Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/rdft/scalar/r2cf/r2cf_128.c @ 10:37bf6b4a2645
Add FFTW3
author | Chris Cannam |
---|---|
date | Wed, 20 Mar 2013 15:35:50 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
9:c0fb53affa76 | 10:37bf6b4a2645 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-11 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Sun Nov 25 07:39:47 EST 2012 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 128 -name r2cf_128 -include r2cf.h */ | |
29 | |
30 /* | |
31 * This function contains 956 FP additions, 516 FP multiplications, | |
32 * (or, 440 additions, 0 multiplications, 516 fused multiply/add), | |
33 * 229 stack variables, 31 constants, and 256 memory accesses | |
34 */ | |
35 #include "r2cf.h" | |
36 | |
37 static void r2cf_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP989176509, +0.989176509964780973451673738016243063983689533); | |
40 DK(KP803207531, +0.803207531480644909806676512963141923879569427); | |
41 DK(KP148335987, +0.148335987538347428753676511486911367000625355); | |
42 DK(KP741650546, +0.741650546272035369581266691172079863842265220); | |
43 DK(KP998795456, +0.998795456205172392714771604759100694443203615); | |
44 DK(KP740951125, +0.740951125354959091175616897495162729728955309); | |
45 DK(KP049126849, +0.049126849769467254105343321271313617079695752); | |
46 DK(KP906347169, +0.906347169019147157946142717268914412664134293); | |
47 DK(KP857728610, +0.857728610000272069902269984284770137042490799); | |
48 DK(KP970031253, +0.970031253194543992603984207286100251456865962); | |
49 DK(KP599376933, +0.599376933681923766271389869014404232837890546); | |
50 DK(KP250486960, +0.250486960191305461595702160124721208578685568); | |
51 DK(KP941544065, +0.941544065183020778412509402599502357185589796); | |
52 DK(KP903989293, +0.903989293123443331586200297230537048710132025); | |
53 DK(KP472964775, +0.472964775891319928124438237972992463904131113); | |
54 DK(KP357805721, +0.357805721314524104672487743774474392487532769); | |
55 DK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
56 DK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
57 DK(KP098491403, +0.098491403357164253077197521291327432293052451); | |
58 DK(KP820678790, +0.820678790828660330972281985331011598767386482); | |
59 DK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
60 DK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
61 DK(KP534511135, +0.534511135950791641089685961295362908582039528); | |
62 DK(KP303346683, +0.303346683607342391675883946941299872384187453); | |
63 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
64 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
65 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
66 DK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
67 DK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
68 DK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
69 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
70 { | |
71 INT i; | |
72 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) { | |
73 E T95, T96; | |
74 { | |
75 E TcD, TdR, T5P, T8v, T27, T7r, Tf, Ta5, T7s, T5S, T8w, T2e, TdS, TcG, Tbn; | |
76 E Tu, TcK, TdU, TK, Ta6, T7w, T8y, T2o, T5U, TcN, TdV, TZ, Ta7, T7z, T8z; | |
77 E T2x, T5V, T1g, Taa, Tab, T1v, Tew, TcX, Tex, TcU, T6A, T2M, T9b, T7E, T9a; | |
78 E T7H, T6z, T2T, TeO, TdK, TeL, Tdz, T9p, T8d, T6O, T5G, T6L, T4X, Tc3, TaV; | |
79 E Tc4, Tbi, T9s, T8o, TeH, Tdp, TeE, Tde, T9i, T7U, T6H, T4r, T6E, T3I, TbW; | |
80 E Tao, TbX, TaL, T9l, T85, T1L, Tad, Tae, T20, Tez, Td6, TeA, Td3, T6x, T37; | |
81 E T9e, T7L, T9d, T7O, T6w, T3e, TbZ, T3Z, T4s, Tc0, TeF, Tds, T4t, T4g, T87; | |
82 E T80, TeI, Tdl, T86, T7X, TaM, TaD, Tb2, Tc6, T8e, T8f, T5e, T5H, Tb9, Tc7; | |
83 E TeM, TdN, T5I, T5v, T8q, T8j, TeP, TdG; | |
84 { | |
85 E T7G, T2S, T2P, T7F; | |
86 { | |
87 E T28, Ti, Tn, T2c, Ts, T29, Tl, To; | |
88 { | |
89 E T4, T23, T3, T25, Td, T5, T8, T9; | |
90 { | |
91 E T1, T2, Tb, Tc; | |
92 T1 = R0[0]; | |
93 T2 = R0[WS(rs, 32)]; | |
94 Tb = R0[WS(rs, 56)]; | |
95 Tc = R0[WS(rs, 24)]; | |
96 T4 = R0[WS(rs, 16)]; | |
97 T23 = T1 - T2; | |
98 T3 = T1 + T2; | |
99 T25 = Tb - Tc; | |
100 Td = Tb + Tc; | |
101 T5 = R0[WS(rs, 48)]; | |
102 T8 = R0[WS(rs, 8)]; | |
103 T9 = R0[WS(rs, 40)]; | |
104 } | |
105 { | |
106 E Tq, Tr, Tj, Tk; | |
107 { | |
108 E Tg, T5N, T6, T24, Ta, Th; | |
109 Tg = R0[WS(rs, 4)]; | |
110 T5N = T4 - T5; | |
111 T6 = T4 + T5; | |
112 T24 = T8 - T9; | |
113 Ta = T8 + T9; | |
114 Th = R0[WS(rs, 36)]; | |
115 { | |
116 E T7, T26, T5O, Te; | |
117 TcD = T3 - T6; | |
118 T7 = T3 + T6; | |
119 T26 = T24 + T25; | |
120 T5O = T25 - T24; | |
121 TdR = Td - Ta; | |
122 Te = Ta + Td; | |
123 T5P = FNMS(KP707106781, T5O, T5N); | |
124 T8v = FMA(KP707106781, T5O, T5N); | |
125 T27 = FMA(KP707106781, T26, T23); | |
126 T7r = FNMS(KP707106781, T26, T23); | |
127 Tf = T7 + Te; | |
128 Ta5 = T7 - Te; | |
129 T28 = Tg - Th; | |
130 Ti = Tg + Th; | |
131 } | |
132 } | |
133 Tq = R0[WS(rs, 12)]; | |
134 Tr = R0[WS(rs, 44)]; | |
135 Tj = R0[WS(rs, 20)]; | |
136 Tk = R0[WS(rs, 52)]; | |
137 Tn = R0[WS(rs, 60)]; | |
138 T2c = Tq - Tr; | |
139 Ts = Tq + Tr; | |
140 T29 = Tj - Tk; | |
141 Tl = Tj + Tk; | |
142 To = R0[WS(rs, 28)]; | |
143 } | |
144 } | |
145 { | |
146 E T2g, T2l, T2h, TF, TcI, TC, T2i, TI; | |
147 { | |
148 E Ty, TG, TB, TH; | |
149 { | |
150 E Tw, T5Q, T2a, TcE, Tm, T2b, Tp, Tx; | |
151 Tw = R0[WS(rs, 2)]; | |
152 T5Q = FMA(KP414213562, T28, T29); | |
153 T2a = FNMS(KP414213562, T29, T28); | |
154 TcE = Ti - Tl; | |
155 Tm = Ti + Tl; | |
156 T2b = Tn - To; | |
157 Tp = Tn + To; | |
158 Tx = R0[WS(rs, 34)]; | |
159 { | |
160 E Tz, TA, TD, TE; | |
161 Tz = R0[WS(rs, 18)]; | |
162 { | |
163 E T5R, T2d, TcF, Tt; | |
164 T5R = FNMS(KP414213562, T2b, T2c); | |
165 T2d = FMA(KP414213562, T2c, T2b); | |
166 TcF = Tp - Ts; | |
167 Tt = Tp + Ts; | |
168 T2g = Tw - Tx; | |
169 Ty = Tw + Tx; | |
170 T7s = T5Q - T5R; | |
171 T5S = T5Q + T5R; | |
172 T8w = T2d - T2a; | |
173 T2e = T2a + T2d; | |
174 TdS = TcF - TcE; | |
175 TcG = TcE + TcF; | |
176 Tbn = Tt - Tm; | |
177 Tu = Tm + Tt; | |
178 TA = R0[WS(rs, 50)]; | |
179 } | |
180 TD = R0[WS(rs, 10)]; | |
181 TE = R0[WS(rs, 42)]; | |
182 TG = R0[WS(rs, 58)]; | |
183 T2l = Tz - TA; | |
184 TB = Tz + TA; | |
185 T2h = TD - TE; | |
186 TF = TD + TE; | |
187 TH = R0[WS(rs, 26)]; | |
188 } | |
189 } | |
190 TcI = Ty - TB; | |
191 TC = Ty + TB; | |
192 T2i = TG - TH; | |
193 TI = TG + TH; | |
194 } | |
195 { | |
196 E T2p, T2u, T2q, TU, TcL, TR, T2r, TX; | |
197 { | |
198 E TN, TV, TQ, TW; | |
199 { | |
200 E T2k, T7u, T2n, T7v, TL, TM; | |
201 TL = R0[WS(rs, 62)]; | |
202 TM = R0[WS(rs, 30)]; | |
203 { | |
204 E TJ, TcJ, T2m, T2j; | |
205 TJ = TF + TI; | |
206 TcJ = TI - TF; | |
207 T2m = T2h - T2i; | |
208 T2j = T2h + T2i; | |
209 TcK = FMA(KP414213562, TcJ, TcI); | |
210 TdU = FNMS(KP414213562, TcI, TcJ); | |
211 TK = TC + TJ; | |
212 Ta6 = TC - TJ; | |
213 T2k = FMA(KP707106781, T2j, T2g); | |
214 T7u = FNMS(KP707106781, T2j, T2g); | |
215 T2n = FMA(KP707106781, T2m, T2l); | |
216 T7v = FNMS(KP707106781, T2m, T2l); | |
217 T2p = TL - TM; | |
218 TN = TL + TM; | |
219 } | |
220 T7w = FMA(KP668178637, T7v, T7u); | |
221 T8y = FNMS(KP668178637, T7u, T7v); | |
222 T2o = FNMS(KP198912367, T2n, T2k); | |
223 T5U = FMA(KP198912367, T2k, T2n); | |
224 { | |
225 E TO, TP, TS, TT; | |
226 TO = R0[WS(rs, 14)]; | |
227 TP = R0[WS(rs, 46)]; | |
228 TS = R0[WS(rs, 6)]; | |
229 TT = R0[WS(rs, 38)]; | |
230 TV = R0[WS(rs, 54)]; | |
231 T2u = TO - TP; | |
232 TQ = TO + TP; | |
233 T2q = TS - TT; | |
234 TU = TS + TT; | |
235 TW = R0[WS(rs, 22)]; | |
236 } | |
237 } | |
238 TcL = TN - TQ; | |
239 TR = TN + TQ; | |
240 T2r = TV - TW; | |
241 TX = TV + TW; | |
242 } | |
243 { | |
244 E T2A, T14, T2N, T17, T1b, T1e, T2D, T2O, T1r, T2I, T1q, T2Q, T2H, TcR, T1n; | |
245 E T1s, T15, T16; | |
246 { | |
247 E T2t, T7x, T2w, T7y, T12, T13; | |
248 T12 = R0[WS(rs, 1)]; | |
249 T13 = R0[WS(rs, 33)]; | |
250 { | |
251 E TY, TcM, T2v, T2s; | |
252 TY = TU + TX; | |
253 TcM = TX - TU; | |
254 T2v = T2q - T2r; | |
255 T2s = T2q + T2r; | |
256 TcN = FNMS(KP414213562, TcM, TcL); | |
257 TdV = FMA(KP414213562, TcL, TcM); | |
258 TZ = TR + TY; | |
259 Ta7 = TR - TY; | |
260 T2t = FMA(KP707106781, T2s, T2p); | |
261 T7x = FNMS(KP707106781, T2s, T2p); | |
262 T2w = FMA(KP707106781, T2v, T2u); | |
263 T7y = FNMS(KP707106781, T2v, T2u); | |
264 T2A = T12 - T13; | |
265 T14 = T12 + T13; | |
266 } | |
267 T7z = FNMS(KP668178637, T7y, T7x); | |
268 T8z = FMA(KP668178637, T7x, T7y); | |
269 T2x = FMA(KP198912367, T2w, T2t); | |
270 T5V = FNMS(KP198912367, T2t, T2w); | |
271 T15 = R0[WS(rs, 17)]; | |
272 T16 = R0[WS(rs, 49)]; | |
273 } | |
274 { | |
275 E T1c, T2B, T1d, T19, T1a; | |
276 T19 = R0[WS(rs, 9)]; | |
277 T1a = R0[WS(rs, 41)]; | |
278 T1c = R0[WS(rs, 57)]; | |
279 T2N = T15 - T16; | |
280 T17 = T15 + T16; | |
281 T2B = T19 - T1a; | |
282 T1b = T19 + T1a; | |
283 T1d = R0[WS(rs, 25)]; | |
284 { | |
285 E T1k, T2F, T1j, T1l, T1h, T1i, T2C; | |
286 T1h = R0[WS(rs, 5)]; | |
287 T1i = R0[WS(rs, 37)]; | |
288 T2C = T1c - T1d; | |
289 T1e = T1c + T1d; | |
290 T1k = R0[WS(rs, 21)]; | |
291 T2F = T1h - T1i; | |
292 T1j = T1h + T1i; | |
293 T2D = T2B + T2C; | |
294 T2O = T2B - T2C; | |
295 T1l = R0[WS(rs, 53)]; | |
296 { | |
297 E T1o, T1p, T2G, T1m; | |
298 T1o = R0[WS(rs, 61)]; | |
299 T1p = R0[WS(rs, 29)]; | |
300 T1r = R0[WS(rs, 13)]; | |
301 T2G = T1k - T1l; | |
302 T1m = T1k + T1l; | |
303 T2I = T1o - T1p; | |
304 T1q = T1o + T1p; | |
305 T2Q = FMA(KP414213562, T2F, T2G); | |
306 T2H = FNMS(KP414213562, T2G, T2F); | |
307 TcR = T1j - T1m; | |
308 T1n = T1j + T1m; | |
309 T1s = R0[WS(rs, 45)]; | |
310 } | |
311 } | |
312 } | |
313 { | |
314 E TcQ, TcV, T2K, T2R, T1u, TcT, TcW, TcS; | |
315 { | |
316 E T18, T1f, T1t, T2J; | |
317 T18 = T14 + T17; | |
318 TcQ = T14 - T17; | |
319 TcV = T1e - T1b; | |
320 T1f = T1b + T1e; | |
321 T1t = T1r + T1s; | |
322 T2J = T1r - T1s; | |
323 T1g = T18 + T1f; | |
324 Taa = T18 - T1f; | |
325 T2K = FMA(KP414213562, T2J, T2I); | |
326 T2R = FNMS(KP414213562, T2I, T2J); | |
327 T1u = T1q + T1t; | |
328 TcS = T1q - T1t; | |
329 } | |
330 TcT = TcR + TcS; | |
331 TcW = TcS - TcR; | |
332 { | |
333 E T7C, T2E, T2L, T7D; | |
334 T7C = FNMS(KP707106781, T2D, T2A); | |
335 T2E = FMA(KP707106781, T2D, T2A); | |
336 Tab = T1u - T1n; | |
337 T1v = T1n + T1u; | |
338 Tew = FNMS(KP707106781, TcW, TcV); | |
339 TcX = FMA(KP707106781, TcW, TcV); | |
340 Tex = FNMS(KP707106781, TcT, TcQ); | |
341 TcU = FMA(KP707106781, TcT, TcQ); | |
342 T2L = T2H + T2K; | |
343 T7G = T2K - T2H; | |
344 T7D = T2Q - T2R; | |
345 T2S = T2Q + T2R; | |
346 T2P = FMA(KP707106781, T2O, T2N); | |
347 T7F = FNMS(KP707106781, T2O, T2N); | |
348 T6A = FNMS(KP923879532, T2L, T2E); | |
349 T2M = FMA(KP923879532, T2L, T2E); | |
350 T9b = FNMS(KP923879532, T7D, T7C); | |
351 T7E = FMA(KP923879532, T7D, T7C); | |
352 } | |
353 } | |
354 } | |
355 } | |
356 } | |
357 } | |
358 { | |
359 E T83, T84, T8m, T8n; | |
360 { | |
361 E TaP, T4z, TaQ, T5A, TaS, TaT, T4G, T5B, T4O, T5D, Tbh, Tdw, T4R, Tbc, T4S; | |
362 E T4T; | |
363 { | |
364 E T4x, T4y, T5y, T5z; | |
365 T4x = R1[WS(rs, 63)]; | |
366 T9a = FNMS(KP923879532, T7G, T7F); | |
367 T7H = FMA(KP923879532, T7G, T7F); | |
368 T6z = FNMS(KP923879532, T2S, T2P); | |
369 T2T = FMA(KP923879532, T2S, T2P); | |
370 T4y = R1[WS(rs, 31)]; | |
371 T5y = R1[WS(rs, 47)]; | |
372 T5z = R1[WS(rs, 15)]; | |
373 { | |
374 E T4A, T4B, T4D, T4E; | |
375 T4A = R1[WS(rs, 7)]; | |
376 TaP = T4x + T4y; | |
377 T4z = T4x - T4y; | |
378 TaQ = T5z + T5y; | |
379 T5A = T5y - T5z; | |
380 T4B = R1[WS(rs, 39)]; | |
381 T4D = R1[WS(rs, 55)]; | |
382 T4E = R1[WS(rs, 23)]; | |
383 { | |
384 E T4K, Tbf, Tbg, T4N, T4P, T4Q; | |
385 { | |
386 E T4I, T4C, T4F, T4J, T4L, T4M; | |
387 T4I = R1[WS(rs, 3)]; | |
388 TaS = T4A + T4B; | |
389 T4C = T4A - T4B; | |
390 TaT = T4D + T4E; | |
391 T4F = T4D - T4E; | |
392 T4J = R1[WS(rs, 35)]; | |
393 T4L = R1[WS(rs, 51)]; | |
394 T4M = R1[WS(rs, 19)]; | |
395 T4G = T4C + T4F; | |
396 T5B = T4F - T4C; | |
397 T4K = T4I - T4J; | |
398 Tbf = T4I + T4J; | |
399 Tbg = T4M + T4L; | |
400 T4N = T4L - T4M; | |
401 } | |
402 T4P = R1[WS(rs, 59)]; | |
403 T4Q = R1[WS(rs, 27)]; | |
404 T4O = FMA(KP414213562, T4N, T4K); | |
405 T5D = FNMS(KP414213562, T4K, T4N); | |
406 Tbh = Tbf + Tbg; | |
407 Tdw = Tbf - Tbg; | |
408 T4R = T4P - T4Q; | |
409 Tbc = T4P + T4Q; | |
410 T4S = R1[WS(rs, 43)]; | |
411 T4T = R1[WS(rs, 11)]; | |
412 } | |
413 } | |
414 } | |
415 { | |
416 E T4H, T8b, TaR, Tdv, TdI, TaU, T4U, Tbd, T5C; | |
417 T4H = FMA(KP707106781, T4G, T4z); | |
418 T8b = FNMS(KP707106781, T4G, T4z); | |
419 TaR = TaP + TaQ; | |
420 Tdv = TaP - TaQ; | |
421 TdI = TaT - TaS; | |
422 TaU = TaS + TaT; | |
423 T4U = T4S - T4T; | |
424 Tbd = T4T + T4S; | |
425 T8m = FNMS(KP707106781, T5B, T5A); | |
426 T5C = FMA(KP707106781, T5B, T5A); | |
427 { | |
428 E Tbe, Tdx, T5E, T4V; | |
429 Tbe = Tbc + Tbd; | |
430 Tdx = Tbc - Tbd; | |
431 T5E = FMA(KP414213562, T4R, T4U); | |
432 T4V = FNMS(KP414213562, T4U, T4R); | |
433 { | |
434 E Tdy, TdJ, T5F, T8c, T4W; | |
435 Tdy = Tdw + Tdx; | |
436 TdJ = Tdx - Tdw; | |
437 T5F = T5D + T5E; | |
438 T8c = T5E - T5D; | |
439 T8n = T4V - T4O; | |
440 T4W = T4O + T4V; | |
441 TeO = FNMS(KP707106781, TdJ, TdI); | |
442 TdK = FMA(KP707106781, TdJ, TdI); | |
443 TeL = FNMS(KP707106781, Tdy, Tdv); | |
444 Tdz = FMA(KP707106781, Tdy, Tdv); | |
445 T9p = FNMS(KP923879532, T8c, T8b); | |
446 T8d = FMA(KP923879532, T8c, T8b); | |
447 T6O = FNMS(KP923879532, T5F, T5C); | |
448 T5G = FMA(KP923879532, T5F, T5C); | |
449 T6L = FNMS(KP923879532, T4W, T4H); | |
450 T4X = FMA(KP923879532, T4W, T4H); | |
451 } | |
452 Tc3 = TaR + TaU; | |
453 TaV = TaR - TaU; | |
454 Tc4 = Tbh + Tbe; | |
455 Tbi = Tbe - Tbh; | |
456 } | |
457 } | |
458 } | |
459 { | |
460 E Tai, T3k, Taj, T4l, Tal, Tam, T4m, T3r, T3D, TaF, T3C, Tdb, TaK, T3z, T4o; | |
461 E T3E; | |
462 { | |
463 E T4j, T4k, T3i, T3j; | |
464 T3i = R1[0]; | |
465 T3j = R1[WS(rs, 32)]; | |
466 T4j = R1[WS(rs, 16)]; | |
467 T9s = FMA(KP923879532, T8n, T8m); | |
468 T8o = FNMS(KP923879532, T8n, T8m); | |
469 Tai = T3i + T3j; | |
470 T3k = T3i - T3j; | |
471 T4k = R1[WS(rs, 48)]; | |
472 { | |
473 E T3o, T3n, T3p, T3l, T3m; | |
474 T3l = R1[WS(rs, 8)]; | |
475 T3m = R1[WS(rs, 40)]; | |
476 T3o = R1[WS(rs, 56)]; | |
477 Taj = T4j + T4k; | |
478 T4l = T4j - T4k; | |
479 T3n = T3l - T3m; | |
480 Tal = T3l + T3m; | |
481 T3p = R1[WS(rs, 24)]; | |
482 { | |
483 E T3w, TaI, T3v, T3x, T3t, T3u, T3q; | |
484 T3t = R1[WS(rs, 4)]; | |
485 T3u = R1[WS(rs, 36)]; | |
486 T3q = T3o - T3p; | |
487 Tam = T3o + T3p; | |
488 T3w = R1[WS(rs, 20)]; | |
489 TaI = T3t + T3u; | |
490 T3v = T3t - T3u; | |
491 T4m = T3n - T3q; | |
492 T3r = T3n + T3q; | |
493 T3x = R1[WS(rs, 52)]; | |
494 { | |
495 E T3A, T3B, TaJ, T3y; | |
496 T3A = R1[WS(rs, 60)]; | |
497 T3B = R1[WS(rs, 28)]; | |
498 T3D = R1[WS(rs, 12)]; | |
499 TaJ = T3w + T3x; | |
500 T3y = T3w - T3x; | |
501 TaF = T3A + T3B; | |
502 T3C = T3A - T3B; | |
503 Tdb = TaI - TaJ; | |
504 TaK = TaI + TaJ; | |
505 T3z = FNMS(KP414213562, T3y, T3v); | |
506 T4o = FMA(KP414213562, T3v, T3y); | |
507 T3E = R1[WS(rs, 44)]; | |
508 } | |
509 } | |
510 } | |
511 } | |
512 { | |
513 E T3s, T7S, Tak, Tda, Tdn, Tan, T3F, TaG, T4n; | |
514 T3s = FMA(KP707106781, T3r, T3k); | |
515 T7S = FNMS(KP707106781, T3r, T3k); | |
516 Tak = Tai + Taj; | |
517 Tda = Tai - Taj; | |
518 Tdn = Tam - Tal; | |
519 Tan = Tal + Tam; | |
520 T3F = T3D - T3E; | |
521 TaG = T3D + T3E; | |
522 T83 = FNMS(KP707106781, T4m, T4l); | |
523 T4n = FMA(KP707106781, T4m, T4l); | |
524 { | |
525 E TaH, Tdc, T4p, T3G; | |
526 TaH = TaF + TaG; | |
527 Tdc = TaF - TaG; | |
528 T4p = FNMS(KP414213562, T3C, T3F); | |
529 T3G = FMA(KP414213562, T3F, T3C); | |
530 { | |
531 E Tdd, Tdo, T4q, T7T, T3H; | |
532 Tdd = Tdb + Tdc; | |
533 Tdo = Tdc - Tdb; | |
534 T4q = T4o + T4p; | |
535 T7T = T4o - T4p; | |
536 T84 = T3G - T3z; | |
537 T3H = T3z + T3G; | |
538 TeH = FNMS(KP707106781, Tdo, Tdn); | |
539 Tdp = FMA(KP707106781, Tdo, Tdn); | |
540 TeE = FNMS(KP707106781, Tdd, Tda); | |
541 Tde = FMA(KP707106781, Tdd, Tda); | |
542 T9i = FNMS(KP923879532, T7T, T7S); | |
543 T7U = FMA(KP923879532, T7T, T7S); | |
544 T6H = FNMS(KP923879532, T4q, T4n); | |
545 T4r = FMA(KP923879532, T4q, T4n); | |
546 T6E = FNMS(KP923879532, T3H, T3s); | |
547 T3I = FMA(KP923879532, T3H, T3s); | |
548 } | |
549 TbW = Tak + Tan; | |
550 Tao = Tak - Tan; | |
551 TbX = TaK + TaH; | |
552 TaL = TaH - TaK; | |
553 } | |
554 } | |
555 } | |
556 { | |
557 E T7N, T3d, T3a, T7M; | |
558 { | |
559 E T2V, T1z, T38, T1C, T1G, T1J, T2Y, T39, T1W, T33, T1V, T3b, T32, Td0, T1S; | |
560 E T1X; | |
561 { | |
562 E T1A, T1B, T1x, T1y; | |
563 T1x = R0[WS(rs, 63)]; | |
564 T1y = R0[WS(rs, 31)]; | |
565 T1A = R0[WS(rs, 15)]; | |
566 T9l = FNMS(KP923879532, T84, T83); | |
567 T85 = FMA(KP923879532, T84, T83); | |
568 T2V = T1x - T1y; | |
569 T1z = T1x + T1y; | |
570 T1B = R0[WS(rs, 47)]; | |
571 { | |
572 E T1H, T2W, T1I, T1E, T1F; | |
573 T1E = R0[WS(rs, 7)]; | |
574 T1F = R0[WS(rs, 39)]; | |
575 T1H = R0[WS(rs, 55)]; | |
576 T38 = T1A - T1B; | |
577 T1C = T1A + T1B; | |
578 T2W = T1E - T1F; | |
579 T1G = T1E + T1F; | |
580 T1I = R0[WS(rs, 23)]; | |
581 { | |
582 E T1P, T30, T1O, T1Q, T1M, T1N, T2X; | |
583 T1M = R0[WS(rs, 3)]; | |
584 T1N = R0[WS(rs, 35)]; | |
585 T2X = T1H - T1I; | |
586 T1J = T1H + T1I; | |
587 T1P = R0[WS(rs, 19)]; | |
588 T30 = T1M - T1N; | |
589 T1O = T1M + T1N; | |
590 T2Y = T2W + T2X; | |
591 T39 = T2W - T2X; | |
592 T1Q = R0[WS(rs, 51)]; | |
593 { | |
594 E T1T, T1U, T31, T1R; | |
595 T1T = R0[WS(rs, 59)]; | |
596 T1U = R0[WS(rs, 27)]; | |
597 T1W = R0[WS(rs, 11)]; | |
598 T31 = T1P - T1Q; | |
599 T1R = T1P + T1Q; | |
600 T33 = T1T - T1U; | |
601 T1V = T1T + T1U; | |
602 T3b = FMA(KP414213562, T30, T31); | |
603 T32 = FNMS(KP414213562, T31, T30); | |
604 Td0 = T1O - T1R; | |
605 T1S = T1O + T1R; | |
606 T1X = R0[WS(rs, 43)]; | |
607 } | |
608 } | |
609 } | |
610 } | |
611 { | |
612 E TcZ, Td4, T35, T3c, T1Z, Td2, Td5, Td1; | |
613 { | |
614 E T1D, T1K, T1Y, T34; | |
615 T1D = T1z + T1C; | |
616 TcZ = T1z - T1C; | |
617 Td4 = T1J - T1G; | |
618 T1K = T1G + T1J; | |
619 T1Y = T1W + T1X; | |
620 T34 = T1W - T1X; | |
621 T1L = T1D + T1K; | |
622 Tad = T1D - T1K; | |
623 T35 = FMA(KP414213562, T34, T33); | |
624 T3c = FNMS(KP414213562, T33, T34); | |
625 T1Z = T1V + T1Y; | |
626 Td1 = T1V - T1Y; | |
627 } | |
628 Td2 = Td0 + Td1; | |
629 Td5 = Td1 - Td0; | |
630 { | |
631 E T7J, T2Z, T36, T7K; | |
632 T7J = FNMS(KP707106781, T2Y, T2V); | |
633 T2Z = FMA(KP707106781, T2Y, T2V); | |
634 Tae = T1Z - T1S; | |
635 T20 = T1S + T1Z; | |
636 Tez = FNMS(KP707106781, Td5, Td4); | |
637 Td6 = FMA(KP707106781, Td5, Td4); | |
638 TeA = FNMS(KP707106781, Td2, TcZ); | |
639 Td3 = FMA(KP707106781, Td2, TcZ); | |
640 T36 = T32 + T35; | |
641 T7N = T35 - T32; | |
642 T7K = T3b - T3c; | |
643 T3d = T3b + T3c; | |
644 T3a = FMA(KP707106781, T39, T38); | |
645 T7M = FNMS(KP707106781, T39, T38); | |
646 T6x = FNMS(KP923879532, T36, T2Z); | |
647 T37 = FMA(KP923879532, T36, T2Z); | |
648 T9e = FNMS(KP923879532, T7K, T7J); | |
649 T7L = FMA(KP923879532, T7K, T7J); | |
650 } | |
651 } | |
652 } | |
653 { | |
654 E Tav, T7V, T7W, TaC; | |
655 { | |
656 E T3L, T3W, Tdf, Tar, T42, T4d, Tay, Tdi, T46, Tau, Tdg, T3X, T3S, Taz, T45; | |
657 E T47, Taw, Tax; | |
658 { | |
659 E T3J, T3K, T3U, T3V; | |
660 T3J = R1[WS(rs, 2)]; | |
661 T9d = FNMS(KP923879532, T7N, T7M); | |
662 T7O = FMA(KP923879532, T7N, T7M); | |
663 T6w = FNMS(KP923879532, T3d, T3a); | |
664 T3e = FMA(KP923879532, T3d, T3a); | |
665 T3K = R1[WS(rs, 34)]; | |
666 T3U = R1[WS(rs, 18)]; | |
667 T3V = R1[WS(rs, 50)]; | |
668 { | |
669 E T40, Tap, Taq, T41, T4b, T4c; | |
670 T40 = R1[WS(rs, 62)]; | |
671 T3L = T3J - T3K; | |
672 Tap = T3J + T3K; | |
673 T3W = T3U - T3V; | |
674 Taq = T3U + T3V; | |
675 T41 = R1[WS(rs, 30)]; | |
676 T4b = R1[WS(rs, 14)]; | |
677 T4c = R1[WS(rs, 46)]; | |
678 Tdf = Tap - Taq; | |
679 Tar = Tap + Taq; | |
680 T42 = T40 - T41; | |
681 Taw = T40 + T41; | |
682 Tax = T4b + T4c; | |
683 T4d = T4b - T4c; | |
684 } | |
685 } | |
686 { | |
687 E T3M, T3N, T3P, T3Q; | |
688 T3M = R1[WS(rs, 10)]; | |
689 Tay = Taw + Tax; | |
690 Tdi = Taw - Tax; | |
691 T3N = R1[WS(rs, 42)]; | |
692 T3P = R1[WS(rs, 58)]; | |
693 T3Q = R1[WS(rs, 26)]; | |
694 { | |
695 E T43, Tas, T3O, Tat, T3R, T44; | |
696 T43 = R1[WS(rs, 6)]; | |
697 Tas = T3M + T3N; | |
698 T3O = T3M - T3N; | |
699 Tat = T3P + T3Q; | |
700 T3R = T3P - T3Q; | |
701 T44 = R1[WS(rs, 38)]; | |
702 T46 = R1[WS(rs, 54)]; | |
703 Tau = Tas + Tat; | |
704 Tdg = Tat - Tas; | |
705 T3X = T3O - T3R; | |
706 T3S = T3O + T3R; | |
707 Taz = T43 + T44; | |
708 T45 = T43 - T44; | |
709 T47 = R1[WS(rs, 22)]; | |
710 } | |
711 } | |
712 { | |
713 E Tdq, Tdh, T49, T4e, Tdr, Tdk; | |
714 Tav = Tar - Tau; | |
715 TbZ = Tar + Tau; | |
716 { | |
717 E T3T, T3Y, TaA, T48, Tdj, TaB; | |
718 T3T = FMA(KP707106781, T3S, T3L); | |
719 T7V = FNMS(KP707106781, T3S, T3L); | |
720 T7W = FNMS(KP707106781, T3X, T3W); | |
721 T3Y = FMA(KP707106781, T3X, T3W); | |
722 TaA = T46 + T47; | |
723 T48 = T46 - T47; | |
724 Tdq = FNMS(KP414213562, Tdf, Tdg); | |
725 Tdh = FMA(KP414213562, Tdg, Tdf); | |
726 T3Z = FNMS(KP198912367, T3Y, T3T); | |
727 T4s = FMA(KP198912367, T3T, T3Y); | |
728 Tdj = TaA - Taz; | |
729 TaB = Taz + TaA; | |
730 T49 = T45 + T48; | |
731 T4e = T45 - T48; | |
732 TaC = Tay - TaB; | |
733 Tc0 = Tay + TaB; | |
734 Tdr = FMA(KP414213562, Tdi, Tdj); | |
735 Tdk = FNMS(KP414213562, Tdj, Tdi); | |
736 } | |
737 { | |
738 E T7Z, T7Y, T4f, T4a; | |
739 T7Z = FNMS(KP707106781, T4e, T4d); | |
740 T4f = FMA(KP707106781, T4e, T4d); | |
741 T4a = FMA(KP707106781, T49, T42); | |
742 T7Y = FNMS(KP707106781, T49, T42); | |
743 TeF = Tdr - Tdq; | |
744 Tds = Tdq + Tdr; | |
745 T4t = FNMS(KP198912367, T4a, T4f); | |
746 T4g = FMA(KP198912367, T4f, T4a); | |
747 T87 = FMA(KP668178637, T7Y, T7Z); | |
748 T80 = FNMS(KP668178637, T7Z, T7Y); | |
749 TeI = Tdh - Tdk; | |
750 Tdl = Tdh + Tdk; | |
751 } | |
752 } | |
753 } | |
754 { | |
755 E T50, T5b, TdA, TaY, T5h, T5s, Tb5, TdD, T5l, Tb1, TdB, T5c, T57, Tb6, T5k; | |
756 E T5m, Tb3, Tb4; | |
757 { | |
758 E T4Y, T4Z, T59, T5a; | |
759 T4Y = R1[WS(rs, 1)]; | |
760 T86 = FNMS(KP668178637, T7V, T7W); | |
761 T7X = FMA(KP668178637, T7W, T7V); | |
762 TaM = TaC - Tav; | |
763 TaD = Tav + TaC; | |
764 T4Z = R1[WS(rs, 33)]; | |
765 T59 = R1[WS(rs, 49)]; | |
766 T5a = R1[WS(rs, 17)]; | |
767 { | |
768 E T5f, TaW, TaX, T5g, T5q, T5r; | |
769 T5f = R1[WS(rs, 61)]; | |
770 T50 = T4Y - T4Z; | |
771 TaW = T4Y + T4Z; | |
772 T5b = T59 - T5a; | |
773 TaX = T5a + T59; | |
774 T5g = R1[WS(rs, 29)]; | |
775 T5q = R1[WS(rs, 45)]; | |
776 T5r = R1[WS(rs, 13)]; | |
777 TdA = TaW - TaX; | |
778 TaY = TaW + TaX; | |
779 T5h = T5f - T5g; | |
780 Tb3 = T5f + T5g; | |
781 Tb4 = T5r + T5q; | |
782 T5s = T5q - T5r; | |
783 } | |
784 } | |
785 { | |
786 E T51, T52, T54, T55; | |
787 T51 = R1[WS(rs, 9)]; | |
788 Tb5 = Tb3 + Tb4; | |
789 TdD = Tb3 - Tb4; | |
790 T52 = R1[WS(rs, 41)]; | |
791 T54 = R1[WS(rs, 57)]; | |
792 T55 = R1[WS(rs, 25)]; | |
793 { | |
794 E T5i, TaZ, T53, Tb0, T56, T5j; | |
795 T5i = R1[WS(rs, 5)]; | |
796 TaZ = T51 + T52; | |
797 T53 = T51 - T52; | |
798 Tb0 = T54 + T55; | |
799 T56 = T54 - T55; | |
800 T5j = R1[WS(rs, 37)]; | |
801 T5l = R1[WS(rs, 53)]; | |
802 Tb1 = TaZ + Tb0; | |
803 TdB = Tb0 - TaZ; | |
804 T5c = T56 - T53; | |
805 T57 = T53 + T56; | |
806 Tb6 = T5i + T5j; | |
807 T5k = T5i - T5j; | |
808 T5m = R1[WS(rs, 21)]; | |
809 } | |
810 } | |
811 { | |
812 E TdL, TdC, T5o, T5t, TdM, TdF; | |
813 Tb2 = TaY - Tb1; | |
814 Tc6 = TaY + Tb1; | |
815 { | |
816 E T58, T5d, Tb7, T5n, TdE, Tb8; | |
817 T58 = FMA(KP707106781, T57, T50); | |
818 T8e = FNMS(KP707106781, T57, T50); | |
819 T8f = FNMS(KP707106781, T5c, T5b); | |
820 T5d = FMA(KP707106781, T5c, T5b); | |
821 Tb7 = T5l + T5m; | |
822 T5n = T5l - T5m; | |
823 TdL = FNMS(KP414213562, TdA, TdB); | |
824 TdC = FMA(KP414213562, TdB, TdA); | |
825 T5e = FMA(KP198912367, T5d, T58); | |
826 T5H = FNMS(KP198912367, T58, T5d); | |
827 TdE = Tb7 - Tb6; | |
828 Tb8 = Tb6 + Tb7; | |
829 T5o = T5k + T5n; | |
830 T5t = T5n - T5k; | |
831 Tb9 = Tb5 - Tb8; | |
832 Tc7 = Tb5 + Tb8; | |
833 TdM = FMA(KP414213562, TdD, TdE); | |
834 TdF = FNMS(KP414213562, TdE, TdD); | |
835 } | |
836 { | |
837 E T8i, T8h, T5u, T5p; | |
838 T8i = FNMS(KP707106781, T5t, T5s); | |
839 T5u = FMA(KP707106781, T5t, T5s); | |
840 T5p = FMA(KP707106781, T5o, T5h); | |
841 T8h = FNMS(KP707106781, T5o, T5h); | |
842 TeM = TdM - TdL; | |
843 TdN = TdL + TdM; | |
844 T5I = FMA(KP198912367, T5p, T5u); | |
845 T5v = FNMS(KP198912367, T5u, T5p); | |
846 T8q = FNMS(KP668178637, T8h, T8i); | |
847 T8j = FMA(KP668178637, T8i, T8h); | |
848 TeP = TdF - TdC; | |
849 TdG = TdC + TdF; | |
850 } | |
851 } | |
852 } | |
853 } | |
854 } | |
855 } | |
856 } | |
857 { | |
858 E T8p, T8g, TcH, TdW, TdT, TcO, Tfp, Tfk, Tfj, Tfq; | |
859 { | |
860 E Tbj, Tba, Tcy, Tco, TcB, Tcl, Tcx, Tcv, Tcz, Tcr; | |
861 { | |
862 E Tch, Tct, Tcp, Tcq, Tci, T1w, TbV, T11, Tcf, Tc9, T21, Tcj, Tcm, TbY, Tc1; | |
863 E Tcn, Tcu, Tck; | |
864 { | |
865 E Tv, T10, Tc5, Tc8; | |
866 Tch = Tf - Tu; | |
867 Tv = Tf + Tu; | |
868 T8p = FMA(KP668178637, T8e, T8f); | |
869 T8g = FNMS(KP668178637, T8f, T8e); | |
870 Tbj = Tb9 - Tb2; | |
871 Tba = Tb2 + Tb9; | |
872 T10 = TK + TZ; | |
873 Tct = TZ - TK; | |
874 Tcp = Tc3 - Tc4; | |
875 Tc5 = Tc3 + Tc4; | |
876 Tc8 = Tc6 + Tc7; | |
877 Tcq = Tc7 - Tc6; | |
878 Tci = T1g - T1v; | |
879 T1w = T1g + T1v; | |
880 TbV = Tv - T10; | |
881 T11 = Tv + T10; | |
882 Tcf = Tc5 + Tc8; | |
883 Tc9 = Tc5 - Tc8; | |
884 T21 = T1L + T20; | |
885 Tcj = T1L - T20; | |
886 Tcm = TbW - TbX; | |
887 TbY = TbW + TbX; | |
888 Tc1 = TbZ + Tc0; | |
889 Tcn = Tc0 - TbZ; | |
890 } | |
891 { | |
892 E Tcb, T22, Tce, Tc2; | |
893 Tcb = T21 - T1w; | |
894 T22 = T1w + T21; | |
895 Tce = TbY + Tc1; | |
896 Tc2 = TbY - Tc1; | |
897 { | |
898 E Tcd, Tcg, Tca, Tcc; | |
899 Tcd = T11 + T22; | |
900 Cr[WS(csr, 32)] = T11 - T22; | |
901 Tcg = Tce + Tcf; | |
902 Ci[WS(csi, 32)] = Tcf - Tce; | |
903 Tca = Tc2 + Tc9; | |
904 Tcc = Tc9 - Tc2; | |
905 Cr[0] = Tcd + Tcg; | |
906 Cr[WS(csr, 64)] = Tcd - Tcg; | |
907 Ci[WS(csi, 48)] = FMS(KP707106781, Tcc, Tcb); | |
908 Ci[WS(csi, 16)] = FMA(KP707106781, Tcc, Tcb); | |
909 Cr[WS(csr, 16)] = FMA(KP707106781, Tca, TbV); | |
910 Cr[WS(csr, 48)] = FNMS(KP707106781, Tca, TbV); | |
911 Tcu = Tcj - Tci; | |
912 Tck = Tci + Tcj; | |
913 Tcy = FNMS(KP414213562, Tcm, Tcn); | |
914 Tco = FMA(KP414213562, Tcn, Tcm); | |
915 } | |
916 } | |
917 TcB = FNMS(KP707106781, Tck, Tch); | |
918 Tcl = FMA(KP707106781, Tck, Tch); | |
919 Tcx = FMA(KP707106781, Tcu, Tct); | |
920 Tcv = FNMS(KP707106781, Tcu, Tct); | |
921 Tcz = FMA(KP414213562, Tcp, Tcq); | |
922 Tcr = FNMS(KP414213562, Tcq, Tcp); | |
923 } | |
924 { | |
925 E TbT, TbO, TbN, TbU; | |
926 { | |
927 E Ta9, TbB, Tbb, TbL, Tbp, TbM, Tag, Tbk, TbR, TbJ, Tbw, TaO, TbC, Tbs, TbQ; | |
928 E TbG; | |
929 { | |
930 E Tbq, Tbr, TbH, TbI; | |
931 { | |
932 E Tbo, Ta8, Tac, Taf; | |
933 Tbo = Ta7 - Ta6; | |
934 Ta8 = Ta6 + Ta7; | |
935 { | |
936 E TcC, TcA, Tcw, Tcs; | |
937 TcC = Tcz - Tcy; | |
938 TcA = Tcy + Tcz; | |
939 Tcw = Tcr - Tco; | |
940 Tcs = Tco + Tcr; | |
941 Cr[WS(csr, 24)] = FMA(KP923879532, TcC, TcB); | |
942 Cr[WS(csr, 40)] = FNMS(KP923879532, TcC, TcB); | |
943 Ci[WS(csi, 56)] = FMS(KP923879532, TcA, Tcx); | |
944 Ci[WS(csi, 8)] = FMA(KP923879532, TcA, Tcx); | |
945 Ci[WS(csi, 40)] = FMA(KP923879532, Tcw, Tcv); | |
946 Ci[WS(csi, 24)] = FMS(KP923879532, Tcw, Tcv); | |
947 Cr[WS(csr, 8)] = FMA(KP923879532, Tcs, Tcl); | |
948 Cr[WS(csr, 56)] = FNMS(KP923879532, Tcs, Tcl); | |
949 Ta9 = FMA(KP707106781, Ta8, Ta5); | |
950 TbB = FNMS(KP707106781, Ta8, Ta5); | |
951 } | |
952 Tbq = FNMS(KP414213562, Taa, Tab); | |
953 Tac = FMA(KP414213562, Tab, Taa); | |
954 Taf = FNMS(KP414213562, Tae, Tad); | |
955 Tbr = FMA(KP414213562, Tad, Tae); | |
956 Tbb = FMA(KP707106781, Tba, TaV); | |
957 TbH = FNMS(KP707106781, Tba, TaV); | |
958 TbL = FNMS(KP707106781, Tbo, Tbn); | |
959 Tbp = FMA(KP707106781, Tbo, Tbn); | |
960 TbM = Taf - Tac; | |
961 Tag = Tac + Taf; | |
962 TbI = FNMS(KP707106781, Tbj, Tbi); | |
963 Tbk = FMA(KP707106781, Tbj, Tbi); | |
964 } | |
965 { | |
966 E TbE, TbF, TaE, TaN; | |
967 TbE = FNMS(KP707106781, TaD, Tao); | |
968 TaE = FMA(KP707106781, TaD, Tao); | |
969 TaN = FMA(KP707106781, TaM, TaL); | |
970 TbF = FNMS(KP707106781, TaM, TaL); | |
971 TbR = FNMS(KP668178637, TbH, TbI); | |
972 TbJ = FMA(KP668178637, TbI, TbH); | |
973 Tbw = FNMS(KP198912367, TaE, TaN); | |
974 TaO = FMA(KP198912367, TaN, TaE); | |
975 TbC = Tbr - Tbq; | |
976 Tbs = Tbq + Tbr; | |
977 TbQ = FMA(KP668178637, TbE, TbF); | |
978 TbG = FNMS(KP668178637, TbF, TbE); | |
979 } | |
980 } | |
981 { | |
982 E Tbz, Tah, Tbv, Tbt, Tbx, Tbl; | |
983 Tbz = FNMS(KP923879532, Tag, Ta9); | |
984 Tah = FMA(KP923879532, Tag, Ta9); | |
985 Tbv = FMA(KP923879532, Tbs, Tbp); | |
986 Tbt = FNMS(KP923879532, Tbs, Tbp); | |
987 Tbx = FMA(KP198912367, Tbb, Tbk); | |
988 Tbl = FNMS(KP198912367, Tbk, Tbb); | |
989 { | |
990 E TbD, TbK, TbP, TbS; | |
991 TbT = FNMS(KP923879532, TbC, TbB); | |
992 TbD = FMA(KP923879532, TbC, TbB); | |
993 { | |
994 E TbA, Tby, Tbu, Tbm; | |
995 TbA = Tbx - Tbw; | |
996 Tby = Tbw + Tbx; | |
997 Tbu = Tbl - TaO; | |
998 Tbm = TaO + Tbl; | |
999 Cr[WS(csr, 28)] = FMA(KP980785280, TbA, Tbz); | |
1000 Cr[WS(csr, 36)] = FNMS(KP980785280, TbA, Tbz); | |
1001 Ci[WS(csi, 60)] = FMS(KP980785280, Tby, Tbv); | |
1002 Ci[WS(csi, 4)] = FMA(KP980785280, Tby, Tbv); | |
1003 Ci[WS(csi, 36)] = FMA(KP980785280, Tbu, Tbt); | |
1004 Ci[WS(csi, 28)] = FMS(KP980785280, Tbu, Tbt); | |
1005 Cr[WS(csr, 4)] = FMA(KP980785280, Tbm, Tah); | |
1006 Cr[WS(csr, 60)] = FNMS(KP980785280, Tbm, Tah); | |
1007 TbK = TbG + TbJ; | |
1008 TbO = TbJ - TbG; | |
1009 } | |
1010 TbN = FMA(KP923879532, TbM, TbL); | |
1011 TbP = FNMS(KP923879532, TbM, TbL); | |
1012 TbS = TbQ + TbR; | |
1013 TbU = TbQ - TbR; | |
1014 Cr[WS(csr, 12)] = FMA(KP831469612, TbK, TbD); | |
1015 Cr[WS(csr, 52)] = FNMS(KP831469612, TbK, TbD); | |
1016 Ci[WS(csi, 52)] = FNMS(KP831469612, TbS, TbP); | |
1017 Ci[WS(csi, 12)] = -(FMA(KP831469612, TbS, TbP)); | |
1018 } | |
1019 } | |
1020 } | |
1021 { | |
1022 E TeN, Tf7, Tev, Tfm, Tfc, TeQ, TeX, TeW, Tfn, Tff, Tfi, TeC, Tf2, TeK, Tfh; | |
1023 E TeV, Tf8; | |
1024 { | |
1025 E TeG, TeJ, Tfd, Tfe, Tey, TeB, TeT, TeU; | |
1026 { | |
1027 E Tet, Teu, Tfa, Tfb; | |
1028 TcH = FMA(KP707106781, TcG, TcD); | |
1029 Tet = FNMS(KP707106781, TcG, TcD); | |
1030 Ci[WS(csi, 44)] = FMS(KP831469612, TbO, TbN); | |
1031 Ci[WS(csi, 20)] = FMA(KP831469612, TbO, TbN); | |
1032 Cr[WS(csr, 20)] = FMA(KP831469612, TbU, TbT); | |
1033 Cr[WS(csr, 44)] = FNMS(KP831469612, TbU, TbT); | |
1034 Teu = TdV - TdU; | |
1035 TdW = TdU + TdV; | |
1036 TeG = FNMS(KP923879532, TeF, TeE); | |
1037 Tfa = FMA(KP923879532, TeF, TeE); | |
1038 Tfb = FMA(KP923879532, TeI, TeH); | |
1039 TeJ = FNMS(KP923879532, TeI, TeH); | |
1040 TeN = FNMS(KP923879532, TeM, TeL); | |
1041 Tfd = FMA(KP923879532, TeM, TeL); | |
1042 Tf7 = FMA(KP923879532, Teu, Tet); | |
1043 Tev = FNMS(KP923879532, Teu, Tet); | |
1044 Tfm = FMA(KP303346683, Tfa, Tfb); | |
1045 Tfc = FNMS(KP303346683, Tfb, Tfa); | |
1046 Tfe = FNMS(KP923879532, TeP, TeO); | |
1047 TeQ = FMA(KP923879532, TeP, TeO); | |
1048 TeX = FNMS(KP668178637, Tew, Tex); | |
1049 Tey = FMA(KP668178637, Tex, Tew); | |
1050 TeB = FNMS(KP668178637, TeA, Tez); | |
1051 TeW = FMA(KP668178637, Tez, TeA); | |
1052 } | |
1053 Tfn = FNMS(KP303346683, Tfd, Tfe); | |
1054 Tff = FMA(KP303346683, Tfe, Tfd); | |
1055 Tfi = Tey + TeB; | |
1056 TeC = Tey - TeB; | |
1057 TdT = FMA(KP707106781, TdS, TdR); | |
1058 TeT = FNMS(KP707106781, TdS, TdR); | |
1059 TeU = TcN - TcK; | |
1060 TcO = TcK + TcN; | |
1061 Tf2 = FNMS(KP534511135, TeG, TeJ); | |
1062 TeK = FMA(KP534511135, TeJ, TeG); | |
1063 Tfh = FNMS(KP923879532, TeU, TeT); | |
1064 TeV = FMA(KP923879532, TeU, TeT); | |
1065 } | |
1066 { | |
1067 E Tf5, TeD, TeY, Tf3, TeR; | |
1068 Tf5 = FNMS(KP831469612, TeC, Tev); | |
1069 TeD = FMA(KP831469612, TeC, Tev); | |
1070 Tf8 = TeX + TeW; | |
1071 TeY = TeW - TeX; | |
1072 Tf3 = FMA(KP534511135, TeN, TeQ); | |
1073 TeR = FNMS(KP534511135, TeQ, TeN); | |
1074 { | |
1075 E Tf1, TeZ, Tf6, Tf4, Tf0, TeS; | |
1076 Tf1 = FMA(KP831469612, TeY, TeV); | |
1077 TeZ = FNMS(KP831469612, TeY, TeV); | |
1078 Tf6 = Tf3 - Tf2; | |
1079 Tf4 = Tf2 + Tf3; | |
1080 Tf0 = TeR - TeK; | |
1081 TeS = TeK + TeR; | |
1082 Ci[WS(csi, 54)] = FMS(KP881921264, Tf4, Tf1); | |
1083 Ci[WS(csi, 10)] = FMA(KP881921264, Tf4, Tf1); | |
1084 Ci[WS(csi, 42)] = FMA(KP881921264, Tf0, TeZ); | |
1085 Ci[WS(csi, 22)] = FMS(KP881921264, Tf0, TeZ); | |
1086 Cr[WS(csr, 10)] = FMA(KP881921264, TeS, TeD); | |
1087 Cr[WS(csr, 54)] = FNMS(KP881921264, TeS, TeD); | |
1088 Cr[WS(csr, 42)] = FNMS(KP881921264, Tf6, Tf5); | |
1089 Cr[WS(csr, 22)] = FMA(KP881921264, Tf6, Tf5); | |
1090 } | |
1091 } | |
1092 { | |
1093 E Tf9, Tfg, Tfl, Tfo; | |
1094 Tfp = FNMS(KP831469612, Tf8, Tf7); | |
1095 Tf9 = FMA(KP831469612, Tf8, Tf7); | |
1096 Tfg = Tfc + Tff; | |
1097 Tfk = Tff - Tfc; | |
1098 Tfj = FNMS(KP831469612, Tfi, Tfh); | |
1099 Tfl = FMA(KP831469612, Tfi, Tfh); | |
1100 Tfo = Tfm + Tfn; | |
1101 Tfq = Tfm - Tfn; | |
1102 Cr[WS(csr, 6)] = FMA(KP956940335, Tfg, Tf9); | |
1103 Cr[WS(csr, 58)] = FNMS(KP956940335, Tfg, Tf9); | |
1104 Ci[WS(csi, 58)] = FNMS(KP956940335, Tfo, Tfl); | |
1105 Ci[WS(csi, 6)] = -(FMA(KP956940335, Tfo, Tfl)); | |
1106 } | |
1107 } | |
1108 } | |
1109 } | |
1110 { | |
1111 E T2f, T5W, T5T, T2y, T5J, T5w, T4u, T4h, T7p, T7q; | |
1112 { | |
1113 E Ter, Tem, Tel, Tes; | |
1114 { | |
1115 E TdH, Te9, TcP, Teo, Tee, TdO, TdY, TdZ, Tep, Teh, Tek, Td8, Te4, Tdu, Tej; | |
1116 E TdX, Tea; | |
1117 { | |
1118 E Tdm, Tdt, Tef, Teg, TcY, Td7, Tec, Ted; | |
1119 Ci[WS(csi, 38)] = FMS(KP956940335, Tfk, Tfj); | |
1120 Ci[WS(csi, 26)] = FMA(KP956940335, Tfk, Tfj); | |
1121 Cr[WS(csr, 26)] = FMA(KP956940335, Tfq, Tfp); | |
1122 Cr[WS(csr, 38)] = FNMS(KP956940335, Tfq, Tfp); | |
1123 Tdm = FMA(KP923879532, Tdl, Tde); | |
1124 Tec = FNMS(KP923879532, Tdl, Tde); | |
1125 Ted = FNMS(KP923879532, Tds, Tdp); | |
1126 Tdt = FMA(KP923879532, Tds, Tdp); | |
1127 TdH = FMA(KP923879532, TdG, Tdz); | |
1128 Tef = FNMS(KP923879532, TdG, Tdz); | |
1129 Te9 = FNMS(KP923879532, TcO, TcH); | |
1130 TcP = FMA(KP923879532, TcO, TcH); | |
1131 Teo = FMA(KP820678790, Tec, Ted); | |
1132 Tee = FNMS(KP820678790, Ted, Tec); | |
1133 Teg = FNMS(KP923879532, TdN, TdK); | |
1134 TdO = FMA(KP923879532, TdN, TdK); | |
1135 TdY = FNMS(KP198912367, TcU, TcX); | |
1136 TcY = FMA(KP198912367, TcX, TcU); | |
1137 Td7 = FNMS(KP198912367, Td6, Td3); | |
1138 TdZ = FMA(KP198912367, Td3, Td6); | |
1139 Tep = FNMS(KP820678790, Tef, Teg); | |
1140 Teh = FMA(KP820678790, Teg, Tef); | |
1141 Tek = Td7 - TcY; | |
1142 Td8 = TcY + Td7; | |
1143 Te4 = FNMS(KP098491403, Tdm, Tdt); | |
1144 Tdu = FMA(KP098491403, Tdt, Tdm); | |
1145 Tej = FNMS(KP923879532, TdW, TdT); | |
1146 TdX = FMA(KP923879532, TdW, TdT); | |
1147 } | |
1148 { | |
1149 E Te7, Td9, Te0, Te5, TdP; | |
1150 Te7 = FNMS(KP980785280, Td8, TcP); | |
1151 Td9 = FMA(KP980785280, Td8, TcP); | |
1152 Tea = TdZ - TdY; | |
1153 Te0 = TdY + TdZ; | |
1154 Te5 = FMA(KP098491403, TdH, TdO); | |
1155 TdP = FNMS(KP098491403, TdO, TdH); | |
1156 { | |
1157 E Te3, Te1, Te8, Te6, Te2, TdQ; | |
1158 Te3 = FMA(KP980785280, Te0, TdX); | |
1159 Te1 = FNMS(KP980785280, Te0, TdX); | |
1160 Te8 = Te5 - Te4; | |
1161 Te6 = Te4 + Te5; | |
1162 Te2 = TdP - Tdu; | |
1163 TdQ = Tdu + TdP; | |
1164 Ci[WS(csi, 62)] = FMS(KP995184726, Te6, Te3); | |
1165 Ci[WS(csi, 2)] = FMA(KP995184726, Te6, Te3); | |
1166 Ci[WS(csi, 34)] = FMA(KP995184726, Te2, Te1); | |
1167 Ci[WS(csi, 30)] = FMS(KP995184726, Te2, Te1); | |
1168 Cr[WS(csr, 2)] = FMA(KP995184726, TdQ, Td9); | |
1169 Cr[WS(csr, 62)] = FNMS(KP995184726, TdQ, Td9); | |
1170 Cr[WS(csr, 34)] = FNMS(KP995184726, Te8, Te7); | |
1171 Cr[WS(csr, 30)] = FMA(KP995184726, Te8, Te7); | |
1172 } | |
1173 } | |
1174 { | |
1175 E Teb, Tei, Ten, Teq; | |
1176 Ter = FNMS(KP980785280, Tea, Te9); | |
1177 Teb = FMA(KP980785280, Tea, Te9); | |
1178 Tei = Tee + Teh; | |
1179 Tem = Teh - Tee; | |
1180 Tel = FMA(KP980785280, Tek, Tej); | |
1181 Ten = FNMS(KP980785280, Tek, Tej); | |
1182 Teq = Teo + Tep; | |
1183 Tes = Teo - Tep; | |
1184 Cr[WS(csr, 14)] = FMA(KP773010453, Tei, Teb); | |
1185 Cr[WS(csr, 50)] = FNMS(KP773010453, Tei, Teb); | |
1186 Ci[WS(csi, 50)] = FNMS(KP773010453, Teq, Ten); | |
1187 Ci[WS(csi, 14)] = -(FMA(KP773010453, Teq, Ten)); | |
1188 } | |
1189 } | |
1190 { | |
1191 E T77, T6v, T7i, T6C, T78, T6Y, T7h, T6V, T6N, T7d, T6P, T6F, T6I; | |
1192 { | |
1193 E T6W, T6X, T6T, T6U, T6M; | |
1194 { | |
1195 E T6t, T6u, T6y, T6B; | |
1196 T2f = FMA(KP923879532, T2e, T27); | |
1197 T6t = FNMS(KP923879532, T2e, T27); | |
1198 Ci[WS(csi, 46)] = FMS(KP773010453, Tem, Tel); | |
1199 Ci[WS(csi, 18)] = FMA(KP773010453, Tem, Tel); | |
1200 Cr[WS(csr, 18)] = FMA(KP773010453, Tes, Ter); | |
1201 Cr[WS(csr, 46)] = FNMS(KP773010453, Tes, Ter); | |
1202 T6u = T5U - T5V; | |
1203 T5W = T5U + T5V; | |
1204 T6W = FNMS(KP820678790, T6w, T6x); | |
1205 T6y = FMA(KP820678790, T6x, T6w); | |
1206 T6B = FNMS(KP820678790, T6A, T6z); | |
1207 T6X = FMA(KP820678790, T6z, T6A); | |
1208 T77 = FMA(KP980785280, T6u, T6t); | |
1209 T6v = FNMS(KP980785280, T6u, T6t); | |
1210 T7i = T6B + T6y; | |
1211 T6C = T6y - T6B; | |
1212 } | |
1213 T5T = FMA(KP923879532, T5S, T5P); | |
1214 T6T = FNMS(KP923879532, T5S, T5P); | |
1215 T6U = T2x - T2o; | |
1216 T2y = T2o + T2x; | |
1217 T5J = T5H + T5I; | |
1218 T6M = T5I - T5H; | |
1219 T78 = T6X + T6W; | |
1220 T6Y = T6W - T6X; | |
1221 T7h = FMA(KP980785280, T6U, T6T); | |
1222 T6V = FNMS(KP980785280, T6U, T6T); | |
1223 T6N = FNMS(KP980785280, T6M, T6L); | |
1224 T7d = FMA(KP980785280, T6M, T6L); | |
1225 T6P = T5v - T5e; | |
1226 T5w = T5e + T5v; | |
1227 T4u = T4s + T4t; | |
1228 T6F = T4s - T4t; | |
1229 T6I = T4g - T3Z; | |
1230 T4h = T3Z + T4g; | |
1231 } | |
1232 { | |
1233 E T75, T7f, T7n, T7c, T7m, T76; | |
1234 { | |
1235 E T6D, T72, T6R, T73, T6K, T71, T6Z, T7e, T6Q, T74, T70, T6S; | |
1236 T75 = FNMS(KP773010453, T6C, T6v); | |
1237 T6D = FMA(KP773010453, T6C, T6v); | |
1238 T7e = FNMS(KP980785280, T6P, T6O); | |
1239 T6Q = FMA(KP980785280, T6P, T6O); | |
1240 { | |
1241 E T7a, T6G, T7b, T6J; | |
1242 T7a = FMA(KP980785280, T6F, T6E); | |
1243 T6G = FNMS(KP980785280, T6F, T6E); | |
1244 T7b = FMA(KP980785280, T6I, T6H); | |
1245 T6J = FNMS(KP980785280, T6I, T6H); | |
1246 T7f = FMA(KP357805721, T7e, T7d); | |
1247 T7n = FNMS(KP357805721, T7d, T7e); | |
1248 T72 = FMA(KP472964775, T6N, T6Q); | |
1249 T6R = FNMS(KP472964775, T6Q, T6N); | |
1250 T7c = FMA(KP357805721, T7b, T7a); | |
1251 T7m = FNMS(KP357805721, T7a, T7b); | |
1252 T73 = FMA(KP472964775, T6G, T6J); | |
1253 T6K = FNMS(KP472964775, T6J, T6G); | |
1254 } | |
1255 T71 = FNMS(KP773010453, T6Y, T6V); | |
1256 T6Z = FMA(KP773010453, T6Y, T6V); | |
1257 T74 = T72 - T73; | |
1258 T76 = T73 + T72; | |
1259 T70 = T6R - T6K; | |
1260 T6S = T6K + T6R; | |
1261 Ci[WS(csi, 55)] = FMA(KP903989293, T74, T71); | |
1262 Ci[WS(csi, 9)] = FMS(KP903989293, T74, T71); | |
1263 Cr[WS(csr, 9)] = FMA(KP903989293, T6S, T6D); | |
1264 Cr[WS(csr, 55)] = FNMS(KP903989293, T6S, T6D); | |
1265 Ci[WS(csi, 41)] = FMS(KP903989293, T70, T6Z); | |
1266 Ci[WS(csi, 23)] = FMA(KP903989293, T70, T6Z); | |
1267 } | |
1268 { | |
1269 E T7k, T7j, T7l, T7o, T79, T7g; | |
1270 T7p = FNMS(KP773010453, T78, T77); | |
1271 T79 = FMA(KP773010453, T78, T77); | |
1272 T7g = T7c + T7f; | |
1273 T7k = T7f - T7c; | |
1274 T7j = FNMS(KP773010453, T7i, T7h); | |
1275 T7l = FMA(KP773010453, T7i, T7h); | |
1276 Cr[WS(csr, 23)] = FMA(KP903989293, T76, T75); | |
1277 Cr[WS(csr, 41)] = FNMS(KP903989293, T76, T75); | |
1278 Cr[WS(csr, 7)] = FMA(KP941544065, T7g, T79); | |
1279 Cr[WS(csr, 57)] = FNMS(KP941544065, T7g, T79); | |
1280 T7o = T7m - T7n; | |
1281 T7q = T7m + T7n; | |
1282 Ci[WS(csi, 57)] = FMS(KP941544065, T7o, T7l); | |
1283 Ci[WS(csi, 7)] = FMA(KP941544065, T7o, T7l); | |
1284 Ci[WS(csi, 39)] = FMA(KP941544065, T7k, T7j); | |
1285 Ci[WS(csi, 25)] = FMS(KP941544065, T7k, T7j); | |
1286 } | |
1287 } | |
1288 } | |
1289 } | |
1290 { | |
1291 E T7t, T8A, T8x, T7A, T8r, T8k, T88, T81, Ta3, Ta4, T6r, T6s; | |
1292 { | |
1293 E T9L, T99, T9W, T9g, T9M, T9C, T9V, T9z, T9k, T9O, T9T, Ta0, T9H, T9v, T9m; | |
1294 { | |
1295 E T9B, T9c, T9f, T9A, T97, T98; | |
1296 T7t = FMA(KP923879532, T7s, T7r); | |
1297 T97 = FNMS(KP923879532, T7s, T7r); | |
1298 T98 = T8z - T8y; | |
1299 T8A = T8y + T8z; | |
1300 T9B = FNMS(KP534511135, T9a, T9b); | |
1301 T9c = FMA(KP534511135, T9b, T9a); | |
1302 Cr[WS(csr, 25)] = FNMS(KP941544065, T7q, T7p); | |
1303 Cr[WS(csr, 39)] = FMA(KP941544065, T7q, T7p); | |
1304 T9L = FMA(KP831469612, T98, T97); | |
1305 T99 = FNMS(KP831469612, T98, T97); | |
1306 T9f = FNMS(KP534511135, T9e, T9d); | |
1307 T9A = FMA(KP534511135, T9d, T9e); | |
1308 { | |
1309 E T9x, T9y, T9q, T9t; | |
1310 T8x = FMA(KP923879532, T8w, T8v); | |
1311 T9x = FNMS(KP923879532, T8w, T8v); | |
1312 T9W = T9c + T9f; | |
1313 T9g = T9c - T9f; | |
1314 T9M = T9B + T9A; | |
1315 T9C = T9A - T9B; | |
1316 T9y = T7z - T7w; | |
1317 T7A = T7w + T7z; | |
1318 T8r = T8p + T8q; | |
1319 T9q = T8p - T8q; | |
1320 T9t = T8j - T8g; | |
1321 T8k = T8g + T8j; | |
1322 { | |
1323 E T9R, T9r, T9S, T9u, T9j; | |
1324 T88 = T86 + T87; | |
1325 T9j = T87 - T86; | |
1326 T9V = FNMS(KP831469612, T9y, T9x); | |
1327 T9z = FMA(KP831469612, T9y, T9x); | |
1328 T9R = FMA(KP831469612, T9q, T9p); | |
1329 T9r = FNMS(KP831469612, T9q, T9p); | |
1330 T9S = FMA(KP831469612, T9t, T9s); | |
1331 T9u = FNMS(KP831469612, T9t, T9s); | |
1332 T9k = FNMS(KP831469612, T9j, T9i); | |
1333 T9O = FMA(KP831469612, T9j, T9i); | |
1334 T9T = FNMS(KP250486960, T9S, T9R); | |
1335 Ta0 = FMA(KP250486960, T9R, T9S); | |
1336 T9H = FNMS(KP599376933, T9r, T9u); | |
1337 T9v = FMA(KP599376933, T9u, T9r); | |
1338 T9m = T7X - T80; | |
1339 T81 = T7X + T80; | |
1340 } | |
1341 } | |
1342 } | |
1343 { | |
1344 E T9J, T9h, T9F, T9D, T9P, T9n; | |
1345 T9J = FNMS(KP881921264, T9g, T99); | |
1346 T9h = FMA(KP881921264, T9g, T99); | |
1347 T9F = FMA(KP881921264, T9C, T9z); | |
1348 T9D = FNMS(KP881921264, T9C, T9z); | |
1349 T9P = FMA(KP831469612, T9m, T9l); | |
1350 T9n = FNMS(KP831469612, T9m, T9l); | |
1351 { | |
1352 E T9Y, T9X, T9Z, Ta2; | |
1353 { | |
1354 E T9N, Ta1, T9G, T9o, T9U, T9Q; | |
1355 Ta3 = FNMS(KP881921264, T9M, T9L); | |
1356 T9N = FMA(KP881921264, T9M, T9L); | |
1357 T9Q = FNMS(KP250486960, T9P, T9O); | |
1358 Ta1 = FMA(KP250486960, T9O, T9P); | |
1359 T9G = FNMS(KP599376933, T9k, T9n); | |
1360 T9o = FMA(KP599376933, T9n, T9k); | |
1361 T9U = T9Q + T9T; | |
1362 T9Y = T9T - T9Q; | |
1363 T9X = FNMS(KP881921264, T9W, T9V); | |
1364 T9Z = FMA(KP881921264, T9W, T9V); | |
1365 { | |
1366 E T9K, T9I, T9E, T9w; | |
1367 T9K = T9G + T9H; | |
1368 T9I = T9G - T9H; | |
1369 T9E = T9v - T9o; | |
1370 T9w = T9o + T9v; | |
1371 Cr[WS(csr, 5)] = FMA(KP970031253, T9U, T9N); | |
1372 Cr[WS(csr, 59)] = FNMS(KP970031253, T9U, T9N); | |
1373 Cr[WS(csr, 21)] = FNMS(KP857728610, T9K, T9J); | |
1374 Cr[WS(csr, 43)] = FMA(KP857728610, T9K, T9J); | |
1375 Ci[WS(csi, 53)] = FMS(KP857728610, T9I, T9F); | |
1376 Ci[WS(csi, 11)] = FMA(KP857728610, T9I, T9F); | |
1377 Ci[WS(csi, 43)] = FMA(KP857728610, T9E, T9D); | |
1378 Ci[WS(csi, 21)] = FMS(KP857728610, T9E, T9D); | |
1379 Cr[WS(csr, 11)] = FMA(KP857728610, T9w, T9h); | |
1380 Cr[WS(csr, 53)] = FNMS(KP857728610, T9w, T9h); | |
1381 Ta2 = Ta0 - Ta1; | |
1382 Ta4 = Ta1 + Ta0; | |
1383 } | |
1384 } | |
1385 Ci[WS(csi, 59)] = FMA(KP970031253, Ta2, T9Z); | |
1386 Ci[WS(csi, 5)] = FMS(KP970031253, Ta2, T9Z); | |
1387 Ci[WS(csi, 37)] = FMS(KP970031253, T9Y, T9X); | |
1388 Ci[WS(csi, 27)] = FMA(KP970031253, T9Y, T9X); | |
1389 } | |
1390 } | |
1391 } | |
1392 { | |
1393 E T69, T2z, T6k, T3g, T6a, T60, T6j, T5X, T4i, T6c, T6h, T6p, T64, T5L; | |
1394 { | |
1395 E T5Y, T2U, T3f, T5Z; | |
1396 T5Y = FMA(KP098491403, T2M, T2T); | |
1397 T2U = FNMS(KP098491403, T2T, T2M); | |
1398 Cr[WS(csr, 27)] = FMA(KP970031253, Ta4, Ta3); | |
1399 Cr[WS(csr, 37)] = FNMS(KP970031253, Ta4, Ta3); | |
1400 T69 = FNMS(KP980785280, T2y, T2f); | |
1401 T2z = FMA(KP980785280, T2y, T2f); | |
1402 T3f = FMA(KP098491403, T3e, T37); | |
1403 T5Z = FNMS(KP098491403, T37, T3e); | |
1404 T6k = T3f - T2U; | |
1405 T3g = T2U + T3f; | |
1406 T6a = T5Y - T5Z; | |
1407 T60 = T5Y + T5Z; | |
1408 { | |
1409 E T6f, T5x, T6g, T5K; | |
1410 T6j = FNMS(KP980785280, T5W, T5T); | |
1411 T5X = FMA(KP980785280, T5W, T5T); | |
1412 T6f = FNMS(KP980785280, T5w, T4X); | |
1413 T5x = FMA(KP980785280, T5w, T4X); | |
1414 T6g = FNMS(KP980785280, T5J, T5G); | |
1415 T5K = FMA(KP980785280, T5J, T5G); | |
1416 T4i = FMA(KP980785280, T4h, T3I); | |
1417 T6c = FNMS(KP980785280, T4h, T3I); | |
1418 T6h = FMA(KP906347169, T6g, T6f); | |
1419 T6p = FNMS(KP906347169, T6f, T6g); | |
1420 T64 = FMA(KP049126849, T5x, T5K); | |
1421 T5L = FNMS(KP049126849, T5K, T5x); | |
1422 } | |
1423 } | |
1424 { | |
1425 E T67, T3h, T63, T61, T6d, T4v; | |
1426 T67 = FNMS(KP995184726, T3g, T2z); | |
1427 T3h = FMA(KP995184726, T3g, T2z); | |
1428 T63 = FMA(KP995184726, T60, T5X); | |
1429 T61 = FNMS(KP995184726, T60, T5X); | |
1430 T6d = FNMS(KP980785280, T4u, T4r); | |
1431 T4v = FMA(KP980785280, T4u, T4r); | |
1432 { | |
1433 E T6m, T6l, T6n, T6q; | |
1434 { | |
1435 E T6b, T6o, T65, T4w, T6i, T6e; | |
1436 T6r = FNMS(KP995184726, T6a, T69); | |
1437 T6b = FMA(KP995184726, T6a, T69); | |
1438 T6e = FMA(KP906347169, T6d, T6c); | |
1439 T6o = FNMS(KP906347169, T6c, T6d); | |
1440 T65 = FMA(KP049126849, T4i, T4v); | |
1441 T4w = FNMS(KP049126849, T4v, T4i); | |
1442 T6i = T6e + T6h; | |
1443 T6m = T6h - T6e; | |
1444 T6l = FNMS(KP995184726, T6k, T6j); | |
1445 T6n = FMA(KP995184726, T6k, T6j); | |
1446 { | |
1447 E T68, T66, T62, T5M; | |
1448 T68 = T65 + T64; | |
1449 T66 = T64 - T65; | |
1450 T62 = T5L - T4w; | |
1451 T5M = T4w + T5L; | |
1452 Cr[WS(csr, 15)] = FMA(KP740951125, T6i, T6b); | |
1453 Cr[WS(csr, 49)] = FNMS(KP740951125, T6i, T6b); | |
1454 Cr[WS(csr, 31)] = FMA(KP998795456, T68, T67); | |
1455 Cr[WS(csr, 33)] = FNMS(KP998795456, T68, T67); | |
1456 Ci[WS(csi, 63)] = FMA(KP998795456, T66, T63); | |
1457 Ci[WS(csi, 1)] = FMS(KP998795456, T66, T63); | |
1458 Ci[WS(csi, 33)] = FMS(KP998795456, T62, T61); | |
1459 Ci[WS(csi, 31)] = FMA(KP998795456, T62, T61); | |
1460 Cr[WS(csr, 1)] = FMA(KP998795456, T5M, T3h); | |
1461 Cr[WS(csr, 63)] = FNMS(KP998795456, T5M, T3h); | |
1462 T6q = T6o - T6p; | |
1463 T6s = T6o + T6p; | |
1464 } | |
1465 } | |
1466 Ci[WS(csi, 49)] = FMS(KP740951125, T6q, T6n); | |
1467 Ci[WS(csi, 15)] = FMA(KP740951125, T6q, T6n); | |
1468 Ci[WS(csi, 47)] = FMA(KP740951125, T6m, T6l); | |
1469 Ci[WS(csi, 17)] = FMS(KP740951125, T6m, T6l); | |
1470 } | |
1471 } | |
1472 } | |
1473 { | |
1474 E T8N, T7B, T8Y, T7Q, T8O, T8E, T8X, T8B, T82, T8Q, T8V, T92, T8J, T8t; | |
1475 { | |
1476 E T8C, T7I, T7P, T8D; | |
1477 T8C = FNMS(KP303346683, T7E, T7H); | |
1478 T7I = FMA(KP303346683, T7H, T7E); | |
1479 Cr[WS(csr, 17)] = FNMS(KP740951125, T6s, T6r); | |
1480 Cr[WS(csr, 47)] = FMA(KP740951125, T6s, T6r); | |
1481 T8N = FNMS(KP831469612, T7A, T7t); | |
1482 T7B = FMA(KP831469612, T7A, T7t); | |
1483 T7P = FNMS(KP303346683, T7O, T7L); | |
1484 T8D = FMA(KP303346683, T7L, T7O); | |
1485 T8Y = T7P - T7I; | |
1486 T7Q = T7I + T7P; | |
1487 T8O = T8D - T8C; | |
1488 T8E = T8C + T8D; | |
1489 { | |
1490 E T8T, T8l, T8U, T8s; | |
1491 T8X = FNMS(KP831469612, T8A, T8x); | |
1492 T8B = FMA(KP831469612, T8A, T8x); | |
1493 T8T = FNMS(KP831469612, T8k, T8d); | |
1494 T8l = FMA(KP831469612, T8k, T8d); | |
1495 T8U = FNMS(KP831469612, T8r, T8o); | |
1496 T8s = FMA(KP831469612, T8r, T8o); | |
1497 T82 = FMA(KP831469612, T81, T7U); | |
1498 T8Q = FNMS(KP831469612, T81, T7U); | |
1499 T8V = FNMS(KP741650546, T8U, T8T); | |
1500 T92 = FMA(KP741650546, T8T, T8U); | |
1501 T8J = FNMS(KP148335987, T8l, T8s); | |
1502 T8t = FMA(KP148335987, T8s, T8l); | |
1503 } | |
1504 } | |
1505 { | |
1506 E T8L, T7R, T8H, T8F, T8R, T89; | |
1507 T8L = FNMS(KP956940335, T7Q, T7B); | |
1508 T7R = FMA(KP956940335, T7Q, T7B); | |
1509 T8H = FMA(KP956940335, T8E, T8B); | |
1510 T8F = FNMS(KP956940335, T8E, T8B); | |
1511 T8R = FNMS(KP831469612, T88, T85); | |
1512 T89 = FMA(KP831469612, T88, T85); | |
1513 { | |
1514 E T90, T8Z, T91, T94; | |
1515 { | |
1516 E T8P, T93, T8I, T8a, T8W, T8S; | |
1517 T95 = FNMS(KP956940335, T8O, T8N); | |
1518 T8P = FMA(KP956940335, T8O, T8N); | |
1519 T8S = FNMS(KP741650546, T8R, T8Q); | |
1520 T93 = FMA(KP741650546, T8Q, T8R); | |
1521 T8I = FNMS(KP148335987, T82, T89); | |
1522 T8a = FMA(KP148335987, T89, T82); | |
1523 T8W = T8S + T8V; | |
1524 T90 = T8V - T8S; | |
1525 T8Z = FMA(KP956940335, T8Y, T8X); | |
1526 T91 = FNMS(KP956940335, T8Y, T8X); | |
1527 { | |
1528 E T8M, T8K, T8G, T8u; | |
1529 T8M = T8I + T8J; | |
1530 T8K = T8I - T8J; | |
1531 T8G = T8t - T8a; | |
1532 T8u = T8a + T8t; | |
1533 Cr[WS(csr, 13)] = FMA(KP803207531, T8W, T8P); | |
1534 Cr[WS(csr, 51)] = FNMS(KP803207531, T8W, T8P); | |
1535 Cr[WS(csr, 29)] = FNMS(KP989176509, T8M, T8L); | |
1536 Cr[WS(csr, 35)] = FMA(KP989176509, T8M, T8L); | |
1537 Ci[WS(csi, 61)] = FMS(KP989176509, T8K, T8H); | |
1538 Ci[WS(csi, 3)] = FMA(KP989176509, T8K, T8H); | |
1539 Ci[WS(csi, 35)] = FMA(KP989176509, T8G, T8F); | |
1540 Ci[WS(csi, 29)] = FMS(KP989176509, T8G, T8F); | |
1541 Cr[WS(csr, 3)] = FMA(KP989176509, T8u, T7R); | |
1542 Cr[WS(csr, 61)] = FNMS(KP989176509, T8u, T7R); | |
1543 T94 = T92 - T93; | |
1544 T96 = T93 + T92; | |
1545 } | |
1546 } | |
1547 Ci[WS(csi, 51)] = FMA(KP803207531, T94, T91); | |
1548 Ci[WS(csi, 13)] = FMS(KP803207531, T94, T91); | |
1549 Ci[WS(csi, 45)] = FMS(KP803207531, T90, T8Z); | |
1550 Ci[WS(csi, 19)] = FMA(KP803207531, T90, T8Z); | |
1551 } | |
1552 } | |
1553 } | |
1554 } | |
1555 } | |
1556 } | |
1557 } | |
1558 Cr[WS(csr, 19)] = FMA(KP803207531, T96, T95); | |
1559 Cr[WS(csr, 45)] = FNMS(KP803207531, T96, T95); | |
1560 } | |
1561 } | |
1562 } | |
1563 | |
1564 static const kr2c_desc desc = { 128, "r2cf_128", {440, 0, 516, 0}, &GENUS }; | |
1565 | |
1566 void X(codelet_r2cf_128) (planner *p) { | |
1567 X(kr2c_register) (p, r2cf_128, &desc); | |
1568 } | |
1569 | |
1570 #else /* HAVE_FMA */ | |
1571 | |
1572 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 128 -name r2cf_128 -include r2cf.h */ | |
1573 | |
1574 /* | |
1575 * This function contains 956 FP additions, 330 FP multiplications, | |
1576 * (or, 812 additions, 186 multiplications, 144 fused multiply/add), | |
1577 * 186 stack variables, 31 constants, and 256 memory accesses | |
1578 */ | |
1579 #include "r2cf.h" | |
1580 | |
1581 static void r2cf_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
1582 { | |
1583 DK(KP803207531, +0.803207531480644909806676512963141923879569427); | |
1584 DK(KP595699304, +0.595699304492433343467036528829969889511926338); | |
1585 DK(KP146730474, +0.146730474455361751658850129646717819706215317); | |
1586 DK(KP989176509, +0.989176509964780973451673738016243063983689533); | |
1587 DK(KP740951125, +0.740951125354959091175616897495162729728955309); | |
1588 DK(KP671558954, +0.671558954847018400625376850427421803228750632); | |
1589 DK(KP049067674, +0.049067674327418014254954976942682658314745363); | |
1590 DK(KP998795456, +0.998795456205172392714771604759100694443203615); | |
1591 DK(KP242980179, +0.242980179903263889948274162077471118320990783); | |
1592 DK(KP970031253, +0.970031253194543992603984207286100251456865962); | |
1593 DK(KP514102744, +0.514102744193221726593693838968815772608049120); | |
1594 DK(KP857728610, +0.857728610000272069902269984284770137042490799); | |
1595 DK(KP336889853, +0.336889853392220050689253212619147570477766780); | |
1596 DK(KP941544065, +0.941544065183020778412509402599502357185589796); | |
1597 DK(KP427555093, +0.427555093430282094320966856888798534304578629); | |
1598 DK(KP903989293, +0.903989293123443331586200297230537048710132025); | |
1599 DK(KP098017140, +0.098017140329560601994195563888641845861136673); | |
1600 DK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
1601 DK(KP634393284, +0.634393284163645498215171613225493370675687095); | |
1602 DK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
1603 DK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
1604 DK(KP471396736, +0.471396736825997648556387625905254377657460319); | |
1605 DK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
1606 DK(KP290284677, +0.290284677254462367636192375817395274691476278); | |
1607 DK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
1608 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
1609 DK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
1610 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
1611 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
1612 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
1613 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
1614 { | |
1615 INT i; | |
1616 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) { | |
1617 E TcD, TdU, T27, T7r, T5S, T8y, Tf, Ta5, Tu, Tbq, TcG, TdV, T2e, T8z, T5V; | |
1618 E T7s, TK, Ta6, TcK, TdX, T2o, T5X, T7w, T8B, TZ, Ta7, TcN, TdY, T2x, T5Y; | |
1619 E T7z, T8C, T1g, Taa, TcU, TeA, TcX, Tez, T1v, Tab, T2M, T6z, T7E, T9e, T7H; | |
1620 E T9d, T2T, T6A, T4X, T6L, Tdz, TeL, TdK, TeP, T5G, T6P, T8d, T9p, TaV, Tc3; | |
1621 E Tbi, Tc4, T8o, T9t, T3I, T6H, Tde, TeH, Tdp, TeF, T4r, T6F, T7U, T9l, Tao; | |
1622 E TbW, TaL, TbX, T85, T9j, T1L, Tad, Td3, Tew, Td6, Tex, T20, Tae, T37, T6x; | |
1623 E T7L, T9a, T7O, T9b, T3e, T6w, TbZ, Tc0, T3Z, T4s, Tds, TeI, T4g, T4t, T80; | |
1624 E T87, Tdl, TeE, T7X, T86, TaD, TaM, Tc6, Tc7, T5e, T5H, TdN, TeM, T5v, T5I; | |
1625 E T8j, T8q, TdG, TeO, T8g, T8p, Tba, Tbj; | |
1626 { | |
1627 E T3, T23, Td, T25, T6, T5R, Ta, T24; | |
1628 { | |
1629 E T1, T2, Tb, Tc; | |
1630 T1 = R0[0]; | |
1631 T2 = R0[WS(rs, 32)]; | |
1632 T3 = T1 + T2; | |
1633 T23 = T1 - T2; | |
1634 Tb = R0[WS(rs, 56)]; | |
1635 Tc = R0[WS(rs, 24)]; | |
1636 Td = Tb + Tc; | |
1637 T25 = Tb - Tc; | |
1638 } | |
1639 { | |
1640 E T4, T5, T8, T9; | |
1641 T4 = R0[WS(rs, 16)]; | |
1642 T5 = R0[WS(rs, 48)]; | |
1643 T6 = T4 + T5; | |
1644 T5R = T4 - T5; | |
1645 T8 = R0[WS(rs, 8)]; | |
1646 T9 = R0[WS(rs, 40)]; | |
1647 Ta = T8 + T9; | |
1648 T24 = T8 - T9; | |
1649 } | |
1650 TcD = T3 - T6; | |
1651 TdU = Td - Ta; | |
1652 { | |
1653 E T26, T5Q, T7, Te; | |
1654 T26 = KP707106781 * (T24 + T25); | |
1655 T27 = T23 + T26; | |
1656 T7r = T23 - T26; | |
1657 T5Q = KP707106781 * (T25 - T24); | |
1658 T5S = T5Q - T5R; | |
1659 T8y = T5R + T5Q; | |
1660 T7 = T3 + T6; | |
1661 Te = Ta + Td; | |
1662 Tf = T7 + Te; | |
1663 Ta5 = T7 - Te; | |
1664 } | |
1665 } | |
1666 { | |
1667 E Ti, T28, Ts, T2c, Tl, T29, Tp, T2b; | |
1668 { | |
1669 E Tg, Th, Tq, Tr; | |
1670 Tg = R0[WS(rs, 4)]; | |
1671 Th = R0[WS(rs, 36)]; | |
1672 Ti = Tg + Th; | |
1673 T28 = Tg - Th; | |
1674 Tq = R0[WS(rs, 12)]; | |
1675 Tr = R0[WS(rs, 44)]; | |
1676 Ts = Tq + Tr; | |
1677 T2c = Tq - Tr; | |
1678 } | |
1679 { | |
1680 E Tj, Tk, Tn, To; | |
1681 Tj = R0[WS(rs, 20)]; | |
1682 Tk = R0[WS(rs, 52)]; | |
1683 Tl = Tj + Tk; | |
1684 T29 = Tj - Tk; | |
1685 Tn = R0[WS(rs, 60)]; | |
1686 To = R0[WS(rs, 28)]; | |
1687 Tp = Tn + To; | |
1688 T2b = Tn - To; | |
1689 } | |
1690 { | |
1691 E Tm, Tt, TcE, TcF; | |
1692 Tm = Ti + Tl; | |
1693 Tt = Tp + Ts; | |
1694 Tu = Tm + Tt; | |
1695 Tbq = Tt - Tm; | |
1696 TcE = Ti - Tl; | |
1697 TcF = Tp - Ts; | |
1698 TcG = KP707106781 * (TcE + TcF); | |
1699 TdV = KP707106781 * (TcF - TcE); | |
1700 } | |
1701 { | |
1702 E T2a, T2d, T5T, T5U; | |
1703 T2a = FNMS(KP382683432, T29, KP923879532 * T28); | |
1704 T2d = FMA(KP923879532, T2b, KP382683432 * T2c); | |
1705 T2e = T2a + T2d; | |
1706 T8z = T2d - T2a; | |
1707 T5T = FNMS(KP923879532, T2c, KP382683432 * T2b); | |
1708 T5U = FMA(KP382683432, T28, KP923879532 * T29); | |
1709 T5V = T5T - T5U; | |
1710 T7s = T5U + T5T; | |
1711 } | |
1712 } | |
1713 { | |
1714 E Ty, T2g, TB, T2m, TF, T2l, TI, T2j; | |
1715 { | |
1716 E Tw, Tx, Tz, TA; | |
1717 Tw = R0[WS(rs, 2)]; | |
1718 Tx = R0[WS(rs, 34)]; | |
1719 Ty = Tw + Tx; | |
1720 T2g = Tw - Tx; | |
1721 Tz = R0[WS(rs, 18)]; | |
1722 TA = R0[WS(rs, 50)]; | |
1723 TB = Tz + TA; | |
1724 T2m = Tz - TA; | |
1725 { | |
1726 E TD, TE, T2h, TG, TH, T2i; | |
1727 TD = R0[WS(rs, 10)]; | |
1728 TE = R0[WS(rs, 42)]; | |
1729 T2h = TD - TE; | |
1730 TG = R0[WS(rs, 58)]; | |
1731 TH = R0[WS(rs, 26)]; | |
1732 T2i = TG - TH; | |
1733 TF = TD + TE; | |
1734 T2l = KP707106781 * (T2i - T2h); | |
1735 TI = TG + TH; | |
1736 T2j = KP707106781 * (T2h + T2i); | |
1737 } | |
1738 } | |
1739 { | |
1740 E TC, TJ, TcI, TcJ; | |
1741 TC = Ty + TB; | |
1742 TJ = TF + TI; | |
1743 TK = TC + TJ; | |
1744 Ta6 = TC - TJ; | |
1745 TcI = Ty - TB; | |
1746 TcJ = TI - TF; | |
1747 TcK = FMA(KP923879532, TcI, KP382683432 * TcJ); | |
1748 TdX = FNMS(KP382683432, TcI, KP923879532 * TcJ); | |
1749 } | |
1750 { | |
1751 E T2k, T2n, T7u, T7v; | |
1752 T2k = T2g + T2j; | |
1753 T2n = T2l - T2m; | |
1754 T2o = FMA(KP980785280, T2k, KP195090322 * T2n); | |
1755 T5X = FNMS(KP195090322, T2k, KP980785280 * T2n); | |
1756 T7u = T2g - T2j; | |
1757 T7v = T2m + T2l; | |
1758 T7w = FMA(KP831469612, T7u, KP555570233 * T7v); | |
1759 T8B = FNMS(KP555570233, T7u, KP831469612 * T7v); | |
1760 } | |
1761 } | |
1762 { | |
1763 E TN, T2p, TQ, T2v, TU, T2u, TX, T2s; | |
1764 { | |
1765 E TL, TM, TO, TP; | |
1766 TL = R0[WS(rs, 62)]; | |
1767 TM = R0[WS(rs, 30)]; | |
1768 TN = TL + TM; | |
1769 T2p = TL - TM; | |
1770 TO = R0[WS(rs, 14)]; | |
1771 TP = R0[WS(rs, 46)]; | |
1772 TQ = TO + TP; | |
1773 T2v = TO - TP; | |
1774 { | |
1775 E TS, TT, T2q, TV, TW, T2r; | |
1776 TS = R0[WS(rs, 6)]; | |
1777 TT = R0[WS(rs, 38)]; | |
1778 T2q = TS - TT; | |
1779 TV = R0[WS(rs, 54)]; | |
1780 TW = R0[WS(rs, 22)]; | |
1781 T2r = TV - TW; | |
1782 TU = TS + TT; | |
1783 T2u = KP707106781 * (T2r - T2q); | |
1784 TX = TV + TW; | |
1785 T2s = KP707106781 * (T2q + T2r); | |
1786 } | |
1787 } | |
1788 { | |
1789 E TR, TY, TcL, TcM; | |
1790 TR = TN + TQ; | |
1791 TY = TU + TX; | |
1792 TZ = TR + TY; | |
1793 Ta7 = TR - TY; | |
1794 TcL = TN - TQ; | |
1795 TcM = TX - TU; | |
1796 TcN = FNMS(KP382683432, TcM, KP923879532 * TcL); | |
1797 TdY = FMA(KP382683432, TcL, KP923879532 * TcM); | |
1798 } | |
1799 { | |
1800 E T2t, T2w, T7x, T7y; | |
1801 T2t = T2p + T2s; | |
1802 T2w = T2u - T2v; | |
1803 T2x = FNMS(KP195090322, T2w, KP980785280 * T2t); | |
1804 T5Y = FMA(KP195090322, T2t, KP980785280 * T2w); | |
1805 T7x = T2p - T2s; | |
1806 T7y = T2v + T2u; | |
1807 T7z = FNMS(KP555570233, T7y, KP831469612 * T7x); | |
1808 T8C = FMA(KP555570233, T7x, KP831469612 * T7y); | |
1809 } | |
1810 } | |
1811 { | |
1812 E T14, T2N, T17, T2D, T1b, T2O, T1e, T2C, T1j, T1m, T2K, TcR, T2Q, T1q, T1t; | |
1813 E T2H, TcS, T2R; | |
1814 { | |
1815 E T12, T13, T15, T16; | |
1816 T12 = R0[WS(rs, 1)]; | |
1817 T13 = R0[WS(rs, 33)]; | |
1818 T14 = T12 + T13; | |
1819 T2N = T12 - T13; | |
1820 T15 = R0[WS(rs, 17)]; | |
1821 T16 = R0[WS(rs, 49)]; | |
1822 T17 = T15 + T16; | |
1823 T2D = T15 - T16; | |
1824 } | |
1825 { | |
1826 E T19, T1a, T2B, T1c, T1d, T2A; | |
1827 T19 = R0[WS(rs, 9)]; | |
1828 T1a = R0[WS(rs, 41)]; | |
1829 T2B = T19 - T1a; | |
1830 T1c = R0[WS(rs, 57)]; | |
1831 T1d = R0[WS(rs, 25)]; | |
1832 T2A = T1c - T1d; | |
1833 T1b = T19 + T1a; | |
1834 T2O = KP707106781 * (T2B + T2A); | |
1835 T1e = T1c + T1d; | |
1836 T2C = KP707106781 * (T2A - T2B); | |
1837 } | |
1838 { | |
1839 E T2I, T2J, T2F, T2G; | |
1840 { | |
1841 E T1h, T1i, T1k, T1l; | |
1842 T1h = R0[WS(rs, 5)]; | |
1843 T1i = R0[WS(rs, 37)]; | |
1844 T1j = T1h + T1i; | |
1845 T2I = T1h - T1i; | |
1846 T1k = R0[WS(rs, 21)]; | |
1847 T1l = R0[WS(rs, 53)]; | |
1848 T1m = T1k + T1l; | |
1849 T2J = T1k - T1l; | |
1850 } | |
1851 T2K = FMA(KP382683432, T2I, KP923879532 * T2J); | |
1852 TcR = T1j - T1m; | |
1853 T2Q = FNMS(KP382683432, T2J, KP923879532 * T2I); | |
1854 { | |
1855 E T1o, T1p, T1r, T1s; | |
1856 T1o = R0[WS(rs, 61)]; | |
1857 T1p = R0[WS(rs, 29)]; | |
1858 T1q = T1o + T1p; | |
1859 T2F = T1o - T1p; | |
1860 T1r = R0[WS(rs, 13)]; | |
1861 T1s = R0[WS(rs, 45)]; | |
1862 T1t = T1r + T1s; | |
1863 T2G = T1r - T1s; | |
1864 } | |
1865 T2H = FNMS(KP923879532, T2G, KP382683432 * T2F); | |
1866 TcS = T1q - T1t; | |
1867 T2R = FMA(KP923879532, T2F, KP382683432 * T2G); | |
1868 } | |
1869 { | |
1870 E T18, T1f, TcQ, TcT; | |
1871 T18 = T14 + T17; | |
1872 T1f = T1b + T1e; | |
1873 T1g = T18 + T1f; | |
1874 Taa = T18 - T1f; | |
1875 TcQ = T14 - T17; | |
1876 TcT = KP707106781 * (TcR + TcS); | |
1877 TcU = TcQ + TcT; | |
1878 TeA = TcQ - TcT; | |
1879 } | |
1880 { | |
1881 E TcV, TcW, T1n, T1u; | |
1882 TcV = T1e - T1b; | |
1883 TcW = KP707106781 * (TcS - TcR); | |
1884 TcX = TcV + TcW; | |
1885 Tez = TcW - TcV; | |
1886 T1n = T1j + T1m; | |
1887 T1u = T1q + T1t; | |
1888 T1v = T1n + T1u; | |
1889 Tab = T1u - T1n; | |
1890 } | |
1891 { | |
1892 E T2E, T2L, T7C, T7D; | |
1893 T2E = T2C - T2D; | |
1894 T2L = T2H - T2K; | |
1895 T2M = T2E + T2L; | |
1896 T6z = T2L - T2E; | |
1897 T7C = T2N - T2O; | |
1898 T7D = T2K + T2H; | |
1899 T7E = T7C + T7D; | |
1900 T9e = T7C - T7D; | |
1901 } | |
1902 { | |
1903 E T7F, T7G, T2P, T2S; | |
1904 T7F = T2D + T2C; | |
1905 T7G = T2R - T2Q; | |
1906 T7H = T7F + T7G; | |
1907 T9d = T7G - T7F; | |
1908 T2P = T2N + T2O; | |
1909 T2S = T2Q + T2R; | |
1910 T2T = T2P + T2S; | |
1911 T6A = T2P - T2S; | |
1912 } | |
1913 } | |
1914 { | |
1915 E T4z, TaP, T5B, TaQ, T4G, TaT, T5y, TaS, Tbf, Tbg, T4O, Tdw, T5E, Tbc, Tbd; | |
1916 E T4V, Tdx, T5D; | |
1917 { | |
1918 E T4x, T4y, T5z, T5A; | |
1919 T4x = R1[WS(rs, 63)]; | |
1920 T4y = R1[WS(rs, 31)]; | |
1921 T4z = T4x - T4y; | |
1922 TaP = T4x + T4y; | |
1923 T5z = R1[WS(rs, 15)]; | |
1924 T5A = R1[WS(rs, 47)]; | |
1925 T5B = T5z - T5A; | |
1926 TaQ = T5z + T5A; | |
1927 } | |
1928 { | |
1929 E T4A, T4B, T4C, T4D, T4E, T4F; | |
1930 T4A = R1[WS(rs, 7)]; | |
1931 T4B = R1[WS(rs, 39)]; | |
1932 T4C = T4A - T4B; | |
1933 T4D = R1[WS(rs, 55)]; | |
1934 T4E = R1[WS(rs, 23)]; | |
1935 T4F = T4D - T4E; | |
1936 T4G = KP707106781 * (T4C + T4F); | |
1937 TaT = T4D + T4E; | |
1938 T5y = KP707106781 * (T4F - T4C); | |
1939 TaS = T4A + T4B; | |
1940 } | |
1941 { | |
1942 E T4K, T4N, T4R, T4U; | |
1943 { | |
1944 E T4I, T4J, T4L, T4M; | |
1945 T4I = R1[WS(rs, 3)]; | |
1946 T4J = R1[WS(rs, 35)]; | |
1947 T4K = T4I - T4J; | |
1948 Tbf = T4I + T4J; | |
1949 T4L = R1[WS(rs, 19)]; | |
1950 T4M = R1[WS(rs, 51)]; | |
1951 T4N = T4L - T4M; | |
1952 Tbg = T4L + T4M; | |
1953 } | |
1954 T4O = FNMS(KP382683432, T4N, KP923879532 * T4K); | |
1955 Tdw = Tbf - Tbg; | |
1956 T5E = FMA(KP382683432, T4K, KP923879532 * T4N); | |
1957 { | |
1958 E T4P, T4Q, T4S, T4T; | |
1959 T4P = R1[WS(rs, 59)]; | |
1960 T4Q = R1[WS(rs, 27)]; | |
1961 T4R = T4P - T4Q; | |
1962 Tbc = T4P + T4Q; | |
1963 T4S = R1[WS(rs, 11)]; | |
1964 T4T = R1[WS(rs, 43)]; | |
1965 T4U = T4S - T4T; | |
1966 Tbd = T4S + T4T; | |
1967 } | |
1968 T4V = FMA(KP923879532, T4R, KP382683432 * T4U); | |
1969 Tdx = Tbc - Tbd; | |
1970 T5D = FNMS(KP923879532, T4U, KP382683432 * T4R); | |
1971 } | |
1972 { | |
1973 E T4H, T4W, Tdv, Tdy; | |
1974 T4H = T4z + T4G; | |
1975 T4W = T4O + T4V; | |
1976 T4X = T4H + T4W; | |
1977 T6L = T4H - T4W; | |
1978 Tdv = TaP - TaQ; | |
1979 Tdy = KP707106781 * (Tdw + Tdx); | |
1980 Tdz = Tdv + Tdy; | |
1981 TeL = Tdv - Tdy; | |
1982 } | |
1983 { | |
1984 E TdI, TdJ, T5C, T5F; | |
1985 TdI = TaT - TaS; | |
1986 TdJ = KP707106781 * (Tdx - Tdw); | |
1987 TdK = TdI + TdJ; | |
1988 TeP = TdJ - TdI; | |
1989 T5C = T5y - T5B; | |
1990 T5F = T5D - T5E; | |
1991 T5G = T5C + T5F; | |
1992 T6P = T5F - T5C; | |
1993 } | |
1994 { | |
1995 E T8b, T8c, TaR, TaU; | |
1996 T8b = T4z - T4G; | |
1997 T8c = T5E + T5D; | |
1998 T8d = T8b + T8c; | |
1999 T9p = T8b - T8c; | |
2000 TaR = TaP + TaQ; | |
2001 TaU = TaS + TaT; | |
2002 TaV = TaR - TaU; | |
2003 Tc3 = TaR + TaU; | |
2004 } | |
2005 { | |
2006 E Tbe, Tbh, T8m, T8n; | |
2007 Tbe = Tbc + Tbd; | |
2008 Tbh = Tbf + Tbg; | |
2009 Tbi = Tbe - Tbh; | |
2010 Tc4 = Tbh + Tbe; | |
2011 T8m = T5B + T5y; | |
2012 T8n = T4V - T4O; | |
2013 T8o = T8m + T8n; | |
2014 T9t = T8n - T8m; | |
2015 } | |
2016 } | |
2017 { | |
2018 E T3k, Tai, T4m, Taj, T3r, Tam, T4j, Tal, TaI, TaJ, T3z, Tdb, T4p, TaF, TaG; | |
2019 E T3G, Tdc, T4o; | |
2020 { | |
2021 E T3i, T3j, T4k, T4l; | |
2022 T3i = R1[0]; | |
2023 T3j = R1[WS(rs, 32)]; | |
2024 T3k = T3i - T3j; | |
2025 Tai = T3i + T3j; | |
2026 T4k = R1[WS(rs, 16)]; | |
2027 T4l = R1[WS(rs, 48)]; | |
2028 T4m = T4k - T4l; | |
2029 Taj = T4k + T4l; | |
2030 } | |
2031 { | |
2032 E T3l, T3m, T3n, T3o, T3p, T3q; | |
2033 T3l = R1[WS(rs, 8)]; | |
2034 T3m = R1[WS(rs, 40)]; | |
2035 T3n = T3l - T3m; | |
2036 T3o = R1[WS(rs, 56)]; | |
2037 T3p = R1[WS(rs, 24)]; | |
2038 T3q = T3o - T3p; | |
2039 T3r = KP707106781 * (T3n + T3q); | |
2040 Tam = T3o + T3p; | |
2041 T4j = KP707106781 * (T3q - T3n); | |
2042 Tal = T3l + T3m; | |
2043 } | |
2044 { | |
2045 E T3v, T3y, T3C, T3F; | |
2046 { | |
2047 E T3t, T3u, T3w, T3x; | |
2048 T3t = R1[WS(rs, 4)]; | |
2049 T3u = R1[WS(rs, 36)]; | |
2050 T3v = T3t - T3u; | |
2051 TaI = T3t + T3u; | |
2052 T3w = R1[WS(rs, 20)]; | |
2053 T3x = R1[WS(rs, 52)]; | |
2054 T3y = T3w - T3x; | |
2055 TaJ = T3w + T3x; | |
2056 } | |
2057 T3z = FNMS(KP382683432, T3y, KP923879532 * T3v); | |
2058 Tdb = TaI - TaJ; | |
2059 T4p = FMA(KP382683432, T3v, KP923879532 * T3y); | |
2060 { | |
2061 E T3A, T3B, T3D, T3E; | |
2062 T3A = R1[WS(rs, 60)]; | |
2063 T3B = R1[WS(rs, 28)]; | |
2064 T3C = T3A - T3B; | |
2065 TaF = T3A + T3B; | |
2066 T3D = R1[WS(rs, 12)]; | |
2067 T3E = R1[WS(rs, 44)]; | |
2068 T3F = T3D - T3E; | |
2069 TaG = T3D + T3E; | |
2070 } | |
2071 T3G = FMA(KP923879532, T3C, KP382683432 * T3F); | |
2072 Tdc = TaF - TaG; | |
2073 T4o = FNMS(KP923879532, T3F, KP382683432 * T3C); | |
2074 } | |
2075 { | |
2076 E T3s, T3H, Tda, Tdd; | |
2077 T3s = T3k + T3r; | |
2078 T3H = T3z + T3G; | |
2079 T3I = T3s + T3H; | |
2080 T6H = T3s - T3H; | |
2081 Tda = Tai - Taj; | |
2082 Tdd = KP707106781 * (Tdb + Tdc); | |
2083 Tde = Tda + Tdd; | |
2084 TeH = Tda - Tdd; | |
2085 } | |
2086 { | |
2087 E Tdn, Tdo, T4n, T4q; | |
2088 Tdn = Tam - Tal; | |
2089 Tdo = KP707106781 * (Tdc - Tdb); | |
2090 Tdp = Tdn + Tdo; | |
2091 TeF = Tdo - Tdn; | |
2092 T4n = T4j - T4m; | |
2093 T4q = T4o - T4p; | |
2094 T4r = T4n + T4q; | |
2095 T6F = T4q - T4n; | |
2096 } | |
2097 { | |
2098 E T7S, T7T, Tak, Tan; | |
2099 T7S = T3k - T3r; | |
2100 T7T = T4p + T4o; | |
2101 T7U = T7S + T7T; | |
2102 T9l = T7S - T7T; | |
2103 Tak = Tai + Taj; | |
2104 Tan = Tal + Tam; | |
2105 Tao = Tak - Tan; | |
2106 TbW = Tak + Tan; | |
2107 } | |
2108 { | |
2109 E TaH, TaK, T83, T84; | |
2110 TaH = TaF + TaG; | |
2111 TaK = TaI + TaJ; | |
2112 TaL = TaH - TaK; | |
2113 TbX = TaK + TaH; | |
2114 T83 = T4m + T4j; | |
2115 T84 = T3G - T3z; | |
2116 T85 = T83 + T84; | |
2117 T9j = T84 - T83; | |
2118 } | |
2119 } | |
2120 { | |
2121 E T1z, T2V, T1C, T39, T1G, T38, T1J, T2Y, T1O, T1R, T32, Td0, T3c, T1V, T1Y; | |
2122 E T35, Td1, T3b; | |
2123 { | |
2124 E T1x, T1y, T1A, T1B; | |
2125 T1x = R0[WS(rs, 63)]; | |
2126 T1y = R0[WS(rs, 31)]; | |
2127 T1z = T1x + T1y; | |
2128 T2V = T1x - T1y; | |
2129 T1A = R0[WS(rs, 15)]; | |
2130 T1B = R0[WS(rs, 47)]; | |
2131 T1C = T1A + T1B; | |
2132 T39 = T1A - T1B; | |
2133 } | |
2134 { | |
2135 E T1E, T1F, T2W, T1H, T1I, T2X; | |
2136 T1E = R0[WS(rs, 7)]; | |
2137 T1F = R0[WS(rs, 39)]; | |
2138 T2W = T1E - T1F; | |
2139 T1H = R0[WS(rs, 55)]; | |
2140 T1I = R0[WS(rs, 23)]; | |
2141 T2X = T1H - T1I; | |
2142 T1G = T1E + T1F; | |
2143 T38 = KP707106781 * (T2X - T2W); | |
2144 T1J = T1H + T1I; | |
2145 T2Y = KP707106781 * (T2W + T2X); | |
2146 } | |
2147 { | |
2148 E T30, T31, T33, T34; | |
2149 { | |
2150 E T1M, T1N, T1P, T1Q; | |
2151 T1M = R0[WS(rs, 3)]; | |
2152 T1N = R0[WS(rs, 35)]; | |
2153 T1O = T1M + T1N; | |
2154 T30 = T1M - T1N; | |
2155 T1P = R0[WS(rs, 19)]; | |
2156 T1Q = R0[WS(rs, 51)]; | |
2157 T1R = T1P + T1Q; | |
2158 T31 = T1P - T1Q; | |
2159 } | |
2160 T32 = FNMS(KP382683432, T31, KP923879532 * T30); | |
2161 Td0 = T1O - T1R; | |
2162 T3c = FMA(KP382683432, T30, KP923879532 * T31); | |
2163 { | |
2164 E T1T, T1U, T1W, T1X; | |
2165 T1T = R0[WS(rs, 59)]; | |
2166 T1U = R0[WS(rs, 27)]; | |
2167 T1V = T1T + T1U; | |
2168 T33 = T1T - T1U; | |
2169 T1W = R0[WS(rs, 11)]; | |
2170 T1X = R0[WS(rs, 43)]; | |
2171 T1Y = T1W + T1X; | |
2172 T34 = T1W - T1X; | |
2173 } | |
2174 T35 = FMA(KP923879532, T33, KP382683432 * T34); | |
2175 Td1 = T1V - T1Y; | |
2176 T3b = FNMS(KP923879532, T34, KP382683432 * T33); | |
2177 } | |
2178 { | |
2179 E T1D, T1K, TcZ, Td2; | |
2180 T1D = T1z + T1C; | |
2181 T1K = T1G + T1J; | |
2182 T1L = T1D + T1K; | |
2183 Tad = T1D - T1K; | |
2184 TcZ = T1z - T1C; | |
2185 Td2 = KP707106781 * (Td0 + Td1); | |
2186 Td3 = TcZ + Td2; | |
2187 Tew = TcZ - Td2; | |
2188 } | |
2189 { | |
2190 E Td4, Td5, T1S, T1Z; | |
2191 Td4 = T1J - T1G; | |
2192 Td5 = KP707106781 * (Td1 - Td0); | |
2193 Td6 = Td4 + Td5; | |
2194 Tex = Td5 - Td4; | |
2195 T1S = T1O + T1R; | |
2196 T1Z = T1V + T1Y; | |
2197 T20 = T1S + T1Z; | |
2198 Tae = T1Z - T1S; | |
2199 } | |
2200 { | |
2201 E T2Z, T36, T7J, T7K; | |
2202 T2Z = T2V + T2Y; | |
2203 T36 = T32 + T35; | |
2204 T37 = T2Z + T36; | |
2205 T6x = T2Z - T36; | |
2206 T7J = T2V - T2Y; | |
2207 T7K = T3c + T3b; | |
2208 T7L = T7J + T7K; | |
2209 T9a = T7J - T7K; | |
2210 } | |
2211 { | |
2212 E T7M, T7N, T3a, T3d; | |
2213 T7M = T39 + T38; | |
2214 T7N = T35 - T32; | |
2215 T7O = T7M + T7N; | |
2216 T9b = T7N - T7M; | |
2217 T3a = T38 - T39; | |
2218 T3d = T3b - T3c; | |
2219 T3e = T3a + T3d; | |
2220 T6w = T3d - T3a; | |
2221 } | |
2222 } | |
2223 { | |
2224 E T3L, Tdf, T3X, Tar, T42, Tdi, T4e, Tay, T3S, Tdg, T3U, Tau, T49, Tdj, T4b; | |
2225 E TaB, Tdh, Tdk; | |
2226 { | |
2227 E T3J, T3K, Tap, T3V, T3W, Taq; | |
2228 T3J = R1[WS(rs, 2)]; | |
2229 T3K = R1[WS(rs, 34)]; | |
2230 Tap = T3J + T3K; | |
2231 T3V = R1[WS(rs, 18)]; | |
2232 T3W = R1[WS(rs, 50)]; | |
2233 Taq = T3V + T3W; | |
2234 T3L = T3J - T3K; | |
2235 Tdf = Tap - Taq; | |
2236 T3X = T3V - T3W; | |
2237 Tar = Tap + Taq; | |
2238 } | |
2239 { | |
2240 E T40, T41, Taw, T4c, T4d, Tax; | |
2241 T40 = R1[WS(rs, 62)]; | |
2242 T41 = R1[WS(rs, 30)]; | |
2243 Taw = T40 + T41; | |
2244 T4c = R1[WS(rs, 14)]; | |
2245 T4d = R1[WS(rs, 46)]; | |
2246 Tax = T4c + T4d; | |
2247 T42 = T40 - T41; | |
2248 Tdi = Taw - Tax; | |
2249 T4e = T4c - T4d; | |
2250 Tay = Taw + Tax; | |
2251 } | |
2252 { | |
2253 E T3O, Tas, T3R, Tat; | |
2254 { | |
2255 E T3M, T3N, T3P, T3Q; | |
2256 T3M = R1[WS(rs, 10)]; | |
2257 T3N = R1[WS(rs, 42)]; | |
2258 T3O = T3M - T3N; | |
2259 Tas = T3M + T3N; | |
2260 T3P = R1[WS(rs, 58)]; | |
2261 T3Q = R1[WS(rs, 26)]; | |
2262 T3R = T3P - T3Q; | |
2263 Tat = T3P + T3Q; | |
2264 } | |
2265 T3S = KP707106781 * (T3O + T3R); | |
2266 Tdg = Tat - Tas; | |
2267 T3U = KP707106781 * (T3R - T3O); | |
2268 Tau = Tas + Tat; | |
2269 } | |
2270 { | |
2271 E T45, Taz, T48, TaA; | |
2272 { | |
2273 E T43, T44, T46, T47; | |
2274 T43 = R1[WS(rs, 6)]; | |
2275 T44 = R1[WS(rs, 38)]; | |
2276 T45 = T43 - T44; | |
2277 Taz = T43 + T44; | |
2278 T46 = R1[WS(rs, 54)]; | |
2279 T47 = R1[WS(rs, 22)]; | |
2280 T48 = T46 - T47; | |
2281 TaA = T46 + T47; | |
2282 } | |
2283 T49 = KP707106781 * (T45 + T48); | |
2284 Tdj = TaA - Taz; | |
2285 T4b = KP707106781 * (T48 - T45); | |
2286 TaB = Taz + TaA; | |
2287 } | |
2288 TbZ = Tar + Tau; | |
2289 Tc0 = Tay + TaB; | |
2290 { | |
2291 E T3T, T3Y, Tdq, Tdr; | |
2292 T3T = T3L + T3S; | |
2293 T3Y = T3U - T3X; | |
2294 T3Z = FMA(KP980785280, T3T, KP195090322 * T3Y); | |
2295 T4s = FNMS(KP195090322, T3T, KP980785280 * T3Y); | |
2296 Tdq = FNMS(KP382683432, Tdf, KP923879532 * Tdg); | |
2297 Tdr = FMA(KP382683432, Tdi, KP923879532 * Tdj); | |
2298 Tds = Tdq + Tdr; | |
2299 TeI = Tdr - Tdq; | |
2300 } | |
2301 { | |
2302 E T4a, T4f, T7Y, T7Z; | |
2303 T4a = T42 + T49; | |
2304 T4f = T4b - T4e; | |
2305 T4g = FNMS(KP195090322, T4f, KP980785280 * T4a); | |
2306 T4t = FMA(KP195090322, T4a, KP980785280 * T4f); | |
2307 T7Y = T42 - T49; | |
2308 T7Z = T4e + T4b; | |
2309 T80 = FNMS(KP555570233, T7Z, KP831469612 * T7Y); | |
2310 T87 = FMA(KP555570233, T7Y, KP831469612 * T7Z); | |
2311 } | |
2312 Tdh = FMA(KP923879532, Tdf, KP382683432 * Tdg); | |
2313 Tdk = FNMS(KP382683432, Tdj, KP923879532 * Tdi); | |
2314 Tdl = Tdh + Tdk; | |
2315 TeE = Tdk - Tdh; | |
2316 { | |
2317 E T7V, T7W, Tav, TaC; | |
2318 T7V = T3L - T3S; | |
2319 T7W = T3X + T3U; | |
2320 T7X = FMA(KP831469612, T7V, KP555570233 * T7W); | |
2321 T86 = FNMS(KP555570233, T7V, KP831469612 * T7W); | |
2322 Tav = Tar - Tau; | |
2323 TaC = Tay - TaB; | |
2324 TaD = KP707106781 * (Tav + TaC); | |
2325 TaM = KP707106781 * (TaC - Tav); | |
2326 } | |
2327 } | |
2328 { | |
2329 E T50, TdA, T5c, TaY, T5h, TdD, T5t, Tb5, T57, TdB, T59, Tb1, T5o, TdE, T5q; | |
2330 E Tb8, TdC, TdF; | |
2331 { | |
2332 E T4Y, T4Z, TaW, T5a, T5b, TaX; | |
2333 T4Y = R1[WS(rs, 1)]; | |
2334 T4Z = R1[WS(rs, 33)]; | |
2335 TaW = T4Y + T4Z; | |
2336 T5a = R1[WS(rs, 17)]; | |
2337 T5b = R1[WS(rs, 49)]; | |
2338 TaX = T5a + T5b; | |
2339 T50 = T4Y - T4Z; | |
2340 TdA = TaW - TaX; | |
2341 T5c = T5a - T5b; | |
2342 TaY = TaW + TaX; | |
2343 } | |
2344 { | |
2345 E T5f, T5g, Tb3, T5r, T5s, Tb4; | |
2346 T5f = R1[WS(rs, 61)]; | |
2347 T5g = R1[WS(rs, 29)]; | |
2348 Tb3 = T5f + T5g; | |
2349 T5r = R1[WS(rs, 13)]; | |
2350 T5s = R1[WS(rs, 45)]; | |
2351 Tb4 = T5r + T5s; | |
2352 T5h = T5f - T5g; | |
2353 TdD = Tb3 - Tb4; | |
2354 T5t = T5r - T5s; | |
2355 Tb5 = Tb3 + Tb4; | |
2356 } | |
2357 { | |
2358 E T53, TaZ, T56, Tb0; | |
2359 { | |
2360 E T51, T52, T54, T55; | |
2361 T51 = R1[WS(rs, 9)]; | |
2362 T52 = R1[WS(rs, 41)]; | |
2363 T53 = T51 - T52; | |
2364 TaZ = T51 + T52; | |
2365 T54 = R1[WS(rs, 57)]; | |
2366 T55 = R1[WS(rs, 25)]; | |
2367 T56 = T54 - T55; | |
2368 Tb0 = T54 + T55; | |
2369 } | |
2370 T57 = KP707106781 * (T53 + T56); | |
2371 TdB = Tb0 - TaZ; | |
2372 T59 = KP707106781 * (T56 - T53); | |
2373 Tb1 = TaZ + Tb0; | |
2374 } | |
2375 { | |
2376 E T5k, Tb6, T5n, Tb7; | |
2377 { | |
2378 E T5i, T5j, T5l, T5m; | |
2379 T5i = R1[WS(rs, 5)]; | |
2380 T5j = R1[WS(rs, 37)]; | |
2381 T5k = T5i - T5j; | |
2382 Tb6 = T5i + T5j; | |
2383 T5l = R1[WS(rs, 53)]; | |
2384 T5m = R1[WS(rs, 21)]; | |
2385 T5n = T5l - T5m; | |
2386 Tb7 = T5l + T5m; | |
2387 } | |
2388 T5o = KP707106781 * (T5k + T5n); | |
2389 TdE = Tb7 - Tb6; | |
2390 T5q = KP707106781 * (T5n - T5k); | |
2391 Tb8 = Tb6 + Tb7; | |
2392 } | |
2393 Tc6 = TaY + Tb1; | |
2394 Tc7 = Tb5 + Tb8; | |
2395 { | |
2396 E T58, T5d, TdL, TdM; | |
2397 T58 = T50 + T57; | |
2398 T5d = T59 - T5c; | |
2399 T5e = FMA(KP980785280, T58, KP195090322 * T5d); | |
2400 T5H = FNMS(KP195090322, T58, KP980785280 * T5d); | |
2401 TdL = FNMS(KP382683432, TdA, KP923879532 * TdB); | |
2402 TdM = FMA(KP382683432, TdD, KP923879532 * TdE); | |
2403 TdN = TdL + TdM; | |
2404 TeM = TdM - TdL; | |
2405 } | |
2406 { | |
2407 E T5p, T5u, T8h, T8i; | |
2408 T5p = T5h + T5o; | |
2409 T5u = T5q - T5t; | |
2410 T5v = FNMS(KP195090322, T5u, KP980785280 * T5p); | |
2411 T5I = FMA(KP195090322, T5p, KP980785280 * T5u); | |
2412 T8h = T5h - T5o; | |
2413 T8i = T5t + T5q; | |
2414 T8j = FNMS(KP555570233, T8i, KP831469612 * T8h); | |
2415 T8q = FMA(KP555570233, T8h, KP831469612 * T8i); | |
2416 } | |
2417 TdC = FMA(KP923879532, TdA, KP382683432 * TdB); | |
2418 TdF = FNMS(KP382683432, TdE, KP923879532 * TdD); | |
2419 TdG = TdC + TdF; | |
2420 TeO = TdF - TdC; | |
2421 { | |
2422 E T8e, T8f, Tb2, Tb9; | |
2423 T8e = T50 - T57; | |
2424 T8f = T5c + T59; | |
2425 T8g = FMA(KP831469612, T8e, KP555570233 * T8f); | |
2426 T8p = FNMS(KP555570233, T8e, KP831469612 * T8f); | |
2427 Tb2 = TaY - Tb1; | |
2428 Tb9 = Tb5 - Tb8; | |
2429 Tba = KP707106781 * (Tb2 + Tb9); | |
2430 Tbj = KP707106781 * (Tb9 - Tb2); | |
2431 } | |
2432 } | |
2433 { | |
2434 E T11, TbV, Tc9, Tcf, T22, Tcb, Tc2, Tce; | |
2435 { | |
2436 E Tv, T10, Tc5, Tc8; | |
2437 Tv = Tf + Tu; | |
2438 T10 = TK + TZ; | |
2439 T11 = Tv + T10; | |
2440 TbV = Tv - T10; | |
2441 Tc5 = Tc3 + Tc4; | |
2442 Tc8 = Tc6 + Tc7; | |
2443 Tc9 = Tc5 - Tc8; | |
2444 Tcf = Tc5 + Tc8; | |
2445 } | |
2446 { | |
2447 E T1w, T21, TbY, Tc1; | |
2448 T1w = T1g + T1v; | |
2449 T21 = T1L + T20; | |
2450 T22 = T1w + T21; | |
2451 Tcb = T21 - T1w; | |
2452 TbY = TbW + TbX; | |
2453 Tc1 = TbZ + Tc0; | |
2454 Tc2 = TbY - Tc1; | |
2455 Tce = TbY + Tc1; | |
2456 } | |
2457 Cr[WS(csr, 32)] = T11 - T22; | |
2458 Ci[WS(csi, 32)] = Tcf - Tce; | |
2459 { | |
2460 E Tca, Tcc, Tcd, Tcg; | |
2461 Tca = KP707106781 * (Tc2 + Tc9); | |
2462 Cr[WS(csr, 48)] = TbV - Tca; | |
2463 Cr[WS(csr, 16)] = TbV + Tca; | |
2464 Tcc = KP707106781 * (Tc9 - Tc2); | |
2465 Ci[WS(csi, 16)] = Tcb + Tcc; | |
2466 Ci[WS(csi, 48)] = Tcc - Tcb; | |
2467 Tcd = T11 + T22; | |
2468 Tcg = Tce + Tcf; | |
2469 Cr[WS(csr, 64)] = Tcd - Tcg; | |
2470 Cr[0] = Tcd + Tcg; | |
2471 } | |
2472 } | |
2473 { | |
2474 E Tch, Tcu, Tck, Tct, Tco, Tcy, Tcr, Tcz, Tci, Tcj; | |
2475 Tch = Tf - Tu; | |
2476 Tcu = TZ - TK; | |
2477 Tci = T1g - T1v; | |
2478 Tcj = T1L - T20; | |
2479 Tck = KP707106781 * (Tci + Tcj); | |
2480 Tct = KP707106781 * (Tcj - Tci); | |
2481 { | |
2482 E Tcm, Tcn, Tcp, Tcq; | |
2483 Tcm = TbW - TbX; | |
2484 Tcn = Tc0 - TbZ; | |
2485 Tco = FMA(KP923879532, Tcm, KP382683432 * Tcn); | |
2486 Tcy = FNMS(KP382683432, Tcm, KP923879532 * Tcn); | |
2487 Tcp = Tc3 - Tc4; | |
2488 Tcq = Tc7 - Tc6; | |
2489 Tcr = FNMS(KP382683432, Tcq, KP923879532 * Tcp); | |
2490 Tcz = FMA(KP382683432, Tcp, KP923879532 * Tcq); | |
2491 } | |
2492 { | |
2493 E Tcl, Tcs, Tcx, TcA; | |
2494 Tcl = Tch + Tck; | |
2495 Tcs = Tco + Tcr; | |
2496 Cr[WS(csr, 56)] = Tcl - Tcs; | |
2497 Cr[WS(csr, 8)] = Tcl + Tcs; | |
2498 Tcx = Tcu + Tct; | |
2499 TcA = Tcy + Tcz; | |
2500 Ci[WS(csi, 8)] = Tcx + TcA; | |
2501 Ci[WS(csi, 56)] = TcA - Tcx; | |
2502 } | |
2503 { | |
2504 E Tcv, Tcw, TcB, TcC; | |
2505 Tcv = Tct - Tcu; | |
2506 Tcw = Tcr - Tco; | |
2507 Ci[WS(csi, 24)] = Tcv + Tcw; | |
2508 Ci[WS(csi, 40)] = Tcw - Tcv; | |
2509 TcB = Tch - Tck; | |
2510 TcC = Tcz - Tcy; | |
2511 Cr[WS(csr, 40)] = TcB - TcC; | |
2512 Cr[WS(csr, 24)] = TcB + TcC; | |
2513 } | |
2514 } | |
2515 { | |
2516 E Ta9, TbB, Tbs, TbM, Tag, TbL, TbJ, TbR, TaO, Tbw, Tbp, TbC, TbG, TbQ, Tbl; | |
2517 E Tbx, Ta8, Tbr; | |
2518 Ta8 = KP707106781 * (Ta6 + Ta7); | |
2519 Ta9 = Ta5 + Ta8; | |
2520 TbB = Ta5 - Ta8; | |
2521 Tbr = KP707106781 * (Ta7 - Ta6); | |
2522 Tbs = Tbq + Tbr; | |
2523 TbM = Tbr - Tbq; | |
2524 { | |
2525 E Tac, Taf, TbH, TbI; | |
2526 Tac = FMA(KP923879532, Taa, KP382683432 * Tab); | |
2527 Taf = FNMS(KP382683432, Tae, KP923879532 * Tad); | |
2528 Tag = Tac + Taf; | |
2529 TbL = Taf - Tac; | |
2530 TbH = TaV - Tba; | |
2531 TbI = Tbj - Tbi; | |
2532 TbJ = FNMS(KP555570233, TbI, KP831469612 * TbH); | |
2533 TbR = FMA(KP555570233, TbH, KP831469612 * TbI); | |
2534 } | |
2535 { | |
2536 E TaE, TaN, Tbn, Tbo; | |
2537 TaE = Tao + TaD; | |
2538 TaN = TaL + TaM; | |
2539 TaO = FMA(KP980785280, TaE, KP195090322 * TaN); | |
2540 Tbw = FNMS(KP195090322, TaE, KP980785280 * TaN); | |
2541 Tbn = FNMS(KP382683432, Taa, KP923879532 * Tab); | |
2542 Tbo = FMA(KP382683432, Tad, KP923879532 * Tae); | |
2543 Tbp = Tbn + Tbo; | |
2544 TbC = Tbo - Tbn; | |
2545 } | |
2546 { | |
2547 E TbE, TbF, Tbb, Tbk; | |
2548 TbE = Tao - TaD; | |
2549 TbF = TaM - TaL; | |
2550 TbG = FMA(KP831469612, TbE, KP555570233 * TbF); | |
2551 TbQ = FNMS(KP555570233, TbE, KP831469612 * TbF); | |
2552 Tbb = TaV + Tba; | |
2553 Tbk = Tbi + Tbj; | |
2554 Tbl = FNMS(KP195090322, Tbk, KP980785280 * Tbb); | |
2555 Tbx = FMA(KP195090322, Tbb, KP980785280 * Tbk); | |
2556 } | |
2557 { | |
2558 E Tah, Tbm, Tbv, Tby; | |
2559 Tah = Ta9 + Tag; | |
2560 Tbm = TaO + Tbl; | |
2561 Cr[WS(csr, 60)] = Tah - Tbm; | |
2562 Cr[WS(csr, 4)] = Tah + Tbm; | |
2563 Tbv = Tbs + Tbp; | |
2564 Tby = Tbw + Tbx; | |
2565 Ci[WS(csi, 4)] = Tbv + Tby; | |
2566 Ci[WS(csi, 60)] = Tby - Tbv; | |
2567 } | |
2568 { | |
2569 E Tbt, Tbu, Tbz, TbA; | |
2570 Tbt = Tbp - Tbs; | |
2571 Tbu = Tbl - TaO; | |
2572 Ci[WS(csi, 28)] = Tbt + Tbu; | |
2573 Ci[WS(csi, 36)] = Tbu - Tbt; | |
2574 Tbz = Ta9 - Tag; | |
2575 TbA = Tbx - Tbw; | |
2576 Cr[WS(csr, 36)] = Tbz - TbA; | |
2577 Cr[WS(csr, 28)] = Tbz + TbA; | |
2578 } | |
2579 { | |
2580 E TbD, TbK, TbP, TbS; | |
2581 TbD = TbB + TbC; | |
2582 TbK = TbG + TbJ; | |
2583 Cr[WS(csr, 52)] = TbD - TbK; | |
2584 Cr[WS(csr, 12)] = TbD + TbK; | |
2585 TbP = TbM + TbL; | |
2586 TbS = TbQ + TbR; | |
2587 Ci[WS(csi, 12)] = TbP + TbS; | |
2588 Ci[WS(csi, 52)] = TbS - TbP; | |
2589 } | |
2590 { | |
2591 E TbN, TbO, TbT, TbU; | |
2592 TbN = TbL - TbM; | |
2593 TbO = TbJ - TbG; | |
2594 Ci[WS(csi, 20)] = TbN + TbO; | |
2595 Ci[WS(csi, 44)] = TbO - TbN; | |
2596 TbT = TbB - TbC; | |
2597 TbU = TbR - TbQ; | |
2598 Cr[WS(csr, 44)] = TbT - TbU; | |
2599 Cr[WS(csr, 20)] = TbT + TbU; | |
2600 } | |
2601 } | |
2602 { | |
2603 E Tev, Tf7, Tfc, Tfm, Tff, Tfn, TeC, Tfh, TeK, Tf2, TeV, Tf8, TeY, Tfi, TeR; | |
2604 E Tf3; | |
2605 { | |
2606 E Tet, Teu, Tfa, Tfb; | |
2607 Tet = TcD - TcG; | |
2608 Teu = TdY - TdX; | |
2609 Tev = Tet - Teu; | |
2610 Tf7 = Tet + Teu; | |
2611 Tfa = TeF + TeE; | |
2612 Tfb = TeH + TeI; | |
2613 Tfc = FMA(KP290284677, Tfa, KP956940335 * Tfb); | |
2614 Tfm = FNMS(KP290284677, Tfb, KP956940335 * Tfa); | |
2615 } | |
2616 { | |
2617 E Tfd, Tfe, Tey, TeB; | |
2618 Tfd = TeL + TeM; | |
2619 Tfe = TeP + TeO; | |
2620 Tff = FNMS(KP290284677, Tfe, KP956940335 * Tfd); | |
2621 Tfn = FMA(KP956940335, Tfe, KP290284677 * Tfd); | |
2622 Tey = FMA(KP555570233, Tew, KP831469612 * Tex); | |
2623 TeB = FNMS(KP555570233, TeA, KP831469612 * Tez); | |
2624 TeC = Tey - TeB; | |
2625 Tfh = TeB + Tey; | |
2626 } | |
2627 { | |
2628 E TeG, TeJ, TeT, TeU; | |
2629 TeG = TeE - TeF; | |
2630 TeJ = TeH - TeI; | |
2631 TeK = FMA(KP471396736, TeG, KP881921264 * TeJ); | |
2632 Tf2 = FNMS(KP471396736, TeJ, KP881921264 * TeG); | |
2633 TeT = FNMS(KP555570233, Tex, KP831469612 * Tew); | |
2634 TeU = FMA(KP831469612, TeA, KP555570233 * Tez); | |
2635 TeV = TeT - TeU; | |
2636 Tf8 = TeU + TeT; | |
2637 } | |
2638 { | |
2639 E TeW, TeX, TeN, TeQ; | |
2640 TeW = TcN - TcK; | |
2641 TeX = TdV - TdU; | |
2642 TeY = TeW - TeX; | |
2643 Tfi = TeX + TeW; | |
2644 TeN = TeL - TeM; | |
2645 TeQ = TeO - TeP; | |
2646 TeR = FNMS(KP471396736, TeQ, KP881921264 * TeN); | |
2647 Tf3 = FMA(KP881921264, TeQ, KP471396736 * TeN); | |
2648 } | |
2649 { | |
2650 E TeD, TeS, Tf1, Tf4; | |
2651 TeD = Tev + TeC; | |
2652 TeS = TeK + TeR; | |
2653 Cr[WS(csr, 54)] = TeD - TeS; | |
2654 Cr[WS(csr, 10)] = TeD + TeS; | |
2655 Tf1 = TeY + TeV; | |
2656 Tf4 = Tf2 + Tf3; | |
2657 Ci[WS(csi, 10)] = Tf1 + Tf4; | |
2658 Ci[WS(csi, 54)] = Tf4 - Tf1; | |
2659 } | |
2660 { | |
2661 E TeZ, Tf0, Tf5, Tf6; | |
2662 TeZ = TeV - TeY; | |
2663 Tf0 = TeR - TeK; | |
2664 Ci[WS(csi, 22)] = TeZ + Tf0; | |
2665 Ci[WS(csi, 42)] = Tf0 - TeZ; | |
2666 Tf5 = Tev - TeC; | |
2667 Tf6 = Tf3 - Tf2; | |
2668 Cr[WS(csr, 42)] = Tf5 - Tf6; | |
2669 Cr[WS(csr, 22)] = Tf5 + Tf6; | |
2670 } | |
2671 { | |
2672 E Tf9, Tfg, Tfl, Tfo; | |
2673 Tf9 = Tf7 + Tf8; | |
2674 Tfg = Tfc + Tff; | |
2675 Cr[WS(csr, 58)] = Tf9 - Tfg; | |
2676 Cr[WS(csr, 6)] = Tf9 + Tfg; | |
2677 Tfl = Tfi + Tfh; | |
2678 Tfo = Tfm + Tfn; | |
2679 Ci[WS(csi, 6)] = Tfl + Tfo; | |
2680 Ci[WS(csi, 58)] = Tfo - Tfl; | |
2681 } | |
2682 { | |
2683 E Tfj, Tfk, Tfp, Tfq; | |
2684 Tfj = Tfh - Tfi; | |
2685 Tfk = Tff - Tfc; | |
2686 Ci[WS(csi, 26)] = Tfj + Tfk; | |
2687 Ci[WS(csi, 38)] = Tfk - Tfj; | |
2688 Tfp = Tf7 - Tf8; | |
2689 Tfq = Tfn - Tfm; | |
2690 Cr[WS(csr, 38)] = Tfp - Tfq; | |
2691 Cr[WS(csr, 26)] = Tfp + Tfq; | |
2692 } | |
2693 } | |
2694 { | |
2695 E TcP, Te9, Tee, Teo, Teh, Tep, Td8, Tej, Tdu, Te4, TdT, Tea, Te0, Tek, TdP; | |
2696 E Te5; | |
2697 { | |
2698 E TcH, TcO, Tec, Ted; | |
2699 TcH = TcD + TcG; | |
2700 TcO = TcK + TcN; | |
2701 TcP = TcH + TcO; | |
2702 Te9 = TcH - TcO; | |
2703 Tec = Tde - Tdl; | |
2704 Ted = Tds - Tdp; | |
2705 Tee = FMA(KP773010453, Tec, KP634393284 * Ted); | |
2706 Teo = FNMS(KP634393284, Tec, KP773010453 * Ted); | |
2707 } | |
2708 { | |
2709 E Tef, Teg, TcY, Td7; | |
2710 Tef = Tdz - TdG; | |
2711 Teg = TdN - TdK; | |
2712 Teh = FNMS(KP634393284, Teg, KP773010453 * Tef); | |
2713 Tep = FMA(KP634393284, Tef, KP773010453 * Teg); | |
2714 TcY = FMA(KP980785280, TcU, KP195090322 * TcX); | |
2715 Td7 = FNMS(KP195090322, Td6, KP980785280 * Td3); | |
2716 Td8 = TcY + Td7; | |
2717 Tej = Td7 - TcY; | |
2718 } | |
2719 { | |
2720 E Tdm, Tdt, TdR, TdS; | |
2721 Tdm = Tde + Tdl; | |
2722 Tdt = Tdp + Tds; | |
2723 Tdu = FMA(KP995184726, Tdm, KP098017140 * Tdt); | |
2724 Te4 = FNMS(KP098017140, Tdm, KP995184726 * Tdt); | |
2725 TdR = FNMS(KP195090322, TcU, KP980785280 * TcX); | |
2726 TdS = FMA(KP195090322, Td3, KP980785280 * Td6); | |
2727 TdT = TdR + TdS; | |
2728 Tea = TdS - TdR; | |
2729 } | |
2730 { | |
2731 E TdW, TdZ, TdH, TdO; | |
2732 TdW = TdU + TdV; | |
2733 TdZ = TdX + TdY; | |
2734 Te0 = TdW + TdZ; | |
2735 Tek = TdZ - TdW; | |
2736 TdH = Tdz + TdG; | |
2737 TdO = TdK + TdN; | |
2738 TdP = FNMS(KP098017140, TdO, KP995184726 * TdH); | |
2739 Te5 = FMA(KP098017140, TdH, KP995184726 * TdO); | |
2740 } | |
2741 { | |
2742 E Td9, TdQ, Te3, Te6; | |
2743 Td9 = TcP + Td8; | |
2744 TdQ = Tdu + TdP; | |
2745 Cr[WS(csr, 62)] = Td9 - TdQ; | |
2746 Cr[WS(csr, 2)] = Td9 + TdQ; | |
2747 Te3 = Te0 + TdT; | |
2748 Te6 = Te4 + Te5; | |
2749 Ci[WS(csi, 2)] = Te3 + Te6; | |
2750 Ci[WS(csi, 62)] = Te6 - Te3; | |
2751 } | |
2752 { | |
2753 E Te1, Te2, Te7, Te8; | |
2754 Te1 = TdT - Te0; | |
2755 Te2 = TdP - Tdu; | |
2756 Ci[WS(csi, 30)] = Te1 + Te2; | |
2757 Ci[WS(csi, 34)] = Te2 - Te1; | |
2758 Te7 = TcP - Td8; | |
2759 Te8 = Te5 - Te4; | |
2760 Cr[WS(csr, 34)] = Te7 - Te8; | |
2761 Cr[WS(csr, 30)] = Te7 + Te8; | |
2762 } | |
2763 { | |
2764 E Teb, Tei, Ten, Teq; | |
2765 Teb = Te9 + Tea; | |
2766 Tei = Tee + Teh; | |
2767 Cr[WS(csr, 50)] = Teb - Tei; | |
2768 Cr[WS(csr, 14)] = Teb + Tei; | |
2769 Ten = Tek + Tej; | |
2770 Teq = Teo + Tep; | |
2771 Ci[WS(csi, 14)] = Ten + Teq; | |
2772 Ci[WS(csi, 50)] = Teq - Ten; | |
2773 } | |
2774 { | |
2775 E Tel, Tem, Ter, Tes; | |
2776 Tel = Tej - Tek; | |
2777 Tem = Teh - Tee; | |
2778 Ci[WS(csi, 18)] = Tel + Tem; | |
2779 Ci[WS(csi, 46)] = Tem - Tel; | |
2780 Ter = Te9 - Tea; | |
2781 Tes = Tep - Teo; | |
2782 Cr[WS(csr, 46)] = Ter - Tes; | |
2783 Cr[WS(csr, 18)] = Ter + Tes; | |
2784 } | |
2785 } | |
2786 { | |
2787 E T6v, T77, T6C, T7h, T6Y, T7i, T6V, T78, T6R, T7n, T73, T7f, T6K, T7m, T72; | |
2788 E T7c; | |
2789 { | |
2790 E T6t, T6u, T6T, T6U; | |
2791 T6t = T27 - T2e; | |
2792 T6u = T5Y - T5X; | |
2793 T6v = T6t - T6u; | |
2794 T77 = T6t + T6u; | |
2795 { | |
2796 E T6y, T6B, T6W, T6X; | |
2797 T6y = FMA(KP773010453, T6w, KP634393284 * T6x); | |
2798 T6B = FNMS(KP634393284, T6A, KP773010453 * T6z); | |
2799 T6C = T6y - T6B; | |
2800 T7h = T6B + T6y; | |
2801 T6W = T2x - T2o; | |
2802 T6X = T5V - T5S; | |
2803 T6Y = T6W - T6X; | |
2804 T7i = T6X + T6W; | |
2805 } | |
2806 T6T = FNMS(KP634393284, T6w, KP773010453 * T6x); | |
2807 T6U = FMA(KP634393284, T6z, KP773010453 * T6A); | |
2808 T6V = T6T - T6U; | |
2809 T78 = T6U + T6T; | |
2810 { | |
2811 E T6N, T7d, T6Q, T7e, T6M, T6O; | |
2812 T6M = T5I - T5H; | |
2813 T6N = T6L - T6M; | |
2814 T7d = T6L + T6M; | |
2815 T6O = T5v - T5e; | |
2816 T6Q = T6O - T6P; | |
2817 T7e = T6P + T6O; | |
2818 T6R = FNMS(KP427555093, T6Q, KP903989293 * T6N); | |
2819 T7n = FMA(KP941544065, T7e, KP336889853 * T7d); | |
2820 T73 = FMA(KP903989293, T6Q, KP427555093 * T6N); | |
2821 T7f = FNMS(KP336889853, T7e, KP941544065 * T7d); | |
2822 } | |
2823 { | |
2824 E T6G, T7a, T6J, T7b, T6E, T6I; | |
2825 T6E = T4g - T3Z; | |
2826 T6G = T6E - T6F; | |
2827 T7a = T6F + T6E; | |
2828 T6I = T4t - T4s; | |
2829 T6J = T6H - T6I; | |
2830 T7b = T6H + T6I; | |
2831 T6K = FMA(KP427555093, T6G, KP903989293 * T6J); | |
2832 T7m = FNMS(KP336889853, T7b, KP941544065 * T7a); | |
2833 T72 = FNMS(KP427555093, T6J, KP903989293 * T6G); | |
2834 T7c = FMA(KP336889853, T7a, KP941544065 * T7b); | |
2835 } | |
2836 } | |
2837 { | |
2838 E T6D, T6S, T71, T74; | |
2839 T6D = T6v + T6C; | |
2840 T6S = T6K + T6R; | |
2841 Cr[WS(csr, 55)] = T6D - T6S; | |
2842 Cr[WS(csr, 9)] = T6D + T6S; | |
2843 T71 = T6Y + T6V; | |
2844 T74 = T72 + T73; | |
2845 Ci[WS(csi, 9)] = T71 + T74; | |
2846 Ci[WS(csi, 55)] = T74 - T71; | |
2847 } | |
2848 { | |
2849 E T6Z, T70, T75, T76; | |
2850 T6Z = T6V - T6Y; | |
2851 T70 = T6R - T6K; | |
2852 Ci[WS(csi, 23)] = T6Z + T70; | |
2853 Ci[WS(csi, 41)] = T70 - T6Z; | |
2854 T75 = T6v - T6C; | |
2855 T76 = T73 - T72; | |
2856 Cr[WS(csr, 41)] = T75 - T76; | |
2857 Cr[WS(csr, 23)] = T75 + T76; | |
2858 } | |
2859 { | |
2860 E T79, T7g, T7l, T7o; | |
2861 T79 = T77 + T78; | |
2862 T7g = T7c + T7f; | |
2863 Cr[WS(csr, 57)] = T79 - T7g; | |
2864 Cr[WS(csr, 7)] = T79 + T7g; | |
2865 T7l = T7i + T7h; | |
2866 T7o = T7m + T7n; | |
2867 Ci[WS(csi, 7)] = T7l + T7o; | |
2868 Ci[WS(csi, 57)] = T7o - T7l; | |
2869 } | |
2870 { | |
2871 E T7j, T7k, T7p, T7q; | |
2872 T7j = T7h - T7i; | |
2873 T7k = T7f - T7c; | |
2874 Ci[WS(csi, 25)] = T7j + T7k; | |
2875 Ci[WS(csi, 39)] = T7k - T7j; | |
2876 T7p = T77 - T78; | |
2877 T7q = T7n - T7m; | |
2878 Cr[WS(csr, 39)] = T7p - T7q; | |
2879 Cr[WS(csr, 25)] = T7p + T7q; | |
2880 } | |
2881 } | |
2882 { | |
2883 E T99, T9L, T9g, T9V, T9C, T9W, T9z, T9M, T9v, Ta1, T9H, T9T, T9o, Ta0, T9G; | |
2884 E T9Q; | |
2885 { | |
2886 E T97, T98, T9x, T9y; | |
2887 T97 = T7r - T7s; | |
2888 T98 = T8C - T8B; | |
2889 T99 = T97 - T98; | |
2890 T9L = T97 + T98; | |
2891 { | |
2892 E T9c, T9f, T9A, T9B; | |
2893 T9c = FMA(KP471396736, T9a, KP881921264 * T9b); | |
2894 T9f = FNMS(KP471396736, T9e, KP881921264 * T9d); | |
2895 T9g = T9c - T9f; | |
2896 T9V = T9f + T9c; | |
2897 T9A = T7z - T7w; | |
2898 T9B = T8z - T8y; | |
2899 T9C = T9A - T9B; | |
2900 T9W = T9B + T9A; | |
2901 } | |
2902 T9x = FNMS(KP471396736, T9b, KP881921264 * T9a); | |
2903 T9y = FMA(KP881921264, T9e, KP471396736 * T9d); | |
2904 T9z = T9x - T9y; | |
2905 T9M = T9y + T9x; | |
2906 { | |
2907 E T9r, T9R, T9u, T9S, T9q, T9s; | |
2908 T9q = T8q - T8p; | |
2909 T9r = T9p - T9q; | |
2910 T9R = T9p + T9q; | |
2911 T9s = T8j - T8g; | |
2912 T9u = T9s - T9t; | |
2913 T9S = T9t + T9s; | |
2914 T9v = FNMS(KP514102744, T9u, KP857728610 * T9r); | |
2915 Ta1 = FMA(KP970031253, T9S, KP242980179 * T9R); | |
2916 T9H = FMA(KP857728610, T9u, KP514102744 * T9r); | |
2917 T9T = FNMS(KP242980179, T9S, KP970031253 * T9R); | |
2918 } | |
2919 { | |
2920 E T9k, T9O, T9n, T9P, T9i, T9m; | |
2921 T9i = T80 - T7X; | |
2922 T9k = T9i - T9j; | |
2923 T9O = T9j + T9i; | |
2924 T9m = T87 - T86; | |
2925 T9n = T9l - T9m; | |
2926 T9P = T9l + T9m; | |
2927 T9o = FMA(KP514102744, T9k, KP857728610 * T9n); | |
2928 Ta0 = FNMS(KP242980179, T9P, KP970031253 * T9O); | |
2929 T9G = FNMS(KP514102744, T9n, KP857728610 * T9k); | |
2930 T9Q = FMA(KP242980179, T9O, KP970031253 * T9P); | |
2931 } | |
2932 } | |
2933 { | |
2934 E T9h, T9w, T9F, T9I; | |
2935 T9h = T99 + T9g; | |
2936 T9w = T9o + T9v; | |
2937 Cr[WS(csr, 53)] = T9h - T9w; | |
2938 Cr[WS(csr, 11)] = T9h + T9w; | |
2939 T9F = T9C + T9z; | |
2940 T9I = T9G + T9H; | |
2941 Ci[WS(csi, 11)] = T9F + T9I; | |
2942 Ci[WS(csi, 53)] = T9I - T9F; | |
2943 } | |
2944 { | |
2945 E T9D, T9E, T9J, T9K; | |
2946 T9D = T9z - T9C; | |
2947 T9E = T9v - T9o; | |
2948 Ci[WS(csi, 21)] = T9D + T9E; | |
2949 Ci[WS(csi, 43)] = T9E - T9D; | |
2950 T9J = T99 - T9g; | |
2951 T9K = T9H - T9G; | |
2952 Cr[WS(csr, 43)] = T9J - T9K; | |
2953 Cr[WS(csr, 21)] = T9J + T9K; | |
2954 } | |
2955 { | |
2956 E T9N, T9U, T9Z, Ta2; | |
2957 T9N = T9L + T9M; | |
2958 T9U = T9Q + T9T; | |
2959 Cr[WS(csr, 59)] = T9N - T9U; | |
2960 Cr[WS(csr, 5)] = T9N + T9U; | |
2961 T9Z = T9W + T9V; | |
2962 Ta2 = Ta0 + Ta1; | |
2963 Ci[WS(csi, 5)] = T9Z + Ta2; | |
2964 Ci[WS(csi, 59)] = Ta2 - T9Z; | |
2965 } | |
2966 { | |
2967 E T9X, T9Y, Ta3, Ta4; | |
2968 T9X = T9V - T9W; | |
2969 T9Y = T9T - T9Q; | |
2970 Ci[WS(csi, 27)] = T9X + T9Y; | |
2971 Ci[WS(csi, 37)] = T9Y - T9X; | |
2972 Ta3 = T9L - T9M; | |
2973 Ta4 = Ta1 - Ta0; | |
2974 Cr[WS(csr, 37)] = Ta3 - Ta4; | |
2975 Cr[WS(csr, 27)] = Ta3 + Ta4; | |
2976 } | |
2977 } | |
2978 { | |
2979 E T2z, T69, T3g, T6j, T60, T6k, T5P, T6a, T5L, T6p, T65, T6h, T4w, T6o, T64; | |
2980 E T6e; | |
2981 { | |
2982 E T2f, T2y, T5N, T5O; | |
2983 T2f = T27 + T2e; | |
2984 T2y = T2o + T2x; | |
2985 T2z = T2f + T2y; | |
2986 T69 = T2f - T2y; | |
2987 { | |
2988 E T2U, T3f, T5W, T5Z; | |
2989 T2U = FMA(KP098017140, T2M, KP995184726 * T2T); | |
2990 T3f = FNMS(KP098017140, T3e, KP995184726 * T37); | |
2991 T3g = T2U + T3f; | |
2992 T6j = T3f - T2U; | |
2993 T5W = T5S + T5V; | |
2994 T5Z = T5X + T5Y; | |
2995 T60 = T5W + T5Z; | |
2996 T6k = T5Z - T5W; | |
2997 } | |
2998 T5N = FNMS(KP098017140, T2T, KP995184726 * T2M); | |
2999 T5O = FMA(KP995184726, T3e, KP098017140 * T37); | |
3000 T5P = T5N + T5O; | |
3001 T6a = T5O - T5N; | |
3002 { | |
3003 E T5x, T6f, T5K, T6g, T5w, T5J; | |
3004 T5w = T5e + T5v; | |
3005 T5x = T4X + T5w; | |
3006 T6f = T4X - T5w; | |
3007 T5J = T5H + T5I; | |
3008 T5K = T5G + T5J; | |
3009 T6g = T5J - T5G; | |
3010 T5L = FNMS(KP049067674, T5K, KP998795456 * T5x); | |
3011 T6p = FMA(KP671558954, T6f, KP740951125 * T6g); | |
3012 T65 = FMA(KP049067674, T5x, KP998795456 * T5K); | |
3013 T6h = FNMS(KP671558954, T6g, KP740951125 * T6f); | |
3014 } | |
3015 { | |
3016 E T4i, T6c, T4v, T6d, T4h, T4u; | |
3017 T4h = T3Z + T4g; | |
3018 T4i = T3I + T4h; | |
3019 T6c = T3I - T4h; | |
3020 T4u = T4s + T4t; | |
3021 T4v = T4r + T4u; | |
3022 T6d = T4u - T4r; | |
3023 T4w = FMA(KP998795456, T4i, KP049067674 * T4v); | |
3024 T6o = FNMS(KP671558954, T6c, KP740951125 * T6d); | |
3025 T64 = FNMS(KP049067674, T4i, KP998795456 * T4v); | |
3026 T6e = FMA(KP740951125, T6c, KP671558954 * T6d); | |
3027 } | |
3028 } | |
3029 { | |
3030 E T3h, T5M, T63, T66; | |
3031 T3h = T2z + T3g; | |
3032 T5M = T4w + T5L; | |
3033 Cr[WS(csr, 63)] = T3h - T5M; | |
3034 Cr[WS(csr, 1)] = T3h + T5M; | |
3035 T63 = T60 + T5P; | |
3036 T66 = T64 + T65; | |
3037 Ci[WS(csi, 1)] = T63 + T66; | |
3038 Ci[WS(csi, 63)] = T66 - T63; | |
3039 } | |
3040 { | |
3041 E T61, T62, T67, T68; | |
3042 T61 = T5P - T60; | |
3043 T62 = T5L - T4w; | |
3044 Ci[WS(csi, 31)] = T61 + T62; | |
3045 Ci[WS(csi, 33)] = T62 - T61; | |
3046 T67 = T2z - T3g; | |
3047 T68 = T65 - T64; | |
3048 Cr[WS(csr, 33)] = T67 - T68; | |
3049 Cr[WS(csr, 31)] = T67 + T68; | |
3050 } | |
3051 { | |
3052 E T6b, T6i, T6n, T6q; | |
3053 T6b = T69 + T6a; | |
3054 T6i = T6e + T6h; | |
3055 Cr[WS(csr, 49)] = T6b - T6i; | |
3056 Cr[WS(csr, 15)] = T6b + T6i; | |
3057 T6n = T6k + T6j; | |
3058 T6q = T6o + T6p; | |
3059 Ci[WS(csi, 15)] = T6n + T6q; | |
3060 Ci[WS(csi, 49)] = T6q - T6n; | |
3061 } | |
3062 { | |
3063 E T6l, T6m, T6r, T6s; | |
3064 T6l = T6j - T6k; | |
3065 T6m = T6h - T6e; | |
3066 Ci[WS(csi, 17)] = T6l + T6m; | |
3067 Ci[WS(csi, 47)] = T6m - T6l; | |
3068 T6r = T69 - T6a; | |
3069 T6s = T6p - T6o; | |
3070 Cr[WS(csr, 47)] = T6r - T6s; | |
3071 Cr[WS(csr, 17)] = T6r + T6s; | |
3072 } | |
3073 } | |
3074 { | |
3075 E T7B, T8N, T7Q, T8X, T8E, T8Y, T8x, T8O, T8t, T93, T8J, T8V, T8a, T92, T8I; | |
3076 E T8S; | |
3077 { | |
3078 E T7t, T7A, T8v, T8w; | |
3079 T7t = T7r + T7s; | |
3080 T7A = T7w + T7z; | |
3081 T7B = T7t + T7A; | |
3082 T8N = T7t - T7A; | |
3083 { | |
3084 E T7I, T7P, T8A, T8D; | |
3085 T7I = FMA(KP956940335, T7E, KP290284677 * T7H); | |
3086 T7P = FNMS(KP290284677, T7O, KP956940335 * T7L); | |
3087 T7Q = T7I + T7P; | |
3088 T8X = T7P - T7I; | |
3089 T8A = T8y + T8z; | |
3090 T8D = T8B + T8C; | |
3091 T8E = T8A + T8D; | |
3092 T8Y = T8D - T8A; | |
3093 } | |
3094 T8v = FNMS(KP290284677, T7E, KP956940335 * T7H); | |
3095 T8w = FMA(KP290284677, T7L, KP956940335 * T7O); | |
3096 T8x = T8v + T8w; | |
3097 T8O = T8w - T8v; | |
3098 { | |
3099 E T8l, T8T, T8s, T8U, T8k, T8r; | |
3100 T8k = T8g + T8j; | |
3101 T8l = T8d + T8k; | |
3102 T8T = T8d - T8k; | |
3103 T8r = T8p + T8q; | |
3104 T8s = T8o + T8r; | |
3105 T8U = T8r - T8o; | |
3106 T8t = FNMS(KP146730474, T8s, KP989176509 * T8l); | |
3107 T93 = FMA(KP595699304, T8T, KP803207531 * T8U); | |
3108 T8J = FMA(KP146730474, T8l, KP989176509 * T8s); | |
3109 T8V = FNMS(KP595699304, T8U, KP803207531 * T8T); | |
3110 } | |
3111 { | |
3112 E T82, T8Q, T89, T8R, T81, T88; | |
3113 T81 = T7X + T80; | |
3114 T82 = T7U + T81; | |
3115 T8Q = T7U - T81; | |
3116 T88 = T86 + T87; | |
3117 T89 = T85 + T88; | |
3118 T8R = T88 - T85; | |
3119 T8a = FMA(KP989176509, T82, KP146730474 * T89); | |
3120 T92 = FNMS(KP595699304, T8Q, KP803207531 * T8R); | |
3121 T8I = FNMS(KP146730474, T82, KP989176509 * T89); | |
3122 T8S = FMA(KP803207531, T8Q, KP595699304 * T8R); | |
3123 } | |
3124 } | |
3125 { | |
3126 E T7R, T8u, T8H, T8K; | |
3127 T7R = T7B + T7Q; | |
3128 T8u = T8a + T8t; | |
3129 Cr[WS(csr, 61)] = T7R - T8u; | |
3130 Cr[WS(csr, 3)] = T7R + T8u; | |
3131 T8H = T8E + T8x; | |
3132 T8K = T8I + T8J; | |
3133 Ci[WS(csi, 3)] = T8H + T8K; | |
3134 Ci[WS(csi, 61)] = T8K - T8H; | |
3135 } | |
3136 { | |
3137 E T8F, T8G, T8L, T8M; | |
3138 T8F = T8x - T8E; | |
3139 T8G = T8t - T8a; | |
3140 Ci[WS(csi, 29)] = T8F + T8G; | |
3141 Ci[WS(csi, 35)] = T8G - T8F; | |
3142 T8L = T7B - T7Q; | |
3143 T8M = T8J - T8I; | |
3144 Cr[WS(csr, 35)] = T8L - T8M; | |
3145 Cr[WS(csr, 29)] = T8L + T8M; | |
3146 } | |
3147 { | |
3148 E T8P, T8W, T91, T94; | |
3149 T8P = T8N + T8O; | |
3150 T8W = T8S + T8V; | |
3151 Cr[WS(csr, 51)] = T8P - T8W; | |
3152 Cr[WS(csr, 13)] = T8P + T8W; | |
3153 T91 = T8Y + T8X; | |
3154 T94 = T92 + T93; | |
3155 Ci[WS(csi, 13)] = T91 + T94; | |
3156 Ci[WS(csi, 51)] = T94 - T91; | |
3157 } | |
3158 { | |
3159 E T8Z, T90, T95, T96; | |
3160 T8Z = T8X - T8Y; | |
3161 T90 = T8V - T8S; | |
3162 Ci[WS(csi, 19)] = T8Z + T90; | |
3163 Ci[WS(csi, 45)] = T90 - T8Z; | |
3164 T95 = T8N - T8O; | |
3165 T96 = T93 - T92; | |
3166 Cr[WS(csr, 45)] = T95 - T96; | |
3167 Cr[WS(csr, 19)] = T95 + T96; | |
3168 } | |
3169 } | |
3170 } | |
3171 } | |
3172 } | |
3173 | |
3174 static const kr2c_desc desc = { 128, "r2cf_128", {812, 186, 144, 0}, &GENUS }; | |
3175 | |
3176 void X(codelet_r2cf_128) (planner *p) { | |
3177 X(kr2c_register) (p, r2cf_128, &desc); | |
3178 } | |
3179 | |
3180 #endif /* HAVE_FMA */ |