Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/rdft/scalar/r2cf/r2cf_128.c @ 167:bd3cc4d1df30
Add FFTW 3.3.8 source, and a Linux build
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Tue, 19 Nov 2019 14:52:55 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
166:cbd6d7e562c7 | 167:bd3cc4d1df30 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-14 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Thu May 24 08:06:27 EDT 2018 */ | |
23 | |
24 #include "rdft/codelet-rdft.h" | |
25 | |
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA) | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 128 -name r2cf_128 -include rdft/scalar/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 * 186 stack variables, 31 constants, and 256 memory accesses | |
34 */ | |
35 #include "rdft/scalar/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(KP803207531, +0.803207531480644909806676512963141923879569427); | |
40 DK(KP989176509, +0.989176509964780973451673738016243063983689533); | |
41 DK(KP741650546, +0.741650546272035369581266691172079863842265220); | |
42 DK(KP148335987, +0.148335987538347428753676511486911367000625355); | |
43 DK(KP740951125, +0.740951125354959091175616897495162729728955309); | |
44 DK(KP998795456, +0.998795456205172392714771604759100694443203615); | |
45 DK(KP906347169, +0.906347169019147157946142717268914412664134293); | |
46 DK(KP049126849, +0.049126849769467254105343321271313617079695752); | |
47 DK(KP970031253, +0.970031253194543992603984207286100251456865962); | |
48 DK(KP857728610, +0.857728610000272069902269984284770137042490799); | |
49 DK(KP250486960, +0.250486960191305461595702160124721208578685568); | |
50 DK(KP599376933, +0.599376933681923766271389869014404232837890546); | |
51 DK(KP941544065, +0.941544065183020778412509402599502357185589796); | |
52 DK(KP903989293, +0.903989293123443331586200297230537048710132025); | |
53 DK(KP357805721, +0.357805721314524104672487743774474392487532769); | |
54 DK(KP472964775, +0.472964775891319928124438237972992463904131113); | |
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(KP668178637, +0.668178637919298919997757686523080761552472251); | |
67 DK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
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 TcD, TdR, T27, T7r, T5P, T8v, Tf, Ta5, Tu, Tbn, TcG, TdS, T2e, T8w, T5S; | |
74 E T7s, TK, Ta6, TcK, TdU, T2o, T5U, T7w, T8y, TZ, Ta7, TcN, TdV, T2x, T5V; | |
75 E T7z, T8z, T1g, Taa, TcU, Tex, TcX, Tew, T1v, Tab, T2M, T6A, T7E, T9b, T7H; | |
76 E T9a, T2T, T6z, T4X, T6L, Tdz, TeL, TdK, TeO, T5G, T6O, T8d, T9p, TaV, Tc3; | |
77 E Tbi, Tc4, T8o, T9s, T3I, T6E, Tde, TeE, Tdp, TeH, T4r, T6H, T7U, T9i, Tao; | |
78 E TbW, TaL, TbX, T85, T9l, T1L, Tad, Td3, TeA, Td6, Tez, T20, Tae, T37, T6x; | |
79 E T7L, T9e, T7O, T9d, T3e, T6w, TbZ, Tc0, T3Z, T4s, Tds, TeF, T4g, T4t, T80; | |
80 E T87, Tdl, TeI, T7X, T86, TaD, TaM, Tc6, Tc7, T5e, T5H, TdN, TeM, T5v, T5I; | |
81 E T8j, T8q, TdG, TeP, T8g, T8p, Tba, Tbj; | |
82 { | |
83 E T3, T23, Td, T25, T6, T5N, Ta, T24; | |
84 { | |
85 E T1, T2, Tb, Tc; | |
86 T1 = R0[0]; | |
87 T2 = R0[WS(rs, 32)]; | |
88 T3 = T1 + T2; | |
89 T23 = T1 - T2; | |
90 Tb = R0[WS(rs, 56)]; | |
91 Tc = R0[WS(rs, 24)]; | |
92 Td = Tb + Tc; | |
93 T25 = Tb - Tc; | |
94 } | |
95 { | |
96 E T4, T5, T8, T9; | |
97 T4 = R0[WS(rs, 16)]; | |
98 T5 = R0[WS(rs, 48)]; | |
99 T6 = T4 + T5; | |
100 T5N = T4 - T5; | |
101 T8 = R0[WS(rs, 8)]; | |
102 T9 = R0[WS(rs, 40)]; | |
103 Ta = T8 + T9; | |
104 T24 = T8 - T9; | |
105 } | |
106 TcD = T3 - T6; | |
107 TdR = Td - Ta; | |
108 { | |
109 E T26, T5O, T7, Te; | |
110 T26 = T24 + T25; | |
111 T27 = FMA(KP707106781, T26, T23); | |
112 T7r = FNMS(KP707106781, T26, T23); | |
113 T5O = T25 - T24; | |
114 T5P = FNMS(KP707106781, T5O, T5N); | |
115 T8v = FMA(KP707106781, T5O, T5N); | |
116 T7 = T3 + T6; | |
117 Te = Ta + Td; | |
118 Tf = T7 + Te; | |
119 Ta5 = T7 - Te; | |
120 } | |
121 } | |
122 { | |
123 E Ti, T28, Ts, T2c, Tl, T29, Tp, T2b; | |
124 { | |
125 E Tg, Th, Tq, Tr; | |
126 Tg = R0[WS(rs, 4)]; | |
127 Th = R0[WS(rs, 36)]; | |
128 Ti = Tg + Th; | |
129 T28 = Tg - Th; | |
130 Tq = R0[WS(rs, 12)]; | |
131 Tr = R0[WS(rs, 44)]; | |
132 Ts = Tq + Tr; | |
133 T2c = Tq - Tr; | |
134 } | |
135 { | |
136 E Tj, Tk, Tn, To; | |
137 Tj = R0[WS(rs, 20)]; | |
138 Tk = R0[WS(rs, 52)]; | |
139 Tl = Tj + Tk; | |
140 T29 = Tj - Tk; | |
141 Tn = R0[WS(rs, 60)]; | |
142 To = R0[WS(rs, 28)]; | |
143 Tp = Tn + To; | |
144 T2b = Tn - To; | |
145 } | |
146 { | |
147 E Tm, Tt, TcE, TcF; | |
148 Tm = Ti + Tl; | |
149 Tt = Tp + Ts; | |
150 Tu = Tm + Tt; | |
151 Tbn = Tt - Tm; | |
152 TcE = Ti - Tl; | |
153 TcF = Tp - Ts; | |
154 TcG = TcE + TcF; | |
155 TdS = TcF - TcE; | |
156 } | |
157 { | |
158 E T2a, T2d, T5Q, T5R; | |
159 T2a = FNMS(KP414213562, T29, T28); | |
160 T2d = FMA(KP414213562, T2c, T2b); | |
161 T2e = T2a + T2d; | |
162 T8w = T2d - T2a; | |
163 T5Q = FMA(KP414213562, T28, T29); | |
164 T5R = FNMS(KP414213562, T2b, T2c); | |
165 T5S = T5Q + T5R; | |
166 T7s = T5Q - T5R; | |
167 } | |
168 } | |
169 { | |
170 E Ty, T2g, TB, T2l, TF, T2m, TI, T2j; | |
171 { | |
172 E Tw, Tx, Tz, TA; | |
173 Tw = R0[WS(rs, 2)]; | |
174 Tx = R0[WS(rs, 34)]; | |
175 Ty = Tw + Tx; | |
176 T2g = Tw - Tx; | |
177 Tz = R0[WS(rs, 18)]; | |
178 TA = R0[WS(rs, 50)]; | |
179 TB = Tz + TA; | |
180 T2l = Tz - TA; | |
181 { | |
182 E TD, TE, T2h, TG, TH, T2i; | |
183 TD = R0[WS(rs, 10)]; | |
184 TE = R0[WS(rs, 42)]; | |
185 T2h = TD - TE; | |
186 TG = R0[WS(rs, 58)]; | |
187 TH = R0[WS(rs, 26)]; | |
188 T2i = TG - TH; | |
189 TF = TD + TE; | |
190 T2m = T2h - T2i; | |
191 TI = TG + TH; | |
192 T2j = T2h + T2i; | |
193 } | |
194 } | |
195 { | |
196 E TC, TJ, TcI, TcJ; | |
197 TC = Ty + TB; | |
198 TJ = TF + TI; | |
199 TK = TC + TJ; | |
200 Ta6 = TC - TJ; | |
201 TcI = Ty - TB; | |
202 TcJ = TI - TF; | |
203 TcK = FMA(KP414213562, TcJ, TcI); | |
204 TdU = FNMS(KP414213562, TcI, TcJ); | |
205 } | |
206 { | |
207 E T2k, T2n, T7u, T7v; | |
208 T2k = FMA(KP707106781, T2j, T2g); | |
209 T2n = FMA(KP707106781, T2m, T2l); | |
210 T2o = FNMS(KP198912367, T2n, T2k); | |
211 T5U = FMA(KP198912367, T2k, T2n); | |
212 T7u = FNMS(KP707106781, T2j, T2g); | |
213 T7v = FNMS(KP707106781, T2m, T2l); | |
214 T7w = FMA(KP668178637, T7v, T7u); | |
215 T8y = FNMS(KP668178637, T7u, T7v); | |
216 } | |
217 } | |
218 { | |
219 E TN, T2p, TQ, T2u, TU, T2v, TX, T2s; | |
220 { | |
221 E TL, TM, TO, TP; | |
222 TL = R0[WS(rs, 62)]; | |
223 TM = R0[WS(rs, 30)]; | |
224 TN = TL + TM; | |
225 T2p = TL - TM; | |
226 TO = R0[WS(rs, 14)]; | |
227 TP = R0[WS(rs, 46)]; | |
228 TQ = TO + TP; | |
229 T2u = TO - TP; | |
230 { | |
231 E TS, TT, T2q, TV, TW, T2r; | |
232 TS = R0[WS(rs, 6)]; | |
233 TT = R0[WS(rs, 38)]; | |
234 T2q = TS - TT; | |
235 TV = R0[WS(rs, 54)]; | |
236 TW = R0[WS(rs, 22)]; | |
237 T2r = TV - TW; | |
238 TU = TS + TT; | |
239 T2v = T2q - T2r; | |
240 TX = TV + TW; | |
241 T2s = T2q + T2r; | |
242 } | |
243 } | |
244 { | |
245 E TR, TY, TcL, TcM; | |
246 TR = TN + TQ; | |
247 TY = TU + TX; | |
248 TZ = TR + TY; | |
249 Ta7 = TR - TY; | |
250 TcL = TN - TQ; | |
251 TcM = TX - TU; | |
252 TcN = FNMS(KP414213562, TcM, TcL); | |
253 TdV = FMA(KP414213562, TcL, TcM); | |
254 } | |
255 { | |
256 E T2t, T2w, T7x, T7y; | |
257 T2t = FMA(KP707106781, T2s, T2p); | |
258 T2w = FMA(KP707106781, T2v, T2u); | |
259 T2x = FMA(KP198912367, T2w, T2t); | |
260 T5V = FNMS(KP198912367, T2t, T2w); | |
261 T7x = FNMS(KP707106781, T2s, T2p); | |
262 T7y = FNMS(KP707106781, T2v, T2u); | |
263 T7z = FNMS(KP668178637, T7y, T7x); | |
264 T8z = FMA(KP668178637, T7x, T7y); | |
265 } | |
266 } | |
267 { | |
268 E T14, T2A, T17, T2N, T1b, T2O, T1e, T2D, T1j, T1m, T2H, TcR, T2Q, T1q, T1t; | |
269 E T2K, TcS, T2R; | |
270 { | |
271 E T12, T13, T15, T16; | |
272 T12 = R0[WS(rs, 1)]; | |
273 T13 = R0[WS(rs, 33)]; | |
274 T14 = T12 + T13; | |
275 T2A = T12 - T13; | |
276 T15 = R0[WS(rs, 17)]; | |
277 T16 = R0[WS(rs, 49)]; | |
278 T17 = T15 + T16; | |
279 T2N = T15 - T16; | |
280 } | |
281 { | |
282 E T19, T1a, T2B, T1c, T1d, T2C; | |
283 T19 = R0[WS(rs, 9)]; | |
284 T1a = R0[WS(rs, 41)]; | |
285 T2B = T19 - T1a; | |
286 T1c = R0[WS(rs, 57)]; | |
287 T1d = R0[WS(rs, 25)]; | |
288 T2C = T1c - T1d; | |
289 T1b = T19 + T1a; | |
290 T2O = T2B - T2C; | |
291 T1e = T1c + T1d; | |
292 T2D = T2B + T2C; | |
293 } | |
294 { | |
295 E T2F, T2G, T2I, T2J; | |
296 { | |
297 E T1h, T1i, T1k, T1l; | |
298 T1h = R0[WS(rs, 5)]; | |
299 T1i = R0[WS(rs, 37)]; | |
300 T1j = T1h + T1i; | |
301 T2F = T1h - T1i; | |
302 T1k = R0[WS(rs, 21)]; | |
303 T1l = R0[WS(rs, 53)]; | |
304 T1m = T1k + T1l; | |
305 T2G = T1k - T1l; | |
306 } | |
307 T2H = FNMS(KP414213562, T2G, T2F); | |
308 TcR = T1j - T1m; | |
309 T2Q = FMA(KP414213562, T2F, T2G); | |
310 { | |
311 E T1o, T1p, T1r, T1s; | |
312 T1o = R0[WS(rs, 61)]; | |
313 T1p = R0[WS(rs, 29)]; | |
314 T1q = T1o + T1p; | |
315 T2I = T1o - T1p; | |
316 T1r = R0[WS(rs, 13)]; | |
317 T1s = R0[WS(rs, 45)]; | |
318 T1t = T1r + T1s; | |
319 T2J = T1r - T1s; | |
320 } | |
321 T2K = FMA(KP414213562, T2J, T2I); | |
322 TcS = T1q - T1t; | |
323 T2R = FNMS(KP414213562, T2I, T2J); | |
324 } | |
325 { | |
326 E T18, T1f, TcQ, TcT; | |
327 T18 = T14 + T17; | |
328 T1f = T1b + T1e; | |
329 T1g = T18 + T1f; | |
330 Taa = T18 - T1f; | |
331 TcQ = T14 - T17; | |
332 TcT = TcR + TcS; | |
333 TcU = FMA(KP707106781, TcT, TcQ); | |
334 Tex = FNMS(KP707106781, TcT, TcQ); | |
335 } | |
336 { | |
337 E TcV, TcW, T1n, T1u; | |
338 TcV = T1e - T1b; | |
339 TcW = TcS - TcR; | |
340 TcX = FMA(KP707106781, TcW, TcV); | |
341 Tew = FNMS(KP707106781, TcW, TcV); | |
342 T1n = T1j + T1m; | |
343 T1u = T1q + T1t; | |
344 T1v = T1n + T1u; | |
345 Tab = T1u - T1n; | |
346 } | |
347 { | |
348 E T2E, T2L, T7C, T7D; | |
349 T2E = FMA(KP707106781, T2D, T2A); | |
350 T2L = T2H + T2K; | |
351 T2M = FMA(KP923879532, T2L, T2E); | |
352 T6A = FNMS(KP923879532, T2L, T2E); | |
353 T7C = FNMS(KP707106781, T2D, T2A); | |
354 T7D = T2Q - T2R; | |
355 T7E = FMA(KP923879532, T7D, T7C); | |
356 T9b = FNMS(KP923879532, T7D, T7C); | |
357 } | |
358 { | |
359 E T7F, T7G, T2P, T2S; | |
360 T7F = FNMS(KP707106781, T2O, T2N); | |
361 T7G = T2K - T2H; | |
362 T7H = FMA(KP923879532, T7G, T7F); | |
363 T9a = FNMS(KP923879532, T7G, T7F); | |
364 T2P = FMA(KP707106781, T2O, T2N); | |
365 T2S = T2Q + T2R; | |
366 T2T = FMA(KP923879532, T2S, T2P); | |
367 T6z = FNMS(KP923879532, T2S, T2P); | |
368 } | |
369 } | |
370 { | |
371 E T4z, TaP, T5A, TaQ, T4G, TaT, T5B, TaS, Tbf, Tbg, T4O, Tdw, T5D, Tbc, Tbd; | |
372 E T4V, Tdx, T5E; | |
373 { | |
374 E T4x, T4y, T5y, T5z; | |
375 T4x = R1[WS(rs, 63)]; | |
376 T4y = R1[WS(rs, 31)]; | |
377 T4z = T4x - T4y; | |
378 TaP = T4x + T4y; | |
379 T5y = R1[WS(rs, 47)]; | |
380 T5z = R1[WS(rs, 15)]; | |
381 T5A = T5y - T5z; | |
382 TaQ = T5z + T5y; | |
383 } | |
384 { | |
385 E T4A, T4B, T4C, T4D, T4E, T4F; | |
386 T4A = R1[WS(rs, 7)]; | |
387 T4B = R1[WS(rs, 39)]; | |
388 T4C = T4A - T4B; | |
389 T4D = R1[WS(rs, 55)]; | |
390 T4E = R1[WS(rs, 23)]; | |
391 T4F = T4D - T4E; | |
392 T4G = T4C + T4F; | |
393 TaT = T4D + T4E; | |
394 T5B = T4F - T4C; | |
395 TaS = T4A + T4B; | |
396 } | |
397 { | |
398 E T4K, T4N, T4R, T4U; | |
399 { | |
400 E T4I, T4J, T4L, T4M; | |
401 T4I = R1[WS(rs, 3)]; | |
402 T4J = R1[WS(rs, 35)]; | |
403 T4K = T4I - T4J; | |
404 Tbf = T4I + T4J; | |
405 T4L = R1[WS(rs, 51)]; | |
406 T4M = R1[WS(rs, 19)]; | |
407 T4N = T4L - T4M; | |
408 Tbg = T4M + T4L; | |
409 } | |
410 T4O = FMA(KP414213562, T4N, T4K); | |
411 Tdw = Tbf - Tbg; | |
412 T5D = FNMS(KP414213562, T4K, T4N); | |
413 { | |
414 E T4P, T4Q, T4S, T4T; | |
415 T4P = R1[WS(rs, 59)]; | |
416 T4Q = R1[WS(rs, 27)]; | |
417 T4R = T4P - T4Q; | |
418 Tbc = T4P + T4Q; | |
419 T4S = R1[WS(rs, 43)]; | |
420 T4T = R1[WS(rs, 11)]; | |
421 T4U = T4S - T4T; | |
422 Tbd = T4T + T4S; | |
423 } | |
424 T4V = FNMS(KP414213562, T4U, T4R); | |
425 Tdx = Tbc - Tbd; | |
426 T5E = FMA(KP414213562, T4R, T4U); | |
427 } | |
428 { | |
429 E T4H, T4W, Tdv, Tdy; | |
430 T4H = FMA(KP707106781, T4G, T4z); | |
431 T4W = T4O + T4V; | |
432 T4X = FMA(KP923879532, T4W, T4H); | |
433 T6L = FNMS(KP923879532, T4W, T4H); | |
434 Tdv = TaP - TaQ; | |
435 Tdy = Tdw + Tdx; | |
436 Tdz = FMA(KP707106781, Tdy, Tdv); | |
437 TeL = FNMS(KP707106781, Tdy, Tdv); | |
438 } | |
439 { | |
440 E TdI, TdJ, T5C, T5F; | |
441 TdI = TaT - TaS; | |
442 TdJ = Tdx - Tdw; | |
443 TdK = FMA(KP707106781, TdJ, TdI); | |
444 TeO = FNMS(KP707106781, TdJ, TdI); | |
445 T5C = FMA(KP707106781, T5B, T5A); | |
446 T5F = T5D + T5E; | |
447 T5G = FMA(KP923879532, T5F, T5C); | |
448 T6O = FNMS(KP923879532, T5F, T5C); | |
449 } | |
450 { | |
451 E T8b, T8c, TaR, TaU; | |
452 T8b = FNMS(KP707106781, T4G, T4z); | |
453 T8c = T5E - T5D; | |
454 T8d = FMA(KP923879532, T8c, T8b); | |
455 T9p = FNMS(KP923879532, T8c, T8b); | |
456 TaR = TaP + TaQ; | |
457 TaU = TaS + TaT; | |
458 TaV = TaR - TaU; | |
459 Tc3 = TaR + TaU; | |
460 } | |
461 { | |
462 E Tbe, Tbh, T8m, T8n; | |
463 Tbe = Tbc + Tbd; | |
464 Tbh = Tbf + Tbg; | |
465 Tbi = Tbe - Tbh; | |
466 Tc4 = Tbh + Tbe; | |
467 T8m = FNMS(KP707106781, T5B, T5A); | |
468 T8n = T4V - T4O; | |
469 T8o = FNMS(KP923879532, T8n, T8m); | |
470 T9s = FMA(KP923879532, T8n, T8m); | |
471 } | |
472 } | |
473 { | |
474 E T3k, Tai, T4l, Taj, T3r, Tam, T4m, Tal, TaI, TaJ, T3z, Tdb, T4o, TaF, TaG; | |
475 E T3G, Tdc, T4p; | |
476 { | |
477 E T3i, T3j, T4j, T4k; | |
478 T3i = R1[0]; | |
479 T3j = R1[WS(rs, 32)]; | |
480 T3k = T3i - T3j; | |
481 Tai = T3i + T3j; | |
482 T4j = R1[WS(rs, 16)]; | |
483 T4k = R1[WS(rs, 48)]; | |
484 T4l = T4j - T4k; | |
485 Taj = T4j + T4k; | |
486 } | |
487 { | |
488 E T3l, T3m, T3n, T3o, T3p, T3q; | |
489 T3l = R1[WS(rs, 8)]; | |
490 T3m = R1[WS(rs, 40)]; | |
491 T3n = T3l - T3m; | |
492 T3o = R1[WS(rs, 56)]; | |
493 T3p = R1[WS(rs, 24)]; | |
494 T3q = T3o - T3p; | |
495 T3r = T3n + T3q; | |
496 Tam = T3o + T3p; | |
497 T4m = T3n - T3q; | |
498 Tal = T3l + T3m; | |
499 } | |
500 { | |
501 E T3v, T3y, T3C, T3F; | |
502 { | |
503 E T3t, T3u, T3w, T3x; | |
504 T3t = R1[WS(rs, 4)]; | |
505 T3u = R1[WS(rs, 36)]; | |
506 T3v = T3t - T3u; | |
507 TaI = T3t + T3u; | |
508 T3w = R1[WS(rs, 20)]; | |
509 T3x = R1[WS(rs, 52)]; | |
510 T3y = T3w - T3x; | |
511 TaJ = T3w + T3x; | |
512 } | |
513 T3z = FNMS(KP414213562, T3y, T3v); | |
514 Tdb = TaI - TaJ; | |
515 T4o = FMA(KP414213562, T3v, T3y); | |
516 { | |
517 E T3A, T3B, T3D, T3E; | |
518 T3A = R1[WS(rs, 60)]; | |
519 T3B = R1[WS(rs, 28)]; | |
520 T3C = T3A - T3B; | |
521 TaF = T3A + T3B; | |
522 T3D = R1[WS(rs, 12)]; | |
523 T3E = R1[WS(rs, 44)]; | |
524 T3F = T3D - T3E; | |
525 TaG = T3D + T3E; | |
526 } | |
527 T3G = FMA(KP414213562, T3F, T3C); | |
528 Tdc = TaF - TaG; | |
529 T4p = FNMS(KP414213562, T3C, T3F); | |
530 } | |
531 { | |
532 E T3s, T3H, Tda, Tdd; | |
533 T3s = FMA(KP707106781, T3r, T3k); | |
534 T3H = T3z + T3G; | |
535 T3I = FMA(KP923879532, T3H, T3s); | |
536 T6E = FNMS(KP923879532, T3H, T3s); | |
537 Tda = Tai - Taj; | |
538 Tdd = Tdb + Tdc; | |
539 Tde = FMA(KP707106781, Tdd, Tda); | |
540 TeE = FNMS(KP707106781, Tdd, Tda); | |
541 } | |
542 { | |
543 E Tdn, Tdo, T4n, T4q; | |
544 Tdn = Tam - Tal; | |
545 Tdo = Tdc - Tdb; | |
546 Tdp = FMA(KP707106781, Tdo, Tdn); | |
547 TeH = FNMS(KP707106781, Tdo, Tdn); | |
548 T4n = FMA(KP707106781, T4m, T4l); | |
549 T4q = T4o + T4p; | |
550 T4r = FMA(KP923879532, T4q, T4n); | |
551 T6H = FNMS(KP923879532, T4q, T4n); | |
552 } | |
553 { | |
554 E T7S, T7T, Tak, Tan; | |
555 T7S = FNMS(KP707106781, T3r, T3k); | |
556 T7T = T4o - T4p; | |
557 T7U = FMA(KP923879532, T7T, T7S); | |
558 T9i = FNMS(KP923879532, T7T, T7S); | |
559 Tak = Tai + Taj; | |
560 Tan = Tal + Tam; | |
561 Tao = Tak - Tan; | |
562 TbW = Tak + Tan; | |
563 } | |
564 { | |
565 E TaH, TaK, T83, T84; | |
566 TaH = TaF + TaG; | |
567 TaK = TaI + TaJ; | |
568 TaL = TaH - TaK; | |
569 TbX = TaK + TaH; | |
570 T83 = FNMS(KP707106781, T4m, T4l); | |
571 T84 = T3G - T3z; | |
572 T85 = FMA(KP923879532, T84, T83); | |
573 T9l = FNMS(KP923879532, T84, T83); | |
574 } | |
575 } | |
576 { | |
577 E T1z, T2V, T1C, T38, T1G, T39, T1J, T2Y, T1O, T1R, T32, Td0, T3b, T1V, T1Y; | |
578 E T35, Td1, T3c; | |
579 { | |
580 E T1x, T1y, T1A, T1B; | |
581 T1x = R0[WS(rs, 63)]; | |
582 T1y = R0[WS(rs, 31)]; | |
583 T1z = T1x + T1y; | |
584 T2V = T1x - T1y; | |
585 T1A = R0[WS(rs, 15)]; | |
586 T1B = R0[WS(rs, 47)]; | |
587 T1C = T1A + T1B; | |
588 T38 = T1A - T1B; | |
589 } | |
590 { | |
591 E T1E, T1F, T2W, T1H, T1I, T2X; | |
592 T1E = R0[WS(rs, 7)]; | |
593 T1F = R0[WS(rs, 39)]; | |
594 T2W = T1E - T1F; | |
595 T1H = R0[WS(rs, 55)]; | |
596 T1I = R0[WS(rs, 23)]; | |
597 T2X = T1H - T1I; | |
598 T1G = T1E + T1F; | |
599 T39 = T2W - T2X; | |
600 T1J = T1H + T1I; | |
601 T2Y = T2W + T2X; | |
602 } | |
603 { | |
604 E T30, T31, T33, T34; | |
605 { | |
606 E T1M, T1N, T1P, T1Q; | |
607 T1M = R0[WS(rs, 3)]; | |
608 T1N = R0[WS(rs, 35)]; | |
609 T1O = T1M + T1N; | |
610 T30 = T1M - T1N; | |
611 T1P = R0[WS(rs, 19)]; | |
612 T1Q = R0[WS(rs, 51)]; | |
613 T1R = T1P + T1Q; | |
614 T31 = T1P - T1Q; | |
615 } | |
616 T32 = FNMS(KP414213562, T31, T30); | |
617 Td0 = T1O - T1R; | |
618 T3b = FMA(KP414213562, T30, T31); | |
619 { | |
620 E T1T, T1U, T1W, T1X; | |
621 T1T = R0[WS(rs, 59)]; | |
622 T1U = R0[WS(rs, 27)]; | |
623 T1V = T1T + T1U; | |
624 T33 = T1T - T1U; | |
625 T1W = R0[WS(rs, 11)]; | |
626 T1X = R0[WS(rs, 43)]; | |
627 T1Y = T1W + T1X; | |
628 T34 = T1W - T1X; | |
629 } | |
630 T35 = FMA(KP414213562, T34, T33); | |
631 Td1 = T1V - T1Y; | |
632 T3c = FNMS(KP414213562, T33, T34); | |
633 } | |
634 { | |
635 E T1D, T1K, TcZ, Td2; | |
636 T1D = T1z + T1C; | |
637 T1K = T1G + T1J; | |
638 T1L = T1D + T1K; | |
639 Tad = T1D - T1K; | |
640 TcZ = T1z - T1C; | |
641 Td2 = Td0 + Td1; | |
642 Td3 = FMA(KP707106781, Td2, TcZ); | |
643 TeA = FNMS(KP707106781, Td2, TcZ); | |
644 } | |
645 { | |
646 E Td4, Td5, T1S, T1Z; | |
647 Td4 = T1J - T1G; | |
648 Td5 = Td1 - Td0; | |
649 Td6 = FMA(KP707106781, Td5, Td4); | |
650 Tez = FNMS(KP707106781, Td5, Td4); | |
651 T1S = T1O + T1R; | |
652 T1Z = T1V + T1Y; | |
653 T20 = T1S + T1Z; | |
654 Tae = T1Z - T1S; | |
655 } | |
656 { | |
657 E T2Z, T36, T7J, T7K; | |
658 T2Z = FMA(KP707106781, T2Y, T2V); | |
659 T36 = T32 + T35; | |
660 T37 = FMA(KP923879532, T36, T2Z); | |
661 T6x = FNMS(KP923879532, T36, T2Z); | |
662 T7J = FNMS(KP707106781, T2Y, T2V); | |
663 T7K = T3b - T3c; | |
664 T7L = FMA(KP923879532, T7K, T7J); | |
665 T9e = FNMS(KP923879532, T7K, T7J); | |
666 } | |
667 { | |
668 E T7M, T7N, T3a, T3d; | |
669 T7M = FNMS(KP707106781, T39, T38); | |
670 T7N = T35 - T32; | |
671 T7O = FMA(KP923879532, T7N, T7M); | |
672 T9d = FNMS(KP923879532, T7N, T7M); | |
673 T3a = FMA(KP707106781, T39, T38); | |
674 T3d = T3b + T3c; | |
675 T3e = FMA(KP923879532, T3d, T3a); | |
676 T6w = FNMS(KP923879532, T3d, T3a); | |
677 } | |
678 } | |
679 { | |
680 E T3L, Tdf, T3W, Tar, T42, Tdi, T4d, Tay, T3S, Tdg, T3X, Tau, T49, Tdj, T4e; | |
681 E TaB, Tdh, Tdk; | |
682 { | |
683 E T3J, T3K, Tap, T3U, T3V, Taq; | |
684 T3J = R1[WS(rs, 2)]; | |
685 T3K = R1[WS(rs, 34)]; | |
686 Tap = T3J + T3K; | |
687 T3U = R1[WS(rs, 18)]; | |
688 T3V = R1[WS(rs, 50)]; | |
689 Taq = T3U + T3V; | |
690 T3L = T3J - T3K; | |
691 Tdf = Tap - Taq; | |
692 T3W = T3U - T3V; | |
693 Tar = Tap + Taq; | |
694 } | |
695 { | |
696 E T40, T41, Taw, T4b, T4c, Tax; | |
697 T40 = R1[WS(rs, 62)]; | |
698 T41 = R1[WS(rs, 30)]; | |
699 Taw = T40 + T41; | |
700 T4b = R1[WS(rs, 14)]; | |
701 T4c = R1[WS(rs, 46)]; | |
702 Tax = T4b + T4c; | |
703 T42 = T40 - T41; | |
704 Tdi = Taw - Tax; | |
705 T4d = T4b - T4c; | |
706 Tay = Taw + Tax; | |
707 } | |
708 { | |
709 E T3O, Tas, T3R, Tat; | |
710 { | |
711 E T3M, T3N, T3P, T3Q; | |
712 T3M = R1[WS(rs, 10)]; | |
713 T3N = R1[WS(rs, 42)]; | |
714 T3O = T3M - T3N; | |
715 Tas = T3M + T3N; | |
716 T3P = R1[WS(rs, 58)]; | |
717 T3Q = R1[WS(rs, 26)]; | |
718 T3R = T3P - T3Q; | |
719 Tat = T3P + T3Q; | |
720 } | |
721 T3S = T3O + T3R; | |
722 Tdg = Tat - Tas; | |
723 T3X = T3O - T3R; | |
724 Tau = Tas + Tat; | |
725 } | |
726 { | |
727 E T45, Taz, T48, TaA; | |
728 { | |
729 E T43, T44, T46, T47; | |
730 T43 = R1[WS(rs, 6)]; | |
731 T44 = R1[WS(rs, 38)]; | |
732 T45 = T43 - T44; | |
733 Taz = T43 + T44; | |
734 T46 = R1[WS(rs, 54)]; | |
735 T47 = R1[WS(rs, 22)]; | |
736 T48 = T46 - T47; | |
737 TaA = T46 + T47; | |
738 } | |
739 T49 = T45 + T48; | |
740 Tdj = TaA - Taz; | |
741 T4e = T45 - T48; | |
742 TaB = Taz + TaA; | |
743 } | |
744 TbZ = Tar + Tau; | |
745 Tc0 = Tay + TaB; | |
746 { | |
747 E T3T, T3Y, Tdq, Tdr; | |
748 T3T = FMA(KP707106781, T3S, T3L); | |
749 T3Y = FMA(KP707106781, T3X, T3W); | |
750 T3Z = FNMS(KP198912367, T3Y, T3T); | |
751 T4s = FMA(KP198912367, T3T, T3Y); | |
752 Tdq = FNMS(KP414213562, Tdf, Tdg); | |
753 Tdr = FMA(KP414213562, Tdi, Tdj); | |
754 Tds = Tdq + Tdr; | |
755 TeF = Tdr - Tdq; | |
756 } | |
757 { | |
758 E T4a, T4f, T7Y, T7Z; | |
759 T4a = FMA(KP707106781, T49, T42); | |
760 T4f = FMA(KP707106781, T4e, T4d); | |
761 T4g = FMA(KP198912367, T4f, T4a); | |
762 T4t = FNMS(KP198912367, T4a, T4f); | |
763 T7Y = FNMS(KP707106781, T49, T42); | |
764 T7Z = FNMS(KP707106781, T4e, T4d); | |
765 T80 = FNMS(KP668178637, T7Z, T7Y); | |
766 T87 = FMA(KP668178637, T7Y, T7Z); | |
767 } | |
768 Tdh = FMA(KP414213562, Tdg, Tdf); | |
769 Tdk = FNMS(KP414213562, Tdj, Tdi); | |
770 Tdl = Tdh + Tdk; | |
771 TeI = Tdh - Tdk; | |
772 { | |
773 E T7V, T7W, Tav, TaC; | |
774 T7V = FNMS(KP707106781, T3S, T3L); | |
775 T7W = FNMS(KP707106781, T3X, T3W); | |
776 T7X = FMA(KP668178637, T7W, T7V); | |
777 T86 = FNMS(KP668178637, T7V, T7W); | |
778 Tav = Tar - Tau; | |
779 TaC = Tay - TaB; | |
780 TaD = Tav + TaC; | |
781 TaM = TaC - Tav; | |
782 } | |
783 } | |
784 { | |
785 E T50, TdA, T5b, TaY, T5h, TdD, T5s, Tb5, T57, TdB, T5c, Tb1, T5o, TdE, T5t; | |
786 E Tb8, TdC, TdF; | |
787 { | |
788 E T4Y, T4Z, TaW, T59, T5a, TaX; | |
789 T4Y = R1[WS(rs, 1)]; | |
790 T4Z = R1[WS(rs, 33)]; | |
791 TaW = T4Y + T4Z; | |
792 T59 = R1[WS(rs, 49)]; | |
793 T5a = R1[WS(rs, 17)]; | |
794 TaX = T5a + T59; | |
795 T50 = T4Y - T4Z; | |
796 TdA = TaW - TaX; | |
797 T5b = T59 - T5a; | |
798 TaY = TaW + TaX; | |
799 } | |
800 { | |
801 E T5f, T5g, Tb3, T5q, T5r, Tb4; | |
802 T5f = R1[WS(rs, 61)]; | |
803 T5g = R1[WS(rs, 29)]; | |
804 Tb3 = T5f + T5g; | |
805 T5q = R1[WS(rs, 45)]; | |
806 T5r = R1[WS(rs, 13)]; | |
807 Tb4 = T5r + T5q; | |
808 T5h = T5f - T5g; | |
809 TdD = Tb3 - Tb4; | |
810 T5s = T5q - T5r; | |
811 Tb5 = Tb3 + Tb4; | |
812 } | |
813 { | |
814 E T53, TaZ, T56, Tb0; | |
815 { | |
816 E T51, T52, T54, T55; | |
817 T51 = R1[WS(rs, 9)]; | |
818 T52 = R1[WS(rs, 41)]; | |
819 T53 = T51 - T52; | |
820 TaZ = T51 + T52; | |
821 T54 = R1[WS(rs, 57)]; | |
822 T55 = R1[WS(rs, 25)]; | |
823 T56 = T54 - T55; | |
824 Tb0 = T54 + T55; | |
825 } | |
826 T57 = T53 + T56; | |
827 TdB = Tb0 - TaZ; | |
828 T5c = T56 - T53; | |
829 Tb1 = TaZ + Tb0; | |
830 } | |
831 { | |
832 E T5k, Tb6, T5n, Tb7; | |
833 { | |
834 E T5i, T5j, T5l, T5m; | |
835 T5i = R1[WS(rs, 5)]; | |
836 T5j = R1[WS(rs, 37)]; | |
837 T5k = T5i - T5j; | |
838 Tb6 = T5i + T5j; | |
839 T5l = R1[WS(rs, 53)]; | |
840 T5m = R1[WS(rs, 21)]; | |
841 T5n = T5l - T5m; | |
842 Tb7 = T5l + T5m; | |
843 } | |
844 T5o = T5k + T5n; | |
845 TdE = Tb7 - Tb6; | |
846 T5t = T5n - T5k; | |
847 Tb8 = Tb6 + Tb7; | |
848 } | |
849 Tc6 = TaY + Tb1; | |
850 Tc7 = Tb5 + Tb8; | |
851 { | |
852 E T58, T5d, TdL, TdM; | |
853 T58 = FMA(KP707106781, T57, T50); | |
854 T5d = FMA(KP707106781, T5c, T5b); | |
855 T5e = FMA(KP198912367, T5d, T58); | |
856 T5H = FNMS(KP198912367, T58, T5d); | |
857 TdL = FNMS(KP414213562, TdA, TdB); | |
858 TdM = FMA(KP414213562, TdD, TdE); | |
859 TdN = TdL + TdM; | |
860 TeM = TdM - TdL; | |
861 } | |
862 { | |
863 E T5p, T5u, T8h, T8i; | |
864 T5p = FMA(KP707106781, T5o, T5h); | |
865 T5u = FMA(KP707106781, T5t, T5s); | |
866 T5v = FNMS(KP198912367, T5u, T5p); | |
867 T5I = FMA(KP198912367, T5p, T5u); | |
868 T8h = FNMS(KP707106781, T5o, T5h); | |
869 T8i = FNMS(KP707106781, T5t, T5s); | |
870 T8j = FMA(KP668178637, T8i, T8h); | |
871 T8q = FNMS(KP668178637, T8h, T8i); | |
872 } | |
873 TdC = FMA(KP414213562, TdB, TdA); | |
874 TdF = FNMS(KP414213562, TdE, TdD); | |
875 TdG = TdC + TdF; | |
876 TeP = TdF - TdC; | |
877 { | |
878 E T8e, T8f, Tb2, Tb9; | |
879 T8e = FNMS(KP707106781, T57, T50); | |
880 T8f = FNMS(KP707106781, T5c, T5b); | |
881 T8g = FNMS(KP668178637, T8f, T8e); | |
882 T8p = FMA(KP668178637, T8e, T8f); | |
883 Tb2 = TaY - Tb1; | |
884 Tb9 = Tb5 - Tb8; | |
885 Tba = Tb2 + Tb9; | |
886 Tbj = Tb9 - Tb2; | |
887 } | |
888 } | |
889 { | |
890 E T11, TbV, Tc9, Tcf, T22, Tcb, Tc2, Tce; | |
891 { | |
892 E Tv, T10, Tc5, Tc8; | |
893 Tv = Tf + Tu; | |
894 T10 = TK + TZ; | |
895 T11 = Tv + T10; | |
896 TbV = Tv - T10; | |
897 Tc5 = Tc3 + Tc4; | |
898 Tc8 = Tc6 + Tc7; | |
899 Tc9 = Tc5 - Tc8; | |
900 Tcf = Tc5 + Tc8; | |
901 } | |
902 { | |
903 E T1w, T21, TbY, Tc1; | |
904 T1w = T1g + T1v; | |
905 T21 = T1L + T20; | |
906 T22 = T1w + T21; | |
907 Tcb = T21 - T1w; | |
908 TbY = TbW + TbX; | |
909 Tc1 = TbZ + Tc0; | |
910 Tc2 = TbY - Tc1; | |
911 Tce = TbY + Tc1; | |
912 } | |
913 Cr[WS(csr, 32)] = T11 - T22; | |
914 Ci[WS(csi, 32)] = Tcf - Tce; | |
915 { | |
916 E Tca, Tcc, Tcd, Tcg; | |
917 Tca = Tc2 + Tc9; | |
918 Cr[WS(csr, 48)] = FNMS(KP707106781, Tca, TbV); | |
919 Cr[WS(csr, 16)] = FMA(KP707106781, Tca, TbV); | |
920 Tcc = Tc9 - Tc2; | |
921 Ci[WS(csi, 16)] = FMA(KP707106781, Tcc, Tcb); | |
922 Ci[WS(csi, 48)] = FMS(KP707106781, Tcc, Tcb); | |
923 Tcd = T11 + T22; | |
924 Tcg = Tce + Tcf; | |
925 Cr[WS(csr, 64)] = Tcd - Tcg; | |
926 Cr[0] = Tcd + Tcg; | |
927 } | |
928 } | |
929 { | |
930 E Tch, Tct, Tck, Tcu, Tco, Tcy, Tcr, Tcz, Tci, Tcj; | |
931 Tch = Tf - Tu; | |
932 Tct = TZ - TK; | |
933 Tci = T1g - T1v; | |
934 Tcj = T1L - T20; | |
935 Tck = Tci + Tcj; | |
936 Tcu = Tcj - Tci; | |
937 { | |
938 E Tcm, Tcn, Tcp, Tcq; | |
939 Tcm = TbW - TbX; | |
940 Tcn = Tc0 - TbZ; | |
941 Tco = FMA(KP414213562, Tcn, Tcm); | |
942 Tcy = FNMS(KP414213562, Tcm, Tcn); | |
943 Tcp = Tc3 - Tc4; | |
944 Tcq = Tc7 - Tc6; | |
945 Tcr = FNMS(KP414213562, Tcq, Tcp); | |
946 Tcz = FMA(KP414213562, Tcp, Tcq); | |
947 } | |
948 { | |
949 E Tcl, Tcs, Tcx, TcA; | |
950 Tcl = FMA(KP707106781, Tck, Tch); | |
951 Tcs = Tco + Tcr; | |
952 Cr[WS(csr, 56)] = FNMS(KP923879532, Tcs, Tcl); | |
953 Cr[WS(csr, 8)] = FMA(KP923879532, Tcs, Tcl); | |
954 Tcx = FMA(KP707106781, Tcu, Tct); | |
955 TcA = Tcy + Tcz; | |
956 Ci[WS(csi, 8)] = FMA(KP923879532, TcA, Tcx); | |
957 Ci[WS(csi, 56)] = FMS(KP923879532, TcA, Tcx); | |
958 } | |
959 { | |
960 E Tcv, Tcw, TcB, TcC; | |
961 Tcv = FNMS(KP707106781, Tcu, Tct); | |
962 Tcw = Tcr - Tco; | |
963 Ci[WS(csi, 24)] = FMS(KP923879532, Tcw, Tcv); | |
964 Ci[WS(csi, 40)] = FMA(KP923879532, Tcw, Tcv); | |
965 TcB = FNMS(KP707106781, Tck, Tch); | |
966 TcC = Tcz - Tcy; | |
967 Cr[WS(csr, 40)] = FNMS(KP923879532, TcC, TcB); | |
968 Cr[WS(csr, 24)] = FMA(KP923879532, TcC, TcB); | |
969 } | |
970 } | |
971 { | |
972 E Ta9, TbB, Tbp, TbL, Tag, TbM, TbJ, TbR, TaO, Tbw, Tbs, TbC, TbG, TbQ, Tbl; | |
973 E Tbx, Ta8, Tbo; | |
974 Ta8 = Ta6 + Ta7; | |
975 Ta9 = FMA(KP707106781, Ta8, Ta5); | |
976 TbB = FNMS(KP707106781, Ta8, Ta5); | |
977 Tbo = Ta7 - Ta6; | |
978 Tbp = FMA(KP707106781, Tbo, Tbn); | |
979 TbL = FNMS(KP707106781, Tbo, Tbn); | |
980 { | |
981 E Tac, Taf, TbH, TbI; | |
982 Tac = FMA(KP414213562, Tab, Taa); | |
983 Taf = FNMS(KP414213562, Tae, Tad); | |
984 Tag = Tac + Taf; | |
985 TbM = Taf - Tac; | |
986 TbH = FNMS(KP707106781, Tba, TaV); | |
987 TbI = FNMS(KP707106781, Tbj, Tbi); | |
988 TbJ = FMA(KP668178637, TbI, TbH); | |
989 TbR = FNMS(KP668178637, TbH, TbI); | |
990 } | |
991 { | |
992 E TaE, TaN, Tbq, Tbr; | |
993 TaE = FMA(KP707106781, TaD, Tao); | |
994 TaN = FMA(KP707106781, TaM, TaL); | |
995 TaO = FMA(KP198912367, TaN, TaE); | |
996 Tbw = FNMS(KP198912367, TaE, TaN); | |
997 Tbq = FNMS(KP414213562, Taa, Tab); | |
998 Tbr = FMA(KP414213562, Tad, Tae); | |
999 Tbs = Tbq + Tbr; | |
1000 TbC = Tbr - Tbq; | |
1001 } | |
1002 { | |
1003 E TbE, TbF, Tbb, Tbk; | |
1004 TbE = FNMS(KP707106781, TaD, Tao); | |
1005 TbF = FNMS(KP707106781, TaM, TaL); | |
1006 TbG = FNMS(KP668178637, TbF, TbE); | |
1007 TbQ = FMA(KP668178637, TbE, TbF); | |
1008 Tbb = FMA(KP707106781, Tba, TaV); | |
1009 Tbk = FMA(KP707106781, Tbj, Tbi); | |
1010 Tbl = FNMS(KP198912367, Tbk, Tbb); | |
1011 Tbx = FMA(KP198912367, Tbb, Tbk); | |
1012 } | |
1013 { | |
1014 E Tah, Tbm, Tbv, Tby; | |
1015 Tah = FMA(KP923879532, Tag, Ta9); | |
1016 Tbm = TaO + Tbl; | |
1017 Cr[WS(csr, 60)] = FNMS(KP980785280, Tbm, Tah); | |
1018 Cr[WS(csr, 4)] = FMA(KP980785280, Tbm, Tah); | |
1019 Tbv = FMA(KP923879532, Tbs, Tbp); | |
1020 Tby = Tbw + Tbx; | |
1021 Ci[WS(csi, 4)] = FMA(KP980785280, Tby, Tbv); | |
1022 Ci[WS(csi, 60)] = FMS(KP980785280, Tby, Tbv); | |
1023 } | |
1024 { | |
1025 E Tbt, Tbu, Tbz, TbA; | |
1026 Tbt = FNMS(KP923879532, Tbs, Tbp); | |
1027 Tbu = Tbl - TaO; | |
1028 Ci[WS(csi, 28)] = FMS(KP980785280, Tbu, Tbt); | |
1029 Ci[WS(csi, 36)] = FMA(KP980785280, Tbu, Tbt); | |
1030 Tbz = FNMS(KP923879532, Tag, Ta9); | |
1031 TbA = Tbx - Tbw; | |
1032 Cr[WS(csr, 36)] = FNMS(KP980785280, TbA, Tbz); | |
1033 Cr[WS(csr, 28)] = FMA(KP980785280, TbA, Tbz); | |
1034 } | |
1035 { | |
1036 E TbD, TbK, TbP, TbS; | |
1037 TbD = FMA(KP923879532, TbC, TbB); | |
1038 TbK = TbG + TbJ; | |
1039 Cr[WS(csr, 52)] = FNMS(KP831469612, TbK, TbD); | |
1040 Cr[WS(csr, 12)] = FMA(KP831469612, TbK, TbD); | |
1041 TbP = FNMS(KP923879532, TbM, TbL); | |
1042 TbS = TbQ + TbR; | |
1043 Ci[WS(csi, 12)] = -(FMA(KP831469612, TbS, TbP)); | |
1044 Ci[WS(csi, 52)] = FNMS(KP831469612, TbS, TbP); | |
1045 } | |
1046 { | |
1047 E TbN, TbO, TbT, TbU; | |
1048 TbN = FMA(KP923879532, TbM, TbL); | |
1049 TbO = TbJ - TbG; | |
1050 Ci[WS(csi, 20)] = FMA(KP831469612, TbO, TbN); | |
1051 Ci[WS(csi, 44)] = FMS(KP831469612, TbO, TbN); | |
1052 TbT = FNMS(KP923879532, TbC, TbB); | |
1053 TbU = TbQ - TbR; | |
1054 Cr[WS(csr, 44)] = FNMS(KP831469612, TbU, TbT); | |
1055 Cr[WS(csr, 20)] = FMA(KP831469612, TbU, TbT); | |
1056 } | |
1057 } | |
1058 { | |
1059 E Tev, Tf7, Tfc, Tfm, Tff, Tfn, TeC, Tfi, TeK, Tf2, TeV, Tfh, TeY, Tf8, TeR; | |
1060 E Tf3; | |
1061 { | |
1062 E Tet, Teu, Tfa, Tfb; | |
1063 Tet = FNMS(KP707106781, TcG, TcD); | |
1064 Teu = TdV - TdU; | |
1065 Tev = FNMS(KP923879532, Teu, Tet); | |
1066 Tf7 = FMA(KP923879532, Teu, Tet); | |
1067 Tfa = FMA(KP923879532, TeF, TeE); | |
1068 Tfb = FMA(KP923879532, TeI, TeH); | |
1069 Tfc = FNMS(KP303346683, Tfb, Tfa); | |
1070 Tfm = FMA(KP303346683, Tfa, Tfb); | |
1071 } | |
1072 { | |
1073 E Tfd, Tfe, Tey, TeB; | |
1074 Tfd = FMA(KP923879532, TeM, TeL); | |
1075 Tfe = FNMS(KP923879532, TeP, TeO); | |
1076 Tff = FMA(KP303346683, Tfe, Tfd); | |
1077 Tfn = FNMS(KP303346683, Tfd, Tfe); | |
1078 Tey = FMA(KP668178637, Tex, Tew); | |
1079 TeB = FNMS(KP668178637, TeA, Tez); | |
1080 TeC = Tey - TeB; | |
1081 Tfi = Tey + TeB; | |
1082 } | |
1083 { | |
1084 E TeG, TeJ, TeT, TeU; | |
1085 TeG = FNMS(KP923879532, TeF, TeE); | |
1086 TeJ = FNMS(KP923879532, TeI, TeH); | |
1087 TeK = FMA(KP534511135, TeJ, TeG); | |
1088 Tf2 = FNMS(KP534511135, TeG, TeJ); | |
1089 TeT = FNMS(KP707106781, TdS, TdR); | |
1090 TeU = TcN - TcK; | |
1091 TeV = FMA(KP923879532, TeU, TeT); | |
1092 Tfh = FNMS(KP923879532, TeU, TeT); | |
1093 } | |
1094 { | |
1095 E TeW, TeX, TeN, TeQ; | |
1096 TeW = FMA(KP668178637, Tez, TeA); | |
1097 TeX = FNMS(KP668178637, Tew, Tex); | |
1098 TeY = TeW - TeX; | |
1099 Tf8 = TeX + TeW; | |
1100 TeN = FNMS(KP923879532, TeM, TeL); | |
1101 TeQ = FMA(KP923879532, TeP, TeO); | |
1102 TeR = FNMS(KP534511135, TeQ, TeN); | |
1103 Tf3 = FMA(KP534511135, TeN, TeQ); | |
1104 } | |
1105 { | |
1106 E TeD, TeS, Tf1, Tf4; | |
1107 TeD = FMA(KP831469612, TeC, Tev); | |
1108 TeS = TeK + TeR; | |
1109 Cr[WS(csr, 54)] = FNMS(KP881921264, TeS, TeD); | |
1110 Cr[WS(csr, 10)] = FMA(KP881921264, TeS, TeD); | |
1111 Tf1 = FMA(KP831469612, TeY, TeV); | |
1112 Tf4 = Tf2 + Tf3; | |
1113 Ci[WS(csi, 10)] = FMA(KP881921264, Tf4, Tf1); | |
1114 Ci[WS(csi, 54)] = FMS(KP881921264, Tf4, Tf1); | |
1115 } | |
1116 { | |
1117 E TeZ, Tf0, Tf5, Tf6; | |
1118 TeZ = FNMS(KP831469612, TeY, TeV); | |
1119 Tf0 = TeR - TeK; | |
1120 Ci[WS(csi, 22)] = FMS(KP881921264, Tf0, TeZ); | |
1121 Ci[WS(csi, 42)] = FMA(KP881921264, Tf0, TeZ); | |
1122 Tf5 = FNMS(KP831469612, TeC, Tev); | |
1123 Tf6 = Tf3 - Tf2; | |
1124 Cr[WS(csr, 42)] = FNMS(KP881921264, Tf6, Tf5); | |
1125 Cr[WS(csr, 22)] = FMA(KP881921264, Tf6, Tf5); | |
1126 } | |
1127 { | |
1128 E Tf9, Tfg, Tfl, Tfo; | |
1129 Tf9 = FMA(KP831469612, Tf8, Tf7); | |
1130 Tfg = Tfc + Tff; | |
1131 Cr[WS(csr, 58)] = FNMS(KP956940335, Tfg, Tf9); | |
1132 Cr[WS(csr, 6)] = FMA(KP956940335, Tfg, Tf9); | |
1133 Tfl = FMA(KP831469612, Tfi, Tfh); | |
1134 Tfo = Tfm + Tfn; | |
1135 Ci[WS(csi, 6)] = -(FMA(KP956940335, Tfo, Tfl)); | |
1136 Ci[WS(csi, 58)] = FNMS(KP956940335, Tfo, Tfl); | |
1137 } | |
1138 { | |
1139 E Tfj, Tfk, Tfp, Tfq; | |
1140 Tfj = FNMS(KP831469612, Tfi, Tfh); | |
1141 Tfk = Tff - Tfc; | |
1142 Ci[WS(csi, 26)] = FMA(KP956940335, Tfk, Tfj); | |
1143 Ci[WS(csi, 38)] = FMS(KP956940335, Tfk, Tfj); | |
1144 Tfp = FNMS(KP831469612, Tf8, Tf7); | |
1145 Tfq = Tfm - Tfn; | |
1146 Cr[WS(csr, 38)] = FNMS(KP956940335, Tfq, Tfp); | |
1147 Cr[WS(csr, 26)] = FMA(KP956940335, Tfq, Tfp); | |
1148 } | |
1149 } | |
1150 { | |
1151 E TcP, Te9, Tee, Teo, Teh, Tep, Td8, Tek, Tdu, Te4, TdX, Tej, Te0, Tea, TdP; | |
1152 E Te5; | |
1153 { | |
1154 E TcH, TcO, Tec, Ted; | |
1155 TcH = FMA(KP707106781, TcG, TcD); | |
1156 TcO = TcK + TcN; | |
1157 TcP = FMA(KP923879532, TcO, TcH); | |
1158 Te9 = FNMS(KP923879532, TcO, TcH); | |
1159 Tec = FNMS(KP923879532, Tdl, Tde); | |
1160 Ted = FNMS(KP923879532, Tds, Tdp); | |
1161 Tee = FNMS(KP820678790, Ted, Tec); | |
1162 Teo = FMA(KP820678790, Tec, Ted); | |
1163 } | |
1164 { | |
1165 E Tef, Teg, TcY, Td7; | |
1166 Tef = FNMS(KP923879532, TdG, Tdz); | |
1167 Teg = FNMS(KP923879532, TdN, TdK); | |
1168 Teh = FMA(KP820678790, Teg, Tef); | |
1169 Tep = FNMS(KP820678790, Tef, Teg); | |
1170 TcY = FMA(KP198912367, TcX, TcU); | |
1171 Td7 = FNMS(KP198912367, Td6, Td3); | |
1172 Td8 = TcY + Td7; | |
1173 Tek = Td7 - TcY; | |
1174 } | |
1175 { | |
1176 E Tdm, Tdt, TdT, TdW; | |
1177 Tdm = FMA(KP923879532, Tdl, Tde); | |
1178 Tdt = FMA(KP923879532, Tds, Tdp); | |
1179 Tdu = FMA(KP098491403, Tdt, Tdm); | |
1180 Te4 = FNMS(KP098491403, Tdm, Tdt); | |
1181 TdT = FMA(KP707106781, TdS, TdR); | |
1182 TdW = TdU + TdV; | |
1183 TdX = FMA(KP923879532, TdW, TdT); | |
1184 Tej = FNMS(KP923879532, TdW, TdT); | |
1185 } | |
1186 { | |
1187 E TdY, TdZ, TdH, TdO; | |
1188 TdY = FNMS(KP198912367, TcU, TcX); | |
1189 TdZ = FMA(KP198912367, Td3, Td6); | |
1190 Te0 = TdY + TdZ; | |
1191 Tea = TdZ - TdY; | |
1192 TdH = FMA(KP923879532, TdG, Tdz); | |
1193 TdO = FMA(KP923879532, TdN, TdK); | |
1194 TdP = FNMS(KP098491403, TdO, TdH); | |
1195 Te5 = FMA(KP098491403, TdH, TdO); | |
1196 } | |
1197 { | |
1198 E Td9, TdQ, Te3, Te6; | |
1199 Td9 = FMA(KP980785280, Td8, TcP); | |
1200 TdQ = Tdu + TdP; | |
1201 Cr[WS(csr, 62)] = FNMS(KP995184726, TdQ, Td9); | |
1202 Cr[WS(csr, 2)] = FMA(KP995184726, TdQ, Td9); | |
1203 Te3 = FMA(KP980785280, Te0, TdX); | |
1204 Te6 = Te4 + Te5; | |
1205 Ci[WS(csi, 2)] = FMA(KP995184726, Te6, Te3); | |
1206 Ci[WS(csi, 62)] = FMS(KP995184726, Te6, Te3); | |
1207 } | |
1208 { | |
1209 E Te1, Te2, Te7, Te8; | |
1210 Te1 = FNMS(KP980785280, Te0, TdX); | |
1211 Te2 = TdP - Tdu; | |
1212 Ci[WS(csi, 30)] = FMS(KP995184726, Te2, Te1); | |
1213 Ci[WS(csi, 34)] = FMA(KP995184726, Te2, Te1); | |
1214 Te7 = FNMS(KP980785280, Td8, TcP); | |
1215 Te8 = Te5 - Te4; | |
1216 Cr[WS(csr, 34)] = FNMS(KP995184726, Te8, Te7); | |
1217 Cr[WS(csr, 30)] = FMA(KP995184726, Te8, Te7); | |
1218 } | |
1219 { | |
1220 E Teb, Tei, Ten, Teq; | |
1221 Teb = FMA(KP980785280, Tea, Te9); | |
1222 Tei = Tee + Teh; | |
1223 Cr[WS(csr, 50)] = FNMS(KP773010453, Tei, Teb); | |
1224 Cr[WS(csr, 14)] = FMA(KP773010453, Tei, Teb); | |
1225 Ten = FNMS(KP980785280, Tek, Tej); | |
1226 Teq = Teo + Tep; | |
1227 Ci[WS(csi, 14)] = -(FMA(KP773010453, Teq, Ten)); | |
1228 Ci[WS(csi, 50)] = FNMS(KP773010453, Teq, Ten); | |
1229 } | |
1230 { | |
1231 E Tel, Tem, Ter, Tes; | |
1232 Tel = FMA(KP980785280, Tek, Tej); | |
1233 Tem = Teh - Tee; | |
1234 Ci[WS(csi, 18)] = FMA(KP773010453, Tem, Tel); | |
1235 Ci[WS(csi, 46)] = FMS(KP773010453, Tem, Tel); | |
1236 Ter = FNMS(KP980785280, Tea, Te9); | |
1237 Tes = Teo - Tep; | |
1238 Cr[WS(csr, 46)] = FNMS(KP773010453, Tes, Ter); | |
1239 Cr[WS(csr, 18)] = FMA(KP773010453, Tes, Ter); | |
1240 } | |
1241 } | |
1242 { | |
1243 E T6v, T77, T6C, T7i, T6Y, T78, T6V, T7h, T6R, T7n, T72, T7f, T6K, T7m, T73; | |
1244 E T7c; | |
1245 { | |
1246 E T6t, T6u, T6T, T6U; | |
1247 T6t = FNMS(KP923879532, T2e, T27); | |
1248 T6u = T5U - T5V; | |
1249 T6v = FNMS(KP980785280, T6u, T6t); | |
1250 T77 = FMA(KP980785280, T6u, T6t); | |
1251 { | |
1252 E T6y, T6B, T6W, T6X; | |
1253 T6y = FMA(KP820678790, T6x, T6w); | |
1254 T6B = FNMS(KP820678790, T6A, T6z); | |
1255 T6C = T6y - T6B; | |
1256 T7i = T6B + T6y; | |
1257 T6W = FNMS(KP820678790, T6w, T6x); | |
1258 T6X = FMA(KP820678790, T6z, T6A); | |
1259 T6Y = T6W - T6X; | |
1260 T78 = T6X + T6W; | |
1261 } | |
1262 T6T = FNMS(KP923879532, T5S, T5P); | |
1263 T6U = T2x - T2o; | |
1264 T6V = FNMS(KP980785280, T6U, T6T); | |
1265 T7h = FMA(KP980785280, T6U, T6T); | |
1266 { | |
1267 E T6N, T7d, T6Q, T7e, T6M, T6P; | |
1268 T6M = T5I - T5H; | |
1269 T6N = FNMS(KP980785280, T6M, T6L); | |
1270 T7d = FMA(KP980785280, T6M, T6L); | |
1271 T6P = T5v - T5e; | |
1272 T6Q = FMA(KP980785280, T6P, T6O); | |
1273 T7e = FNMS(KP980785280, T6P, T6O); | |
1274 T6R = FNMS(KP472964775, T6Q, T6N); | |
1275 T7n = FNMS(KP357805721, T7d, T7e); | |
1276 T72 = FMA(KP472964775, T6N, T6Q); | |
1277 T7f = FMA(KP357805721, T7e, T7d); | |
1278 } | |
1279 { | |
1280 E T6G, T7a, T6J, T7b, T6F, T6I; | |
1281 T6F = T4s - T4t; | |
1282 T6G = FNMS(KP980785280, T6F, T6E); | |
1283 T7a = FMA(KP980785280, T6F, T6E); | |
1284 T6I = T4g - T3Z; | |
1285 T6J = FNMS(KP980785280, T6I, T6H); | |
1286 T7b = FMA(KP980785280, T6I, T6H); | |
1287 T6K = FNMS(KP472964775, T6J, T6G); | |
1288 T7m = FNMS(KP357805721, T7a, T7b); | |
1289 T73 = FMA(KP472964775, T6G, T6J); | |
1290 T7c = FMA(KP357805721, T7b, T7a); | |
1291 } | |
1292 } | |
1293 { | |
1294 E T6D, T6S, T71, T74; | |
1295 T6D = FMA(KP773010453, T6C, T6v); | |
1296 T6S = T6K + T6R; | |
1297 Cr[WS(csr, 55)] = FNMS(KP903989293, T6S, T6D); | |
1298 Cr[WS(csr, 9)] = FMA(KP903989293, T6S, T6D); | |
1299 T71 = FNMS(KP773010453, T6Y, T6V); | |
1300 T74 = T72 - T73; | |
1301 Ci[WS(csi, 9)] = FMS(KP903989293, T74, T71); | |
1302 Ci[WS(csi, 55)] = FMA(KP903989293, T74, T71); | |
1303 } | |
1304 { | |
1305 E T6Z, T70, T75, T76; | |
1306 T6Z = FMA(KP773010453, T6Y, T6V); | |
1307 T70 = T6R - T6K; | |
1308 Ci[WS(csi, 23)] = FMA(KP903989293, T70, T6Z); | |
1309 Ci[WS(csi, 41)] = FMS(KP903989293, T70, T6Z); | |
1310 T75 = FNMS(KP773010453, T6C, T6v); | |
1311 T76 = T73 + T72; | |
1312 Cr[WS(csr, 41)] = FNMS(KP903989293, T76, T75); | |
1313 Cr[WS(csr, 23)] = FMA(KP903989293, T76, T75); | |
1314 } | |
1315 { | |
1316 E T79, T7g, T7l, T7o; | |
1317 T79 = FMA(KP773010453, T78, T77); | |
1318 T7g = T7c + T7f; | |
1319 Cr[WS(csr, 57)] = FNMS(KP941544065, T7g, T79); | |
1320 Cr[WS(csr, 7)] = FMA(KP941544065, T7g, T79); | |
1321 T7l = FMA(KP773010453, T7i, T7h); | |
1322 T7o = T7m - T7n; | |
1323 Ci[WS(csi, 7)] = FMA(KP941544065, T7o, T7l); | |
1324 Ci[WS(csi, 57)] = FMS(KP941544065, T7o, T7l); | |
1325 } | |
1326 { | |
1327 E T7j, T7k, T7p, T7q; | |
1328 T7j = FNMS(KP773010453, T7i, T7h); | |
1329 T7k = T7f - T7c; | |
1330 Ci[WS(csi, 25)] = FMS(KP941544065, T7k, T7j); | |
1331 Ci[WS(csi, 39)] = FMA(KP941544065, T7k, T7j); | |
1332 T7p = FNMS(KP773010453, T78, T77); | |
1333 T7q = T7m + T7n; | |
1334 Cr[WS(csr, 39)] = FMA(KP941544065, T7q, T7p); | |
1335 Cr[WS(csr, 25)] = FNMS(KP941544065, T7q, T7p); | |
1336 } | |
1337 } | |
1338 { | |
1339 E T99, T9L, T9g, T9W, T9C, T9M, T9z, T9V, T9v, Ta0, T9H, T9T, T9o, Ta1, T9G; | |
1340 E T9Q; | |
1341 { | |
1342 E T97, T98, T9x, T9y; | |
1343 T97 = FNMS(KP923879532, T7s, T7r); | |
1344 T98 = T8z - T8y; | |
1345 T99 = FNMS(KP831469612, T98, T97); | |
1346 T9L = FMA(KP831469612, T98, T97); | |
1347 { | |
1348 E T9c, T9f, T9A, T9B; | |
1349 T9c = FMA(KP534511135, T9b, T9a); | |
1350 T9f = FNMS(KP534511135, T9e, T9d); | |
1351 T9g = T9c - T9f; | |
1352 T9W = T9c + T9f; | |
1353 T9A = FMA(KP534511135, T9d, T9e); | |
1354 T9B = FNMS(KP534511135, T9a, T9b); | |
1355 T9C = T9A - T9B; | |
1356 T9M = T9B + T9A; | |
1357 } | |
1358 T9x = FNMS(KP923879532, T8w, T8v); | |
1359 T9y = T7z - T7w; | |
1360 T9z = FMA(KP831469612, T9y, T9x); | |
1361 T9V = FNMS(KP831469612, T9y, T9x); | |
1362 { | |
1363 E T9r, T9R, T9u, T9S, T9q, T9t; | |
1364 T9q = T8p - T8q; | |
1365 T9r = FNMS(KP831469612, T9q, T9p); | |
1366 T9R = FMA(KP831469612, T9q, T9p); | |
1367 T9t = T8j - T8g; | |
1368 T9u = FNMS(KP831469612, T9t, T9s); | |
1369 T9S = FMA(KP831469612, T9t, T9s); | |
1370 T9v = FMA(KP599376933, T9u, T9r); | |
1371 Ta0 = FMA(KP250486960, T9R, T9S); | |
1372 T9H = FNMS(KP599376933, T9r, T9u); | |
1373 T9T = FNMS(KP250486960, T9S, T9R); | |
1374 } | |
1375 { | |
1376 E T9k, T9O, T9n, T9P, T9j, T9m; | |
1377 T9j = T87 - T86; | |
1378 T9k = FNMS(KP831469612, T9j, T9i); | |
1379 T9O = FMA(KP831469612, T9j, T9i); | |
1380 T9m = T7X - T80; | |
1381 T9n = FNMS(KP831469612, T9m, T9l); | |
1382 T9P = FMA(KP831469612, T9m, T9l); | |
1383 T9o = FMA(KP599376933, T9n, T9k); | |
1384 Ta1 = FMA(KP250486960, T9O, T9P); | |
1385 T9G = FNMS(KP599376933, T9k, T9n); | |
1386 T9Q = FNMS(KP250486960, T9P, T9O); | |
1387 } | |
1388 } | |
1389 { | |
1390 E T9h, T9w, T9F, T9I; | |
1391 T9h = FMA(KP881921264, T9g, T99); | |
1392 T9w = T9o + T9v; | |
1393 Cr[WS(csr, 53)] = FNMS(KP857728610, T9w, T9h); | |
1394 Cr[WS(csr, 11)] = FMA(KP857728610, T9w, T9h); | |
1395 T9F = FMA(KP881921264, T9C, T9z); | |
1396 T9I = T9G - T9H; | |
1397 Ci[WS(csi, 11)] = FMA(KP857728610, T9I, T9F); | |
1398 Ci[WS(csi, 53)] = FMS(KP857728610, T9I, T9F); | |
1399 } | |
1400 { | |
1401 E T9D, T9E, T9J, T9K; | |
1402 T9D = FNMS(KP881921264, T9C, T9z); | |
1403 T9E = T9v - T9o; | |
1404 Ci[WS(csi, 21)] = FMS(KP857728610, T9E, T9D); | |
1405 Ci[WS(csi, 43)] = FMA(KP857728610, T9E, T9D); | |
1406 T9J = FNMS(KP881921264, T9g, T99); | |
1407 T9K = T9G + T9H; | |
1408 Cr[WS(csr, 43)] = FMA(KP857728610, T9K, T9J); | |
1409 Cr[WS(csr, 21)] = FNMS(KP857728610, T9K, T9J); | |
1410 } | |
1411 { | |
1412 E T9N, T9U, T9Z, Ta2; | |
1413 T9N = FMA(KP881921264, T9M, T9L); | |
1414 T9U = T9Q + T9T; | |
1415 Cr[WS(csr, 59)] = FNMS(KP970031253, T9U, T9N); | |
1416 Cr[WS(csr, 5)] = FMA(KP970031253, T9U, T9N); | |
1417 T9Z = FMA(KP881921264, T9W, T9V); | |
1418 Ta2 = Ta0 - Ta1; | |
1419 Ci[WS(csi, 5)] = FMS(KP970031253, Ta2, T9Z); | |
1420 Ci[WS(csi, 59)] = FMA(KP970031253, Ta2, T9Z); | |
1421 } | |
1422 { | |
1423 E T9X, T9Y, Ta3, Ta4; | |
1424 T9X = FNMS(KP881921264, T9W, T9V); | |
1425 T9Y = T9T - T9Q; | |
1426 Ci[WS(csi, 27)] = FMA(KP970031253, T9Y, T9X); | |
1427 Ci[WS(csi, 37)] = FMS(KP970031253, T9Y, T9X); | |
1428 Ta3 = FNMS(KP881921264, T9M, T9L); | |
1429 Ta4 = Ta1 + Ta0; | |
1430 Cr[WS(csr, 37)] = FNMS(KP970031253, Ta4, Ta3); | |
1431 Cr[WS(csr, 27)] = FMA(KP970031253, Ta4, Ta3); | |
1432 } | |
1433 } | |
1434 { | |
1435 E T2z, T69, T3g, T6k, T60, T6a, T5X, T6j, T5L, T6p, T64, T6h, T4w, T6o, T65; | |
1436 E T6e; | |
1437 { | |
1438 E T2f, T2y, T5T, T5W; | |
1439 T2f = FMA(KP923879532, T2e, T27); | |
1440 T2y = T2o + T2x; | |
1441 T2z = FMA(KP980785280, T2y, T2f); | |
1442 T69 = FNMS(KP980785280, T2y, T2f); | |
1443 { | |
1444 E T2U, T3f, T5Y, T5Z; | |
1445 T2U = FNMS(KP098491403, T2T, T2M); | |
1446 T3f = FMA(KP098491403, T3e, T37); | |
1447 T3g = T2U + T3f; | |
1448 T6k = T3f - T2U; | |
1449 T5Y = FMA(KP098491403, T2M, T2T); | |
1450 T5Z = FNMS(KP098491403, T37, T3e); | |
1451 T60 = T5Y + T5Z; | |
1452 T6a = T5Y - T5Z; | |
1453 } | |
1454 T5T = FMA(KP923879532, T5S, T5P); | |
1455 T5W = T5U + T5V; | |
1456 T5X = FMA(KP980785280, T5W, T5T); | |
1457 T6j = FNMS(KP980785280, T5W, T5T); | |
1458 { | |
1459 E T5x, T6f, T5K, T6g, T5w, T5J; | |
1460 T5w = T5e + T5v; | |
1461 T5x = FMA(KP980785280, T5w, T4X); | |
1462 T6f = FNMS(KP980785280, T5w, T4X); | |
1463 T5J = T5H + T5I; | |
1464 T5K = FMA(KP980785280, T5J, T5G); | |
1465 T6g = FNMS(KP980785280, T5J, T5G); | |
1466 T5L = FNMS(KP049126849, T5K, T5x); | |
1467 T6p = FNMS(KP906347169, T6f, T6g); | |
1468 T64 = FMA(KP049126849, T5x, T5K); | |
1469 T6h = FMA(KP906347169, T6g, T6f); | |
1470 } | |
1471 { | |
1472 E T4i, T6c, T4v, T6d, T4h, T4u; | |
1473 T4h = T3Z + T4g; | |
1474 T4i = FMA(KP980785280, T4h, T3I); | |
1475 T6c = FNMS(KP980785280, T4h, T3I); | |
1476 T4u = T4s + T4t; | |
1477 T4v = FMA(KP980785280, T4u, T4r); | |
1478 T6d = FNMS(KP980785280, T4u, T4r); | |
1479 T4w = FNMS(KP049126849, T4v, T4i); | |
1480 T6o = FNMS(KP906347169, T6c, T6d); | |
1481 T65 = FMA(KP049126849, T4i, T4v); | |
1482 T6e = FMA(KP906347169, T6d, T6c); | |
1483 } | |
1484 } | |
1485 { | |
1486 E T3h, T5M, T63, T66; | |
1487 T3h = FMA(KP995184726, T3g, T2z); | |
1488 T5M = T4w + T5L; | |
1489 Cr[WS(csr, 63)] = FNMS(KP998795456, T5M, T3h); | |
1490 Cr[WS(csr, 1)] = FMA(KP998795456, T5M, T3h); | |
1491 T63 = FMA(KP995184726, T60, T5X); | |
1492 T66 = T64 - T65; | |
1493 Ci[WS(csi, 1)] = FMS(KP998795456, T66, T63); | |
1494 Ci[WS(csi, 63)] = FMA(KP998795456, T66, T63); | |
1495 } | |
1496 { | |
1497 E T61, T62, T67, T68; | |
1498 T61 = FNMS(KP995184726, T60, T5X); | |
1499 T62 = T5L - T4w; | |
1500 Ci[WS(csi, 31)] = FMA(KP998795456, T62, T61); | |
1501 Ci[WS(csi, 33)] = FMS(KP998795456, T62, T61); | |
1502 T67 = FNMS(KP995184726, T3g, T2z); | |
1503 T68 = T65 + T64; | |
1504 Cr[WS(csr, 33)] = FNMS(KP998795456, T68, T67); | |
1505 Cr[WS(csr, 31)] = FMA(KP998795456, T68, T67); | |
1506 } | |
1507 { | |
1508 E T6b, T6i, T6n, T6q; | |
1509 T6b = FMA(KP995184726, T6a, T69); | |
1510 T6i = T6e + T6h; | |
1511 Cr[WS(csr, 49)] = FNMS(KP740951125, T6i, T6b); | |
1512 Cr[WS(csr, 15)] = FMA(KP740951125, T6i, T6b); | |
1513 T6n = FMA(KP995184726, T6k, T6j); | |
1514 T6q = T6o - T6p; | |
1515 Ci[WS(csi, 15)] = FMA(KP740951125, T6q, T6n); | |
1516 Ci[WS(csi, 49)] = FMS(KP740951125, T6q, T6n); | |
1517 } | |
1518 { | |
1519 E T6l, T6m, T6r, T6s; | |
1520 T6l = FNMS(KP995184726, T6k, T6j); | |
1521 T6m = T6h - T6e; | |
1522 Ci[WS(csi, 17)] = FMS(KP740951125, T6m, T6l); | |
1523 Ci[WS(csi, 47)] = FMA(KP740951125, T6m, T6l); | |
1524 T6r = FNMS(KP995184726, T6a, T69); | |
1525 T6s = T6o + T6p; | |
1526 Cr[WS(csr, 47)] = FMA(KP740951125, T6s, T6r); | |
1527 Cr[WS(csr, 17)] = FNMS(KP740951125, T6s, T6r); | |
1528 } | |
1529 } | |
1530 { | |
1531 E T7B, T8N, T7Q, T8Y, T8E, T8O, T8B, T8X, T8t, T92, T8J, T8V, T8a, T93, T8I; | |
1532 E T8S; | |
1533 { | |
1534 E T7t, T7A, T8x, T8A; | |
1535 T7t = FMA(KP923879532, T7s, T7r); | |
1536 T7A = T7w + T7z; | |
1537 T7B = FMA(KP831469612, T7A, T7t); | |
1538 T8N = FNMS(KP831469612, T7A, T7t); | |
1539 { | |
1540 E T7I, T7P, T8C, T8D; | |
1541 T7I = FMA(KP303346683, T7H, T7E); | |
1542 T7P = FNMS(KP303346683, T7O, T7L); | |
1543 T7Q = T7I + T7P; | |
1544 T8Y = T7P - T7I; | |
1545 T8C = FNMS(KP303346683, T7E, T7H); | |
1546 T8D = FMA(KP303346683, T7L, T7O); | |
1547 T8E = T8C + T8D; | |
1548 T8O = T8D - T8C; | |
1549 } | |
1550 T8x = FMA(KP923879532, T8w, T8v); | |
1551 T8A = T8y + T8z; | |
1552 T8B = FMA(KP831469612, T8A, T8x); | |
1553 T8X = FNMS(KP831469612, T8A, T8x); | |
1554 { | |
1555 E T8l, T8T, T8s, T8U, T8k, T8r; | |
1556 T8k = T8g + T8j; | |
1557 T8l = FMA(KP831469612, T8k, T8d); | |
1558 T8T = FNMS(KP831469612, T8k, T8d); | |
1559 T8r = T8p + T8q; | |
1560 T8s = FMA(KP831469612, T8r, T8o); | |
1561 T8U = FNMS(KP831469612, T8r, T8o); | |
1562 T8t = FMA(KP148335987, T8s, T8l); | |
1563 T92 = FMA(KP741650546, T8T, T8U); | |
1564 T8J = FNMS(KP148335987, T8l, T8s); | |
1565 T8V = FNMS(KP741650546, T8U, T8T); | |
1566 } | |
1567 { | |
1568 E T82, T8Q, T89, T8R, T81, T88; | |
1569 T81 = T7X + T80; | |
1570 T82 = FMA(KP831469612, T81, T7U); | |
1571 T8Q = FNMS(KP831469612, T81, T7U); | |
1572 T88 = T86 + T87; | |
1573 T89 = FMA(KP831469612, T88, T85); | |
1574 T8R = FNMS(KP831469612, T88, T85); | |
1575 T8a = FMA(KP148335987, T89, T82); | |
1576 T93 = FMA(KP741650546, T8Q, T8R); | |
1577 T8I = FNMS(KP148335987, T82, T89); | |
1578 T8S = FNMS(KP741650546, T8R, T8Q); | |
1579 } | |
1580 } | |
1581 { | |
1582 E T7R, T8u, T8H, T8K; | |
1583 T7R = FMA(KP956940335, T7Q, T7B); | |
1584 T8u = T8a + T8t; | |
1585 Cr[WS(csr, 61)] = FNMS(KP989176509, T8u, T7R); | |
1586 Cr[WS(csr, 3)] = FMA(KP989176509, T8u, T7R); | |
1587 T8H = FMA(KP956940335, T8E, T8B); | |
1588 T8K = T8I - T8J; | |
1589 Ci[WS(csi, 3)] = FMA(KP989176509, T8K, T8H); | |
1590 Ci[WS(csi, 61)] = FMS(KP989176509, T8K, T8H); | |
1591 } | |
1592 { | |
1593 E T8F, T8G, T8L, T8M; | |
1594 T8F = FNMS(KP956940335, T8E, T8B); | |
1595 T8G = T8t - T8a; | |
1596 Ci[WS(csi, 29)] = FMS(KP989176509, T8G, T8F); | |
1597 Ci[WS(csi, 35)] = FMA(KP989176509, T8G, T8F); | |
1598 T8L = FNMS(KP956940335, T7Q, T7B); | |
1599 T8M = T8I + T8J; | |
1600 Cr[WS(csr, 35)] = FMA(KP989176509, T8M, T8L); | |
1601 Cr[WS(csr, 29)] = FNMS(KP989176509, T8M, T8L); | |
1602 } | |
1603 { | |
1604 E T8P, T8W, T91, T94; | |
1605 T8P = FMA(KP956940335, T8O, T8N); | |
1606 T8W = T8S + T8V; | |
1607 Cr[WS(csr, 51)] = FNMS(KP803207531, T8W, T8P); | |
1608 Cr[WS(csr, 13)] = FMA(KP803207531, T8W, T8P); | |
1609 T91 = FNMS(KP956940335, T8Y, T8X); | |
1610 T94 = T92 - T93; | |
1611 Ci[WS(csi, 13)] = FMS(KP803207531, T94, T91); | |
1612 Ci[WS(csi, 51)] = FMA(KP803207531, T94, T91); | |
1613 } | |
1614 { | |
1615 E T8Z, T90, T95, T96; | |
1616 T8Z = FMA(KP956940335, T8Y, T8X); | |
1617 T90 = T8V - T8S; | |
1618 Ci[WS(csi, 19)] = FMA(KP803207531, T90, T8Z); | |
1619 Ci[WS(csi, 45)] = FMS(KP803207531, T90, T8Z); | |
1620 T95 = FNMS(KP956940335, T8O, T8N); | |
1621 T96 = T93 + T92; | |
1622 Cr[WS(csr, 45)] = FNMS(KP803207531, T96, T95); | |
1623 Cr[WS(csr, 19)] = FMA(KP803207531, T96, T95); | |
1624 } | |
1625 } | |
1626 } | |
1627 } | |
1628 } | |
1629 | |
1630 static const kr2c_desc desc = { 128, "r2cf_128", {440, 0, 516, 0}, &GENUS }; | |
1631 | |
1632 void X(codelet_r2cf_128) (planner *p) { | |
1633 X(kr2c_register) (p, r2cf_128, &desc); | |
1634 } | |
1635 | |
1636 #else | |
1637 | |
1638 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 128 -name r2cf_128 -include rdft/scalar/r2cf.h */ | |
1639 | |
1640 /* | |
1641 * This function contains 956 FP additions, 330 FP multiplications, | |
1642 * (or, 812 additions, 186 multiplications, 144 fused multiply/add), | |
1643 * 186 stack variables, 31 constants, and 256 memory accesses | |
1644 */ | |
1645 #include "rdft/scalar/r2cf.h" | |
1646 | |
1647 static void r2cf_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
1648 { | |
1649 DK(KP803207531, +0.803207531480644909806676512963141923879569427); | |
1650 DK(KP595699304, +0.595699304492433343467036528829969889511926338); | |
1651 DK(KP146730474, +0.146730474455361751658850129646717819706215317); | |
1652 DK(KP989176509, +0.989176509964780973451673738016243063983689533); | |
1653 DK(KP740951125, +0.740951125354959091175616897495162729728955309); | |
1654 DK(KP671558954, +0.671558954847018400625376850427421803228750632); | |
1655 DK(KP049067674, +0.049067674327418014254954976942682658314745363); | |
1656 DK(KP998795456, +0.998795456205172392714771604759100694443203615); | |
1657 DK(KP242980179, +0.242980179903263889948274162077471118320990783); | |
1658 DK(KP970031253, +0.970031253194543992603984207286100251456865962); | |
1659 DK(KP514102744, +0.514102744193221726593693838968815772608049120); | |
1660 DK(KP857728610, +0.857728610000272069902269984284770137042490799); | |
1661 DK(KP336889853, +0.336889853392220050689253212619147570477766780); | |
1662 DK(KP941544065, +0.941544065183020778412509402599502357185589796); | |
1663 DK(KP427555093, +0.427555093430282094320966856888798534304578629); | |
1664 DK(KP903989293, +0.903989293123443331586200297230537048710132025); | |
1665 DK(KP098017140, +0.098017140329560601994195563888641845861136673); | |
1666 DK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
1667 DK(KP634393284, +0.634393284163645498215171613225493370675687095); | |
1668 DK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
1669 DK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
1670 DK(KP471396736, +0.471396736825997648556387625905254377657460319); | |
1671 DK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
1672 DK(KP290284677, +0.290284677254462367636192375817395274691476278); | |
1673 DK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
1674 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
1675 DK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
1676 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
1677 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
1678 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
1679 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
1680 { | |
1681 INT i; | |
1682 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)) { | |
1683 E TcD, TdU, T27, T7r, T5S, T8y, Tf, Ta5, Tu, Tbq, TcG, TdV, T2e, T8z, T5V; | |
1684 E T7s, TK, Ta6, TcK, TdX, T2o, T5X, T7w, T8B, TZ, Ta7, TcN, TdY, T2x, T5Y; | |
1685 E T7z, T8C, T1g, Taa, TcU, TeA, TcX, Tez, T1v, Tab, T2M, T6z, T7E, T9e, T7H; | |
1686 E T9d, T2T, T6A, T4X, T6L, Tdz, TeL, TdK, TeP, T5G, T6P, T8d, T9p, TaV, Tc3; | |
1687 E Tbi, Tc4, T8o, T9t, T3I, T6H, Tde, TeH, Tdp, TeF, T4r, T6F, T7U, T9l, Tao; | |
1688 E TbW, TaL, TbX, T85, T9j, T1L, Tad, Td3, Tew, Td6, Tex, T20, Tae, T37, T6x; | |
1689 E T7L, T9a, T7O, T9b, T3e, T6w, TbZ, Tc0, T3Z, T4s, Tds, TeI, T4g, T4t, T80; | |
1690 E T87, Tdl, TeE, T7X, T86, TaD, TaM, Tc6, Tc7, T5e, T5H, TdN, TeM, T5v, T5I; | |
1691 E T8j, T8q, TdG, TeO, T8g, T8p, Tba, Tbj; | |
1692 { | |
1693 E T3, T23, Td, T25, T6, T5R, Ta, T24; | |
1694 { | |
1695 E T1, T2, Tb, Tc; | |
1696 T1 = R0[0]; | |
1697 T2 = R0[WS(rs, 32)]; | |
1698 T3 = T1 + T2; | |
1699 T23 = T1 - T2; | |
1700 Tb = R0[WS(rs, 56)]; | |
1701 Tc = R0[WS(rs, 24)]; | |
1702 Td = Tb + Tc; | |
1703 T25 = Tb - Tc; | |
1704 } | |
1705 { | |
1706 E T4, T5, T8, T9; | |
1707 T4 = R0[WS(rs, 16)]; | |
1708 T5 = R0[WS(rs, 48)]; | |
1709 T6 = T4 + T5; | |
1710 T5R = T4 - T5; | |
1711 T8 = R0[WS(rs, 8)]; | |
1712 T9 = R0[WS(rs, 40)]; | |
1713 Ta = T8 + T9; | |
1714 T24 = T8 - T9; | |
1715 } | |
1716 TcD = T3 - T6; | |
1717 TdU = Td - Ta; | |
1718 { | |
1719 E T26, T5Q, T7, Te; | |
1720 T26 = KP707106781 * (T24 + T25); | |
1721 T27 = T23 + T26; | |
1722 T7r = T23 - T26; | |
1723 T5Q = KP707106781 * (T25 - T24); | |
1724 T5S = T5Q - T5R; | |
1725 T8y = T5R + T5Q; | |
1726 T7 = T3 + T6; | |
1727 Te = Ta + Td; | |
1728 Tf = T7 + Te; | |
1729 Ta5 = T7 - Te; | |
1730 } | |
1731 } | |
1732 { | |
1733 E Ti, T28, Ts, T2c, Tl, T29, Tp, T2b; | |
1734 { | |
1735 E Tg, Th, Tq, Tr; | |
1736 Tg = R0[WS(rs, 4)]; | |
1737 Th = R0[WS(rs, 36)]; | |
1738 Ti = Tg + Th; | |
1739 T28 = Tg - Th; | |
1740 Tq = R0[WS(rs, 12)]; | |
1741 Tr = R0[WS(rs, 44)]; | |
1742 Ts = Tq + Tr; | |
1743 T2c = Tq - Tr; | |
1744 } | |
1745 { | |
1746 E Tj, Tk, Tn, To; | |
1747 Tj = R0[WS(rs, 20)]; | |
1748 Tk = R0[WS(rs, 52)]; | |
1749 Tl = Tj + Tk; | |
1750 T29 = Tj - Tk; | |
1751 Tn = R0[WS(rs, 60)]; | |
1752 To = R0[WS(rs, 28)]; | |
1753 Tp = Tn + To; | |
1754 T2b = Tn - To; | |
1755 } | |
1756 { | |
1757 E Tm, Tt, TcE, TcF; | |
1758 Tm = Ti + Tl; | |
1759 Tt = Tp + Ts; | |
1760 Tu = Tm + Tt; | |
1761 Tbq = Tt - Tm; | |
1762 TcE = Ti - Tl; | |
1763 TcF = Tp - Ts; | |
1764 TcG = KP707106781 * (TcE + TcF); | |
1765 TdV = KP707106781 * (TcF - TcE); | |
1766 } | |
1767 { | |
1768 E T2a, T2d, T5T, T5U; | |
1769 T2a = FNMS(KP382683432, T29, KP923879532 * T28); | |
1770 T2d = FMA(KP923879532, T2b, KP382683432 * T2c); | |
1771 T2e = T2a + T2d; | |
1772 T8z = T2d - T2a; | |
1773 T5T = FNMS(KP923879532, T2c, KP382683432 * T2b); | |
1774 T5U = FMA(KP382683432, T28, KP923879532 * T29); | |
1775 T5V = T5T - T5U; | |
1776 T7s = T5U + T5T; | |
1777 } | |
1778 } | |
1779 { | |
1780 E Ty, T2g, TB, T2m, TF, T2l, TI, T2j; | |
1781 { | |
1782 E Tw, Tx, Tz, TA; | |
1783 Tw = R0[WS(rs, 2)]; | |
1784 Tx = R0[WS(rs, 34)]; | |
1785 Ty = Tw + Tx; | |
1786 T2g = Tw - Tx; | |
1787 Tz = R0[WS(rs, 18)]; | |
1788 TA = R0[WS(rs, 50)]; | |
1789 TB = Tz + TA; | |
1790 T2m = Tz - TA; | |
1791 { | |
1792 E TD, TE, T2h, TG, TH, T2i; | |
1793 TD = R0[WS(rs, 10)]; | |
1794 TE = R0[WS(rs, 42)]; | |
1795 T2h = TD - TE; | |
1796 TG = R0[WS(rs, 58)]; | |
1797 TH = R0[WS(rs, 26)]; | |
1798 T2i = TG - TH; | |
1799 TF = TD + TE; | |
1800 T2l = KP707106781 * (T2i - T2h); | |
1801 TI = TG + TH; | |
1802 T2j = KP707106781 * (T2h + T2i); | |
1803 } | |
1804 } | |
1805 { | |
1806 E TC, TJ, TcI, TcJ; | |
1807 TC = Ty + TB; | |
1808 TJ = TF + TI; | |
1809 TK = TC + TJ; | |
1810 Ta6 = TC - TJ; | |
1811 TcI = Ty - TB; | |
1812 TcJ = TI - TF; | |
1813 TcK = FMA(KP923879532, TcI, KP382683432 * TcJ); | |
1814 TdX = FNMS(KP382683432, TcI, KP923879532 * TcJ); | |
1815 } | |
1816 { | |
1817 E T2k, T2n, T7u, T7v; | |
1818 T2k = T2g + T2j; | |
1819 T2n = T2l - T2m; | |
1820 T2o = FMA(KP980785280, T2k, KP195090322 * T2n); | |
1821 T5X = FNMS(KP195090322, T2k, KP980785280 * T2n); | |
1822 T7u = T2g - T2j; | |
1823 T7v = T2m + T2l; | |
1824 T7w = FMA(KP831469612, T7u, KP555570233 * T7v); | |
1825 T8B = FNMS(KP555570233, T7u, KP831469612 * T7v); | |
1826 } | |
1827 } | |
1828 { | |
1829 E TN, T2p, TQ, T2v, TU, T2u, TX, T2s; | |
1830 { | |
1831 E TL, TM, TO, TP; | |
1832 TL = R0[WS(rs, 62)]; | |
1833 TM = R0[WS(rs, 30)]; | |
1834 TN = TL + TM; | |
1835 T2p = TL - TM; | |
1836 TO = R0[WS(rs, 14)]; | |
1837 TP = R0[WS(rs, 46)]; | |
1838 TQ = TO + TP; | |
1839 T2v = TO - TP; | |
1840 { | |
1841 E TS, TT, T2q, TV, TW, T2r; | |
1842 TS = R0[WS(rs, 6)]; | |
1843 TT = R0[WS(rs, 38)]; | |
1844 T2q = TS - TT; | |
1845 TV = R0[WS(rs, 54)]; | |
1846 TW = R0[WS(rs, 22)]; | |
1847 T2r = TV - TW; | |
1848 TU = TS + TT; | |
1849 T2u = KP707106781 * (T2r - T2q); | |
1850 TX = TV + TW; | |
1851 T2s = KP707106781 * (T2q + T2r); | |
1852 } | |
1853 } | |
1854 { | |
1855 E TR, TY, TcL, TcM; | |
1856 TR = TN + TQ; | |
1857 TY = TU + TX; | |
1858 TZ = TR + TY; | |
1859 Ta7 = TR - TY; | |
1860 TcL = TN - TQ; | |
1861 TcM = TX - TU; | |
1862 TcN = FNMS(KP382683432, TcM, KP923879532 * TcL); | |
1863 TdY = FMA(KP382683432, TcL, KP923879532 * TcM); | |
1864 } | |
1865 { | |
1866 E T2t, T2w, T7x, T7y; | |
1867 T2t = T2p + T2s; | |
1868 T2w = T2u - T2v; | |
1869 T2x = FNMS(KP195090322, T2w, KP980785280 * T2t); | |
1870 T5Y = FMA(KP195090322, T2t, KP980785280 * T2w); | |
1871 T7x = T2p - T2s; | |
1872 T7y = T2v + T2u; | |
1873 T7z = FNMS(KP555570233, T7y, KP831469612 * T7x); | |
1874 T8C = FMA(KP555570233, T7x, KP831469612 * T7y); | |
1875 } | |
1876 } | |
1877 { | |
1878 E T14, T2N, T17, T2D, T1b, T2O, T1e, T2C, T1j, T1m, T2K, TcR, T2Q, T1q, T1t; | |
1879 E T2H, TcS, T2R; | |
1880 { | |
1881 E T12, T13, T15, T16; | |
1882 T12 = R0[WS(rs, 1)]; | |
1883 T13 = R0[WS(rs, 33)]; | |
1884 T14 = T12 + T13; | |
1885 T2N = T12 - T13; | |
1886 T15 = R0[WS(rs, 17)]; | |
1887 T16 = R0[WS(rs, 49)]; | |
1888 T17 = T15 + T16; | |
1889 T2D = T15 - T16; | |
1890 } | |
1891 { | |
1892 E T19, T1a, T2B, T1c, T1d, T2A; | |
1893 T19 = R0[WS(rs, 9)]; | |
1894 T1a = R0[WS(rs, 41)]; | |
1895 T2B = T19 - T1a; | |
1896 T1c = R0[WS(rs, 57)]; | |
1897 T1d = R0[WS(rs, 25)]; | |
1898 T2A = T1c - T1d; | |
1899 T1b = T19 + T1a; | |
1900 T2O = KP707106781 * (T2B + T2A); | |
1901 T1e = T1c + T1d; | |
1902 T2C = KP707106781 * (T2A - T2B); | |
1903 } | |
1904 { | |
1905 E T2I, T2J, T2F, T2G; | |
1906 { | |
1907 E T1h, T1i, T1k, T1l; | |
1908 T1h = R0[WS(rs, 5)]; | |
1909 T1i = R0[WS(rs, 37)]; | |
1910 T1j = T1h + T1i; | |
1911 T2I = T1h - T1i; | |
1912 T1k = R0[WS(rs, 21)]; | |
1913 T1l = R0[WS(rs, 53)]; | |
1914 T1m = T1k + T1l; | |
1915 T2J = T1k - T1l; | |
1916 } | |
1917 T2K = FMA(KP382683432, T2I, KP923879532 * T2J); | |
1918 TcR = T1j - T1m; | |
1919 T2Q = FNMS(KP382683432, T2J, KP923879532 * T2I); | |
1920 { | |
1921 E T1o, T1p, T1r, T1s; | |
1922 T1o = R0[WS(rs, 61)]; | |
1923 T1p = R0[WS(rs, 29)]; | |
1924 T1q = T1o + T1p; | |
1925 T2F = T1o - T1p; | |
1926 T1r = R0[WS(rs, 13)]; | |
1927 T1s = R0[WS(rs, 45)]; | |
1928 T1t = T1r + T1s; | |
1929 T2G = T1r - T1s; | |
1930 } | |
1931 T2H = FNMS(KP923879532, T2G, KP382683432 * T2F); | |
1932 TcS = T1q - T1t; | |
1933 T2R = FMA(KP923879532, T2F, KP382683432 * T2G); | |
1934 } | |
1935 { | |
1936 E T18, T1f, TcQ, TcT; | |
1937 T18 = T14 + T17; | |
1938 T1f = T1b + T1e; | |
1939 T1g = T18 + T1f; | |
1940 Taa = T18 - T1f; | |
1941 TcQ = T14 - T17; | |
1942 TcT = KP707106781 * (TcR + TcS); | |
1943 TcU = TcQ + TcT; | |
1944 TeA = TcQ - TcT; | |
1945 } | |
1946 { | |
1947 E TcV, TcW, T1n, T1u; | |
1948 TcV = T1e - T1b; | |
1949 TcW = KP707106781 * (TcS - TcR); | |
1950 TcX = TcV + TcW; | |
1951 Tez = TcW - TcV; | |
1952 T1n = T1j + T1m; | |
1953 T1u = T1q + T1t; | |
1954 T1v = T1n + T1u; | |
1955 Tab = T1u - T1n; | |
1956 } | |
1957 { | |
1958 E T2E, T2L, T7C, T7D; | |
1959 T2E = T2C - T2D; | |
1960 T2L = T2H - T2K; | |
1961 T2M = T2E + T2L; | |
1962 T6z = T2L - T2E; | |
1963 T7C = T2N - T2O; | |
1964 T7D = T2K + T2H; | |
1965 T7E = T7C + T7D; | |
1966 T9e = T7C - T7D; | |
1967 } | |
1968 { | |
1969 E T7F, T7G, T2P, T2S; | |
1970 T7F = T2D + T2C; | |
1971 T7G = T2R - T2Q; | |
1972 T7H = T7F + T7G; | |
1973 T9d = T7G - T7F; | |
1974 T2P = T2N + T2O; | |
1975 T2S = T2Q + T2R; | |
1976 T2T = T2P + T2S; | |
1977 T6A = T2P - T2S; | |
1978 } | |
1979 } | |
1980 { | |
1981 E T4z, TaP, T5B, TaQ, T4G, TaT, T5y, TaS, Tbf, Tbg, T4O, Tdw, T5E, Tbc, Tbd; | |
1982 E T4V, Tdx, T5D; | |
1983 { | |
1984 E T4x, T4y, T5z, T5A; | |
1985 T4x = R1[WS(rs, 63)]; | |
1986 T4y = R1[WS(rs, 31)]; | |
1987 T4z = T4x - T4y; | |
1988 TaP = T4x + T4y; | |
1989 T5z = R1[WS(rs, 15)]; | |
1990 T5A = R1[WS(rs, 47)]; | |
1991 T5B = T5z - T5A; | |
1992 TaQ = T5z + T5A; | |
1993 } | |
1994 { | |
1995 E T4A, T4B, T4C, T4D, T4E, T4F; | |
1996 T4A = R1[WS(rs, 7)]; | |
1997 T4B = R1[WS(rs, 39)]; | |
1998 T4C = T4A - T4B; | |
1999 T4D = R1[WS(rs, 55)]; | |
2000 T4E = R1[WS(rs, 23)]; | |
2001 T4F = T4D - T4E; | |
2002 T4G = KP707106781 * (T4C + T4F); | |
2003 TaT = T4D + T4E; | |
2004 T5y = KP707106781 * (T4F - T4C); | |
2005 TaS = T4A + T4B; | |
2006 } | |
2007 { | |
2008 E T4K, T4N, T4R, T4U; | |
2009 { | |
2010 E T4I, T4J, T4L, T4M; | |
2011 T4I = R1[WS(rs, 3)]; | |
2012 T4J = R1[WS(rs, 35)]; | |
2013 T4K = T4I - T4J; | |
2014 Tbf = T4I + T4J; | |
2015 T4L = R1[WS(rs, 19)]; | |
2016 T4M = R1[WS(rs, 51)]; | |
2017 T4N = T4L - T4M; | |
2018 Tbg = T4L + T4M; | |
2019 } | |
2020 T4O = FNMS(KP382683432, T4N, KP923879532 * T4K); | |
2021 Tdw = Tbf - Tbg; | |
2022 T5E = FMA(KP382683432, T4K, KP923879532 * T4N); | |
2023 { | |
2024 E T4P, T4Q, T4S, T4T; | |
2025 T4P = R1[WS(rs, 59)]; | |
2026 T4Q = R1[WS(rs, 27)]; | |
2027 T4R = T4P - T4Q; | |
2028 Tbc = T4P + T4Q; | |
2029 T4S = R1[WS(rs, 11)]; | |
2030 T4T = R1[WS(rs, 43)]; | |
2031 T4U = T4S - T4T; | |
2032 Tbd = T4S + T4T; | |
2033 } | |
2034 T4V = FMA(KP923879532, T4R, KP382683432 * T4U); | |
2035 Tdx = Tbc - Tbd; | |
2036 T5D = FNMS(KP923879532, T4U, KP382683432 * T4R); | |
2037 } | |
2038 { | |
2039 E T4H, T4W, Tdv, Tdy; | |
2040 T4H = T4z + T4G; | |
2041 T4W = T4O + T4V; | |
2042 T4X = T4H + T4W; | |
2043 T6L = T4H - T4W; | |
2044 Tdv = TaP - TaQ; | |
2045 Tdy = KP707106781 * (Tdw + Tdx); | |
2046 Tdz = Tdv + Tdy; | |
2047 TeL = Tdv - Tdy; | |
2048 } | |
2049 { | |
2050 E TdI, TdJ, T5C, T5F; | |
2051 TdI = TaT - TaS; | |
2052 TdJ = KP707106781 * (Tdx - Tdw); | |
2053 TdK = TdI + TdJ; | |
2054 TeP = TdJ - TdI; | |
2055 T5C = T5y - T5B; | |
2056 T5F = T5D - T5E; | |
2057 T5G = T5C + T5F; | |
2058 T6P = T5F - T5C; | |
2059 } | |
2060 { | |
2061 E T8b, T8c, TaR, TaU; | |
2062 T8b = T4z - T4G; | |
2063 T8c = T5E + T5D; | |
2064 T8d = T8b + T8c; | |
2065 T9p = T8b - T8c; | |
2066 TaR = TaP + TaQ; | |
2067 TaU = TaS + TaT; | |
2068 TaV = TaR - TaU; | |
2069 Tc3 = TaR + TaU; | |
2070 } | |
2071 { | |
2072 E Tbe, Tbh, T8m, T8n; | |
2073 Tbe = Tbc + Tbd; | |
2074 Tbh = Tbf + Tbg; | |
2075 Tbi = Tbe - Tbh; | |
2076 Tc4 = Tbh + Tbe; | |
2077 T8m = T5B + T5y; | |
2078 T8n = T4V - T4O; | |
2079 T8o = T8m + T8n; | |
2080 T9t = T8n - T8m; | |
2081 } | |
2082 } | |
2083 { | |
2084 E T3k, Tai, T4m, Taj, T3r, Tam, T4j, Tal, TaI, TaJ, T3z, Tdb, T4p, TaF, TaG; | |
2085 E T3G, Tdc, T4o; | |
2086 { | |
2087 E T3i, T3j, T4k, T4l; | |
2088 T3i = R1[0]; | |
2089 T3j = R1[WS(rs, 32)]; | |
2090 T3k = T3i - T3j; | |
2091 Tai = T3i + T3j; | |
2092 T4k = R1[WS(rs, 16)]; | |
2093 T4l = R1[WS(rs, 48)]; | |
2094 T4m = T4k - T4l; | |
2095 Taj = T4k + T4l; | |
2096 } | |
2097 { | |
2098 E T3l, T3m, T3n, T3o, T3p, T3q; | |
2099 T3l = R1[WS(rs, 8)]; | |
2100 T3m = R1[WS(rs, 40)]; | |
2101 T3n = T3l - T3m; | |
2102 T3o = R1[WS(rs, 56)]; | |
2103 T3p = R1[WS(rs, 24)]; | |
2104 T3q = T3o - T3p; | |
2105 T3r = KP707106781 * (T3n + T3q); | |
2106 Tam = T3o + T3p; | |
2107 T4j = KP707106781 * (T3q - T3n); | |
2108 Tal = T3l + T3m; | |
2109 } | |
2110 { | |
2111 E T3v, T3y, T3C, T3F; | |
2112 { | |
2113 E T3t, T3u, T3w, T3x; | |
2114 T3t = R1[WS(rs, 4)]; | |
2115 T3u = R1[WS(rs, 36)]; | |
2116 T3v = T3t - T3u; | |
2117 TaI = T3t + T3u; | |
2118 T3w = R1[WS(rs, 20)]; | |
2119 T3x = R1[WS(rs, 52)]; | |
2120 T3y = T3w - T3x; | |
2121 TaJ = T3w + T3x; | |
2122 } | |
2123 T3z = FNMS(KP382683432, T3y, KP923879532 * T3v); | |
2124 Tdb = TaI - TaJ; | |
2125 T4p = FMA(KP382683432, T3v, KP923879532 * T3y); | |
2126 { | |
2127 E T3A, T3B, T3D, T3E; | |
2128 T3A = R1[WS(rs, 60)]; | |
2129 T3B = R1[WS(rs, 28)]; | |
2130 T3C = T3A - T3B; | |
2131 TaF = T3A + T3B; | |
2132 T3D = R1[WS(rs, 12)]; | |
2133 T3E = R1[WS(rs, 44)]; | |
2134 T3F = T3D - T3E; | |
2135 TaG = T3D + T3E; | |
2136 } | |
2137 T3G = FMA(KP923879532, T3C, KP382683432 * T3F); | |
2138 Tdc = TaF - TaG; | |
2139 T4o = FNMS(KP923879532, T3F, KP382683432 * T3C); | |
2140 } | |
2141 { | |
2142 E T3s, T3H, Tda, Tdd; | |
2143 T3s = T3k + T3r; | |
2144 T3H = T3z + T3G; | |
2145 T3I = T3s + T3H; | |
2146 T6H = T3s - T3H; | |
2147 Tda = Tai - Taj; | |
2148 Tdd = KP707106781 * (Tdb + Tdc); | |
2149 Tde = Tda + Tdd; | |
2150 TeH = Tda - Tdd; | |
2151 } | |
2152 { | |
2153 E Tdn, Tdo, T4n, T4q; | |
2154 Tdn = Tam - Tal; | |
2155 Tdo = KP707106781 * (Tdc - Tdb); | |
2156 Tdp = Tdn + Tdo; | |
2157 TeF = Tdo - Tdn; | |
2158 T4n = T4j - T4m; | |
2159 T4q = T4o - T4p; | |
2160 T4r = T4n + T4q; | |
2161 T6F = T4q - T4n; | |
2162 } | |
2163 { | |
2164 E T7S, T7T, Tak, Tan; | |
2165 T7S = T3k - T3r; | |
2166 T7T = T4p + T4o; | |
2167 T7U = T7S + T7T; | |
2168 T9l = T7S - T7T; | |
2169 Tak = Tai + Taj; | |
2170 Tan = Tal + Tam; | |
2171 Tao = Tak - Tan; | |
2172 TbW = Tak + Tan; | |
2173 } | |
2174 { | |
2175 E TaH, TaK, T83, T84; | |
2176 TaH = TaF + TaG; | |
2177 TaK = TaI + TaJ; | |
2178 TaL = TaH - TaK; | |
2179 TbX = TaK + TaH; | |
2180 T83 = T4m + T4j; | |
2181 T84 = T3G - T3z; | |
2182 T85 = T83 + T84; | |
2183 T9j = T84 - T83; | |
2184 } | |
2185 } | |
2186 { | |
2187 E T1z, T2V, T1C, T39, T1G, T38, T1J, T2Y, T1O, T1R, T32, Td0, T3c, T1V, T1Y; | |
2188 E T35, Td1, T3b; | |
2189 { | |
2190 E T1x, T1y, T1A, T1B; | |
2191 T1x = R0[WS(rs, 63)]; | |
2192 T1y = R0[WS(rs, 31)]; | |
2193 T1z = T1x + T1y; | |
2194 T2V = T1x - T1y; | |
2195 T1A = R0[WS(rs, 15)]; | |
2196 T1B = R0[WS(rs, 47)]; | |
2197 T1C = T1A + T1B; | |
2198 T39 = T1A - T1B; | |
2199 } | |
2200 { | |
2201 E T1E, T1F, T2W, T1H, T1I, T2X; | |
2202 T1E = R0[WS(rs, 7)]; | |
2203 T1F = R0[WS(rs, 39)]; | |
2204 T2W = T1E - T1F; | |
2205 T1H = R0[WS(rs, 55)]; | |
2206 T1I = R0[WS(rs, 23)]; | |
2207 T2X = T1H - T1I; | |
2208 T1G = T1E + T1F; | |
2209 T38 = KP707106781 * (T2X - T2W); | |
2210 T1J = T1H + T1I; | |
2211 T2Y = KP707106781 * (T2W + T2X); | |
2212 } | |
2213 { | |
2214 E T30, T31, T33, T34; | |
2215 { | |
2216 E T1M, T1N, T1P, T1Q; | |
2217 T1M = R0[WS(rs, 3)]; | |
2218 T1N = R0[WS(rs, 35)]; | |
2219 T1O = T1M + T1N; | |
2220 T30 = T1M - T1N; | |
2221 T1P = R0[WS(rs, 19)]; | |
2222 T1Q = R0[WS(rs, 51)]; | |
2223 T1R = T1P + T1Q; | |
2224 T31 = T1P - T1Q; | |
2225 } | |
2226 T32 = FNMS(KP382683432, T31, KP923879532 * T30); | |
2227 Td0 = T1O - T1R; | |
2228 T3c = FMA(KP382683432, T30, KP923879532 * T31); | |
2229 { | |
2230 E T1T, T1U, T1W, T1X; | |
2231 T1T = R0[WS(rs, 59)]; | |
2232 T1U = R0[WS(rs, 27)]; | |
2233 T1V = T1T + T1U; | |
2234 T33 = T1T - T1U; | |
2235 T1W = R0[WS(rs, 11)]; | |
2236 T1X = R0[WS(rs, 43)]; | |
2237 T1Y = T1W + T1X; | |
2238 T34 = T1W - T1X; | |
2239 } | |
2240 T35 = FMA(KP923879532, T33, KP382683432 * T34); | |
2241 Td1 = T1V - T1Y; | |
2242 T3b = FNMS(KP923879532, T34, KP382683432 * T33); | |
2243 } | |
2244 { | |
2245 E T1D, T1K, TcZ, Td2; | |
2246 T1D = T1z + T1C; | |
2247 T1K = T1G + T1J; | |
2248 T1L = T1D + T1K; | |
2249 Tad = T1D - T1K; | |
2250 TcZ = T1z - T1C; | |
2251 Td2 = KP707106781 * (Td0 + Td1); | |
2252 Td3 = TcZ + Td2; | |
2253 Tew = TcZ - Td2; | |
2254 } | |
2255 { | |
2256 E Td4, Td5, T1S, T1Z; | |
2257 Td4 = T1J - T1G; | |
2258 Td5 = KP707106781 * (Td1 - Td0); | |
2259 Td6 = Td4 + Td5; | |
2260 Tex = Td5 - Td4; | |
2261 T1S = T1O + T1R; | |
2262 T1Z = T1V + T1Y; | |
2263 T20 = T1S + T1Z; | |
2264 Tae = T1Z - T1S; | |
2265 } | |
2266 { | |
2267 E T2Z, T36, T7J, T7K; | |
2268 T2Z = T2V + T2Y; | |
2269 T36 = T32 + T35; | |
2270 T37 = T2Z + T36; | |
2271 T6x = T2Z - T36; | |
2272 T7J = T2V - T2Y; | |
2273 T7K = T3c + T3b; | |
2274 T7L = T7J + T7K; | |
2275 T9a = T7J - T7K; | |
2276 } | |
2277 { | |
2278 E T7M, T7N, T3a, T3d; | |
2279 T7M = T39 + T38; | |
2280 T7N = T35 - T32; | |
2281 T7O = T7M + T7N; | |
2282 T9b = T7N - T7M; | |
2283 T3a = T38 - T39; | |
2284 T3d = T3b - T3c; | |
2285 T3e = T3a + T3d; | |
2286 T6w = T3d - T3a; | |
2287 } | |
2288 } | |
2289 { | |
2290 E T3L, Tdf, T3X, Tar, T42, Tdi, T4e, Tay, T3S, Tdg, T3U, Tau, T49, Tdj, T4b; | |
2291 E TaB, Tdh, Tdk; | |
2292 { | |
2293 E T3J, T3K, Tap, T3V, T3W, Taq; | |
2294 T3J = R1[WS(rs, 2)]; | |
2295 T3K = R1[WS(rs, 34)]; | |
2296 Tap = T3J + T3K; | |
2297 T3V = R1[WS(rs, 18)]; | |
2298 T3W = R1[WS(rs, 50)]; | |
2299 Taq = T3V + T3W; | |
2300 T3L = T3J - T3K; | |
2301 Tdf = Tap - Taq; | |
2302 T3X = T3V - T3W; | |
2303 Tar = Tap + Taq; | |
2304 } | |
2305 { | |
2306 E T40, T41, Taw, T4c, T4d, Tax; | |
2307 T40 = R1[WS(rs, 62)]; | |
2308 T41 = R1[WS(rs, 30)]; | |
2309 Taw = T40 + T41; | |
2310 T4c = R1[WS(rs, 14)]; | |
2311 T4d = R1[WS(rs, 46)]; | |
2312 Tax = T4c + T4d; | |
2313 T42 = T40 - T41; | |
2314 Tdi = Taw - Tax; | |
2315 T4e = T4c - T4d; | |
2316 Tay = Taw + Tax; | |
2317 } | |
2318 { | |
2319 E T3O, Tas, T3R, Tat; | |
2320 { | |
2321 E T3M, T3N, T3P, T3Q; | |
2322 T3M = R1[WS(rs, 10)]; | |
2323 T3N = R1[WS(rs, 42)]; | |
2324 T3O = T3M - T3N; | |
2325 Tas = T3M + T3N; | |
2326 T3P = R1[WS(rs, 58)]; | |
2327 T3Q = R1[WS(rs, 26)]; | |
2328 T3R = T3P - T3Q; | |
2329 Tat = T3P + T3Q; | |
2330 } | |
2331 T3S = KP707106781 * (T3O + T3R); | |
2332 Tdg = Tat - Tas; | |
2333 T3U = KP707106781 * (T3R - T3O); | |
2334 Tau = Tas + Tat; | |
2335 } | |
2336 { | |
2337 E T45, Taz, T48, TaA; | |
2338 { | |
2339 E T43, T44, T46, T47; | |
2340 T43 = R1[WS(rs, 6)]; | |
2341 T44 = R1[WS(rs, 38)]; | |
2342 T45 = T43 - T44; | |
2343 Taz = T43 + T44; | |
2344 T46 = R1[WS(rs, 54)]; | |
2345 T47 = R1[WS(rs, 22)]; | |
2346 T48 = T46 - T47; | |
2347 TaA = T46 + T47; | |
2348 } | |
2349 T49 = KP707106781 * (T45 + T48); | |
2350 Tdj = TaA - Taz; | |
2351 T4b = KP707106781 * (T48 - T45); | |
2352 TaB = Taz + TaA; | |
2353 } | |
2354 TbZ = Tar + Tau; | |
2355 Tc0 = Tay + TaB; | |
2356 { | |
2357 E T3T, T3Y, Tdq, Tdr; | |
2358 T3T = T3L + T3S; | |
2359 T3Y = T3U - T3X; | |
2360 T3Z = FMA(KP980785280, T3T, KP195090322 * T3Y); | |
2361 T4s = FNMS(KP195090322, T3T, KP980785280 * T3Y); | |
2362 Tdq = FNMS(KP382683432, Tdf, KP923879532 * Tdg); | |
2363 Tdr = FMA(KP382683432, Tdi, KP923879532 * Tdj); | |
2364 Tds = Tdq + Tdr; | |
2365 TeI = Tdr - Tdq; | |
2366 } | |
2367 { | |
2368 E T4a, T4f, T7Y, T7Z; | |
2369 T4a = T42 + T49; | |
2370 T4f = T4b - T4e; | |
2371 T4g = FNMS(KP195090322, T4f, KP980785280 * T4a); | |
2372 T4t = FMA(KP195090322, T4a, KP980785280 * T4f); | |
2373 T7Y = T42 - T49; | |
2374 T7Z = T4e + T4b; | |
2375 T80 = FNMS(KP555570233, T7Z, KP831469612 * T7Y); | |
2376 T87 = FMA(KP555570233, T7Y, KP831469612 * T7Z); | |
2377 } | |
2378 Tdh = FMA(KP923879532, Tdf, KP382683432 * Tdg); | |
2379 Tdk = FNMS(KP382683432, Tdj, KP923879532 * Tdi); | |
2380 Tdl = Tdh + Tdk; | |
2381 TeE = Tdk - Tdh; | |
2382 { | |
2383 E T7V, T7W, Tav, TaC; | |
2384 T7V = T3L - T3S; | |
2385 T7W = T3X + T3U; | |
2386 T7X = FMA(KP831469612, T7V, KP555570233 * T7W); | |
2387 T86 = FNMS(KP555570233, T7V, KP831469612 * T7W); | |
2388 Tav = Tar - Tau; | |
2389 TaC = Tay - TaB; | |
2390 TaD = KP707106781 * (Tav + TaC); | |
2391 TaM = KP707106781 * (TaC - Tav); | |
2392 } | |
2393 } | |
2394 { | |
2395 E T50, TdA, T5c, TaY, T5h, TdD, T5t, Tb5, T57, TdB, T59, Tb1, T5o, TdE, T5q; | |
2396 E Tb8, TdC, TdF; | |
2397 { | |
2398 E T4Y, T4Z, TaW, T5a, T5b, TaX; | |
2399 T4Y = R1[WS(rs, 1)]; | |
2400 T4Z = R1[WS(rs, 33)]; | |
2401 TaW = T4Y + T4Z; | |
2402 T5a = R1[WS(rs, 17)]; | |
2403 T5b = R1[WS(rs, 49)]; | |
2404 TaX = T5a + T5b; | |
2405 T50 = T4Y - T4Z; | |
2406 TdA = TaW - TaX; | |
2407 T5c = T5a - T5b; | |
2408 TaY = TaW + TaX; | |
2409 } | |
2410 { | |
2411 E T5f, T5g, Tb3, T5r, T5s, Tb4; | |
2412 T5f = R1[WS(rs, 61)]; | |
2413 T5g = R1[WS(rs, 29)]; | |
2414 Tb3 = T5f + T5g; | |
2415 T5r = R1[WS(rs, 13)]; | |
2416 T5s = R1[WS(rs, 45)]; | |
2417 Tb4 = T5r + T5s; | |
2418 T5h = T5f - T5g; | |
2419 TdD = Tb3 - Tb4; | |
2420 T5t = T5r - T5s; | |
2421 Tb5 = Tb3 + Tb4; | |
2422 } | |
2423 { | |
2424 E T53, TaZ, T56, Tb0; | |
2425 { | |
2426 E T51, T52, T54, T55; | |
2427 T51 = R1[WS(rs, 9)]; | |
2428 T52 = R1[WS(rs, 41)]; | |
2429 T53 = T51 - T52; | |
2430 TaZ = T51 + T52; | |
2431 T54 = R1[WS(rs, 57)]; | |
2432 T55 = R1[WS(rs, 25)]; | |
2433 T56 = T54 - T55; | |
2434 Tb0 = T54 + T55; | |
2435 } | |
2436 T57 = KP707106781 * (T53 + T56); | |
2437 TdB = Tb0 - TaZ; | |
2438 T59 = KP707106781 * (T56 - T53); | |
2439 Tb1 = TaZ + Tb0; | |
2440 } | |
2441 { | |
2442 E T5k, Tb6, T5n, Tb7; | |
2443 { | |
2444 E T5i, T5j, T5l, T5m; | |
2445 T5i = R1[WS(rs, 5)]; | |
2446 T5j = R1[WS(rs, 37)]; | |
2447 T5k = T5i - T5j; | |
2448 Tb6 = T5i + T5j; | |
2449 T5l = R1[WS(rs, 53)]; | |
2450 T5m = R1[WS(rs, 21)]; | |
2451 T5n = T5l - T5m; | |
2452 Tb7 = T5l + T5m; | |
2453 } | |
2454 T5o = KP707106781 * (T5k + T5n); | |
2455 TdE = Tb7 - Tb6; | |
2456 T5q = KP707106781 * (T5n - T5k); | |
2457 Tb8 = Tb6 + Tb7; | |
2458 } | |
2459 Tc6 = TaY + Tb1; | |
2460 Tc7 = Tb5 + Tb8; | |
2461 { | |
2462 E T58, T5d, TdL, TdM; | |
2463 T58 = T50 + T57; | |
2464 T5d = T59 - T5c; | |
2465 T5e = FMA(KP980785280, T58, KP195090322 * T5d); | |
2466 T5H = FNMS(KP195090322, T58, KP980785280 * T5d); | |
2467 TdL = FNMS(KP382683432, TdA, KP923879532 * TdB); | |
2468 TdM = FMA(KP382683432, TdD, KP923879532 * TdE); | |
2469 TdN = TdL + TdM; | |
2470 TeM = TdM - TdL; | |
2471 } | |
2472 { | |
2473 E T5p, T5u, T8h, T8i; | |
2474 T5p = T5h + T5o; | |
2475 T5u = T5q - T5t; | |
2476 T5v = FNMS(KP195090322, T5u, KP980785280 * T5p); | |
2477 T5I = FMA(KP195090322, T5p, KP980785280 * T5u); | |
2478 T8h = T5h - T5o; | |
2479 T8i = T5t + T5q; | |
2480 T8j = FNMS(KP555570233, T8i, KP831469612 * T8h); | |
2481 T8q = FMA(KP555570233, T8h, KP831469612 * T8i); | |
2482 } | |
2483 TdC = FMA(KP923879532, TdA, KP382683432 * TdB); | |
2484 TdF = FNMS(KP382683432, TdE, KP923879532 * TdD); | |
2485 TdG = TdC + TdF; | |
2486 TeO = TdF - TdC; | |
2487 { | |
2488 E T8e, T8f, Tb2, Tb9; | |
2489 T8e = T50 - T57; | |
2490 T8f = T5c + T59; | |
2491 T8g = FMA(KP831469612, T8e, KP555570233 * T8f); | |
2492 T8p = FNMS(KP555570233, T8e, KP831469612 * T8f); | |
2493 Tb2 = TaY - Tb1; | |
2494 Tb9 = Tb5 - Tb8; | |
2495 Tba = KP707106781 * (Tb2 + Tb9); | |
2496 Tbj = KP707106781 * (Tb9 - Tb2); | |
2497 } | |
2498 } | |
2499 { | |
2500 E T11, TbV, Tc9, Tcf, T22, Tcb, Tc2, Tce; | |
2501 { | |
2502 E Tv, T10, Tc5, Tc8; | |
2503 Tv = Tf + Tu; | |
2504 T10 = TK + TZ; | |
2505 T11 = Tv + T10; | |
2506 TbV = Tv - T10; | |
2507 Tc5 = Tc3 + Tc4; | |
2508 Tc8 = Tc6 + Tc7; | |
2509 Tc9 = Tc5 - Tc8; | |
2510 Tcf = Tc5 + Tc8; | |
2511 } | |
2512 { | |
2513 E T1w, T21, TbY, Tc1; | |
2514 T1w = T1g + T1v; | |
2515 T21 = T1L + T20; | |
2516 T22 = T1w + T21; | |
2517 Tcb = T21 - T1w; | |
2518 TbY = TbW + TbX; | |
2519 Tc1 = TbZ + Tc0; | |
2520 Tc2 = TbY - Tc1; | |
2521 Tce = TbY + Tc1; | |
2522 } | |
2523 Cr[WS(csr, 32)] = T11 - T22; | |
2524 Ci[WS(csi, 32)] = Tcf - Tce; | |
2525 { | |
2526 E Tca, Tcc, Tcd, Tcg; | |
2527 Tca = KP707106781 * (Tc2 + Tc9); | |
2528 Cr[WS(csr, 48)] = TbV - Tca; | |
2529 Cr[WS(csr, 16)] = TbV + Tca; | |
2530 Tcc = KP707106781 * (Tc9 - Tc2); | |
2531 Ci[WS(csi, 16)] = Tcb + Tcc; | |
2532 Ci[WS(csi, 48)] = Tcc - Tcb; | |
2533 Tcd = T11 + T22; | |
2534 Tcg = Tce + Tcf; | |
2535 Cr[WS(csr, 64)] = Tcd - Tcg; | |
2536 Cr[0] = Tcd + Tcg; | |
2537 } | |
2538 } | |
2539 { | |
2540 E Tch, Tcu, Tck, Tct, Tco, Tcy, Tcr, Tcz, Tci, Tcj; | |
2541 Tch = Tf - Tu; | |
2542 Tcu = TZ - TK; | |
2543 Tci = T1g - T1v; | |
2544 Tcj = T1L - T20; | |
2545 Tck = KP707106781 * (Tci + Tcj); | |
2546 Tct = KP707106781 * (Tcj - Tci); | |
2547 { | |
2548 E Tcm, Tcn, Tcp, Tcq; | |
2549 Tcm = TbW - TbX; | |
2550 Tcn = Tc0 - TbZ; | |
2551 Tco = FMA(KP923879532, Tcm, KP382683432 * Tcn); | |
2552 Tcy = FNMS(KP382683432, Tcm, KP923879532 * Tcn); | |
2553 Tcp = Tc3 - Tc4; | |
2554 Tcq = Tc7 - Tc6; | |
2555 Tcr = FNMS(KP382683432, Tcq, KP923879532 * Tcp); | |
2556 Tcz = FMA(KP382683432, Tcp, KP923879532 * Tcq); | |
2557 } | |
2558 { | |
2559 E Tcl, Tcs, Tcx, TcA; | |
2560 Tcl = Tch + Tck; | |
2561 Tcs = Tco + Tcr; | |
2562 Cr[WS(csr, 56)] = Tcl - Tcs; | |
2563 Cr[WS(csr, 8)] = Tcl + Tcs; | |
2564 Tcx = Tcu + Tct; | |
2565 TcA = Tcy + Tcz; | |
2566 Ci[WS(csi, 8)] = Tcx + TcA; | |
2567 Ci[WS(csi, 56)] = TcA - Tcx; | |
2568 } | |
2569 { | |
2570 E Tcv, Tcw, TcB, TcC; | |
2571 Tcv = Tct - Tcu; | |
2572 Tcw = Tcr - Tco; | |
2573 Ci[WS(csi, 24)] = Tcv + Tcw; | |
2574 Ci[WS(csi, 40)] = Tcw - Tcv; | |
2575 TcB = Tch - Tck; | |
2576 TcC = Tcz - Tcy; | |
2577 Cr[WS(csr, 40)] = TcB - TcC; | |
2578 Cr[WS(csr, 24)] = TcB + TcC; | |
2579 } | |
2580 } | |
2581 { | |
2582 E Ta9, TbB, Tbs, TbM, Tag, TbL, TbJ, TbR, TaO, Tbw, Tbp, TbC, TbG, TbQ, Tbl; | |
2583 E Tbx, Ta8, Tbr; | |
2584 Ta8 = KP707106781 * (Ta6 + Ta7); | |
2585 Ta9 = Ta5 + Ta8; | |
2586 TbB = Ta5 - Ta8; | |
2587 Tbr = KP707106781 * (Ta7 - Ta6); | |
2588 Tbs = Tbq + Tbr; | |
2589 TbM = Tbr - Tbq; | |
2590 { | |
2591 E Tac, Taf, TbH, TbI; | |
2592 Tac = FMA(KP923879532, Taa, KP382683432 * Tab); | |
2593 Taf = FNMS(KP382683432, Tae, KP923879532 * Tad); | |
2594 Tag = Tac + Taf; | |
2595 TbL = Taf - Tac; | |
2596 TbH = TaV - Tba; | |
2597 TbI = Tbj - Tbi; | |
2598 TbJ = FNMS(KP555570233, TbI, KP831469612 * TbH); | |
2599 TbR = FMA(KP555570233, TbH, KP831469612 * TbI); | |
2600 } | |
2601 { | |
2602 E TaE, TaN, Tbn, Tbo; | |
2603 TaE = Tao + TaD; | |
2604 TaN = TaL + TaM; | |
2605 TaO = FMA(KP980785280, TaE, KP195090322 * TaN); | |
2606 Tbw = FNMS(KP195090322, TaE, KP980785280 * TaN); | |
2607 Tbn = FNMS(KP382683432, Taa, KP923879532 * Tab); | |
2608 Tbo = FMA(KP382683432, Tad, KP923879532 * Tae); | |
2609 Tbp = Tbn + Tbo; | |
2610 TbC = Tbo - Tbn; | |
2611 } | |
2612 { | |
2613 E TbE, TbF, Tbb, Tbk; | |
2614 TbE = Tao - TaD; | |
2615 TbF = TaM - TaL; | |
2616 TbG = FMA(KP831469612, TbE, KP555570233 * TbF); | |
2617 TbQ = FNMS(KP555570233, TbE, KP831469612 * TbF); | |
2618 Tbb = TaV + Tba; | |
2619 Tbk = Tbi + Tbj; | |
2620 Tbl = FNMS(KP195090322, Tbk, KP980785280 * Tbb); | |
2621 Tbx = FMA(KP195090322, Tbb, KP980785280 * Tbk); | |
2622 } | |
2623 { | |
2624 E Tah, Tbm, Tbv, Tby; | |
2625 Tah = Ta9 + Tag; | |
2626 Tbm = TaO + Tbl; | |
2627 Cr[WS(csr, 60)] = Tah - Tbm; | |
2628 Cr[WS(csr, 4)] = Tah + Tbm; | |
2629 Tbv = Tbs + Tbp; | |
2630 Tby = Tbw + Tbx; | |
2631 Ci[WS(csi, 4)] = Tbv + Tby; | |
2632 Ci[WS(csi, 60)] = Tby - Tbv; | |
2633 } | |
2634 { | |
2635 E Tbt, Tbu, Tbz, TbA; | |
2636 Tbt = Tbp - Tbs; | |
2637 Tbu = Tbl - TaO; | |
2638 Ci[WS(csi, 28)] = Tbt + Tbu; | |
2639 Ci[WS(csi, 36)] = Tbu - Tbt; | |
2640 Tbz = Ta9 - Tag; | |
2641 TbA = Tbx - Tbw; | |
2642 Cr[WS(csr, 36)] = Tbz - TbA; | |
2643 Cr[WS(csr, 28)] = Tbz + TbA; | |
2644 } | |
2645 { | |
2646 E TbD, TbK, TbP, TbS; | |
2647 TbD = TbB + TbC; | |
2648 TbK = TbG + TbJ; | |
2649 Cr[WS(csr, 52)] = TbD - TbK; | |
2650 Cr[WS(csr, 12)] = TbD + TbK; | |
2651 TbP = TbM + TbL; | |
2652 TbS = TbQ + TbR; | |
2653 Ci[WS(csi, 12)] = TbP + TbS; | |
2654 Ci[WS(csi, 52)] = TbS - TbP; | |
2655 } | |
2656 { | |
2657 E TbN, TbO, TbT, TbU; | |
2658 TbN = TbL - TbM; | |
2659 TbO = TbJ - TbG; | |
2660 Ci[WS(csi, 20)] = TbN + TbO; | |
2661 Ci[WS(csi, 44)] = TbO - TbN; | |
2662 TbT = TbB - TbC; | |
2663 TbU = TbR - TbQ; | |
2664 Cr[WS(csr, 44)] = TbT - TbU; | |
2665 Cr[WS(csr, 20)] = TbT + TbU; | |
2666 } | |
2667 } | |
2668 { | |
2669 E Tev, Tf7, Tfc, Tfm, Tff, Tfn, TeC, Tfh, TeK, Tf2, TeV, Tf8, TeY, Tfi, TeR; | |
2670 E Tf3; | |
2671 { | |
2672 E Tet, Teu, Tfa, Tfb; | |
2673 Tet = TcD - TcG; | |
2674 Teu = TdY - TdX; | |
2675 Tev = Tet - Teu; | |
2676 Tf7 = Tet + Teu; | |
2677 Tfa = TeF + TeE; | |
2678 Tfb = TeH + TeI; | |
2679 Tfc = FMA(KP290284677, Tfa, KP956940335 * Tfb); | |
2680 Tfm = FNMS(KP290284677, Tfb, KP956940335 * Tfa); | |
2681 } | |
2682 { | |
2683 E Tfd, Tfe, Tey, TeB; | |
2684 Tfd = TeL + TeM; | |
2685 Tfe = TeP + TeO; | |
2686 Tff = FNMS(KP290284677, Tfe, KP956940335 * Tfd); | |
2687 Tfn = FMA(KP956940335, Tfe, KP290284677 * Tfd); | |
2688 Tey = FMA(KP555570233, Tew, KP831469612 * Tex); | |
2689 TeB = FNMS(KP555570233, TeA, KP831469612 * Tez); | |
2690 TeC = Tey - TeB; | |
2691 Tfh = TeB + Tey; | |
2692 } | |
2693 { | |
2694 E TeG, TeJ, TeT, TeU; | |
2695 TeG = TeE - TeF; | |
2696 TeJ = TeH - TeI; | |
2697 TeK = FMA(KP471396736, TeG, KP881921264 * TeJ); | |
2698 Tf2 = FNMS(KP471396736, TeJ, KP881921264 * TeG); | |
2699 TeT = FNMS(KP555570233, Tex, KP831469612 * Tew); | |
2700 TeU = FMA(KP831469612, TeA, KP555570233 * Tez); | |
2701 TeV = TeT - TeU; | |
2702 Tf8 = TeU + TeT; | |
2703 } | |
2704 { | |
2705 E TeW, TeX, TeN, TeQ; | |
2706 TeW = TcN - TcK; | |
2707 TeX = TdV - TdU; | |
2708 TeY = TeW - TeX; | |
2709 Tfi = TeX + TeW; | |
2710 TeN = TeL - TeM; | |
2711 TeQ = TeO - TeP; | |
2712 TeR = FNMS(KP471396736, TeQ, KP881921264 * TeN); | |
2713 Tf3 = FMA(KP881921264, TeQ, KP471396736 * TeN); | |
2714 } | |
2715 { | |
2716 E TeD, TeS, Tf1, Tf4; | |
2717 TeD = Tev + TeC; | |
2718 TeS = TeK + TeR; | |
2719 Cr[WS(csr, 54)] = TeD - TeS; | |
2720 Cr[WS(csr, 10)] = TeD + TeS; | |
2721 Tf1 = TeY + TeV; | |
2722 Tf4 = Tf2 + Tf3; | |
2723 Ci[WS(csi, 10)] = Tf1 + Tf4; | |
2724 Ci[WS(csi, 54)] = Tf4 - Tf1; | |
2725 } | |
2726 { | |
2727 E TeZ, Tf0, Tf5, Tf6; | |
2728 TeZ = TeV - TeY; | |
2729 Tf0 = TeR - TeK; | |
2730 Ci[WS(csi, 22)] = TeZ + Tf0; | |
2731 Ci[WS(csi, 42)] = Tf0 - TeZ; | |
2732 Tf5 = Tev - TeC; | |
2733 Tf6 = Tf3 - Tf2; | |
2734 Cr[WS(csr, 42)] = Tf5 - Tf6; | |
2735 Cr[WS(csr, 22)] = Tf5 + Tf6; | |
2736 } | |
2737 { | |
2738 E Tf9, Tfg, Tfl, Tfo; | |
2739 Tf9 = Tf7 + Tf8; | |
2740 Tfg = Tfc + Tff; | |
2741 Cr[WS(csr, 58)] = Tf9 - Tfg; | |
2742 Cr[WS(csr, 6)] = Tf9 + Tfg; | |
2743 Tfl = Tfi + Tfh; | |
2744 Tfo = Tfm + Tfn; | |
2745 Ci[WS(csi, 6)] = Tfl + Tfo; | |
2746 Ci[WS(csi, 58)] = Tfo - Tfl; | |
2747 } | |
2748 { | |
2749 E Tfj, Tfk, Tfp, Tfq; | |
2750 Tfj = Tfh - Tfi; | |
2751 Tfk = Tff - Tfc; | |
2752 Ci[WS(csi, 26)] = Tfj + Tfk; | |
2753 Ci[WS(csi, 38)] = Tfk - Tfj; | |
2754 Tfp = Tf7 - Tf8; | |
2755 Tfq = Tfn - Tfm; | |
2756 Cr[WS(csr, 38)] = Tfp - Tfq; | |
2757 Cr[WS(csr, 26)] = Tfp + Tfq; | |
2758 } | |
2759 } | |
2760 { | |
2761 E TcP, Te9, Tee, Teo, Teh, Tep, Td8, Tej, Tdu, Te4, TdT, Tea, Te0, Tek, TdP; | |
2762 E Te5; | |
2763 { | |
2764 E TcH, TcO, Tec, Ted; | |
2765 TcH = TcD + TcG; | |
2766 TcO = TcK + TcN; | |
2767 TcP = TcH + TcO; | |
2768 Te9 = TcH - TcO; | |
2769 Tec = Tde - Tdl; | |
2770 Ted = Tds - Tdp; | |
2771 Tee = FMA(KP773010453, Tec, KP634393284 * Ted); | |
2772 Teo = FNMS(KP634393284, Tec, KP773010453 * Ted); | |
2773 } | |
2774 { | |
2775 E Tef, Teg, TcY, Td7; | |
2776 Tef = Tdz - TdG; | |
2777 Teg = TdN - TdK; | |
2778 Teh = FNMS(KP634393284, Teg, KP773010453 * Tef); | |
2779 Tep = FMA(KP634393284, Tef, KP773010453 * Teg); | |
2780 TcY = FMA(KP980785280, TcU, KP195090322 * TcX); | |
2781 Td7 = FNMS(KP195090322, Td6, KP980785280 * Td3); | |
2782 Td8 = TcY + Td7; | |
2783 Tej = Td7 - TcY; | |
2784 } | |
2785 { | |
2786 E Tdm, Tdt, TdR, TdS; | |
2787 Tdm = Tde + Tdl; | |
2788 Tdt = Tdp + Tds; | |
2789 Tdu = FMA(KP995184726, Tdm, KP098017140 * Tdt); | |
2790 Te4 = FNMS(KP098017140, Tdm, KP995184726 * Tdt); | |
2791 TdR = FNMS(KP195090322, TcU, KP980785280 * TcX); | |
2792 TdS = FMA(KP195090322, Td3, KP980785280 * Td6); | |
2793 TdT = TdR + TdS; | |
2794 Tea = TdS - TdR; | |
2795 } | |
2796 { | |
2797 E TdW, TdZ, TdH, TdO; | |
2798 TdW = TdU + TdV; | |
2799 TdZ = TdX + TdY; | |
2800 Te0 = TdW + TdZ; | |
2801 Tek = TdZ - TdW; | |
2802 TdH = Tdz + TdG; | |
2803 TdO = TdK + TdN; | |
2804 TdP = FNMS(KP098017140, TdO, KP995184726 * TdH); | |
2805 Te5 = FMA(KP098017140, TdH, KP995184726 * TdO); | |
2806 } | |
2807 { | |
2808 E Td9, TdQ, Te3, Te6; | |
2809 Td9 = TcP + Td8; | |
2810 TdQ = Tdu + TdP; | |
2811 Cr[WS(csr, 62)] = Td9 - TdQ; | |
2812 Cr[WS(csr, 2)] = Td9 + TdQ; | |
2813 Te3 = Te0 + TdT; | |
2814 Te6 = Te4 + Te5; | |
2815 Ci[WS(csi, 2)] = Te3 + Te6; | |
2816 Ci[WS(csi, 62)] = Te6 - Te3; | |
2817 } | |
2818 { | |
2819 E Te1, Te2, Te7, Te8; | |
2820 Te1 = TdT - Te0; | |
2821 Te2 = TdP - Tdu; | |
2822 Ci[WS(csi, 30)] = Te1 + Te2; | |
2823 Ci[WS(csi, 34)] = Te2 - Te1; | |
2824 Te7 = TcP - Td8; | |
2825 Te8 = Te5 - Te4; | |
2826 Cr[WS(csr, 34)] = Te7 - Te8; | |
2827 Cr[WS(csr, 30)] = Te7 + Te8; | |
2828 } | |
2829 { | |
2830 E Teb, Tei, Ten, Teq; | |
2831 Teb = Te9 + Tea; | |
2832 Tei = Tee + Teh; | |
2833 Cr[WS(csr, 50)] = Teb - Tei; | |
2834 Cr[WS(csr, 14)] = Teb + Tei; | |
2835 Ten = Tek + Tej; | |
2836 Teq = Teo + Tep; | |
2837 Ci[WS(csi, 14)] = Ten + Teq; | |
2838 Ci[WS(csi, 50)] = Teq - Ten; | |
2839 } | |
2840 { | |
2841 E Tel, Tem, Ter, Tes; | |
2842 Tel = Tej - Tek; | |
2843 Tem = Teh - Tee; | |
2844 Ci[WS(csi, 18)] = Tel + Tem; | |
2845 Ci[WS(csi, 46)] = Tem - Tel; | |
2846 Ter = Te9 - Tea; | |
2847 Tes = Tep - Teo; | |
2848 Cr[WS(csr, 46)] = Ter - Tes; | |
2849 Cr[WS(csr, 18)] = Ter + Tes; | |
2850 } | |
2851 } | |
2852 { | |
2853 E T6v, T77, T6C, T7h, T6Y, T7i, T6V, T78, T6R, T7n, T73, T7f, T6K, T7m, T72; | |
2854 E T7c; | |
2855 { | |
2856 E T6t, T6u, T6T, T6U; | |
2857 T6t = T27 - T2e; | |
2858 T6u = T5Y - T5X; | |
2859 T6v = T6t - T6u; | |
2860 T77 = T6t + T6u; | |
2861 { | |
2862 E T6y, T6B, T6W, T6X; | |
2863 T6y = FMA(KP773010453, T6w, KP634393284 * T6x); | |
2864 T6B = FNMS(KP634393284, T6A, KP773010453 * T6z); | |
2865 T6C = T6y - T6B; | |
2866 T7h = T6B + T6y; | |
2867 T6W = T2x - T2o; | |
2868 T6X = T5V - T5S; | |
2869 T6Y = T6W - T6X; | |
2870 T7i = T6X + T6W; | |
2871 } | |
2872 T6T = FNMS(KP634393284, T6w, KP773010453 * T6x); | |
2873 T6U = FMA(KP634393284, T6z, KP773010453 * T6A); | |
2874 T6V = T6T - T6U; | |
2875 T78 = T6U + T6T; | |
2876 { | |
2877 E T6N, T7d, T6Q, T7e, T6M, T6O; | |
2878 T6M = T5I - T5H; | |
2879 T6N = T6L - T6M; | |
2880 T7d = T6L + T6M; | |
2881 T6O = T5v - T5e; | |
2882 T6Q = T6O - T6P; | |
2883 T7e = T6P + T6O; | |
2884 T6R = FNMS(KP427555093, T6Q, KP903989293 * T6N); | |
2885 T7n = FMA(KP941544065, T7e, KP336889853 * T7d); | |
2886 T73 = FMA(KP903989293, T6Q, KP427555093 * T6N); | |
2887 T7f = FNMS(KP336889853, T7e, KP941544065 * T7d); | |
2888 } | |
2889 { | |
2890 E T6G, T7a, T6J, T7b, T6E, T6I; | |
2891 T6E = T4g - T3Z; | |
2892 T6G = T6E - T6F; | |
2893 T7a = T6F + T6E; | |
2894 T6I = T4t - T4s; | |
2895 T6J = T6H - T6I; | |
2896 T7b = T6H + T6I; | |
2897 T6K = FMA(KP427555093, T6G, KP903989293 * T6J); | |
2898 T7m = FNMS(KP336889853, T7b, KP941544065 * T7a); | |
2899 T72 = FNMS(KP427555093, T6J, KP903989293 * T6G); | |
2900 T7c = FMA(KP336889853, T7a, KP941544065 * T7b); | |
2901 } | |
2902 } | |
2903 { | |
2904 E T6D, T6S, T71, T74; | |
2905 T6D = T6v + T6C; | |
2906 T6S = T6K + T6R; | |
2907 Cr[WS(csr, 55)] = T6D - T6S; | |
2908 Cr[WS(csr, 9)] = T6D + T6S; | |
2909 T71 = T6Y + T6V; | |
2910 T74 = T72 + T73; | |
2911 Ci[WS(csi, 9)] = T71 + T74; | |
2912 Ci[WS(csi, 55)] = T74 - T71; | |
2913 } | |
2914 { | |
2915 E T6Z, T70, T75, T76; | |
2916 T6Z = T6V - T6Y; | |
2917 T70 = T6R - T6K; | |
2918 Ci[WS(csi, 23)] = T6Z + T70; | |
2919 Ci[WS(csi, 41)] = T70 - T6Z; | |
2920 T75 = T6v - T6C; | |
2921 T76 = T73 - T72; | |
2922 Cr[WS(csr, 41)] = T75 - T76; | |
2923 Cr[WS(csr, 23)] = T75 + T76; | |
2924 } | |
2925 { | |
2926 E T79, T7g, T7l, T7o; | |
2927 T79 = T77 + T78; | |
2928 T7g = T7c + T7f; | |
2929 Cr[WS(csr, 57)] = T79 - T7g; | |
2930 Cr[WS(csr, 7)] = T79 + T7g; | |
2931 T7l = T7i + T7h; | |
2932 T7o = T7m + T7n; | |
2933 Ci[WS(csi, 7)] = T7l + T7o; | |
2934 Ci[WS(csi, 57)] = T7o - T7l; | |
2935 } | |
2936 { | |
2937 E T7j, T7k, T7p, T7q; | |
2938 T7j = T7h - T7i; | |
2939 T7k = T7f - T7c; | |
2940 Ci[WS(csi, 25)] = T7j + T7k; | |
2941 Ci[WS(csi, 39)] = T7k - T7j; | |
2942 T7p = T77 - T78; | |
2943 T7q = T7n - T7m; | |
2944 Cr[WS(csr, 39)] = T7p - T7q; | |
2945 Cr[WS(csr, 25)] = T7p + T7q; | |
2946 } | |
2947 } | |
2948 { | |
2949 E T99, T9L, T9g, T9V, T9C, T9W, T9z, T9M, T9v, Ta1, T9H, T9T, T9o, Ta0, T9G; | |
2950 E T9Q; | |
2951 { | |
2952 E T97, T98, T9x, T9y; | |
2953 T97 = T7r - T7s; | |
2954 T98 = T8C - T8B; | |
2955 T99 = T97 - T98; | |
2956 T9L = T97 + T98; | |
2957 { | |
2958 E T9c, T9f, T9A, T9B; | |
2959 T9c = FMA(KP471396736, T9a, KP881921264 * T9b); | |
2960 T9f = FNMS(KP471396736, T9e, KP881921264 * T9d); | |
2961 T9g = T9c - T9f; | |
2962 T9V = T9f + T9c; | |
2963 T9A = T7z - T7w; | |
2964 T9B = T8z - T8y; | |
2965 T9C = T9A - T9B; | |
2966 T9W = T9B + T9A; | |
2967 } | |
2968 T9x = FNMS(KP471396736, T9b, KP881921264 * T9a); | |
2969 T9y = FMA(KP881921264, T9e, KP471396736 * T9d); | |
2970 T9z = T9x - T9y; | |
2971 T9M = T9y + T9x; | |
2972 { | |
2973 E T9r, T9R, T9u, T9S, T9q, T9s; | |
2974 T9q = T8q - T8p; | |
2975 T9r = T9p - T9q; | |
2976 T9R = T9p + T9q; | |
2977 T9s = T8j - T8g; | |
2978 T9u = T9s - T9t; | |
2979 T9S = T9t + T9s; | |
2980 T9v = FNMS(KP514102744, T9u, KP857728610 * T9r); | |
2981 Ta1 = FMA(KP970031253, T9S, KP242980179 * T9R); | |
2982 T9H = FMA(KP857728610, T9u, KP514102744 * T9r); | |
2983 T9T = FNMS(KP242980179, T9S, KP970031253 * T9R); | |
2984 } | |
2985 { | |
2986 E T9k, T9O, T9n, T9P, T9i, T9m; | |
2987 T9i = T80 - T7X; | |
2988 T9k = T9i - T9j; | |
2989 T9O = T9j + T9i; | |
2990 T9m = T87 - T86; | |
2991 T9n = T9l - T9m; | |
2992 T9P = T9l + T9m; | |
2993 T9o = FMA(KP514102744, T9k, KP857728610 * T9n); | |
2994 Ta0 = FNMS(KP242980179, T9P, KP970031253 * T9O); | |
2995 T9G = FNMS(KP514102744, T9n, KP857728610 * T9k); | |
2996 T9Q = FMA(KP242980179, T9O, KP970031253 * T9P); | |
2997 } | |
2998 } | |
2999 { | |
3000 E T9h, T9w, T9F, T9I; | |
3001 T9h = T99 + T9g; | |
3002 T9w = T9o + T9v; | |
3003 Cr[WS(csr, 53)] = T9h - T9w; | |
3004 Cr[WS(csr, 11)] = T9h + T9w; | |
3005 T9F = T9C + T9z; | |
3006 T9I = T9G + T9H; | |
3007 Ci[WS(csi, 11)] = T9F + T9I; | |
3008 Ci[WS(csi, 53)] = T9I - T9F; | |
3009 } | |
3010 { | |
3011 E T9D, T9E, T9J, T9K; | |
3012 T9D = T9z - T9C; | |
3013 T9E = T9v - T9o; | |
3014 Ci[WS(csi, 21)] = T9D + T9E; | |
3015 Ci[WS(csi, 43)] = T9E - T9D; | |
3016 T9J = T99 - T9g; | |
3017 T9K = T9H - T9G; | |
3018 Cr[WS(csr, 43)] = T9J - T9K; | |
3019 Cr[WS(csr, 21)] = T9J + T9K; | |
3020 } | |
3021 { | |
3022 E T9N, T9U, T9Z, Ta2; | |
3023 T9N = T9L + T9M; | |
3024 T9U = T9Q + T9T; | |
3025 Cr[WS(csr, 59)] = T9N - T9U; | |
3026 Cr[WS(csr, 5)] = T9N + T9U; | |
3027 T9Z = T9W + T9V; | |
3028 Ta2 = Ta0 + Ta1; | |
3029 Ci[WS(csi, 5)] = T9Z + Ta2; | |
3030 Ci[WS(csi, 59)] = Ta2 - T9Z; | |
3031 } | |
3032 { | |
3033 E T9X, T9Y, Ta3, Ta4; | |
3034 T9X = T9V - T9W; | |
3035 T9Y = T9T - T9Q; | |
3036 Ci[WS(csi, 27)] = T9X + T9Y; | |
3037 Ci[WS(csi, 37)] = T9Y - T9X; | |
3038 Ta3 = T9L - T9M; | |
3039 Ta4 = Ta1 - Ta0; | |
3040 Cr[WS(csr, 37)] = Ta3 - Ta4; | |
3041 Cr[WS(csr, 27)] = Ta3 + Ta4; | |
3042 } | |
3043 } | |
3044 { | |
3045 E T2z, T69, T3g, T6j, T60, T6k, T5P, T6a, T5L, T6p, T65, T6h, T4w, T6o, T64; | |
3046 E T6e; | |
3047 { | |
3048 E T2f, T2y, T5N, T5O; | |
3049 T2f = T27 + T2e; | |
3050 T2y = T2o + T2x; | |
3051 T2z = T2f + T2y; | |
3052 T69 = T2f - T2y; | |
3053 { | |
3054 E T2U, T3f, T5W, T5Z; | |
3055 T2U = FMA(KP098017140, T2M, KP995184726 * T2T); | |
3056 T3f = FNMS(KP098017140, T3e, KP995184726 * T37); | |
3057 T3g = T2U + T3f; | |
3058 T6j = T3f - T2U; | |
3059 T5W = T5S + T5V; | |
3060 T5Z = T5X + T5Y; | |
3061 T60 = T5W + T5Z; | |
3062 T6k = T5Z - T5W; | |
3063 } | |
3064 T5N = FNMS(KP098017140, T2T, KP995184726 * T2M); | |
3065 T5O = FMA(KP995184726, T3e, KP098017140 * T37); | |
3066 T5P = T5N + T5O; | |
3067 T6a = T5O - T5N; | |
3068 { | |
3069 E T5x, T6f, T5K, T6g, T5w, T5J; | |
3070 T5w = T5e + T5v; | |
3071 T5x = T4X + T5w; | |
3072 T6f = T4X - T5w; | |
3073 T5J = T5H + T5I; | |
3074 T5K = T5G + T5J; | |
3075 T6g = T5J - T5G; | |
3076 T5L = FNMS(KP049067674, T5K, KP998795456 * T5x); | |
3077 T6p = FMA(KP671558954, T6f, KP740951125 * T6g); | |
3078 T65 = FMA(KP049067674, T5x, KP998795456 * T5K); | |
3079 T6h = FNMS(KP671558954, T6g, KP740951125 * T6f); | |
3080 } | |
3081 { | |
3082 E T4i, T6c, T4v, T6d, T4h, T4u; | |
3083 T4h = T3Z + T4g; | |
3084 T4i = T3I + T4h; | |
3085 T6c = T3I - T4h; | |
3086 T4u = T4s + T4t; | |
3087 T4v = T4r + T4u; | |
3088 T6d = T4u - T4r; | |
3089 T4w = FMA(KP998795456, T4i, KP049067674 * T4v); | |
3090 T6o = FNMS(KP671558954, T6c, KP740951125 * T6d); | |
3091 T64 = FNMS(KP049067674, T4i, KP998795456 * T4v); | |
3092 T6e = FMA(KP740951125, T6c, KP671558954 * T6d); | |
3093 } | |
3094 } | |
3095 { | |
3096 E T3h, T5M, T63, T66; | |
3097 T3h = T2z + T3g; | |
3098 T5M = T4w + T5L; | |
3099 Cr[WS(csr, 63)] = T3h - T5M; | |
3100 Cr[WS(csr, 1)] = T3h + T5M; | |
3101 T63 = T60 + T5P; | |
3102 T66 = T64 + T65; | |
3103 Ci[WS(csi, 1)] = T63 + T66; | |
3104 Ci[WS(csi, 63)] = T66 - T63; | |
3105 } | |
3106 { | |
3107 E T61, T62, T67, T68; | |
3108 T61 = T5P - T60; | |
3109 T62 = T5L - T4w; | |
3110 Ci[WS(csi, 31)] = T61 + T62; | |
3111 Ci[WS(csi, 33)] = T62 - T61; | |
3112 T67 = T2z - T3g; | |
3113 T68 = T65 - T64; | |
3114 Cr[WS(csr, 33)] = T67 - T68; | |
3115 Cr[WS(csr, 31)] = T67 + T68; | |
3116 } | |
3117 { | |
3118 E T6b, T6i, T6n, T6q; | |
3119 T6b = T69 + T6a; | |
3120 T6i = T6e + T6h; | |
3121 Cr[WS(csr, 49)] = T6b - T6i; | |
3122 Cr[WS(csr, 15)] = T6b + T6i; | |
3123 T6n = T6k + T6j; | |
3124 T6q = T6o + T6p; | |
3125 Ci[WS(csi, 15)] = T6n + T6q; | |
3126 Ci[WS(csi, 49)] = T6q - T6n; | |
3127 } | |
3128 { | |
3129 E T6l, T6m, T6r, T6s; | |
3130 T6l = T6j - T6k; | |
3131 T6m = T6h - T6e; | |
3132 Ci[WS(csi, 17)] = T6l + T6m; | |
3133 Ci[WS(csi, 47)] = T6m - T6l; | |
3134 T6r = T69 - T6a; | |
3135 T6s = T6p - T6o; | |
3136 Cr[WS(csr, 47)] = T6r - T6s; | |
3137 Cr[WS(csr, 17)] = T6r + T6s; | |
3138 } | |
3139 } | |
3140 { | |
3141 E T7B, T8N, T7Q, T8X, T8E, T8Y, T8x, T8O, T8t, T93, T8J, T8V, T8a, T92, T8I; | |
3142 E T8S; | |
3143 { | |
3144 E T7t, T7A, T8v, T8w; | |
3145 T7t = T7r + T7s; | |
3146 T7A = T7w + T7z; | |
3147 T7B = T7t + T7A; | |
3148 T8N = T7t - T7A; | |
3149 { | |
3150 E T7I, T7P, T8A, T8D; | |
3151 T7I = FMA(KP956940335, T7E, KP290284677 * T7H); | |
3152 T7P = FNMS(KP290284677, T7O, KP956940335 * T7L); | |
3153 T7Q = T7I + T7P; | |
3154 T8X = T7P - T7I; | |
3155 T8A = T8y + T8z; | |
3156 T8D = T8B + T8C; | |
3157 T8E = T8A + T8D; | |
3158 T8Y = T8D - T8A; | |
3159 } | |
3160 T8v = FNMS(KP290284677, T7E, KP956940335 * T7H); | |
3161 T8w = FMA(KP290284677, T7L, KP956940335 * T7O); | |
3162 T8x = T8v + T8w; | |
3163 T8O = T8w - T8v; | |
3164 { | |
3165 E T8l, T8T, T8s, T8U, T8k, T8r; | |
3166 T8k = T8g + T8j; | |
3167 T8l = T8d + T8k; | |
3168 T8T = T8d - T8k; | |
3169 T8r = T8p + T8q; | |
3170 T8s = T8o + T8r; | |
3171 T8U = T8r - T8o; | |
3172 T8t = FNMS(KP146730474, T8s, KP989176509 * T8l); | |
3173 T93 = FMA(KP595699304, T8T, KP803207531 * T8U); | |
3174 T8J = FMA(KP146730474, T8l, KP989176509 * T8s); | |
3175 T8V = FNMS(KP595699304, T8U, KP803207531 * T8T); | |
3176 } | |
3177 { | |
3178 E T82, T8Q, T89, T8R, T81, T88; | |
3179 T81 = T7X + T80; | |
3180 T82 = T7U + T81; | |
3181 T8Q = T7U - T81; | |
3182 T88 = T86 + T87; | |
3183 T89 = T85 + T88; | |
3184 T8R = T88 - T85; | |
3185 T8a = FMA(KP989176509, T82, KP146730474 * T89); | |
3186 T92 = FNMS(KP595699304, T8Q, KP803207531 * T8R); | |
3187 T8I = FNMS(KP146730474, T82, KP989176509 * T89); | |
3188 T8S = FMA(KP803207531, T8Q, KP595699304 * T8R); | |
3189 } | |
3190 } | |
3191 { | |
3192 E T7R, T8u, T8H, T8K; | |
3193 T7R = T7B + T7Q; | |
3194 T8u = T8a + T8t; | |
3195 Cr[WS(csr, 61)] = T7R - T8u; | |
3196 Cr[WS(csr, 3)] = T7R + T8u; | |
3197 T8H = T8E + T8x; | |
3198 T8K = T8I + T8J; | |
3199 Ci[WS(csi, 3)] = T8H + T8K; | |
3200 Ci[WS(csi, 61)] = T8K - T8H; | |
3201 } | |
3202 { | |
3203 E T8F, T8G, T8L, T8M; | |
3204 T8F = T8x - T8E; | |
3205 T8G = T8t - T8a; | |
3206 Ci[WS(csi, 29)] = T8F + T8G; | |
3207 Ci[WS(csi, 35)] = T8G - T8F; | |
3208 T8L = T7B - T7Q; | |
3209 T8M = T8J - T8I; | |
3210 Cr[WS(csr, 35)] = T8L - T8M; | |
3211 Cr[WS(csr, 29)] = T8L + T8M; | |
3212 } | |
3213 { | |
3214 E T8P, T8W, T91, T94; | |
3215 T8P = T8N + T8O; | |
3216 T8W = T8S + T8V; | |
3217 Cr[WS(csr, 51)] = T8P - T8W; | |
3218 Cr[WS(csr, 13)] = T8P + T8W; | |
3219 T91 = T8Y + T8X; | |
3220 T94 = T92 + T93; | |
3221 Ci[WS(csi, 13)] = T91 + T94; | |
3222 Ci[WS(csi, 51)] = T94 - T91; | |
3223 } | |
3224 { | |
3225 E T8Z, T90, T95, T96; | |
3226 T8Z = T8X - T8Y; | |
3227 T90 = T8V - T8S; | |
3228 Ci[WS(csi, 19)] = T8Z + T90; | |
3229 Ci[WS(csi, 45)] = T90 - T8Z; | |
3230 T95 = T8N - T8O; | |
3231 T96 = T93 - T92; | |
3232 Cr[WS(csr, 45)] = T95 - T96; | |
3233 Cr[WS(csr, 19)] = T95 + T96; | |
3234 } | |
3235 } | |
3236 } | |
3237 } | |
3238 } | |
3239 | |
3240 static const kr2c_desc desc = { 128, "r2cf_128", {812, 186, 144, 0}, &GENUS }; | |
3241 | |
3242 void X(codelet_r2cf_128) (planner *p) { | |
3243 X(kr2c_register) (p, r2cf_128, &desc); | |
3244 } | |
3245 | |
3246 #endif |