Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/rdft/scalar/r2cf/r2cfII_64.c @ 10:37bf6b4a2645
Add FFTW3
author | Chris Cannam |
---|---|
date | Wed, 20 Mar 2013 15:35:50 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
9:c0fb53affa76 | 10:37bf6b4a2645 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-11 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Sun Nov 25 07:40:17 EST 2012 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cfII_64 -dft-II -include r2cfII.h */ | |
29 | |
30 /* | |
31 * This function contains 434 FP additions, 320 FP multiplications, | |
32 * (or, 114 additions, 0 multiplications, 320 fused multiply/add), | |
33 * 158 stack variables, 31 constants, and 128 memory accesses | |
34 */ | |
35 #include "r2cfII.h" | |
36 | |
37 static void r2cfII_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP941544065, +0.941544065183020778412509402599502357185589796); | |
40 DK(KP903989293, +0.903989293123443331586200297230537048710132025); | |
41 DK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
42 DK(KP472964775, +0.472964775891319928124438237972992463904131113); | |
43 DK(KP357805721, +0.357805721314524104672487743774474392487532769); | |
44 DK(KP820678790, +0.820678790828660330972281985331011598767386482); | |
45 DK(KP989176509, +0.989176509964780973451673738016243063983689533); | |
46 DK(KP803207531, +0.803207531480644909806676512963141923879569427); | |
47 DK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
48 DK(KP741650546, +0.741650546272035369581266691172079863842265220); | |
49 DK(KP148335987, +0.148335987538347428753676511486911367000625355); | |
50 DK(KP303346683, +0.303346683607342391675883946941299872384187453); | |
51 DK(KP998795456, +0.998795456205172392714771604759100694443203615); | |
52 DK(KP740951125, +0.740951125354959091175616897495162729728955309); | |
53 DK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
54 DK(KP906347169, +0.906347169019147157946142717268914412664134293); | |
55 DK(KP049126849, +0.049126849769467254105343321271313617079695752); | |
56 DK(KP098491403, +0.098491403357164253077197521291327432293052451); | |
57 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
58 DK(KP970031253, +0.970031253194543992603984207286100251456865962); | |
59 DK(KP857728610, +0.857728610000272069902269984284770137042490799); | |
60 DK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
61 DK(KP599376933, +0.599376933681923766271389869014404232837890546); | |
62 DK(KP250486960, +0.250486960191305461595702160124721208578685568); | |
63 DK(KP534511135, +0.534511135950791641089685961295362908582039528); | |
64 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
65 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
66 DK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
67 DK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
68 DK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
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(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) { | |
73 E T5b, T6q, T6p, T5e; | |
74 { | |
75 E T5h, T3Z, T35, Tm, T5g, T3W, T34, Tv, T5f, T3T, T6N, T6z, T6j, T65, T33; | |
76 E Td, T5z, T4D, T3q, T2C, T5C, T4O, T3n, T2b, T5k, T4b, T3c, TR, T5l, T4e; | |
77 E T3b, TK, T5n, T44, T39, T1c, T5o, T47, T38, T15, T5s, T4k, T3j, T1T, T5v; | |
78 E T4v, T3g, T1s, T1t, T1y, T5D, T4K, T5A, T4R, T3o, T2F, T3r, T2u, T1C, T1H; | |
79 E T1D, T1z, T1w, T1E; | |
80 { | |
81 E T2A, T26, T4B, T23, T4M, T2y, T2z, T29; | |
82 { | |
83 E Te, Tj, Tn, Ts, To, Tk, Th, Tp, Tf, Tg; | |
84 Te = R0[WS(rs, 14)]; | |
85 Tj = R0[WS(rs, 30)]; | |
86 Tf = R0[WS(rs, 6)]; | |
87 Tg = R0[WS(rs, 22)]; | |
88 Tn = R0[WS(rs, 18)]; | |
89 Ts = R0[WS(rs, 2)]; | |
90 To = R0[WS(rs, 10)]; | |
91 Tk = Tg - Tf; | |
92 Th = Tf + Tg; | |
93 Tp = R0[WS(rs, 26)]; | |
94 { | |
95 E T3Q, T8, T3P, T5, T6x, T63, T3R, Tb; | |
96 { | |
97 E T1, T61, T9, T62, T4, Ta; | |
98 { | |
99 E T3V, Tu, T3U, Tr, T3Y, Tl; | |
100 T1 = R0[0]; | |
101 T3Y = FMA(KP707106781, Tk, Tj); | |
102 Tl = FNMS(KP707106781, Tk, Tj); | |
103 { | |
104 E T3X, Ti, Tt, Tq; | |
105 T3X = FMA(KP707106781, Th, Te); | |
106 Ti = FNMS(KP707106781, Th, Te); | |
107 Tt = To - Tp; | |
108 Tq = To + Tp; | |
109 T5h = FNMS(KP198912367, T3X, T3Y); | |
110 T3Z = FMA(KP198912367, T3Y, T3X); | |
111 T35 = FMA(KP668178637, Ti, Tl); | |
112 Tm = FNMS(KP668178637, Tl, Ti); | |
113 T3V = FMA(KP707106781, Tt, Ts); | |
114 Tu = FNMS(KP707106781, Tt, Ts); | |
115 T3U = FMA(KP707106781, Tq, Tn); | |
116 Tr = FNMS(KP707106781, Tq, Tn); | |
117 T61 = R0[WS(rs, 16)]; | |
118 } | |
119 { | |
120 E T2, T3, T6, T7; | |
121 T2 = R0[WS(rs, 8)]; | |
122 T5g = FNMS(KP198912367, T3U, T3V); | |
123 T3W = FMA(KP198912367, T3V, T3U); | |
124 T34 = FMA(KP668178637, Tr, Tu); | |
125 Tv = FNMS(KP668178637, Tu, Tr); | |
126 T3 = R0[WS(rs, 24)]; | |
127 T6 = R0[WS(rs, 20)]; | |
128 T7 = R0[WS(rs, 4)]; | |
129 T9 = R0[WS(rs, 12)]; | |
130 T62 = T2 + T3; | |
131 T4 = T2 - T3; | |
132 T3Q = FNMS(KP414213562, T6, T7); | |
133 T8 = FMA(KP414213562, T7, T6); | |
134 Ta = R0[WS(rs, 28)]; | |
135 } | |
136 } | |
137 T3P = FMA(KP707106781, T4, T1); | |
138 T5 = FNMS(KP707106781, T4, T1); | |
139 T6x = FNMS(KP707106781, T62, T61); | |
140 T63 = FMA(KP707106781, T62, T61); | |
141 T3R = FMS(KP414213562, T9, Ta); | |
142 Tb = FMA(KP414213562, Ta, T9); | |
143 } | |
144 { | |
145 E T1Z, T2w, T27, T2x, T22, T28; | |
146 T1Z = R1[WS(rs, 31)]; | |
147 { | |
148 E T3S, T6y, T64, Tc; | |
149 T3S = T3Q + T3R; | |
150 T6y = T3R - T3Q; | |
151 T64 = T8 + Tb; | |
152 Tc = T8 - Tb; | |
153 T5f = FMA(KP923879532, T3S, T3P); | |
154 T3T = FNMS(KP923879532, T3S, T3P); | |
155 T6N = FNMS(KP923879532, T6y, T6x); | |
156 T6z = FMA(KP923879532, T6y, T6x); | |
157 T6j = FNMS(KP923879532, T64, T63); | |
158 T65 = FMA(KP923879532, T64, T63); | |
159 T33 = FMA(KP923879532, Tc, T5); | |
160 Td = FNMS(KP923879532, Tc, T5); | |
161 T2w = R1[WS(rs, 15)]; | |
162 } | |
163 { | |
164 E T20, T21, T24, T25; | |
165 T20 = R1[WS(rs, 7)]; | |
166 T21 = R1[WS(rs, 23)]; | |
167 T24 = R1[WS(rs, 19)]; | |
168 T25 = R1[WS(rs, 3)]; | |
169 T27 = R1[WS(rs, 11)]; | |
170 T2x = T20 + T21; | |
171 T22 = T20 - T21; | |
172 T2A = FNMS(KP414213562, T24, T25); | |
173 T26 = FMA(KP414213562, T25, T24); | |
174 T28 = R1[WS(rs, 27)]; | |
175 } | |
176 T4B = FMS(KP707106781, T22, T1Z); | |
177 T23 = FMA(KP707106781, T22, T1Z); | |
178 T4M = FMA(KP707106781, T2x, T2w); | |
179 T2y = FNMS(KP707106781, T2x, T2w); | |
180 T2z = FMS(KP414213562, T27, T28); | |
181 T29 = FMA(KP414213562, T28, T27); | |
182 } | |
183 } | |
184 } | |
185 { | |
186 E T1a, T10, T42, TX, T45, T18, T19, T13; | |
187 { | |
188 E TP, TF, T49, TC, T4c, TN, TO, TI; | |
189 { | |
190 E Ty, TL, TG, TM, TB, TH; | |
191 Ty = R0[WS(rs, 17)]; | |
192 { | |
193 E T4C, T2B, T4N, T2a; | |
194 T4C = T2A + T2z; | |
195 T2B = T2z - T2A; | |
196 T4N = T26 + T29; | |
197 T2a = T26 - T29; | |
198 T5z = FMA(KP923879532, T4C, T4B); | |
199 T4D = FNMS(KP923879532, T4C, T4B); | |
200 T3q = FMA(KP923879532, T2B, T2y); | |
201 T2C = FNMS(KP923879532, T2B, T2y); | |
202 T5C = FMA(KP923879532, T4N, T4M); | |
203 T4O = FNMS(KP923879532, T4N, T4M); | |
204 T3n = FNMS(KP923879532, T2a, T23); | |
205 T2b = FMA(KP923879532, T2a, T23); | |
206 TL = R0[WS(rs, 1)]; | |
207 } | |
208 { | |
209 E Tz, TA, TD, TE; | |
210 Tz = R0[WS(rs, 9)]; | |
211 TA = R0[WS(rs, 25)]; | |
212 TD = R0[WS(rs, 29)]; | |
213 TE = R0[WS(rs, 13)]; | |
214 TG = R0[WS(rs, 5)]; | |
215 TM = Tz - TA; | |
216 TB = Tz + TA; | |
217 TP = FMA(KP414213562, TD, TE); | |
218 TF = FMS(KP414213562, TE, TD); | |
219 TH = R0[WS(rs, 21)]; | |
220 } | |
221 T49 = FMA(KP707106781, TB, Ty); | |
222 TC = FNMS(KP707106781, TB, Ty); | |
223 T4c = FMA(KP707106781, TM, TL); | |
224 TN = FNMS(KP707106781, TM, TL); | |
225 TO = FMA(KP414213562, TG, TH); | |
226 TI = FNMS(KP414213562, TH, TG); | |
227 } | |
228 { | |
229 E TT, T16, T11, T17, TW, T12; | |
230 TT = R0[WS(rs, 15)]; | |
231 { | |
232 E T4a, TQ, T4d, TJ; | |
233 T4a = TO + TP; | |
234 TQ = TO - TP; | |
235 T4d = TI + TF; | |
236 TJ = TF - TI; | |
237 T5k = FMA(KP923879532, T4a, T49); | |
238 T4b = FNMS(KP923879532, T4a, T49); | |
239 T3c = FMA(KP923879532, TQ, TN); | |
240 TR = FNMS(KP923879532, TQ, TN); | |
241 T5l = FMA(KP923879532, T4d, T4c); | |
242 T4e = FNMS(KP923879532, T4d, T4c); | |
243 T3b = FMA(KP923879532, TJ, TC); | |
244 TK = FNMS(KP923879532, TJ, TC); | |
245 T16 = R0[WS(rs, 31)]; | |
246 } | |
247 { | |
248 E TU, TV, TY, TZ; | |
249 TU = R0[WS(rs, 7)]; | |
250 TV = R0[WS(rs, 23)]; | |
251 TY = R0[WS(rs, 3)]; | |
252 TZ = R0[WS(rs, 19)]; | |
253 T11 = R0[WS(rs, 27)]; | |
254 T17 = TV - TU; | |
255 TW = TU + TV; | |
256 T1a = FMA(KP414213562, TY, TZ); | |
257 T10 = FMS(KP414213562, TZ, TY); | |
258 T12 = R0[WS(rs, 11)]; | |
259 } | |
260 T42 = FMA(KP707106781, TW, TT); | |
261 TX = FNMS(KP707106781, TW, TT); | |
262 T45 = FMA(KP707106781, T17, T16); | |
263 T18 = FNMS(KP707106781, T17, T16); | |
264 T19 = FMA(KP414213562, T11, T12); | |
265 T13 = FNMS(KP414213562, T12, T11); | |
266 } | |
267 } | |
268 { | |
269 E T1R, T1n, T4i, T1k, T4t, T1P, T1Q, T1q; | |
270 { | |
271 E T1g, T1N, T1o, T1O, T1j, T1p; | |
272 T1g = R1[0]; | |
273 { | |
274 E T43, T1b, T46, T14; | |
275 T43 = T1a + T19; | |
276 T1b = T19 - T1a; | |
277 T46 = T10 + T13; | |
278 T14 = T10 - T13; | |
279 T5n = FMA(KP923879532, T43, T42); | |
280 T44 = FNMS(KP923879532, T43, T42); | |
281 T39 = FMA(KP923879532, T1b, T18); | |
282 T1c = FNMS(KP923879532, T1b, T18); | |
283 T5o = FMA(KP923879532, T46, T45); | |
284 T47 = FNMS(KP923879532, T46, T45); | |
285 T38 = FMA(KP923879532, T14, TX); | |
286 T15 = FNMS(KP923879532, T14, TX); | |
287 T1N = R1[WS(rs, 16)]; | |
288 } | |
289 { | |
290 E T1h, T1i, T1l, T1m; | |
291 T1h = R1[WS(rs, 8)]; | |
292 T1i = R1[WS(rs, 24)]; | |
293 T1l = R1[WS(rs, 20)]; | |
294 T1m = R1[WS(rs, 4)]; | |
295 T1o = R1[WS(rs, 12)]; | |
296 T1O = T1h + T1i; | |
297 T1j = T1h - T1i; | |
298 T1R = FNMS(KP414213562, T1l, T1m); | |
299 T1n = FMA(KP414213562, T1m, T1l); | |
300 T1p = R1[WS(rs, 28)]; | |
301 } | |
302 T4i = FMA(KP707106781, T1j, T1g); | |
303 T1k = FNMS(KP707106781, T1j, T1g); | |
304 T4t = FMA(KP707106781, T1O, T1N); | |
305 T1P = FNMS(KP707106781, T1O, T1N); | |
306 T1Q = FMS(KP414213562, T1o, T1p); | |
307 T1q = FMA(KP414213562, T1p, T1o); | |
308 } | |
309 { | |
310 E T2c, T2h, T2l, T2q, T2m, T2i, T2f, T2n, T2d, T2e; | |
311 T2c = R1[WS(rs, 13)]; | |
312 { | |
313 E T4j, T1S, T4u, T1r; | |
314 T4j = T1R + T1Q; | |
315 T1S = T1Q - T1R; | |
316 T4u = T1n + T1q; | |
317 T1r = T1n - T1q; | |
318 T5s = FMA(KP923879532, T4j, T4i); | |
319 T4k = FNMS(KP923879532, T4j, T4i); | |
320 T3j = FMA(KP923879532, T1S, T1P); | |
321 T1T = FNMS(KP923879532, T1S, T1P); | |
322 T5v = FMA(KP923879532, T4u, T4t); | |
323 T4v = FNMS(KP923879532, T4u, T4t); | |
324 T3g = FMA(KP923879532, T1r, T1k); | |
325 T1s = FNMS(KP923879532, T1r, T1k); | |
326 T2h = R1[WS(rs, 29)]; | |
327 T2d = R1[WS(rs, 5)]; | |
328 T2e = R1[WS(rs, 21)]; | |
329 } | |
330 T2l = R1[WS(rs, 17)]; | |
331 T2q = R1[WS(rs, 1)]; | |
332 T2m = R1[WS(rs, 9)]; | |
333 T2i = T2d - T2e; | |
334 T2f = T2d + T2e; | |
335 T2n = R1[WS(rs, 25)]; | |
336 { | |
337 E T1u, T1v, T2j, T4I; | |
338 T1t = R1[WS(rs, 14)]; | |
339 T2j = FMA(KP707106781, T2i, T2h); | |
340 T4I = FMS(KP707106781, T2i, T2h); | |
341 { | |
342 E T4H, T2g, T2r, T2o; | |
343 T4H = FMA(KP707106781, T2f, T2c); | |
344 T2g = FNMS(KP707106781, T2f, T2c); | |
345 T2r = T2m - T2n; | |
346 T2o = T2m + T2n; | |
347 { | |
348 E T4J, T4P, T2E, T2k; | |
349 T4J = FNMS(KP198912367, T4I, T4H); | |
350 T4P = FMA(KP198912367, T4H, T4I); | |
351 T2E = FMA(KP668178637, T2g, T2j); | |
352 T2k = FNMS(KP668178637, T2j, T2g); | |
353 { | |
354 E T2s, T4F, T4E, T2p; | |
355 T2s = FNMS(KP707106781, T2r, T2q); | |
356 T4F = FMA(KP707106781, T2r, T2q); | |
357 T4E = FMA(KP707106781, T2o, T2l); | |
358 T2p = FNMS(KP707106781, T2o, T2l); | |
359 T1y = R1[WS(rs, 30)]; | |
360 T1u = R1[WS(rs, 6)]; | |
361 { | |
362 E T4G, T4Q, T2D, T2t; | |
363 T4G = FMA(KP198912367, T4F, T4E); | |
364 T4Q = FNMS(KP198912367, T4E, T4F); | |
365 T2D = FMA(KP668178637, T2p, T2s); | |
366 T2t = FNMS(KP668178637, T2s, T2p); | |
367 T5D = T4G + T4J; | |
368 T4K = T4G - T4J; | |
369 T5A = T4Q + T4P; | |
370 T4R = T4P - T4Q; | |
371 T3o = T2D - T2E; | |
372 T2F = T2D + T2E; | |
373 T3r = T2t + T2k; | |
374 T2u = T2k - T2t; | |
375 T1v = R1[WS(rs, 22)]; | |
376 } | |
377 } | |
378 } | |
379 } | |
380 T1C = R1[WS(rs, 18)]; | |
381 T1H = R1[WS(rs, 2)]; | |
382 T1D = R1[WS(rs, 10)]; | |
383 T1z = T1u - T1v; | |
384 T1w = T1u + T1v; | |
385 T1E = R1[WS(rs, 26)]; | |
386 } | |
387 } | |
388 } | |
389 } | |
390 } | |
391 { | |
392 E T6A, T4r, T4y, T3h, T3k, T36, T6k, T40, T5X, T6c, T6b, T60; | |
393 { | |
394 E T5w, T5t, T2Z, T6U, T6T, T32; | |
395 { | |
396 E Tx, T2N, T2v, T6V, T6P, T6Q, T1e, T2G, T31, T2X, T2L, T1Y, T6W, T2Q, T30; | |
397 E T2U; | |
398 { | |
399 E T1W, T1L, T2O, T2P, T2V, T2W, T6O, TS, T1d; | |
400 { | |
401 E T4q, T4w, T1V, T1B, T1J, T4m, T4l, T1G, Tw, T1A, T4p; | |
402 T6A = Tv + Tm; | |
403 Tw = Tm - Tv; | |
404 T1A = FMA(KP707106781, T1z, T1y); | |
405 T4p = FMS(KP707106781, T1z, T1y); | |
406 { | |
407 E T4o, T1x, T1I, T1F; | |
408 T4o = FMA(KP707106781, T1w, T1t); | |
409 T1x = FNMS(KP707106781, T1w, T1t); | |
410 T1I = T1D - T1E; | |
411 T1F = T1D + T1E; | |
412 T4q = FNMS(KP198912367, T4p, T4o); | |
413 T4w = FMA(KP198912367, T4o, T4p); | |
414 T1V = FMA(KP668178637, T1x, T1A); | |
415 T1B = FNMS(KP668178637, T1A, T1x); | |
416 T1J = FNMS(KP707106781, T1I, T1H); | |
417 T4m = FMA(KP707106781, T1I, T1H); | |
418 T4l = FMA(KP707106781, T1F, T1C); | |
419 T1G = FNMS(KP707106781, T1F, T1C); | |
420 Tx = FNMS(KP831469612, Tw, Td); | |
421 T2N = FMA(KP831469612, Tw, Td); | |
422 } | |
423 { | |
424 E T4n, T4x, T1U, T1K; | |
425 T4n = FMA(KP198912367, T4m, T4l); | |
426 T4x = FNMS(KP198912367, T4l, T4m); | |
427 T1U = FMA(KP668178637, T1G, T1J); | |
428 T1K = FNMS(KP668178637, T1J, T1G); | |
429 T5w = T4n + T4q; | |
430 T4r = T4n - T4q; | |
431 T5t = T4x + T4w; | |
432 T4y = T4w - T4x; | |
433 T3h = T1U - T1V; | |
434 T1W = T1U + T1V; | |
435 T3k = T1K + T1B; | |
436 T1L = T1B - T1K; | |
437 T6O = T34 + T35; | |
438 T36 = T34 - T35; | |
439 } | |
440 } | |
441 T2O = FNMS(KP534511135, TK, TR); | |
442 TS = FMA(KP534511135, TR, TK); | |
443 T1d = FMA(KP534511135, T1c, T15); | |
444 T2P = FNMS(KP534511135, T15, T1c); | |
445 T2v = FMA(KP831469612, T2u, T2b); | |
446 T2V = FNMS(KP831469612, T2u, T2b); | |
447 T6V = FNMS(KP831469612, T6O, T6N); | |
448 T6P = FMA(KP831469612, T6O, T6N); | |
449 T6Q = TS + T1d; | |
450 T1e = TS - T1d; | |
451 T2W = FMA(KP831469612, T2F, T2C); | |
452 T2G = FNMS(KP831469612, T2F, T2C); | |
453 { | |
454 E T2S, T2T, T1M, T1X; | |
455 T2S = FMA(KP831469612, T1L, T1s); | |
456 T1M = FNMS(KP831469612, T1L, T1s); | |
457 T1X = FNMS(KP831469612, T1W, T1T); | |
458 T2T = FMA(KP831469612, T1W, T1T); | |
459 T31 = FMA(KP250486960, T2V, T2W); | |
460 T2X = FNMS(KP250486960, T2W, T2V); | |
461 T2L = FNMS(KP599376933, T1M, T1X); | |
462 T1Y = FMA(KP599376933, T1X, T1M); | |
463 T6W = T2O + T2P; | |
464 T2Q = T2O - T2P; | |
465 T30 = FMA(KP250486960, T2S, T2T); | |
466 T2U = FNMS(KP250486960, T2T, T2S); | |
467 } | |
468 } | |
469 { | |
470 E T2J, T1f, T6X, T6Z, T2K, T2H; | |
471 T2J = FNMS(KP881921264, T1e, Tx); | |
472 T1f = FMA(KP881921264, T1e, Tx); | |
473 T6X = FNMS(KP881921264, T6W, T6V); | |
474 T6Z = FMA(KP881921264, T6W, T6V); | |
475 T2K = FNMS(KP599376933, T2v, T2G); | |
476 T2H = FMA(KP599376933, T2G, T2v); | |
477 { | |
478 E T2R, T2Y, T6R, T6S; | |
479 T2Z = FNMS(KP881921264, T2Q, T2N); | |
480 T2R = FMA(KP881921264, T2Q, T2N); | |
481 { | |
482 E T2M, T6Y, T70, T2I; | |
483 T2M = T2K - T2L; | |
484 T6Y = T2L + T2K; | |
485 T70 = T1Y + T2H; | |
486 T2I = T1Y - T2H; | |
487 Cr[WS(csr, 10)] = FMA(KP857728610, T2M, T2J); | |
488 Cr[WS(csr, 21)] = FNMS(KP857728610, T2M, T2J); | |
489 Ci[WS(csi, 5)] = FMA(KP857728610, T6Y, T6X); | |
490 Ci[WS(csi, 26)] = FMS(KP857728610, T6Y, T6X); | |
491 Ci[WS(csi, 21)] = FNMS(KP857728610, T70, T6Z); | |
492 Ci[WS(csi, 10)] = -(FMA(KP857728610, T70, T6Z)); | |
493 Cr[WS(csr, 5)] = FMA(KP857728610, T2I, T1f); | |
494 Cr[WS(csr, 26)] = FNMS(KP857728610, T2I, T1f); | |
495 T2Y = T2U - T2X; | |
496 T6U = T2U + T2X; | |
497 } | |
498 T6T = FNMS(KP881921264, T6Q, T6P); | |
499 T6R = FMA(KP881921264, T6Q, T6P); | |
500 T6S = T30 + T31; | |
501 T32 = T30 - T31; | |
502 Cr[WS(csr, 2)] = FMA(KP970031253, T2Y, T2R); | |
503 Cr[WS(csr, 29)] = FNMS(KP970031253, T2Y, T2R); | |
504 Ci[WS(csi, 29)] = FNMS(KP970031253, T6S, T6R); | |
505 Ci[WS(csi, 2)] = -(FMA(KP970031253, T6S, T6R)); | |
506 } | |
507 } | |
508 } | |
509 { | |
510 E T5j, T5L, T5B, T6d, T67, T68, T5q, T5E, T5Z, T5V, T5J, T5y, T6e, T5O, T5Y; | |
511 E T5S; | |
512 { | |
513 E T5M, T5N, T5T, T5U; | |
514 { | |
515 E T66, T5i, T5m, T5p; | |
516 T6k = T5g + T5h; | |
517 T5i = T5g - T5h; | |
518 Cr[WS(csr, 13)] = FMA(KP970031253, T32, T2Z); | |
519 Cr[WS(csr, 18)] = FNMS(KP970031253, T32, T2Z); | |
520 Ci[WS(csi, 13)] = FNMS(KP970031253, T6U, T6T); | |
521 Ci[WS(csi, 18)] = -(FMA(KP970031253, T6U, T6T)); | |
522 T5j = FNMS(KP980785280, T5i, T5f); | |
523 T5L = FMA(KP980785280, T5i, T5f); | |
524 T66 = T3W + T3Z; | |
525 T40 = T3W - T3Z; | |
526 T5M = FNMS(KP098491403, T5k, T5l); | |
527 T5m = FMA(KP098491403, T5l, T5k); | |
528 T5p = FMA(KP098491403, T5o, T5n); | |
529 T5N = FNMS(KP098491403, T5n, T5o); | |
530 T5B = FNMS(KP980785280, T5A, T5z); | |
531 T5T = FMA(KP980785280, T5A, T5z); | |
532 T6d = FNMS(KP980785280, T66, T65); | |
533 T67 = FMA(KP980785280, T66, T65); | |
534 T68 = T5m + T5p; | |
535 T5q = T5m - T5p; | |
536 T5U = FMA(KP980785280, T5D, T5C); | |
537 T5E = FNMS(KP980785280, T5D, T5C); | |
538 } | |
539 { | |
540 E T5Q, T5R, T5u, T5x; | |
541 T5Q = FMA(KP980785280, T5t, T5s); | |
542 T5u = FNMS(KP980785280, T5t, T5s); | |
543 T5x = FNMS(KP980785280, T5w, T5v); | |
544 T5R = FMA(KP980785280, T5w, T5v); | |
545 T5Z = FNMS(KP049126849, T5T, T5U); | |
546 T5V = FMA(KP049126849, T5U, T5T); | |
547 T5J = FNMS(KP906347169, T5u, T5x); | |
548 T5y = FMA(KP906347169, T5x, T5u); | |
549 T6e = T5M + T5N; | |
550 T5O = T5M - T5N; | |
551 T5Y = FMA(KP049126849, T5Q, T5R); | |
552 T5S = FNMS(KP049126849, T5R, T5Q); | |
553 } | |
554 } | |
555 { | |
556 E T5H, T5r, T6f, T6h, T5I, T5F; | |
557 T5H = FNMS(KP995184726, T5q, T5j); | |
558 T5r = FMA(KP995184726, T5q, T5j); | |
559 T6f = FNMS(KP995184726, T6e, T6d); | |
560 T6h = FMA(KP995184726, T6e, T6d); | |
561 T5I = FMA(KP906347169, T5B, T5E); | |
562 T5F = FNMS(KP906347169, T5E, T5B); | |
563 { | |
564 E T5P, T5W, T69, T6a; | |
565 T5X = FNMS(KP995184726, T5O, T5L); | |
566 T5P = FMA(KP995184726, T5O, T5L); | |
567 { | |
568 E T5K, T6g, T6i, T5G; | |
569 T5K = T5I - T5J; | |
570 T6g = T5J + T5I; | |
571 T6i = T5F - T5y; | |
572 T5G = T5y + T5F; | |
573 Cr[WS(csr, 8)] = FMA(KP740951125, T5K, T5H); | |
574 Cr[WS(csr, 23)] = FNMS(KP740951125, T5K, T5H); | |
575 Ci[WS(csi, 7)] = FMA(KP740951125, T6g, T6f); | |
576 Ci[WS(csi, 24)] = FMS(KP740951125, T6g, T6f); | |
577 Ci[WS(csi, 23)] = FMA(KP740951125, T6i, T6h); | |
578 Ci[WS(csi, 8)] = FMS(KP740951125, T6i, T6h); | |
579 Cr[WS(csr, 7)] = FMA(KP740951125, T5G, T5r); | |
580 Cr[WS(csr, 24)] = FNMS(KP740951125, T5G, T5r); | |
581 T5W = T5S + T5V; | |
582 T6c = T5V - T5S; | |
583 } | |
584 T6b = FNMS(KP995184726, T68, T67); | |
585 T69 = FMA(KP995184726, T68, T67); | |
586 T6a = T5Y + T5Z; | |
587 T60 = T5Y - T5Z; | |
588 Cr[0] = FMA(KP998795456, T5W, T5P); | |
589 Cr[WS(csr, 31)] = FNMS(KP998795456, T5W, T5P); | |
590 Ci[WS(csi, 31)] = FNMS(KP998795456, T6a, T69); | |
591 Ci[0] = -(FMA(KP998795456, T6a, T69)); | |
592 } | |
593 } | |
594 } | |
595 } | |
596 { | |
597 E T3L, T6G, T6F, T3O; | |
598 { | |
599 E T37, T3z, T3p, T6H, T6B, T6C, T3e, T3s, T3M, T3J, T3w, T3m, T6I, T3C, T3N; | |
600 E T3G; | |
601 { | |
602 E T3B, T3A, T3H, T3I, T3a, T3d; | |
603 Cr[WS(csr, 15)] = FMA(KP998795456, T60, T5X); | |
604 Cr[WS(csr, 16)] = FNMS(KP998795456, T60, T5X); | |
605 Ci[WS(csi, 15)] = FMA(KP998795456, T6c, T6b); | |
606 Ci[WS(csi, 16)] = FMS(KP998795456, T6c, T6b); | |
607 T37 = FNMS(KP831469612, T36, T33); | |
608 T3z = FMA(KP831469612, T36, T33); | |
609 T3B = FMA(KP303346683, T38, T39); | |
610 T3a = FNMS(KP303346683, T39, T38); | |
611 T3d = FNMS(KP303346683, T3c, T3b); | |
612 T3A = FMA(KP303346683, T3b, T3c); | |
613 T3p = FMA(KP831469612, T3o, T3n); | |
614 T3H = FNMS(KP831469612, T3o, T3n); | |
615 T6H = FNMS(KP831469612, T6A, T6z); | |
616 T6B = FMA(KP831469612, T6A, T6z); | |
617 T6C = T3d + T3a; | |
618 T3e = T3a - T3d; | |
619 T3I = FMA(KP831469612, T3r, T3q); | |
620 T3s = FNMS(KP831469612, T3r, T3q); | |
621 { | |
622 E T3E, T3F, T3i, T3l; | |
623 T3E = FMA(KP831469612, T3h, T3g); | |
624 T3i = FNMS(KP831469612, T3h, T3g); | |
625 T3l = FNMS(KP831469612, T3k, T3j); | |
626 T3F = FMA(KP831469612, T3k, T3j); | |
627 T3M = FNMS(KP148335987, T3H, T3I); | |
628 T3J = FMA(KP148335987, T3I, T3H); | |
629 T3w = FMA(KP741650546, T3i, T3l); | |
630 T3m = FNMS(KP741650546, T3l, T3i); | |
631 T6I = T3A + T3B; | |
632 T3C = T3A - T3B; | |
633 T3N = FNMS(KP148335987, T3E, T3F); | |
634 T3G = FMA(KP148335987, T3F, T3E); | |
635 } | |
636 } | |
637 { | |
638 E T3v, T3f, T6J, T6L, T3x, T3t; | |
639 T3v = FNMS(KP956940335, T3e, T37); | |
640 T3f = FMA(KP956940335, T3e, T37); | |
641 T6J = FMA(KP956940335, T6I, T6H); | |
642 T6L = FNMS(KP956940335, T6I, T6H); | |
643 T3x = FMA(KP741650546, T3p, T3s); | |
644 T3t = FNMS(KP741650546, T3s, T3p); | |
645 { | |
646 E T3D, T3K, T6D, T6E; | |
647 T3L = FNMS(KP956940335, T3C, T3z); | |
648 T3D = FMA(KP956940335, T3C, T3z); | |
649 { | |
650 E T3y, T6K, T6M, T3u; | |
651 T3y = T3w - T3x; | |
652 T6K = T3w + T3x; | |
653 T6M = T3m + T3t; | |
654 T3u = T3m - T3t; | |
655 Cr[WS(csr, 9)] = FMA(KP803207531, T3y, T3v); | |
656 Cr[WS(csr, 22)] = FNMS(KP803207531, T3y, T3v); | |
657 Ci[WS(csi, 25)] = FNMS(KP803207531, T6K, T6J); | |
658 Ci[WS(csi, 6)] = -(FMA(KP803207531, T6K, T6J)); | |
659 Ci[WS(csi, 9)] = FNMS(KP803207531, T6M, T6L); | |
660 Ci[WS(csi, 22)] = -(FMA(KP803207531, T6M, T6L)); | |
661 Cr[WS(csr, 6)] = FMA(KP803207531, T3u, T3f); | |
662 Cr[WS(csr, 25)] = FNMS(KP803207531, T3u, T3f); | |
663 T3K = T3G - T3J; | |
664 T6G = T3G + T3J; | |
665 } | |
666 T6F = FNMS(KP956940335, T6C, T6B); | |
667 T6D = FMA(KP956940335, T6C, T6B); | |
668 T6E = T3N + T3M; | |
669 T3O = T3M - T3N; | |
670 Cr[WS(csr, 1)] = FMA(KP989176509, T3K, T3D); | |
671 Cr[WS(csr, 30)] = FNMS(KP989176509, T3K, T3D); | |
672 Ci[WS(csi, 1)] = FMA(KP989176509, T6E, T6D); | |
673 Ci[WS(csi, 30)] = FMS(KP989176509, T6E, T6D); | |
674 } | |
675 } | |
676 } | |
677 { | |
678 E T41, T4Z, T4L, T6r, T6l, T6m, T4g, T4S, T5c, T59, T4W, T4A, T6s, T52, T5d; | |
679 E T56; | |
680 { | |
681 E T51, T50, T57, T58, T48, T4f; | |
682 Cr[WS(csr, 14)] = FMA(KP989176509, T3O, T3L); | |
683 Cr[WS(csr, 17)] = FNMS(KP989176509, T3O, T3L); | |
684 Ci[WS(csi, 17)] = FNMS(KP989176509, T6G, T6F); | |
685 Ci[WS(csi, 14)] = -(FMA(KP989176509, T6G, T6F)); | |
686 T41 = FNMS(KP980785280, T40, T3T); | |
687 T4Z = FMA(KP980785280, T40, T3T); | |
688 T51 = FMA(KP820678790, T44, T47); | |
689 T48 = FNMS(KP820678790, T47, T44); | |
690 T4f = FNMS(KP820678790, T4e, T4b); | |
691 T50 = FMA(KP820678790, T4b, T4e); | |
692 T4L = FNMS(KP980785280, T4K, T4D); | |
693 T57 = FMA(KP980785280, T4K, T4D); | |
694 T6r = FMA(KP980785280, T6k, T6j); | |
695 T6l = FNMS(KP980785280, T6k, T6j); | |
696 T6m = T4f + T48; | |
697 T4g = T48 - T4f; | |
698 T58 = FMA(KP980785280, T4R, T4O); | |
699 T4S = FNMS(KP980785280, T4R, T4O); | |
700 { | |
701 E T54, T55, T4s, T4z; | |
702 T54 = FMA(KP980785280, T4r, T4k); | |
703 T4s = FNMS(KP980785280, T4r, T4k); | |
704 T4z = FNMS(KP980785280, T4y, T4v); | |
705 T55 = FMA(KP980785280, T4y, T4v); | |
706 T5c = FMA(KP357805721, T57, T58); | |
707 T59 = FNMS(KP357805721, T58, T57); | |
708 T4W = FMA(KP472964775, T4s, T4z); | |
709 T4A = FNMS(KP472964775, T4z, T4s); | |
710 T6s = T50 + T51; | |
711 T52 = T50 - T51; | |
712 T5d = FNMS(KP357805721, T54, T55); | |
713 T56 = FMA(KP357805721, T55, T54); | |
714 } | |
715 } | |
716 { | |
717 E T4V, T4h, T6t, T6v, T4X, T4T; | |
718 T4V = FNMS(KP773010453, T4g, T41); | |
719 T4h = FMA(KP773010453, T4g, T41); | |
720 T6t = FMA(KP773010453, T6s, T6r); | |
721 T6v = FNMS(KP773010453, T6s, T6r); | |
722 T4X = FNMS(KP472964775, T4L, T4S); | |
723 T4T = FMA(KP472964775, T4S, T4L); | |
724 { | |
725 E T53, T5a, T6n, T6o; | |
726 T5b = FNMS(KP773010453, T52, T4Z); | |
727 T53 = FMA(KP773010453, T52, T4Z); | |
728 { | |
729 E T4Y, T6u, T6w, T4U; | |
730 T4Y = T4W - T4X; | |
731 T6u = T4W + T4X; | |
732 T6w = T4T - T4A; | |
733 T4U = T4A + T4T; | |
734 Cr[WS(csr, 11)] = FMA(KP903989293, T4Y, T4V); | |
735 Cr[WS(csr, 20)] = FNMS(KP903989293, T4Y, T4V); | |
736 Ci[WS(csi, 27)] = FNMS(KP903989293, T6u, T6t); | |
737 Ci[WS(csi, 4)] = -(FMA(KP903989293, T6u, T6t)); | |
738 Ci[WS(csi, 11)] = FMA(KP903989293, T6w, T6v); | |
739 Ci[WS(csi, 20)] = FMS(KP903989293, T6w, T6v); | |
740 Cr[WS(csr, 4)] = FMA(KP903989293, T4U, T4h); | |
741 Cr[WS(csr, 27)] = FNMS(KP903989293, T4U, T4h); | |
742 T5a = T56 + T59; | |
743 T6q = T59 - T56; | |
744 } | |
745 T6p = FNMS(KP773010453, T6m, T6l); | |
746 T6n = FMA(KP773010453, T6m, T6l); | |
747 T6o = T5d + T5c; | |
748 T5e = T5c - T5d; | |
749 Cr[WS(csr, 3)] = FMA(KP941544065, T5a, T53); | |
750 Cr[WS(csr, 28)] = FNMS(KP941544065, T5a, T53); | |
751 Ci[WS(csi, 3)] = FMA(KP941544065, T6o, T6n); | |
752 Ci[WS(csi, 28)] = FMS(KP941544065, T6o, T6n); | |
753 } | |
754 } | |
755 } | |
756 } | |
757 } | |
758 } | |
759 Cr[WS(csr, 12)] = FMA(KP941544065, T5e, T5b); | |
760 Cr[WS(csr, 19)] = FNMS(KP941544065, T5e, T5b); | |
761 Ci[WS(csi, 19)] = FMA(KP941544065, T6q, T6p); | |
762 Ci[WS(csi, 12)] = FMS(KP941544065, T6q, T6p); | |
763 } | |
764 } | |
765 } | |
766 | |
767 static const kr2c_desc desc = { 64, "r2cfII_64", {114, 0, 320, 0}, &GENUS }; | |
768 | |
769 void X(codelet_r2cfII_64) (planner *p) { | |
770 X(kr2c_register) (p, r2cfII_64, &desc); | |
771 } | |
772 | |
773 #else /* HAVE_FMA */ | |
774 | |
775 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cfII_64 -dft-II -include r2cfII.h */ | |
776 | |
777 /* | |
778 * This function contains 434 FP additions, 206 FP multiplications, | |
779 * (or, 342 additions, 114 multiplications, 92 fused multiply/add), | |
780 * 118 stack variables, 31 constants, and 128 memory accesses | |
781 */ | |
782 #include "r2cfII.h" | |
783 | |
784 static void r2cfII_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
785 { | |
786 DK(KP242980179, +0.242980179903263889948274162077471118320990783); | |
787 DK(KP970031253, +0.970031253194543992603984207286100251456865962); | |
788 DK(KP857728610, +0.857728610000272069902269984284770137042490799); | |
789 DK(KP514102744, +0.514102744193221726593693838968815772608049120); | |
790 DK(KP471396736, +0.471396736825997648556387625905254377657460319); | |
791 DK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
792 DK(KP427555093, +0.427555093430282094320966856888798534304578629); | |
793 DK(KP903989293, +0.903989293123443331586200297230537048710132025); | |
794 DK(KP336889853, +0.336889853392220050689253212619147570477766780); | |
795 DK(KP941544065, +0.941544065183020778412509402599502357185589796); | |
796 DK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
797 DK(KP634393284, +0.634393284163645498215171613225493370675687095); | |
798 DK(KP595699304, +0.595699304492433343467036528829969889511926338); | |
799 DK(KP803207531, +0.803207531480644909806676512963141923879569427); | |
800 DK(KP146730474, +0.146730474455361751658850129646717819706215317); | |
801 DK(KP989176509, +0.989176509964780973451673738016243063983689533); | |
802 DK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
803 DK(KP290284677, +0.290284677254462367636192375817395274691476278); | |
804 DK(KP049067674, +0.049067674327418014254954976942682658314745363); | |
805 DK(KP998795456, +0.998795456205172392714771604759100694443203615); | |
806 DK(KP671558954, +0.671558954847018400625376850427421803228750632); | |
807 DK(KP740951125, +0.740951125354959091175616897495162729728955309); | |
808 DK(KP098017140, +0.098017140329560601994195563888641845861136673); | |
809 DK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
810 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
811 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
812 DK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
813 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
814 DK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
815 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
816 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
817 { | |
818 INT i; | |
819 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) { | |
820 E Tm, T34, T3Z, T5g, Tv, T35, T3W, T5h, Td, T33, T6B, T6Q, T3T, T5f, T68; | |
821 E T6m, T2b, T3n, T4O, T5D, T2F, T3r, T4K, T5z, TK, T3c, T47, T5n, TR, T3b; | |
822 E T44, T5o, T15, T38, T4e, T5l, T1c, T39, T4b, T5k, T1s, T3g, T4v, T5w, T1W; | |
823 E T3k, T4k, T5s, T2u, T3q, T4R, T5A, T2y, T3o, T4H, T5C, T1L, T3j, T4y, T5t; | |
824 E T1P, T3h, T4r, T5v; | |
825 { | |
826 E Te, Tk, Th, Tj, Tf, Tg; | |
827 Te = R0[WS(rs, 2)]; | |
828 Tk = R0[WS(rs, 18)]; | |
829 Tf = R0[WS(rs, 10)]; | |
830 Tg = R0[WS(rs, 26)]; | |
831 Th = KP707106781 * (Tf - Tg); | |
832 Tj = KP707106781 * (Tf + Tg); | |
833 { | |
834 E Ti, Tl, T3X, T3Y; | |
835 Ti = Te + Th; | |
836 Tl = Tj + Tk; | |
837 Tm = FNMS(KP195090322, Tl, KP980785280 * Ti); | |
838 T34 = FMA(KP195090322, Ti, KP980785280 * Tl); | |
839 T3X = Tk - Tj; | |
840 T3Y = Te - Th; | |
841 T3Z = FNMS(KP555570233, T3Y, KP831469612 * T3X); | |
842 T5g = FMA(KP831469612, T3Y, KP555570233 * T3X); | |
843 } | |
844 } | |
845 { | |
846 E Tq, Tt, Tp, Ts, Tn, To; | |
847 Tq = R0[WS(rs, 30)]; | |
848 Tt = R0[WS(rs, 14)]; | |
849 Tn = R0[WS(rs, 6)]; | |
850 To = R0[WS(rs, 22)]; | |
851 Tp = KP707106781 * (Tn - To); | |
852 Ts = KP707106781 * (Tn + To); | |
853 { | |
854 E Tr, Tu, T3U, T3V; | |
855 Tr = Tp - Tq; | |
856 Tu = Ts + Tt; | |
857 Tv = FMA(KP980785280, Tr, KP195090322 * Tu); | |
858 T35 = FNMS(KP980785280, Tu, KP195090322 * Tr); | |
859 T3U = Tt - Ts; | |
860 T3V = Tp + Tq; | |
861 T3W = FNMS(KP555570233, T3V, KP831469612 * T3U); | |
862 T5h = FMA(KP831469612, T3V, KP555570233 * T3U); | |
863 } | |
864 } | |
865 { | |
866 E T1, T66, T4, T65, T8, T3Q, Tb, T3R, T2, T3; | |
867 T1 = R0[0]; | |
868 T66 = R0[WS(rs, 16)]; | |
869 T2 = R0[WS(rs, 8)]; | |
870 T3 = R0[WS(rs, 24)]; | |
871 T4 = KP707106781 * (T2 - T3); | |
872 T65 = KP707106781 * (T2 + T3); | |
873 { | |
874 E T6, T7, T9, Ta; | |
875 T6 = R0[WS(rs, 4)]; | |
876 T7 = R0[WS(rs, 20)]; | |
877 T8 = FNMS(KP382683432, T7, KP923879532 * T6); | |
878 T3Q = FMA(KP382683432, T6, KP923879532 * T7); | |
879 T9 = R0[WS(rs, 12)]; | |
880 Ta = R0[WS(rs, 28)]; | |
881 Tb = FNMS(KP923879532, Ta, KP382683432 * T9); | |
882 T3R = FMA(KP923879532, T9, KP382683432 * Ta); | |
883 } | |
884 { | |
885 E T5, Tc, T6z, T6A; | |
886 T5 = T1 + T4; | |
887 Tc = T8 + Tb; | |
888 Td = T5 + Tc; | |
889 T33 = T5 - Tc; | |
890 T6z = Tb - T8; | |
891 T6A = T66 - T65; | |
892 T6B = T6z - T6A; | |
893 T6Q = T6z + T6A; | |
894 } | |
895 { | |
896 E T3P, T3S, T64, T67; | |
897 T3P = T1 - T4; | |
898 T3S = T3Q - T3R; | |
899 T3T = T3P - T3S; | |
900 T5f = T3P + T3S; | |
901 T64 = T3Q + T3R; | |
902 T67 = T65 + T66; | |
903 T68 = T64 + T67; | |
904 T6m = T67 - T64; | |
905 } | |
906 } | |
907 { | |
908 E T22, T2D, T21, T2C, T26, T2z, T29, T2A, T1Z, T20; | |
909 T22 = R1[WS(rs, 31)]; | |
910 T2D = R1[WS(rs, 15)]; | |
911 T1Z = R1[WS(rs, 7)]; | |
912 T20 = R1[WS(rs, 23)]; | |
913 T21 = KP707106781 * (T1Z - T20); | |
914 T2C = KP707106781 * (T1Z + T20); | |
915 { | |
916 E T24, T25, T27, T28; | |
917 T24 = R1[WS(rs, 3)]; | |
918 T25 = R1[WS(rs, 19)]; | |
919 T26 = FNMS(KP382683432, T25, KP923879532 * T24); | |
920 T2z = FMA(KP382683432, T24, KP923879532 * T25); | |
921 T27 = R1[WS(rs, 11)]; | |
922 T28 = R1[WS(rs, 27)]; | |
923 T29 = FNMS(KP923879532, T28, KP382683432 * T27); | |
924 T2A = FMA(KP923879532, T27, KP382683432 * T28); | |
925 } | |
926 { | |
927 E T23, T2a, T4M, T4N; | |
928 T23 = T21 - T22; | |
929 T2a = T26 + T29; | |
930 T2b = T23 + T2a; | |
931 T3n = T23 - T2a; | |
932 T4M = T29 - T26; | |
933 T4N = T2D - T2C; | |
934 T4O = T4M - T4N; | |
935 T5D = T4M + T4N; | |
936 } | |
937 { | |
938 E T2B, T2E, T4I, T4J; | |
939 T2B = T2z + T2A; | |
940 T2E = T2C + T2D; | |
941 T2F = T2B + T2E; | |
942 T3r = T2E - T2B; | |
943 T4I = T21 + T22; | |
944 T4J = T2z - T2A; | |
945 T4K = T4I + T4J; | |
946 T5z = T4J - T4I; | |
947 } | |
948 } | |
949 { | |
950 E Ty, TP, TB, TO, TF, TL, TI, TM, Tz, TA; | |
951 Ty = R0[WS(rs, 1)]; | |
952 TP = R0[WS(rs, 17)]; | |
953 Tz = R0[WS(rs, 9)]; | |
954 TA = R0[WS(rs, 25)]; | |
955 TB = KP707106781 * (Tz - TA); | |
956 TO = KP707106781 * (Tz + TA); | |
957 { | |
958 E TD, TE, TG, TH; | |
959 TD = R0[WS(rs, 5)]; | |
960 TE = R0[WS(rs, 21)]; | |
961 TF = FNMS(KP382683432, TE, KP923879532 * TD); | |
962 TL = FMA(KP382683432, TD, KP923879532 * TE); | |
963 TG = R0[WS(rs, 13)]; | |
964 TH = R0[WS(rs, 29)]; | |
965 TI = FNMS(KP923879532, TH, KP382683432 * TG); | |
966 TM = FMA(KP923879532, TG, KP382683432 * TH); | |
967 } | |
968 { | |
969 E TC, TJ, T45, T46; | |
970 TC = Ty + TB; | |
971 TJ = TF + TI; | |
972 TK = TC + TJ; | |
973 T3c = TC - TJ; | |
974 T45 = TI - TF; | |
975 T46 = TP - TO; | |
976 T47 = T45 - T46; | |
977 T5n = T45 + T46; | |
978 } | |
979 { | |
980 E TN, TQ, T42, T43; | |
981 TN = TL + TM; | |
982 TQ = TO + TP; | |
983 TR = TN + TQ; | |
984 T3b = TQ - TN; | |
985 T42 = Ty - TB; | |
986 T43 = TL - TM; | |
987 T44 = T42 - T43; | |
988 T5o = T42 + T43; | |
989 } | |
990 } | |
991 { | |
992 E TW, T1a, TV, T19, T10, T16, T13, T17, TT, TU; | |
993 TW = R0[WS(rs, 31)]; | |
994 T1a = R0[WS(rs, 15)]; | |
995 TT = R0[WS(rs, 7)]; | |
996 TU = R0[WS(rs, 23)]; | |
997 TV = KP707106781 * (TT - TU); | |
998 T19 = KP707106781 * (TT + TU); | |
999 { | |
1000 E TY, TZ, T11, T12; | |
1001 TY = R0[WS(rs, 3)]; | |
1002 TZ = R0[WS(rs, 19)]; | |
1003 T10 = FNMS(KP382683432, TZ, KP923879532 * TY); | |
1004 T16 = FMA(KP382683432, TY, KP923879532 * TZ); | |
1005 T11 = R0[WS(rs, 11)]; | |
1006 T12 = R0[WS(rs, 27)]; | |
1007 T13 = FNMS(KP923879532, T12, KP382683432 * T11); | |
1008 T17 = FMA(KP923879532, T11, KP382683432 * T12); | |
1009 } | |
1010 { | |
1011 E TX, T14, T4c, T4d; | |
1012 TX = TV - TW; | |
1013 T14 = T10 + T13; | |
1014 T15 = TX + T14; | |
1015 T38 = TX - T14; | |
1016 T4c = T13 - T10; | |
1017 T4d = T1a - T19; | |
1018 T4e = T4c - T4d; | |
1019 T5l = T4c + T4d; | |
1020 } | |
1021 { | |
1022 E T18, T1b, T49, T4a; | |
1023 T18 = T16 + T17; | |
1024 T1b = T19 + T1a; | |
1025 T1c = T18 + T1b; | |
1026 T39 = T1b - T18; | |
1027 T49 = TV + TW; | |
1028 T4a = T16 - T17; | |
1029 T4b = T49 + T4a; | |
1030 T5k = T4a - T49; | |
1031 } | |
1032 } | |
1033 { | |
1034 E T1g, T1U, T1j, T1T, T1n, T1Q, T1q, T1R, T1h, T1i; | |
1035 T1g = R1[0]; | |
1036 T1U = R1[WS(rs, 16)]; | |
1037 T1h = R1[WS(rs, 8)]; | |
1038 T1i = R1[WS(rs, 24)]; | |
1039 T1j = KP707106781 * (T1h - T1i); | |
1040 T1T = KP707106781 * (T1h + T1i); | |
1041 { | |
1042 E T1l, T1m, T1o, T1p; | |
1043 T1l = R1[WS(rs, 4)]; | |
1044 T1m = R1[WS(rs, 20)]; | |
1045 T1n = FNMS(KP382683432, T1m, KP923879532 * T1l); | |
1046 T1Q = FMA(KP382683432, T1l, KP923879532 * T1m); | |
1047 T1o = R1[WS(rs, 12)]; | |
1048 T1p = R1[WS(rs, 28)]; | |
1049 T1q = FNMS(KP923879532, T1p, KP382683432 * T1o); | |
1050 T1R = FMA(KP923879532, T1o, KP382683432 * T1p); | |
1051 } | |
1052 { | |
1053 E T1k, T1r, T4t, T4u; | |
1054 T1k = T1g + T1j; | |
1055 T1r = T1n + T1q; | |
1056 T1s = T1k + T1r; | |
1057 T3g = T1k - T1r; | |
1058 T4t = T1q - T1n; | |
1059 T4u = T1U - T1T; | |
1060 T4v = T4t - T4u; | |
1061 T5w = T4t + T4u; | |
1062 } | |
1063 { | |
1064 E T1S, T1V, T4i, T4j; | |
1065 T1S = T1Q + T1R; | |
1066 T1V = T1T + T1U; | |
1067 T1W = T1S + T1V; | |
1068 T3k = T1V - T1S; | |
1069 T4i = T1g - T1j; | |
1070 T4j = T1Q - T1R; | |
1071 T4k = T4i - T4j; | |
1072 T5s = T4i + T4j; | |
1073 } | |
1074 } | |
1075 { | |
1076 E T2g, T4F, T2j, T4E, T2p, T4C, T2s, T4B; | |
1077 { | |
1078 E T2c, T2i, T2f, T2h, T2d, T2e; | |
1079 T2c = R1[WS(rs, 1)]; | |
1080 T2i = R1[WS(rs, 17)]; | |
1081 T2d = R1[WS(rs, 9)]; | |
1082 T2e = R1[WS(rs, 25)]; | |
1083 T2f = KP707106781 * (T2d - T2e); | |
1084 T2h = KP707106781 * (T2d + T2e); | |
1085 T2g = T2c + T2f; | |
1086 T4F = T2c - T2f; | |
1087 T2j = T2h + T2i; | |
1088 T4E = T2i - T2h; | |
1089 } | |
1090 { | |
1091 E T2o, T2r, T2n, T2q, T2l, T2m; | |
1092 T2o = R1[WS(rs, 29)]; | |
1093 T2r = R1[WS(rs, 13)]; | |
1094 T2l = R1[WS(rs, 5)]; | |
1095 T2m = R1[WS(rs, 21)]; | |
1096 T2n = KP707106781 * (T2l - T2m); | |
1097 T2q = KP707106781 * (T2l + T2m); | |
1098 T2p = T2n - T2o; | |
1099 T4C = T2n + T2o; | |
1100 T2s = T2q + T2r; | |
1101 T4B = T2r - T2q; | |
1102 } | |
1103 { | |
1104 E T2k, T2t, T4P, T4Q; | |
1105 T2k = FNMS(KP195090322, T2j, KP980785280 * T2g); | |
1106 T2t = FMA(KP980785280, T2p, KP195090322 * T2s); | |
1107 T2u = T2k + T2t; | |
1108 T3q = T2t - T2k; | |
1109 T4P = FMA(KP831469612, T4F, KP555570233 * T4E); | |
1110 T4Q = FMA(KP831469612, T4C, KP555570233 * T4B); | |
1111 T4R = T4P + T4Q; | |
1112 T5A = T4P - T4Q; | |
1113 } | |
1114 { | |
1115 E T2w, T2x, T4D, T4G; | |
1116 T2w = FNMS(KP980785280, T2s, KP195090322 * T2p); | |
1117 T2x = FMA(KP195090322, T2g, KP980785280 * T2j); | |
1118 T2y = T2w - T2x; | |
1119 T3o = T2x + T2w; | |
1120 T4D = FNMS(KP555570233, T4C, KP831469612 * T4B); | |
1121 T4G = FNMS(KP555570233, T4F, KP831469612 * T4E); | |
1122 T4H = T4D - T4G; | |
1123 T5C = T4G + T4D; | |
1124 } | |
1125 } | |
1126 { | |
1127 E T1x, T4p, T1A, T4o, T1G, T4m, T1J, T4l; | |
1128 { | |
1129 E T1t, T1z, T1w, T1y, T1u, T1v; | |
1130 T1t = R1[WS(rs, 2)]; | |
1131 T1z = R1[WS(rs, 18)]; | |
1132 T1u = R1[WS(rs, 10)]; | |
1133 T1v = R1[WS(rs, 26)]; | |
1134 T1w = KP707106781 * (T1u - T1v); | |
1135 T1y = KP707106781 * (T1u + T1v); | |
1136 T1x = T1t + T1w; | |
1137 T4p = T1t - T1w; | |
1138 T1A = T1y + T1z; | |
1139 T4o = T1z - T1y; | |
1140 } | |
1141 { | |
1142 E T1F, T1I, T1E, T1H, T1C, T1D; | |
1143 T1F = R1[WS(rs, 30)]; | |
1144 T1I = R1[WS(rs, 14)]; | |
1145 T1C = R1[WS(rs, 6)]; | |
1146 T1D = R1[WS(rs, 22)]; | |
1147 T1E = KP707106781 * (T1C - T1D); | |
1148 T1H = KP707106781 * (T1C + T1D); | |
1149 T1G = T1E - T1F; | |
1150 T4m = T1E + T1F; | |
1151 T1J = T1H + T1I; | |
1152 T4l = T1I - T1H; | |
1153 } | |
1154 { | |
1155 E T1B, T1K, T4w, T4x; | |
1156 T1B = FNMS(KP195090322, T1A, KP980785280 * T1x); | |
1157 T1K = FMA(KP980785280, T1G, KP195090322 * T1J); | |
1158 T1L = T1B + T1K; | |
1159 T3j = T1K - T1B; | |
1160 T4w = FMA(KP831469612, T4p, KP555570233 * T4o); | |
1161 T4x = FMA(KP831469612, T4m, KP555570233 * T4l); | |
1162 T4y = T4w + T4x; | |
1163 T5t = T4w - T4x; | |
1164 } | |
1165 { | |
1166 E T1N, T1O, T4n, T4q; | |
1167 T1N = FNMS(KP980785280, T1J, KP195090322 * T1G); | |
1168 T1O = FMA(KP195090322, T1x, KP980785280 * T1A); | |
1169 T1P = T1N - T1O; | |
1170 T3h = T1O + T1N; | |
1171 T4n = FNMS(KP555570233, T4m, KP831469612 * T4l); | |
1172 T4q = FNMS(KP555570233, T4p, KP831469612 * T4o); | |
1173 T4r = T4n - T4q; | |
1174 T5v = T4q + T4n; | |
1175 } | |
1176 } | |
1177 { | |
1178 E Tx, T2N, T69, T6f, T1e, T6e, T2X, T30, T1Y, T2L, T2Q, T62, T2U, T31, T2H; | |
1179 E T2K, Tw, T63; | |
1180 Tw = Tm + Tv; | |
1181 Tx = Td + Tw; | |
1182 T2N = Td - Tw; | |
1183 T63 = T35 - T34; | |
1184 T69 = T63 - T68; | |
1185 T6f = T63 + T68; | |
1186 { | |
1187 E TS, T1d, T2V, T2W; | |
1188 TS = FNMS(KP098017140, TR, KP995184726 * TK); | |
1189 T1d = FMA(KP995184726, T15, KP098017140 * T1c); | |
1190 T1e = TS + T1d; | |
1191 T6e = T1d - TS; | |
1192 T2V = T2b - T2u; | |
1193 T2W = T2y + T2F; | |
1194 T2X = FNMS(KP671558954, T2W, KP740951125 * T2V); | |
1195 T30 = FMA(KP671558954, T2V, KP740951125 * T2W); | |
1196 } | |
1197 { | |
1198 E T1M, T1X, T2O, T2P; | |
1199 T1M = T1s + T1L; | |
1200 T1X = T1P - T1W; | |
1201 T1Y = FMA(KP998795456, T1M, KP049067674 * T1X); | |
1202 T2L = FNMS(KP049067674, T1M, KP998795456 * T1X); | |
1203 T2O = FMA(KP098017140, TK, KP995184726 * TR); | |
1204 T2P = FNMS(KP995184726, T1c, KP098017140 * T15); | |
1205 T2Q = T2O + T2P; | |
1206 T62 = T2P - T2O; | |
1207 } | |
1208 { | |
1209 E T2S, T2T, T2v, T2G; | |
1210 T2S = T1s - T1L; | |
1211 T2T = T1P + T1W; | |
1212 T2U = FMA(KP740951125, T2S, KP671558954 * T2T); | |
1213 T31 = FNMS(KP671558954, T2S, KP740951125 * T2T); | |
1214 T2v = T2b + T2u; | |
1215 T2G = T2y - T2F; | |
1216 T2H = FNMS(KP049067674, T2G, KP998795456 * T2v); | |
1217 T2K = FMA(KP049067674, T2v, KP998795456 * T2G); | |
1218 } | |
1219 { | |
1220 E T1f, T2I, T6b, T6c; | |
1221 T1f = Tx + T1e; | |
1222 T2I = T1Y + T2H; | |
1223 Cr[WS(csr, 31)] = T1f - T2I; | |
1224 Cr[0] = T1f + T2I; | |
1225 T6b = T2L + T2K; | |
1226 T6c = T62 + T69; | |
1227 Ci[WS(csi, 31)] = T6b - T6c; | |
1228 Ci[0] = T6b + T6c; | |
1229 } | |
1230 { | |
1231 E T2J, T2M, T61, T6a; | |
1232 T2J = Tx - T1e; | |
1233 T2M = T2K - T2L; | |
1234 Cr[WS(csr, 16)] = T2J - T2M; | |
1235 Cr[WS(csr, 15)] = T2J + T2M; | |
1236 T61 = T2H - T1Y; | |
1237 T6a = T62 - T69; | |
1238 Ci[WS(csi, 16)] = T61 - T6a; | |
1239 Ci[WS(csi, 15)] = T61 + T6a; | |
1240 } | |
1241 { | |
1242 E T2R, T2Y, T6h, T6i; | |
1243 T2R = T2N + T2Q; | |
1244 T2Y = T2U + T2X; | |
1245 Cr[WS(csr, 24)] = T2R - T2Y; | |
1246 Cr[WS(csr, 7)] = T2R + T2Y; | |
1247 T6h = T31 + T30; | |
1248 T6i = T6e + T6f; | |
1249 Ci[WS(csi, 24)] = T6h - T6i; | |
1250 Ci[WS(csi, 7)] = T6h + T6i; | |
1251 } | |
1252 { | |
1253 E T2Z, T32, T6d, T6g; | |
1254 T2Z = T2N - T2Q; | |
1255 T32 = T30 - T31; | |
1256 Cr[WS(csr, 23)] = T2Z - T32; | |
1257 Cr[WS(csr, 8)] = T2Z + T32; | |
1258 T6d = T2X - T2U; | |
1259 T6g = T6e - T6f; | |
1260 Ci[WS(csi, 23)] = T6d - T6g; | |
1261 Ci[WS(csi, 8)] = T6d + T6g; | |
1262 } | |
1263 } | |
1264 { | |
1265 E T5j, T5L, T6R, T6X, T5q, T6W, T5V, T5Y, T5y, T5J, T5O, T6O, T5S, T5Z, T5F; | |
1266 E T5I, T5i, T6P; | |
1267 T5i = T5g - T5h; | |
1268 T5j = T5f - T5i; | |
1269 T5L = T5f + T5i; | |
1270 T6P = T3Z + T3W; | |
1271 T6R = T6P - T6Q; | |
1272 T6X = T6P + T6Q; | |
1273 { | |
1274 E T5m, T5p, T5T, T5U; | |
1275 T5m = FMA(KP290284677, T5k, KP956940335 * T5l); | |
1276 T5p = FNMS(KP290284677, T5o, KP956940335 * T5n); | |
1277 T5q = T5m - T5p; | |
1278 T6W = T5p + T5m; | |
1279 T5T = T5z + T5A; | |
1280 T5U = T5C + T5D; | |
1281 T5V = FNMS(KP146730474, T5U, KP989176509 * T5T); | |
1282 T5Y = FMA(KP146730474, T5T, KP989176509 * T5U); | |
1283 } | |
1284 { | |
1285 E T5u, T5x, T5M, T5N; | |
1286 T5u = T5s - T5t; | |
1287 T5x = T5v - T5w; | |
1288 T5y = FMA(KP803207531, T5u, KP595699304 * T5x); | |
1289 T5J = FNMS(KP595699304, T5u, KP803207531 * T5x); | |
1290 T5M = FMA(KP956940335, T5o, KP290284677 * T5n); | |
1291 T5N = FNMS(KP290284677, T5l, KP956940335 * T5k); | |
1292 T5O = T5M + T5N; | |
1293 T6O = T5N - T5M; | |
1294 } | |
1295 { | |
1296 E T5Q, T5R, T5B, T5E; | |
1297 T5Q = T5s + T5t; | |
1298 T5R = T5v + T5w; | |
1299 T5S = FMA(KP989176509, T5Q, KP146730474 * T5R); | |
1300 T5Z = FNMS(KP146730474, T5Q, KP989176509 * T5R); | |
1301 T5B = T5z - T5A; | |
1302 T5E = T5C - T5D; | |
1303 T5F = FNMS(KP595699304, T5E, KP803207531 * T5B); | |
1304 T5I = FMA(KP595699304, T5B, KP803207531 * T5E); | |
1305 } | |
1306 { | |
1307 E T5r, T5G, T6T, T6U; | |
1308 T5r = T5j + T5q; | |
1309 T5G = T5y + T5F; | |
1310 Cr[WS(csr, 25)] = T5r - T5G; | |
1311 Cr[WS(csr, 6)] = T5r + T5G; | |
1312 T6T = T5J + T5I; | |
1313 T6U = T6O + T6R; | |
1314 Ci[WS(csi, 25)] = T6T - T6U; | |
1315 Ci[WS(csi, 6)] = T6T + T6U; | |
1316 } | |
1317 { | |
1318 E T5H, T5K, T6N, T6S; | |
1319 T5H = T5j - T5q; | |
1320 T5K = T5I - T5J; | |
1321 Cr[WS(csr, 22)] = T5H - T5K; | |
1322 Cr[WS(csr, 9)] = T5H + T5K; | |
1323 T6N = T5F - T5y; | |
1324 T6S = T6O - T6R; | |
1325 Ci[WS(csi, 22)] = T6N - T6S; | |
1326 Ci[WS(csi, 9)] = T6N + T6S; | |
1327 } | |
1328 { | |
1329 E T5P, T5W, T6Z, T70; | |
1330 T5P = T5L + T5O; | |
1331 T5W = T5S + T5V; | |
1332 Cr[WS(csr, 30)] = T5P - T5W; | |
1333 Cr[WS(csr, 1)] = T5P + T5W; | |
1334 T6Z = T5Z + T5Y; | |
1335 T70 = T6W + T6X; | |
1336 Ci[WS(csi, 30)] = T6Z - T70; | |
1337 Ci[WS(csi, 1)] = T6Z + T70; | |
1338 } | |
1339 { | |
1340 E T5X, T60, T6V, T6Y; | |
1341 T5X = T5L - T5O; | |
1342 T60 = T5Y - T5Z; | |
1343 Cr[WS(csr, 17)] = T5X - T60; | |
1344 Cr[WS(csr, 14)] = T5X + T60; | |
1345 T6V = T5V - T5S; | |
1346 T6Y = T6W - T6X; | |
1347 Ci[WS(csi, 17)] = T6V - T6Y; | |
1348 Ci[WS(csi, 14)] = T6V + T6Y; | |
1349 } | |
1350 } | |
1351 { | |
1352 E T37, T3z, T6n, T6t, T3e, T6s, T3J, T3M, T3m, T3x, T3C, T6k, T3G, T3N, T3t; | |
1353 E T3w, T36, T6l; | |
1354 T36 = T34 + T35; | |
1355 T37 = T33 - T36; | |
1356 T3z = T33 + T36; | |
1357 T6l = Tv - Tm; | |
1358 T6n = T6l - T6m; | |
1359 T6t = T6l + T6m; | |
1360 { | |
1361 E T3a, T3d, T3H, T3I; | |
1362 T3a = FMA(KP634393284, T38, KP773010453 * T39); | |
1363 T3d = FNMS(KP634393284, T3c, KP773010453 * T3b); | |
1364 T3e = T3a - T3d; | |
1365 T6s = T3d + T3a; | |
1366 T3H = T3n + T3o; | |
1367 T3I = T3q + T3r; | |
1368 T3J = FNMS(KP336889853, T3I, KP941544065 * T3H); | |
1369 T3M = FMA(KP336889853, T3H, KP941544065 * T3I); | |
1370 } | |
1371 { | |
1372 E T3i, T3l, T3A, T3B; | |
1373 T3i = T3g - T3h; | |
1374 T3l = T3j - T3k; | |
1375 T3m = FMA(KP903989293, T3i, KP427555093 * T3l); | |
1376 T3x = FNMS(KP427555093, T3i, KP903989293 * T3l); | |
1377 T3A = FMA(KP773010453, T3c, KP634393284 * T3b); | |
1378 T3B = FNMS(KP634393284, T39, KP773010453 * T38); | |
1379 T3C = T3A + T3B; | |
1380 T6k = T3B - T3A; | |
1381 } | |
1382 { | |
1383 E T3E, T3F, T3p, T3s; | |
1384 T3E = T3g + T3h; | |
1385 T3F = T3j + T3k; | |
1386 T3G = FMA(KP941544065, T3E, KP336889853 * T3F); | |
1387 T3N = FNMS(KP336889853, T3E, KP941544065 * T3F); | |
1388 T3p = T3n - T3o; | |
1389 T3s = T3q - T3r; | |
1390 T3t = FNMS(KP427555093, T3s, KP903989293 * T3p); | |
1391 T3w = FMA(KP427555093, T3p, KP903989293 * T3s); | |
1392 } | |
1393 { | |
1394 E T3f, T3u, T6p, T6q; | |
1395 T3f = T37 + T3e; | |
1396 T3u = T3m + T3t; | |
1397 Cr[WS(csr, 27)] = T3f - T3u; | |
1398 Cr[WS(csr, 4)] = T3f + T3u; | |
1399 T6p = T3x + T3w; | |
1400 T6q = T6k + T6n; | |
1401 Ci[WS(csi, 27)] = T6p - T6q; | |
1402 Ci[WS(csi, 4)] = T6p + T6q; | |
1403 } | |
1404 { | |
1405 E T3v, T3y, T6j, T6o; | |
1406 T3v = T37 - T3e; | |
1407 T3y = T3w - T3x; | |
1408 Cr[WS(csr, 20)] = T3v - T3y; | |
1409 Cr[WS(csr, 11)] = T3v + T3y; | |
1410 T6j = T3t - T3m; | |
1411 T6o = T6k - T6n; | |
1412 Ci[WS(csi, 20)] = T6j - T6o; | |
1413 Ci[WS(csi, 11)] = T6j + T6o; | |
1414 } | |
1415 { | |
1416 E T3D, T3K, T6v, T6w; | |
1417 T3D = T3z + T3C; | |
1418 T3K = T3G + T3J; | |
1419 Cr[WS(csr, 28)] = T3D - T3K; | |
1420 Cr[WS(csr, 3)] = T3D + T3K; | |
1421 T6v = T3N + T3M; | |
1422 T6w = T6s + T6t; | |
1423 Ci[WS(csi, 28)] = T6v - T6w; | |
1424 Ci[WS(csi, 3)] = T6v + T6w; | |
1425 } | |
1426 { | |
1427 E T3L, T3O, T6r, T6u; | |
1428 T3L = T3z - T3C; | |
1429 T3O = T3M - T3N; | |
1430 Cr[WS(csr, 19)] = T3L - T3O; | |
1431 Cr[WS(csr, 12)] = T3L + T3O; | |
1432 T6r = T3J - T3G; | |
1433 T6u = T6s - T6t; | |
1434 Ci[WS(csi, 19)] = T6r - T6u; | |
1435 Ci[WS(csi, 12)] = T6r + T6u; | |
1436 } | |
1437 } | |
1438 { | |
1439 E T41, T4Z, T6D, T6J, T4g, T6I, T59, T5d, T4A, T4X, T52, T6y, T56, T5c, T4T; | |
1440 E T4W, T40, T6C; | |
1441 T40 = T3W - T3Z; | |
1442 T41 = T3T + T40; | |
1443 T4Z = T3T - T40; | |
1444 T6C = T5g + T5h; | |
1445 T6D = T6B - T6C; | |
1446 T6J = T6C + T6B; | |
1447 { | |
1448 E T48, T4f, T57, T58; | |
1449 T48 = FMA(KP881921264, T44, KP471396736 * T47); | |
1450 T4f = FMA(KP881921264, T4b, KP471396736 * T4e); | |
1451 T4g = T48 - T4f; | |
1452 T6I = T48 + T4f; | |
1453 T57 = T4K + T4H; | |
1454 T58 = T4R + T4O; | |
1455 T59 = FMA(KP514102744, T57, KP857728610 * T58); | |
1456 T5d = FNMS(KP857728610, T57, KP514102744 * T58); | |
1457 } | |
1458 { | |
1459 E T4s, T4z, T50, T51; | |
1460 T4s = T4k + T4r; | |
1461 T4z = T4v - T4y; | |
1462 T4A = FMA(KP970031253, T4s, KP242980179 * T4z); | |
1463 T4X = FNMS(KP242980179, T4s, KP970031253 * T4z); | |
1464 T50 = FNMS(KP471396736, T4b, KP881921264 * T4e); | |
1465 T51 = FNMS(KP471396736, T44, KP881921264 * T47); | |
1466 T52 = T50 - T51; | |
1467 T6y = T51 + T50; | |
1468 } | |
1469 { | |
1470 E T54, T55, T4L, T4S; | |
1471 T54 = T4k - T4r; | |
1472 T55 = T4y + T4v; | |
1473 T56 = FMA(KP514102744, T54, KP857728610 * T55); | |
1474 T5c = FNMS(KP514102744, T55, KP857728610 * T54); | |
1475 T4L = T4H - T4K; | |
1476 T4S = T4O - T4R; | |
1477 T4T = FNMS(KP242980179, T4S, KP970031253 * T4L); | |
1478 T4W = FMA(KP242980179, T4L, KP970031253 * T4S); | |
1479 } | |
1480 { | |
1481 E T4h, T4U, T6F, T6G; | |
1482 T4h = T41 + T4g; | |
1483 T4U = T4A + T4T; | |
1484 Cr[WS(csr, 29)] = T4h - T4U; | |
1485 Cr[WS(csr, 2)] = T4h + T4U; | |
1486 T6F = T4X + T4W; | |
1487 T6G = T6y + T6D; | |
1488 Ci[WS(csi, 29)] = T6F - T6G; | |
1489 Ci[WS(csi, 2)] = T6F + T6G; | |
1490 } | |
1491 { | |
1492 E T4V, T4Y, T6x, T6E; | |
1493 T4V = T41 - T4g; | |
1494 T4Y = T4W - T4X; | |
1495 Cr[WS(csr, 18)] = T4V - T4Y; | |
1496 Cr[WS(csr, 13)] = T4V + T4Y; | |
1497 T6x = T4T - T4A; | |
1498 T6E = T6y - T6D; | |
1499 Ci[WS(csi, 18)] = T6x - T6E; | |
1500 Ci[WS(csi, 13)] = T6x + T6E; | |
1501 } | |
1502 { | |
1503 E T53, T5a, T6L, T6M; | |
1504 T53 = T4Z - T52; | |
1505 T5a = T56 - T59; | |
1506 Cr[WS(csr, 21)] = T53 - T5a; | |
1507 Cr[WS(csr, 10)] = T53 + T5a; | |
1508 T6L = T5d - T5c; | |
1509 T6M = T6J - T6I; | |
1510 Ci[WS(csi, 21)] = T6L - T6M; | |
1511 Ci[WS(csi, 10)] = T6L + T6M; | |
1512 } | |
1513 { | |
1514 E T5b, T5e, T6H, T6K; | |
1515 T5b = T4Z + T52; | |
1516 T5e = T5c + T5d; | |
1517 Cr[WS(csr, 26)] = T5b - T5e; | |
1518 Cr[WS(csr, 5)] = T5b + T5e; | |
1519 T6H = T56 + T59; | |
1520 T6K = T6I + T6J; | |
1521 Ci[WS(csi, 5)] = -(T6H + T6K); | |
1522 Ci[WS(csi, 26)] = T6K - T6H; | |
1523 } | |
1524 } | |
1525 } | |
1526 } | |
1527 } | |
1528 | |
1529 static const kr2c_desc desc = { 64, "r2cfII_64", {342, 114, 92, 0}, &GENUS }; | |
1530 | |
1531 void X(codelet_r2cfII_64) (planner *p) { | |
1532 X(kr2c_register) (p, r2cfII_64, &desc); | |
1533 } | |
1534 | |
1535 #endif /* HAVE_FMA */ |