v; | |
2326 T3J = Ci[WS(csi, 25)]; | |
2327 T3K = Ci[WS(csi, 39)]; | |
2328 T3L = T3J - T3K; | |
2329 T9X = T3J + T3K; | |
2330 } | |
2331 T3F = T1t - T1w; | |
2332 T3M = T3I - T3L; | |
2333 Tdj = T9Z + Ta0; | |
2334 Tdi = T9W + T9X; | |
2335 Ta1 = T9Z - Ta0; | |
2336 T9Y = T9W - T9X; | |
2337 } | |
2338 { | |
2339 E T1q, T1x, Tdh, Tdk; | |
2340 T1q = T1m + T1p; | |
2341 T1x = T1t + T1w; | |
2342 T1y = T1q + T1x; | |
2343 T75 = T1q - T1x; | |
2344 Tdh = FNMS(KP923879532, Tdg, KP382683432 * Tdf); | |
2345 Tdk = FNMS(KP923879532, Tdj, KP382683432 * Tdi); | |
2346 Tdl = Tdh + Tdk; | |
2347 TeQ = Tdh - Tdk; | |
2348 } | |
2349 { | |
2350 E TdG, TdH, T3E, T3N; | |
2351 TdG = FMA(KP923879532, Tdf, KP382683432 * Tdg); | |
2352 TdH = FMA(KP923879532, Tdi, KP382683432 * Tdj); | |
2353 TdI = TdG - TdH; | |
2354 TeD = TdG + TdH; | |
2355 T3E = T3w - T3D; | |
2356 T3N = T3F + T3M; | |
2357 T3O = KP707106781 * (T3E + T3N); | |
2358 T5N = KP707106781 * (T3E - T3N); | |
2359 } | |
2360 { | |
2361 E T4x, T4y, T9V, Ta2; | |
2362 T4x = T3w + T3D; | |
2363 T4y = T3M - T3F; | |
2364 T4z = KP707106781 * (T4x + T4y); | |
2365 T5A = KP707106781 * (T4y - T4x); | |
2366 T9V = FNMS(KP382683432, T9U, KP923879532 * T9R); | |
2367 Ta2 = FMA(KP923879532, T9Y, KP382683432 * Ta1); | |
2368 Ta3 = T9V + Ta2; | |
2369 Tc0 = T9V - Ta2; | |
2370 } | |
2371 { | |
2372 E TaM, TaN, T6M, T6N; | |
2373 TaM = FMA(KP382683432, T9R, KP923879532 * T9U); | |
2374 TaN = FNMS(KP382683432, T9Y, KP923879532 * Ta1); | |
2375 TaO = TaM + TaN; | |
2376 TbN = TaN - TaM; | |
2377 T6M = T3L + T3I; | |
2378 T6N = T3C + T3z; | |
2379 T6O = T6M - T6N; | |
2380 T80 = T6N + T6M; | |
2381 } | |
2382 } | |
2383 { | |
2384 E TQ, T2R, TT, T2U, T2O, T2V, TcW, TcV, T9g, T9d, TX, T30, T10, T33, T2X; | |
2385 E T34, TcZ, TcY, T9n, T9k; | |
2386 { | |
2387 E T9b, T9f, T9e, T9c; | |
2388 { | |
2389 E TO, TP, T2P, T2Q; | |
2390 TO = Cr[WS(csr, 10)]; | |
2391 TP = Cr[WS(csr, 54)]; | |
2392 TQ = TO + TP; | |
2393 T9b = TO - TP; | |
2394 T2P = Ci[WS(csi, 10)]; | |
2395 T2Q = Ci[WS(csi, 54)]; | |
2396 T2R = T2P - T2Q; | |
2397 T9f = T2P + T2Q; | |
2398 } | |
2399 { | |
2400 E TR, TS, T2S, T2T; | |
2401 TR = Cr[WS(csr, 42)]; | |
2402 TS = Cr[WS(csr, 22)]; | |
2403 TT = TR + TS; | |
2404 T9e = TR - TS; | |
2405 T2S = Ci[WS(csi, 42)]; | |
2406 T2T = Ci[WS(csi, 22)]; | |
2407 T2U = T2S - T2T; | |
2408 T9c = T2S + T2T; | |
2409 } | |
2410 T2O = TQ - TT; | |
2411 T2V = T2R - T2U; | |
2412 TcW = T9f - T9e; | |
2413 TcV = T9b + T9c; | |
2414 T9g = T9e + T9f; | |
2415 T9d = T9b - T9c; | |
2416 } | |
2417 { | |
2418 E T9i, T9m, T9l, T9j; | |
2419 { | |
2420 E TV, TW, T2Y, T2Z; | |
2421 TV = Cr[WS(csr, 6)]; | |
2422 TW = Cr[WS(csr, 58)]; | |
2423 TX = TV + TW; | |
2424 T9i = TV - TW; | |
2425 T2Y = Ci[WS(csi, 58)]; | |
2426 T2Z = Ci[WS(csi, 6)]; | |
2427 T30 = T2Y - T2Z; | |
2428 T9m = T2Z + T2Y; | |
2429 } | |
2430 { | |
2431 E TY, TZ, T31, T32; | |
2432 TY = Cr[WS(csr, 26)]; | |
2433 TZ = Cr[WS(csr, 38)]; | |
2434 T10 = TY + TZ; | |
2435 T9l = TY - TZ; | |
2436 T31 = Ci[WS(csi, 26)]; | |
2437 T32 = Ci[WS(csi, 38)]; | |
2438 T33 = T31 - T32; | |
2439 T9j = T31 + T32; | |
2440 } | |
2441 T2X = TX - T10; | |
2442 T34 = T30 - T33; | |
2443 TcZ = T9l + T9m; | |
2444 TcY = T9i + T9j; | |
2445 T9n = T9l - T9m; | |
2446 T9k = T9i - T9j; | |
2447 } | |
2448 { | |
2449 E TU, T11, TcX, Td0; | |
2450 TU = TQ + TT; | |
2451 T11 = TX + T10; | |
2452 T12 = TU + T11; | |
2453 T6E = TU - T11; | |
2454 TcX = FNMS(KP923879532, TcW, KP382683432 * TcV); | |
2455 Td0 = FNMS(KP923879532, TcZ, KP382683432 * TcY); | |
2456 Td1 = TcX + Td0; | |
2457 Tex = TcX - Td0; | |
2458 } | |
2459 { | |
2460 E Td3, Td4, T2W, T35; | |
2461 Td3 = FMA(KP923879532, TcV, KP382683432 * TcW); | |
2462 Td4 = FMA(KP923879532, TcY, KP382683432 * TcZ); | |
2463 Td5 = Td3 - Td4; | |
2464 Tev = Td3 + Td4; | |
2465 T2W = T2O - T2V; | |
2466 T35 = T2X + T34; | |
2467 T36 = KP707106781 * (T2W + T35); | |
2468 T5u = KP707106781 * (T2W - T35); | |
2469 } | |
2470 { | |
2471 E T38, T39, T9h, T9o; | |
2472 T38 = T2O + T2V; | |
2473 T39 = T34 - T2X; | |
2474 T3a = KP707106781 * (T38 + T39); | |
2475 T5s = KP707106781 * (T39 - T38); | |
2476 T9h = FNMS(KP382683432, T9g, KP923879532 * T9d); | |
2477 T9o = FMA(KP923879532, T9k, KP382683432 * T9n); | |
2478 T9p = T9h + T9o; | |
2479 TbH = T9h - T9o; | |
2480 } | |
2481 { | |
2482 E T9r, T9s, T6A, T6B; | |
2483 T9r = FMA(KP382683432, T9d, KP923879532 * T9g); | |
2484 T9s = FNMS(KP382683432, T9k, KP923879532 * T9n); | |
2485 T9t = T9r + T9s; | |
2486 TbF = T9s - T9r; | |
2487 T6A = T33 + T30; | |
2488 T6B = T2U + T2R; | |
2489 T6C = T6A - T6B; | |
2490 T7P = T6B + T6A; | |
2491 } | |
2492 } | |
2493 { | |
2494 E T13, T8f, Ty, T8e, T25, T8h, T8k, T8p, Ti, T14, T8o; | |
2495 T13 = KP2_000000000 * (TN + T12); | |
2496 T8f = KP2_000000000 * (T7P + T7O); | |
2497 Ti = Ta + Th; | |
2498 Ty = Ti + Tx; | |
2499 T8e = Ti - Tx; | |
2500 { | |
2501 E T1z, T24, T8i, T8j; | |
2502 T1z = T1j + T1y; | |
2503 T24 = T1O + T23; | |
2504 T25 = KP2_000000000 * (T1z + T24); | |
2505 T8h = T1z - T24; | |
2506 T8i = T80 + T7Z; | |
2507 T8j = T7V + T7U; | |
2508 T8k = T8i - T8j; | |
2509 T8p = KP2_000000000 * (T8j + T8i); | |
2510 } | |
2511 T14 = Ty + T13; | |
2512 R0[WS(rs, 32)] = T14 - T25; | |
2513 R0[0] = T14 + T25; | |
2514 T8o = Ty - T13; | |
2515 R0[WS(rs, 16)] = T8o - T8p; | |
2516 R0[WS(rs, 48)] = T8o + T8p; | |
2517 { | |
2518 E T8g, T8l, T8m, T8n; | |
2519 T8g = T8e - T8f; | |
2520 T8l = KP1_414213562 * (T8h - T8k); | |
2521 R0[WS(rs, 40)] = T8g - T8l; | |
2522 R0[WS(rs, 8)] = T8g + T8l; | |
2523 T8m = T8e + T8f; | |
2524 T8n = KP1_414213562 * (T8h + T8k); | |
2525 R0[WS(rs, 24)] = T8m - T8n; | |
2526 R0[WS(rs, 56)] = T8m + T8n; | |
2527 } | |
2528 } | |
2529 { | |
2530 E T7M, T86, T82, T8a, T7R, T87, T7X, T89, T7K, T7Y, T81; | |
2531 T7K = Ta - Th; | |
2532 T7M = T7K - T7L; | |
2533 T86 = T7K + T7L; | |
2534 T7Y = T1O - T23; | |
2535 T81 = T7Z - T80; | |
2536 T82 = T7Y + T81; | |
2537 T8a = T81 - T7Y; | |
2538 { | |
2539 E T7N, T7Q, T7T, T7W; | |
2540 T7N = TN - T12; | |
2541 T7Q = T7O - T7P; | |
2542 T7R = KP1_414213562 * (T7N - T7Q); | |
2543 T87 = KP1_414213562 * (T7N + T7Q); | |
2544 T7T = T1j - T1y; | |
2545 T7W = T7U - T7V; | |
2546 T7X = T7T + T7W; | |
2547 T89 = T7T - T7W; | |
2548 } | |
2549 { | |
2550 E T7S, T83, T8c, T8d; | |
2551 T7S = T7M + T7R; | |
2552 T83 = FNMS(KP765366864, T82, KP1_847759065 * T7X); | |
2553 R0[WS(rs, 36)] = T7S - T83; | |
2554 R0[WS(rs, 4)] = T7S + T83; | |
2555 T8c = T86 + T87; | |
2556 T8d = FMA(KP1_847759065, T89, KP765366864 * T8a); | |
2557 R0[WS(rs, 28)] = T8c - T8d; | |
2558 R0[WS(rs, 60)] = T8c + T8d; | |
2559 } | |
2560 { | |
2561 E T84, T85, T88, T8b; | |
2562 T84 = T7M - T7R; | |
2563 T85 = FMA(KP765366864, T7X, KP1_847759065 * T82); | |
2564 R0[WS(rs, 20)] = T84 - T85; | |
2565 R0[WS(rs, 52)] = T84 + T85; | |
2566 T88 = T86 - T87; | |
2567 T8b = FNMS(KP1_847759065, T8a, KP765366864 * T89); | |
2568 R0[WS(rs, 44)] = T88 - T8b; | |
2569 R0[WS(rs, 12)] = T88 + T8b; | |
2570 } | |
2571 } | |
2572 { | |
2573 E T2E, T4O, T4K, T4S, T3l, T4P, T4t, T4R; | |
2574 { | |
2575 E T2k, T2D, T4w, T4J; | |
2576 T2k = T2a + T2j; | |
2577 T2D = FNMS(KP765366864, T2C, KP1_847759065 * T2t); | |
2578 T2E = T2k + T2D; | |
2579 T4O = T2k - T2D; | |
2580 T4w = T4u + T4v; | |
2581 T4J = T4z + T4I; | |
2582 T4K = T4w + T4J; | |
2583 T4S = T4J - T4w; | |
2584 } | |
2585 { | |
2586 E T37, T3k, T3P, T4s; | |
2587 T37 = T2N + T36; | |
2588 T3k = T3a + T3j; | |
2589 T3l = FNMS(KP390180644, T3k, KP1_961570560 * T37); | |
2590 T4P = FMA(KP390180644, T37, KP1_961570560 * T3k); | |
2591 T3P = T3v + T3O; | |
2592 T4s = T48 + T4r; | |
2593 T4t = T3P + T4s; | |
2594 T4R = T3P - T4s; | |
2595 } | |
2596 { | |
2597 E T3m, T4L, T4U, T4V; | |
2598 T3m = T2E + T3l; | |
2599 T4L = FNMS(KP196034280, T4K, KP1_990369453 * T4t); | |
2600 R0[WS(rs, 33)] = T3m - T4L; | |
2601 R0[WS(rs, 1)] = T3m + T4L; | |
2602 T4U = T4O + T4P; | |
2603 T4V = FMA(KP1_546020906, T4R, KP1_268786568 * T4S); | |
2604 R0[WS(rs, 25)] = T4U - T4V; | |
2605 R0[WS(rs, 57)] = T4U + T4V; | |
2606 } | |
2607 { | |
2608 E T4M, T4N, T4Q, T4T; | |
2609 T4M = T2E - T3l; | |
2610 T4N = FMA(KP196034280, T4t, KP1_990369453 * T4K); | |
2611 R0[WS(rs, 17)] = T4M - T4N; | |
2612 R0[WS(rs, 49)] = T4M + T4N; | |
2613 T4Q = T4O - T4P; | |
2614 T4T = FNMS(KP1_546020906, T4S, KP1_268786568 * T4R); | |
2615 R0[WS(rs, 41)] = T4Q - T4T; | |
2616 R0[WS(rs, 9)] = T4Q + T4T; | |
2617 } | |
2618 } | |
2619 { | |
2620 E T6y, T7e, T7a, T7i, T6J, T7f, T71, T7h; | |
2621 { | |
2622 E T6s, T6x, T74, T79; | |
2623 T6s = T6q - T6r; | |
2624 T6x = KP1_414213562 * (T6t - T6w); | |
2625 T6y = T6s + T6x; | |
2626 T7e = T6s - T6x; | |
2627 T74 = KP707106781 * (T72 + T73); | |
2628 T79 = T75 + T78; | |
2629 T7a = T74 + T79; | |
2630 T7i = T79 - T74; | |
2631 } | |
2632 { | |
2633 E T6D, T6I, T6P, T70; | |
2634 T6D = T6z + T6C; | |
2635 T6I = T6E + T6H; | |
2636 T6J = FNMS(KP765366864, T6I, KP1_847759065 * T6D); | |
2637 T7f = FMA(KP765366864, T6D, KP1_847759065 * T6I); | |
2638 T6P = T6L + T6O; | |
2639 T70 = KP707106781 * (T6U + T6Z); | |
2640 T71 = T6P + T70; | |
2641 T7h = T6P - T70; | |
2642 } | |
2643 { | |
2644 E T6K, T7b, T7k, T7l; | |
2645 T6K = T6y + T6J; | |
2646 T7b = FNMS(KP390180644, T7a, KP1_961570560 * T71); | |
2647 R0[WS(rs, 34)] = T6K - T7b; | |
2648 R0[WS(rs, 2)] = T6K + T7b; | |
2649 T7k = T7e + T7f; | |
2650 T7l = FMA(KP1_662939224, T7h, KP1_111140466 * T7i); | |
2651 R0[WS(rs, 26)] = T7k - T7l; | |
2652 R0[WS(rs, 58)] = T7k + T7l; | |
2653 } | |
2654 { | |
2655 E T7c, T7d, T7g, T7j; | |
2656 T7c = T6y - T6J; | |
2657 T7d = FMA(KP390180644, T71, KP1_961570560 * T7a); | |
2658 R0[WS(rs, 18)] = T7c - T7d; | |
2659 R0[WS(rs, 50)] = T7c + T7d; | |
2660 T7g = T7e - T7f; | |
2661 T7j = FNMS(KP1_662939224, T7i, KP1_111140466 * T7h); | |
2662 R0[WS(rs, 42)] = T7g - T7j; | |
2663 R0[WS(rs, 10)] = T7g + T7j; | |
2664 } | |
2665 } | |
2666 { | |
2667 E T4Y, T5c, T58, T5g, T51, T5d, T55, T5f; | |
2668 { | |
2669 E T4W, T4X, T56, T57; | |
2670 T4W = T2a - T2j; | |
2671 T4X = FMA(KP765366864, T2t, KP1_847759065 * T2C); | |
2672 T4Y = T4W - T4X; | |
2673 T5c = T4W + T4X; | |
2674 T56 = T48 - T4r; | |
2675 T57 = T4I - T4z; | |
2676 T58 = T56 + T57; | |
2677 T5g = T57 - T56; | |
2678 } | |
2679 { | |
2680 E T4Z, T50, T53, T54; | |
2681 T4Z = T2N - T36; | |
2682 T50 = T3j - T3a; | |
2683 T51 = FNMS(KP1_662939224, T50, KP1_111140466 * T4Z); | |
2684 T5d = FMA(KP1_662939224, T4Z, KP1_111140466 * T50); | |
2685 T53 = T3v - T3O; | |
2686 T54 = T4v - T4u; | |
2687 T55 = T53 + T54; | |
2688 T5f = T53 - T54; | |
2689 } | |
2690 { | |
2691 E T52, T59, T5i, T5j; | |
2692 T52 = T4Y + T51; | |
2693 T59 = FNMS(KP942793473, T58, KP1_763842528 * T55); | |
2694 R0[WS(rs, 37)] = T52 - T59; | |
2695 R0[WS(rs, 5)] = T52 + T59; | |
2696 T5i = T5c + T5d; | |
2697 T5j = FMA(KP1_913880671, T5f, KP580569354 * T5g); | |
2698 R0[WS(rs, 29)] = T5i - T5j; | |
2699 R0[WS(rs, 61)] = T5i + T5j; | |
2700 } | |
2701 { | |
2702 E T5a, T5b, T5e, T5h; | |
2703 T5a = T4Y - T51; | |
2704 T5b = FMA(KP942793473, T55, KP1_763842528 * T58); | |
2705 R0[WS(rs, 21)] = T5a - T5b; | |
2706 R0[WS(rs, 53)] = T5a + T5b; | |
2707 T5e = T5c - T5d; | |
2708 T5h = FNMS(KP1_913880671, T5g, KP580569354 * T5f); | |
2709 R0[WS(rs, 45)] = T5e - T5h; | |
2710 R0[WS(rs, 13)] = T5e + T5h; | |
2711 } | |
2712 } | |
2713 { | |
2714 E T7o, T7C, T7y, T7G, T7r, T7D, T7v, T7F; | |
2715 { | |
2716 E T7m, T7n, T7w, T7x; | |
2717 T7m = T6q + T6r; | |
2718 T7n = KP1_414213562 * (T6t + T6w); | |
2719 T7o = T7m - T7n; | |
2720 T7C = T7m + T7n; | |
2721 T7w = KP707106781 * (T6U - T6Z); | |
2722 T7x = T78 - T75; | |
2723 T7y = T7w + T7x; | |
2724 T7G = T7x - T7w; | |
2725 } | |
2726 { | |
2727 E T7p, T7q, T7t, T7u; | |
2728 T7p = T6z - T6C; | |
2729 T7q = T6H - T6E; | |
2730 T7r = FNMS(KP1_847759065, T7q, KP765366864 * T7p); | |
2731 T7D = FMA(KP1_847759065, T7p, KP765366864 * T7q); | |
2732 T7t = T6L - T6O; | |
2733 T7u = KP707106781 * (T73 - T72); | |
2734 T7v = T7t + T7u; | |
2735 T7F = T7t - T7u; | |
2736 } | |
2737 { | |
2738 E T7s, T7z, T7I, T7J; | |
2739 T7s = T7o + T7r; | |
2740 T7z = FNMS(KP1_111140466, T7y, KP1_662939224 * T7v); | |
2741 R0[WS(rs, 38)] = T7s - T7z; | |
2742 R0[WS(rs, 6)] = T7s + T7z; | |
2743 T7I = T7C + T7D; | |
2744 T7J = FMA(KP1_961570560, T7F, KP390180644 * T7G); | |
2745 R0[WS(rs, 30)] = T7I - T7J; | |
2746 R0[WS(rs, 62)] = T7I + T7J; | |
2747 } | |
2748 { | |
2749 E T7A, T7B, T7E, T7H; | |
2750 T7A = T7o - T7r; | |
2751 T7B = FMA(KP1_111140466, T7v, KP1_662939224 * T7y); | |
2752 R0[WS(rs, 22)] = T7A - T7B; | |
2753 R0[WS(rs, 54)] = T7A + T7B; | |
2754 T7E = T7C - T7D; | |
2755 T7H = FNMS(KP1_961570560, T7G, KP390180644 * T7F); | |
2756 R0[WS(rs, 46)] = T7E - T7H; | |
2757 R0[WS(rs, 14)] = T7E + T7H; | |
2758 } | |
2759 } | |
2760 { | |
2761 E T5q, T5U, T5Q, T5Y, T5x, T5V, T5J, T5X; | |
2762 { | |
2763 E T5m, T5p, T5M, T5P; | |
2764 T5m = T5k - T5l; | |
2765 T5p = FNMS(KP1_847759065, T5o, KP765366864 * T5n); | |
2766 T5q = T5m + T5p; | |
2767 T5U = T5m - T5p; | |
2768 T5M = T5K + T5L; | |
2769 T5P = T5N + T5O; | |
2770 T5Q = T5M + T5P; | |
2771 T5Y = T5P - T5M; | |
2772 } | |
2773 { | |
2774 E T5t, T5w, T5B, T5I; | |
2775 T5t = T5r + T5s; | |
2776 T5w = T5u + T5v; | |
2777 T5x = FNMS(KP1_111140466, T5w, KP1_662939224 * T5t); | |
2778 T5V = FMA(KP1_111140466, T5t, KP1_662939224 * T5w); | |
2779 T5B = T5z + T5A; | |
2780 T5I = T5E + T5H; | |
2781 T5J = T5B + T5I; | |
2782 T5X = T5B - T5I; | |
2783 } | |
2784 { | |
2785 E T5y, T5R, T60, T61; | |
2786 T5y = T5q + T5x; | |
2787 T5R = FNMS(KP580569354, T5Q, KP1_913880671 * T5J); | |
2788 R0[WS(rs, 35)] = T5y - T5R; | |
2789 R0[WS(rs, 3)] = T5y + T5R; | |
2790 T60 = T5U + T5V; | |
2791 T61 = FMA(KP1_763842528, T5X, KP942793473 * T5Y); | |
2792 R0[WS(rs, 27)] = T60 - T61; | |
2793 R0[WS(rs, 59)] = T60 + T61; | |
2794 } | |
2795 { | |
2796 E T5S, T5T, T5W, T5Z; | |
2797 T5S = T5q - T5x; | |
2798 T5T = FMA(KP580569354, T5J, KP1_913880671 * T5Q); | |
2799 R0[WS(rs, 19)] = T5S - T5T; | |
2800 R0[WS(rs, 51)] = T5S + T5T; | |
2801 T5W = T5U - T5V; | |
2802 T5Z = FNMS(KP1_763842528, T5Y, KP942793473 * T5X); | |
2803 R0[WS(rs, 43)] = T5W - T5Z; | |
2804 R0[WS(rs, 11)] = T5W + T5Z; | |
2805 } | |
2806 } | |
2807 { | |
2808 E T64, T6i, T6e, T6m, T67, T6j, T6b, T6l; | |
2809 { | |
2810 E T62, T63, T6c, T6d; | |
2811 T62 = T5k + T5l; | |
2812 T63 = FMA(KP1_847759065, T5n, KP765366864 * T5o); | |
2813 T64 = T62 - T63; | |
2814 T6i = T62 + T63; | |
2815 T6c = T5E - T5H; | |
2816 T6d = T5O - T5N; | |
2817 T6e = T6c + T6d; | |
2818 T6m = T6d - T6c; | |
2819 } | |
2820 { | |
2821 E T65, T66, T69, T6a; | |
2822 T65 = T5r - T5s; | |
2823 T66 = T5v - T5u; | |
2824 T67 = FNMS(KP1_961570560, T66, KP390180644 * T65); | |
2825 T6j = FMA(KP1_961570560, T65, KP390180644 * T66); | |
2826 T69 = T5z - T5A; | |
2827 T6a = T5L - T5K; | |
2828 T6b = T69 + T6a; | |
2829 T6l = T69 - T6a; | |
2830 } | |
2831 { | |
2832 E T68, T6f, T6o, T6p; | |
2833 T68 = T64 + T67; | |
2834 T6f = FNMS(KP1_268786568, T6e, KP1_546020906 * T6b); | |
2835 R0[WS(rs, 39)] = T68 - T6f; | |
2836 R0[WS(rs, 7)] = T68 + T6f; | |
2837 T6o = T6i + T6j; | |
2838 T6p = FMA(KP1_990369453, T6l, KP196034280 * T6m); | |
2839 R0[WS(rs, 31)] = T6o - T6p; | |
2840 R0[WS(rs, 63)] = T6o + T6p; | |
2841 } | |
2842 { | |
2843 E T6g, T6h, T6k, T6n; | |
2844 T6g = T64 - T67; | |
2845 T6h = FMA(KP1_268786568, T6b, KP1_546020906 * T6e); | |
2846 R0[WS(rs, 23)] = T6g - T6h; | |
2847 R0[WS(rs, 55)] = T6g + T6h; | |
2848 T6k = T6i - T6j; | |
2849 T6n = FNMS(KP1_990369453, T6m, KP196034280 * T6l); | |
2850 R0[WS(rs, 47)] = T6k - T6n; | |
2851 R0[WS(rs, 15)] = T6k + T6n; | |
2852 } | |
2853 } | |
2854 { | |
2855 E T8Z, Tb1, T9C, Tb2, Tbe, Tbq, Tbb, Tbp, TaX, Tbs, Tb5, Tbi, TaI, Tbt, Tb4; | |
2856 E Tbl; | |
2857 { | |
2858 E T8F, T8Y, Tb9, Tba; | |
2859 T8F = T8x + T8E; | |
2860 T8Y = FNMS(KP390180644, T8X, KP1_961570560 * T8Q); | |
2861 T8Z = T8F + T8Y; | |
2862 Tb1 = T8F - T8Y; | |
2863 { | |
2864 E T9q, T9B, Tbc, Tbd; | |
2865 T9q = T9a + T9p; | |
2866 T9B = T9t + T9A; | |
2867 T9C = FNMS(KP196034280, T9B, KP1_990369453 * T9q); | |
2868 Tb2 = FMA(KP196034280, T9q, KP1_990369453 * T9B); | |
2869 Tbc = T9a - T9p; | |
2870 Tbd = T9A - T9t; | |
2871 Tbe = FNMS(KP1_546020906, Tbd, KP1_268786568 * Tbc); | |
2872 Tbq = FMA(KP1_546020906, Tbc, KP1_268786568 * Tbd); | |
2873 } | |
2874 Tb9 = T8x - T8E; | |
2875 Tba = FMA(KP390180644, T8Q, KP1_961570560 * T8X); | |
2876 Tbb = Tb9 - Tba; | |
2877 Tbp = Tb9 + Tba; | |
2878 { | |
2879 E TaW, Tbg, TaL, Tbh, TaJ, TaK; | |
2880 TaW = TaO + TaV; | |
2881 Tbg = T9O - Ta3; | |
2882 TaJ = FMA(KP195090322, Taf, KP980785280 * Tam); | |
2883 TaK = FNMS(KP195090322, Tay, KP980785280 * TaF); | |
2884 TaL = TaJ + TaK; | |
2885 Tbh = TaK - TaJ; | |
2886 TaX = TaL + TaW; | |
2887 Tbs = Tbg - Tbh; | |
2888 Tb5 = TaW - TaL; | |
2889 Tbi = Tbg + Tbh; | |
2890 } | |
2891 { | |
2892 E Ta4, Tbk, TaH, Tbj, Tan, TaG; | |
2893 Ta4 = T9O + Ta3; | |
2894 Tbk = TaV - TaO; | |
2895 Tan = FNMS(KP195090322, Tam, KP980785280 * Taf); | |
2896 TaG = FMA(KP980785280, Tay, KP195090322 * TaF); | |
2897 TaH = Tan + TaG; | |
2898 Tbj = Tan - TaG; | |
2899 TaI = Ta4 + TaH; | |
2900 Tbt = Tbk - Tbj; | |
2901 Tb4 = Ta4 - TaH; | |
2902 Tbl = Tbj + Tbk; | |
2903 } | |
2904 } | |
2905 { | |
2906 E T9D, TaY, Tbr, Tbu; | |
2907 T9D = T8Z + T9C; | |
2908 TaY = FNMS(KP098135348, TaX, KP1_997590912 * TaI); | |
2909 R1[WS(rs, 32)] = T9D - TaY; | |
2910 R1[0] = T9D + TaY; | |
2911 Tbr = Tbp - Tbq; | |
2912 Tbu = FNMS(KP1_883088130, Tbt, KP673779706 * Tbs); | |
2913 R1[WS(rs, 44)] = Tbr - Tbu; | |
2914 R1[WS(rs, 12)] = Tbr + Tbu; | |
2915 } | |
2916 { | |
2917 E Tbv, Tbw, TaZ, Tb0; | |
2918 Tbv = Tbp + Tbq; | |
2919 Tbw = FMA(KP1_883088130, Tbs, KP673779706 * Tbt); | |
2920 R1[WS(rs, 28)] = Tbv - Tbw; | |
2921 R1[WS(rs, 60)] = Tbv + Tbw; | |
2922 TaZ = T8Z - T9C; | |
2923 Tb0 = FMA(KP098135348, TaI, KP1_997590912 * TaX); | |
2924 R1[WS(rs, 16)] = TaZ - Tb0; | |
2925 R1[WS(rs, 48)] = TaZ + Tb0; | |
2926 } | |
2927 { | |
2928 E Tb3, Tb6, Tbf, Tbm; | |
2929 Tb3 = Tb1 - Tb2; | |
2930 Tb6 = FNMS(KP1_481902250, Tb5, KP1_343117909 * Tb4); | |
2931 R1[WS(rs, 40)] = Tb3 - Tb6; | |
2932 R1[WS(rs, 8)] = Tb3 + Tb6; | |
2933 Tbf = Tbb + Tbe; | |
2934 Tbm = FNMS(KP855110186, Tbl, KP1_807978586 * Tbi); | |
2935 R1[WS(rs, 36)] = Tbf - Tbm; | |
2936 R1[WS(rs, 4)] = Tbf + Tbm; | |
2937 } | |
2938 { | |
2939 E Tbn, Tbo, Tb7, Tb8; | |
2940 Tbn = Tbb - Tbe; | |
2941 Tbo = FMA(KP855110186, Tbi, KP1_807978586 * Tbl); | |
2942 R1[WS(rs, 20)] = Tbn - Tbo; | |
2943 R1[WS(rs, 52)] = Tbn + Tbo; | |
2944 Tb7 = Tb1 + Tb2; | |
2945 Tb8 = FMA(KP1_481902250, Tb4, KP1_343117909 * Tb5); | |
2946 R1[WS(rs, 24)] = Tb7 - Tb8; | |
2947 R1[WS(rs, 56)] = Tb7 + Tb8; | |
2948 } | |
2949 } | |
2950 { | |
2951 E TcR, TdR, Tda, TdS, Te4, Teg, Te1, Tef, TdN, Tei, TdV, Te8, TdC, Tej, TdU; | |
2952 E Teb; | |
2953 { | |
2954 E TcJ, TcQ, TdZ, Te0; | |
2955 TcJ = TcF + TcI; | |
2956 TcQ = FNMS(KP1_111140466, TcP, KP1_662939224 * TcM); | |
2957 TcR = TcJ + TcQ; | |
2958 TdR = TcJ - TcQ; | |
2959 { | |
2960 E Td2, Td9, Te2, Te3; | |
2961 Td2 = TcU + Td1; | |
2962 Td9 = Td5 + Td8; | |
2963 Tda = FNMS(KP580569354, Td9, KP1_913880671 * Td2); | |
2964 TdS = FMA(KP580569354, Td2, KP1_913880671 * Td9); | |
2965 Te2 = TcU - Td1; | |
2966 Te3 = Td8 - Td5; | |
2967 Te4 = FNMS(KP1_763842528, Te3, KP942793473 * Te2); | |
2968 Teg = FMA(KP1_763842528, Te2, KP942793473 * Te3); | |
2969 } | |
2970 TdZ = TcF - TcI; | |
2971 Te0 = FMA(KP1_111140466, TcM, KP1_662939224 * TcP); | |
2972 Te1 = TdZ - Te0; | |
2973 Tef = TdZ + Te0; | |
2974 { | |
2975 E TdM, Te6, TdF, Te7, TdD, TdE; | |
2976 TdM = TdI + TdL; | |
2977 Te6 = Tde - Tdl; | |
2978 TdD = FMA(KP555570233, Tdp, KP831469612 * Tds); | |
2979 TdE = FNMS(KP555570233, Tdw, KP831469612 * Tdz); | |
2980 TdF = TdD + TdE; | |
2981 Te7 = TdE - TdD; | |
2982 TdN = TdF + TdM; | |
2983 Tei = Te6 - Te7; | |
2984 TdV = TdM - TdF; | |
2985 Te8 = Te6 + Te7; | |
2986 } | |
2987 { | |
2988 E Tdm, Tea, TdB, Te9, Tdt, TdA; | |
2989 Tdm = Tde + Tdl; | |
2990 Tea = TdL - TdI; | |
2991 Tdt = FNMS(KP555570233, Tds, KP831469612 * Tdp); | |
2992 TdA = FMA(KP831469612, Tdw, KP555570233 * Tdz); | |
2993 TdB = Tdt + TdA; | |
2994 Te9 = Tdt - TdA; | |
2995 TdC = Tdm + TdB; | |
2996 Tej = Tea - Te9; | |
2997 TdU = Tdm - TdB; | |
2998 Teb = Te9 + Tea; | |
2999 } | |
3000 } | |
3001 { | |
3002 E Tdb, TdO, Teh, Tek; | |
3003 Tdb = TcR + Tda; | |
3004 TdO = FNMS(KP293460948, TdN, KP1_978353019 * TdC); | |
3005 R1[WS(rs, 33)] = Tdb - TdO; | |
3006 R1[WS(rs, 1)] = Tdb + TdO; | |
3007 Teh = Tef - Teg; | |
3008 Tek = FNMS(KP1_940062506, Tej, KP485960359 * Tei); | |
3009 R1[WS(rs, 45)] = Teh - Tek; | |
3010 R1[WS(rs, 13)] = Teh + Tek; | |
3011 } | |
3012 { | |
3013 E Tel, Tem, TdP, TdQ; | |
3014 Tel = Tef + Teg; | |
3015 Tem = FMA(KP1_940062506, Tei, KP485960359 * Tej); | |
3016 R1[WS(rs, 29)] = Tel - Tem; | |
3017 R1[WS(rs, 61)] = Tel + Tem; | |
3018 TdP = TcR - Tda; | |
3019 TdQ = FMA(KP293460948, TdC, KP1_978353019 * TdN); | |
3020 R1[WS(rs, 17)] = TdP - TdQ; | |
3021 R1[WS(rs, 49)] = TdP + TdQ; | |
3022 } | |
3023 { | |
3024 E TdT, TdW, Te5, Tec; | |
3025 TdT = TdR - TdS; | |
3026 TdW = FNMS(KP1_606415062, TdV, KP1_191398608 * TdU); | |
3027 R1[WS(rs, 41)] = TdT - TdW; | |
3028 R1[WS(rs, 9)] = TdT + TdW; | |
3029 Te5 = Te1 + Te4; | |
3030 Tec = FNMS(KP1_028205488, Teb, KP1_715457220 * Te8); | |
3031 R1[WS(rs, 37)] = Te5 - Tec; | |
3032 R1[WS(rs, 5)] = Te5 + Tec; | |
3033 } | |
3034 { | |
3035 E Ted, Tee, TdX, TdY; | |
3036 Ted = Te1 - Te4; | |
3037 Tee = FMA(KP1_028205488, Te8, KP1_715457220 * Teb); | |
3038 R1[WS(rs, 21)] = Ted - Tee; | |
3039 R1[WS(rs, 53)] = Ted + Tee; | |
3040 TdX = TdR + TdS; | |
3041 TdY = FMA(KP1_606415062, TdU, KP1_191398608 * TdV); | |
3042 R1[WS(rs, 25)] = TdX - TdY; | |
3043 R1[WS(rs, 57)] = TdX + TdY; | |
3044 } | |
3045 } | |
3046 { | |
3047 E TbD, Tc7, TbK, Tc8, Tck, Tcw, Tch, Tcv, Tc3, Tcy, Tcb, Tco, TbW, Tcz, Tca; | |
3048 E Tcr; | |
3049 { | |
3050 E Tbz, TbC, Tcf, Tcg; | |
3051 Tbz = Tbx - Tby; | |
3052 TbC = FNMS(KP1_662939224, TbB, KP1_111140466 * TbA); | |
3053 TbD = Tbz + TbC; | |
3054 Tc7 = Tbz - TbC; | |
3055 { | |
3056 E TbG, TbJ, Tci, Tcj; | |
3057 TbG = TbE + TbF; | |
3058 TbJ = TbH + TbI; | |
3059 TbK = FNMS(KP942793473, TbJ, KP1_763842528 * TbG); | |
3060 Tc8 = FMA(KP942793473, TbG, KP1_763842528 * TbJ); | |
3061 Tci = TbE - TbF; | |
3062 Tcj = TbI - TbH; | |
3063 Tck = FNMS(KP1_913880671, Tcj, KP580569354 * Tci); | |
3064 Tcw = FMA(KP1_913880671, Tci, KP580569354 * Tcj); | |
3065 } | |
3066 Tcf = Tbx + Tby; | |
3067 Tcg = FMA(KP1_662939224, TbA, KP1_111140466 * TbB); | |
3068 Tch = Tcf - Tcg; | |
3069 Tcv = Tcf + Tcg; | |
3070 { | |
3071 E Tc2, Tcm, TbZ, Tcn, TbX, TbY; | |
3072 Tc2 = Tc0 + Tc1; | |
3073 Tcm = TbM - TbN; | |
3074 TbX = FMA(KP831469612, TbP, KP555570233 * TbQ); | |
3075 TbY = FNMS(KP831469612, TbS, KP555570233 * TbT); | |
3076 TbZ = TbX + TbY; | |
3077 Tcn = TbY - TbX; | |
3078 Tc3 = TbZ + Tc2; | |
3079 Tcy = Tcm - Tcn; | |
3080 Tcb = Tc2 - TbZ; | |
3081 Tco = Tcm + Tcn; | |
3082 } | |
3083 { | |
3084 E TbO, Tcq, TbV, Tcp, TbR, TbU; | |
3085 TbO = TbM + TbN; | |
3086 Tcq = Tc1 - Tc0; | |
3087 TbR = FNMS(KP831469612, TbQ, KP555570233 * TbP); | |
3088 TbU = FMA(KP555570233, TbS, KP831469612 * TbT); | |
3089 TbV = TbR + TbU; | |
3090 Tcp = TbR - TbU; | |
3091 TbW = TbO + TbV; | |
3092 Tcz = Tcq - Tcp; | |
3093 Tca = TbO - TbV; | |
3094 Tcr = Tcp + Tcq; | |
3095 } | |
3096 } | |
3097 { | |
3098 E TbL, Tc4, Tcx, TcA; | |
3099 TbL = TbD + TbK; | |
3100 Tc4 = FNMS(KP485960359, Tc3, KP1_940062506 * TbW); | |
3101 R1[WS(rs, 34)] = TbL - Tc4; | |
3102 R1[WS(rs, 2)] = TbL + Tc4; | |
3103 Tcx = Tcv - Tcw; | |
3104 TcA = FNMS(KP1_978353019, Tcz, KP293460948 * Tcy); | |
3105 R1[WS(rs, 46)] = Tcx - TcA; | |
3106 R1[WS(rs, 14)] = Tcx + TcA; | |
3107 } | |
3108 { | |
3109 E TcB, TcC, Tc5, Tc6; | |
3110 TcB = Tcv + Tcw; | |
3111 TcC = FMA(KP1_978353019, Tcy, KP293460948 * Tcz); | |
3112 R1[WS(rs, 30)] = TcB - TcC; | |
3113 R1[WS(rs, 62)] = TcB + TcC; | |
3114 Tc5 = TbD - TbK; | |
3115 Tc6 = FMA(KP485960359, TbW, KP1_940062506 * Tc3); | |
3116 R1[WS(rs, 18)] = Tc5 - Tc6; | |
3117 R1[WS(rs, 50)] = Tc5 + Tc6; | |
3118 } | |
3119 { | |
3120 E Tc9, Tcc, Tcl, Tcs; | |
3121 Tc9 = Tc7 - Tc8; | |
3122 Tcc = FNMS(KP1_715457220, Tcb, KP1_028205488 * Tca); | |
3123 R1[WS(rs, 42)] = Tc9 - Tcc; | |
3124 R1[WS(rs, 10)] = Tc9 + Tcc; | |
3125 Tcl = Tch + Tck; | |
3126 Tcs = FNMS(KP1_191398608, Tcr, KP1_606415062 * Tco); | |
3127 R1[WS(rs, 38)] = Tcl - Tcs; | |
3128 R1[WS(rs, 6)] = Tcl + Tcs; | |
3129 } | |
3130 { | |
3131 E Tct, Tcu, Tcd, Tce; | |
3132 Tct = Tch - Tck; | |
3133 Tcu = FMA(KP1_191398608, Tco, KP1_606415062 * Tcr); | |
3134 R1[WS(rs, 22)] = Tct - Tcu; | |
3135 R1[WS(rs, 54)] = Tct + Tcu; | |
3136 Tcd = Tc7 + Tc8; | |
3137 Tce = FMA(KP1_715457220, Tca, KP1_028205488 * Tcb); | |
3138 R1[WS(rs, 26)] = Tcd - Tce; | |
3139 R1[WS(rs, 58)] = Tcd + Tce; | |
3140 } | |
3141 } | |
3142 { | |
3143 E Tet, TeX, TeA, TeY, Tfa, Tfm, Tf7, Tfl, TeT, Tfo, Tf1, Tfe, TeM, Tfp, Tf0; | |
3144 E Tfh; | |
3145 { | |
3146 E Tep, Tes, Tf5, Tf6; | |
3147 Tep = Ten - Teo; | |
3148 Tes = FNMS(KP1_961570560, Ter, KP390180644 * Teq); | |
3149 Tet = Tep + Tes; | |
3150 TeX = Tep - Tes; | |
3151 { | |
3152 E Tew, Tez, Tf8, Tf9; | |
3153 Tew = Teu - Tev; | |
3154 Tez = Tex + Tey; | |
3155 TeA = FNMS(KP1_268786568, Tez, KP1_546020906 * Tew); | |
3156 TeY = FMA(KP1_268786568, Tew, KP1_546020906 * Tez); | |
3157 Tf8 = Teu + Tev; | |
3158 Tf9 = Tey - Tex; | |
3159 Tfa = FNMS(KP1_990369453, Tf9, KP196034280 * Tf8); | |
3160 Tfm = FMA(KP1_990369453, Tf8, KP196034280 * Tf9); | |
3161 } | |
3162 Tf5 = Ten + Teo; | |
3163 Tf6 = FMA(KP1_961570560, Teq, KP390180644 * Ter); | |
3164 Tf7 = Tf5 - Tf6; | |
3165 Tfl = Tf5 + Tf6; | |
3166 { | |
3167 E TeS, Tfc, TeP, Tfd, TeN, TeO; | |
3168 TeS = TeQ + TeR; | |
3169 Tfc = TeC + TeD; | |
3170 TeN = FMA(KP980785280, TeF, KP195090322 * TeG); | |
3171 TeO = FMA(KP980785280, TeI, KP195090322 * TeJ); | |
3172 TeP = TeN - TeO; | |
3173 Tfd = TeN + TeO; | |
3174 TeT = TeP + TeS; | |
3175 Tfo = Tfc + Tfd; | |
3176 Tf1 = TeS - TeP; | |
3177 Tfe = Tfc - Tfd; | |
3178 } | |
3179 { | |
3180 E TeE, Tfg, TeL, Tff, TeH, TeK; | |
3181 TeE = TeC - TeD; | |
3182 Tfg = TeR - TeQ; | |
3183 TeH = FNMS(KP980785280, TeG, KP195090322 * TeF); | |
3184 TeK = FNMS(KP980785280, TeJ, KP195090322 * TeI); | |
3185 TeL = TeH + TeK; | |
3186 Tff = TeH - TeK; | |
3187 TeM = TeE + TeL; | |
3188 Tfp = Tfg - Tff; | |
3189 Tf0 = TeE - TeL; | |
3190 Tfh = Tff + Tfg; | |
3191 } | |
3192 } | |
3193 { | |
3194 E TeB, TeU, Tfn, Tfq; | |
3195 TeB = Tet + TeA; | |
3196 TeU = FNMS(KP673779706, TeT, KP1_883088130 * TeM); | |
3197 R1[WS(rs, 35)] = TeB - TeU; | |
3198 R1[WS(rs, 3)] = TeB + TeU; | |
3199 Tfn = Tfl - Tfm; | |
3200 Tfq = FNMS(KP1_997590912, Tfp, KP098135348 * Tfo); | |
3201 R1[WS(rs, 47)] = Tfn - Tfq; | |
3202 R1[WS(rs, 15)] = Tfn + Tfq; | |
3203 } | |
3204 { | |
3205 E Tfr, Tfs, TeV, TeW; | |
3206 Tfr = Tfl + Tfm; | |
3207 Tfs = FMA(KP1_997590912, Tfo, KP098135348 * Tfp); | |
3208 R1[WS(rs, 31)] = Tfr - Tfs; | |
3209 R1[WS(rs, 63)] = Tfr + Tfs; | |
3210 TeV = Tet - TeA; | |
3211 TeW = FMA(KP673779706, TeM, KP1_883088130 * TeT); | |
3212 R1[WS(rs, 19)] = TeV - TeW; | |
3213 R1[WS(rs, 51)] = TeV + TeW; | |
3214 } | |
3215 { | |
3216 E TeZ, Tf2, Tfb, Tfi; | |
3217 TeZ = TeX - TeY; | |
3218 Tf2 = FNMS(KP1_807978586, Tf1, KP855110186 * Tf0); | |
3219 R1[WS(rs, 43)] = TeZ - Tf2; | |
3220 R1[WS(rs, 11)] = TeZ + Tf2; | |
3221 Tfb = Tf7 + Tfa; | |
3222 Tfi = FNMS(KP1_343117909, Tfh, KP1_481902250 * Tfe); | |
3223 R1[WS(rs, 39)] = Tfb - Tfi; | |
3224 R1[WS(rs, 7)] = Tfb + Tfi; | |
3225 } | |
3226 { | |
3227 E Tfj, Tfk, Tf3, Tf4; | |
3228 Tfj = Tf7 - Tfa; | |
3229 Tfk = FMA(KP1_343117909, Tfe, KP1_481902250 * Tfh); | |
3230 R1[WS(rs, 23)] = Tfj - Tfk; | |
3231 R1[WS(rs, 55)] = Tfj + Tfk; | |
3232 Tf3 = TeX + TeY; | |
3233 Tf4 = FMA(KP1_807978586, Tf0, KP855110186 * Tf1); | |
3234 R1[WS(rs, 27)] = Tf3 - Tf4; | |
3235 R1[WS(rs, 59)] = Tf3 + Tf4; | |
3236 } | |
3237 } | |
3238 } | |
3239 } | |
3240 } | |
3241 | |
3242 static const kr2c_desc desc = { 128, "r2cb_128", {812, 198, 144, 0}, &GENUS }; | |
3243 | |
3244 void X(codelet_r2cb_128) (planner *p) { | |
3245 X(kr2c_register) (p, r2cb_128, &desc); | |
3246 } | |
3247 | |
3248 #endif |