Mercurial > hg > js-dsp-test
comparison fft/fftw/fftw-3.3.4/rdft/scalar/r2cb/r2cb_64.c @ 19:26056e866c29
Add FFTW to comparison table
author | Chris Cannam |
---|---|
date | Tue, 06 Oct 2015 13:08:39 +0100 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
18:8db794ca3e0b | 19:26056e866c29 |
---|---|
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 Tue Mar 4 13:50:24 EST 2014 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cb_64 -include r2cb.h */ | |
29 | |
30 /* | |
31 * This function contains 394 FP additions, 216 FP multiplications, | |
32 * (or, 178 additions, 0 multiplications, 216 fused multiply/add), | |
33 * 143 stack variables, 18 constants, and 128 memory accesses | |
34 */ | |
35 #include "r2cb.h" | |
36 | |
37 static void r2cb_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737); | |
40 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586); | |
41 DK(KP098491403, +0.098491403357164253077197521291327432293052451); | |
42 DK(KP820678790, +0.820678790828660330972281985331011598767386482); | |
43 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411); | |
44 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241); | |
45 DK(KP303346683, +0.303346683607342391675883946941299872384187453); | |
46 DK(KP534511135, +0.534511135950791641089685961295362908582039528); | |
47 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
48 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624); | |
49 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462); | |
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
53 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252); | |
54 DK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
55 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875); | |
56 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
57 { | |
58 INT i; | |
59 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) { | |
60 E T3d, T32, T37, T2Z, T3f, T3b, T3c, T35; | |
61 { | |
62 E T5H, T9, T5j, T4p, T2T, T1b, T3Z, T3j, Tg, T5I, T5k, T4u, T40, T3m, T2U; | |
63 E T1m, T3o, T1s, T1J, T3r, T5K, Tw, T5N, T6c, T4A, T5n, T3s, T1D, T5m, T4F; | |
64 E T3p, T1M, T3w, T1U, T2z, T3H, T5Q, TM, T6f, T5Y, T5q, T4M, T3I, T25, T5t; | |
65 E T53, T3x, T2C, T3A, T5V, T11, T6g, T5T, T55, T4W, T3z, T2E, T2h, T2F, T2s; | |
66 E T3L, T3E, T54, T4R; | |
67 { | |
68 E Td, T1c, Tc, T4r, T1k, Te, T1d, T1e; | |
69 { | |
70 E T3h, T15, T1a, T3i; | |
71 { | |
72 E T4, T14, T17, T13, T3, T16, T8, T18; | |
73 T4 = Cr[WS(csr, 16)]; | |
74 T14 = Ci[WS(csi, 16)]; | |
75 { | |
76 E T1, T2, T6, T7; | |
77 T1 = Cr[0]; | |
78 T2 = Cr[WS(csr, 32)]; | |
79 T6 = Cr[WS(csr, 8)]; | |
80 T7 = Cr[WS(csr, 24)]; | |
81 T17 = Ci[WS(csi, 8)]; | |
82 T13 = T1 - T2; | |
83 T3 = T1 + T2; | |
84 T16 = T6 - T7; | |
85 T8 = T6 + T7; | |
86 T18 = Ci[WS(csi, 24)]; | |
87 } | |
88 { | |
89 E T4n, T5, T4o, T19; | |
90 T4n = FNMS(KP2_000000000, T4, T3); | |
91 T5 = FMA(KP2_000000000, T4, T3); | |
92 T3h = FMA(KP2_000000000, T14, T13); | |
93 T15 = FNMS(KP2_000000000, T14, T13); | |
94 T4o = T17 - T18; | |
95 T19 = T17 + T18; | |
96 T5H = FNMS(KP2_000000000, T8, T5); | |
97 T9 = FMA(KP2_000000000, T8, T5); | |
98 T5j = FMA(KP2_000000000, T4o, T4n); | |
99 T4p = FNMS(KP2_000000000, T4o, T4n); | |
100 T1a = T16 - T19; | |
101 T3i = T16 + T19; | |
102 } | |
103 } | |
104 { | |
105 E Ta, Tb, T1i, T1j; | |
106 Ta = Cr[WS(csr, 4)]; | |
107 T2T = FNMS(KP1_414213562, T1a, T15); | |
108 T1b = FMA(KP1_414213562, T1a, T15); | |
109 T3Z = FMA(KP1_414213562, T3i, T3h); | |
110 T3j = FNMS(KP1_414213562, T3i, T3h); | |
111 Tb = Cr[WS(csr, 28)]; | |
112 T1i = Ci[WS(csi, 4)]; | |
113 T1j = Ci[WS(csi, 28)]; | |
114 Td = Cr[WS(csr, 20)]; | |
115 T1c = Ta - Tb; | |
116 Tc = Ta + Tb; | |
117 T4r = T1i - T1j; | |
118 T1k = T1i + T1j; | |
119 Te = Cr[WS(csr, 12)]; | |
120 T1d = Ci[WS(csi, 20)]; | |
121 T1e = Ci[WS(csi, 12)]; | |
122 } | |
123 } | |
124 { | |
125 E T4B, T4E, T1K, T1L; | |
126 { | |
127 E T1o, Tk, T4C, T1I, T1F, Tn, T4D, T1r, Ts, T1t, Tr, T4y, T1w, Tt, T1z; | |
128 E T1A; | |
129 { | |
130 E Tl, Tm, T1p, T1q; | |
131 { | |
132 E Ti, Tj, T1G, T1H, T1h, Tf; | |
133 Ti = Cr[WS(csr, 2)]; | |
134 T1h = Td - Te; | |
135 Tf = Td + Te; | |
136 { | |
137 E T4s, T1f, T3k, T1l; | |
138 T4s = T1d - T1e; | |
139 T1f = T1d + T1e; | |
140 T3k = T1k - T1h; | |
141 T1l = T1h + T1k; | |
142 { | |
143 E T4q, T4t, T3l, T1g; | |
144 T4q = Tc - Tf; | |
145 Tg = Tc + Tf; | |
146 T4t = T4r - T4s; | |
147 T5I = T4s + T4r; | |
148 T3l = T1c + T1f; | |
149 T1g = T1c - T1f; | |
150 T5k = T4q + T4t; | |
151 T4u = T4q - T4t; | |
152 T40 = FMA(KP414213562, T3k, T3l); | |
153 T3m = FNMS(KP414213562, T3l, T3k); | |
154 T2U = FMA(KP414213562, T1g, T1l); | |
155 T1m = FNMS(KP414213562, T1l, T1g); | |
156 Tj = Cr[WS(csr, 30)]; | |
157 } | |
158 } | |
159 T1G = Ci[WS(csi, 2)]; | |
160 T1H = Ci[WS(csi, 30)]; | |
161 Tl = Cr[WS(csr, 18)]; | |
162 T1o = Ti - Tj; | |
163 Tk = Ti + Tj; | |
164 T4C = T1G - T1H; | |
165 T1I = T1G + T1H; | |
166 Tm = Cr[WS(csr, 14)]; | |
167 T1p = Ci[WS(csi, 18)]; | |
168 T1q = Ci[WS(csi, 14)]; | |
169 } | |
170 { | |
171 E Tp, Tq, T1u, T1v; | |
172 Tp = Cr[WS(csr, 10)]; | |
173 T1F = Tl - Tm; | |
174 Tn = Tl + Tm; | |
175 T4D = T1p - T1q; | |
176 T1r = T1p + T1q; | |
177 Tq = Cr[WS(csr, 22)]; | |
178 T1u = Ci[WS(csi, 10)]; | |
179 T1v = Ci[WS(csi, 22)]; | |
180 Ts = Cr[WS(csr, 6)]; | |
181 T1t = Tp - Tq; | |
182 Tr = Tp + Tq; | |
183 T4y = T1u - T1v; | |
184 T1w = T1u + T1v; | |
185 Tt = Cr[WS(csr, 26)]; | |
186 T1z = Ci[WS(csi, 6)]; | |
187 T1A = Ci[WS(csi, 26)]; | |
188 } | |
189 } | |
190 { | |
191 E T1y, T4x, T1B, T4w, To, Tv, Tu; | |
192 T3o = T1o + T1r; | |
193 T1s = T1o - T1r; | |
194 T1y = Ts - Tt; | |
195 Tu = Ts + Tt; | |
196 T4x = T1A - T1z; | |
197 T1B = T1z + T1A; | |
198 T1J = T1F + T1I; | |
199 T3r = T1I - T1F; | |
200 T4w = Tk - Tn; | |
201 To = Tk + Tn; | |
202 Tv = Tr + Tu; | |
203 T4B = Tr - Tu; | |
204 { | |
205 E T4z, T5L, T5M, T1x, T1C; | |
206 T4E = T4C - T4D; | |
207 T5L = T4D + T4C; | |
208 T5M = T4y + T4x; | |
209 T4z = T4x - T4y; | |
210 T5K = To - Tv; | |
211 Tw = To + Tv; | |
212 T5N = T5L - T5M; | |
213 T6c = T5M + T5L; | |
214 T1K = T1t + T1w; | |
215 T1x = T1t - T1w; | |
216 T1C = T1y - T1B; | |
217 T1L = T1y + T1B; | |
218 T4A = T4w + T4z; | |
219 T5n = T4w - T4z; | |
220 T3s = T1C - T1x; | |
221 T1D = T1x + T1C; | |
222 } | |
223 } | |
224 } | |
225 { | |
226 E T4Z, T52, T2A, T2B; | |
227 { | |
228 E T1Q, TA, T50, T2y, T2v, TD, T51, T1T, TI, T1V, TH, T4K, T1Y, TJ, T21; | |
229 E T22; | |
230 { | |
231 E TB, TC, T1R, T1S; | |
232 { | |
233 E Ty, Tz, T2w, T2x; | |
234 Ty = Cr[WS(csr, 1)]; | |
235 T5m = T4E - T4B; | |
236 T4F = T4B + T4E; | |
237 T3p = T1K + T1L; | |
238 T1M = T1K - T1L; | |
239 Tz = Cr[WS(csr, 31)]; | |
240 T2w = Ci[WS(csi, 1)]; | |
241 T2x = Ci[WS(csi, 31)]; | |
242 TB = Cr[WS(csr, 17)]; | |
243 T1Q = Ty - Tz; | |
244 TA = Ty + Tz; | |
245 T50 = T2w - T2x; | |
246 T2y = T2w + T2x; | |
247 TC = Cr[WS(csr, 15)]; | |
248 T1R = Ci[WS(csi, 17)]; | |
249 T1S = Ci[WS(csi, 15)]; | |
250 } | |
251 { | |
252 E TF, TG, T1W, T1X; | |
253 TF = Cr[WS(csr, 9)]; | |
254 T2v = TB - TC; | |
255 TD = TB + TC; | |
256 T51 = T1R - T1S; | |
257 T1T = T1R + T1S; | |
258 TG = Cr[WS(csr, 23)]; | |
259 T1W = Ci[WS(csi, 9)]; | |
260 T1X = Ci[WS(csi, 23)]; | |
261 TI = Cr[WS(csr, 7)]; | |
262 T1V = TF - TG; | |
263 TH = TF + TG; | |
264 T4K = T1W - T1X; | |
265 T1Y = T1W + T1X; | |
266 TJ = Cr[WS(csr, 25)]; | |
267 T21 = Ci[WS(csi, 7)]; | |
268 T22 = Ci[WS(csi, 25)]; | |
269 } | |
270 } | |
271 { | |
272 E T20, T4J, T23, T4I, TE, TL, TK; | |
273 T3w = T1Q + T1T; | |
274 T1U = T1Q - T1T; | |
275 T20 = TI - TJ; | |
276 TK = TI + TJ; | |
277 T4J = T22 - T21; | |
278 T23 = T21 + T22; | |
279 T2z = T2v + T2y; | |
280 T3H = T2y - T2v; | |
281 T4I = TA - TD; | |
282 TE = TA + TD; | |
283 TL = TH + TK; | |
284 T4Z = TH - TK; | |
285 { | |
286 E T4L, T5W, T5X, T1Z, T24; | |
287 T52 = T50 - T51; | |
288 T5W = T51 + T50; | |
289 T5X = T4K + T4J; | |
290 T4L = T4J - T4K; | |
291 T5Q = TE - TL; | |
292 TM = TE + TL; | |
293 T6f = T5X + T5W; | |
294 T5Y = T5W - T5X; | |
295 T2A = T1V + T1Y; | |
296 T1Z = T1V - T1Y; | |
297 T24 = T20 - T23; | |
298 T2B = T20 + T23; | |
299 T5q = T4I - T4L; | |
300 T4M = T4I + T4L; | |
301 T3I = T24 - T1Z; | |
302 T25 = T1Z + T24; | |
303 } | |
304 } | |
305 } | |
306 { | |
307 E T27, TP, T4O, T2f, T2c, TS, T4P, T2a, TX, T2i, TW, T4T, T2q, TY, T2j; | |
308 E T2k; | |
309 { | |
310 E TQ, TR, T28, T29; | |
311 { | |
312 E TN, TO, T2d, T2e; | |
313 TN = Cr[WS(csr, 5)]; | |
314 T5t = T52 - T4Z; | |
315 T53 = T4Z + T52; | |
316 T3x = T2A + T2B; | |
317 T2C = T2A - T2B; | |
318 TO = Cr[WS(csr, 27)]; | |
319 T2d = Ci[WS(csi, 5)]; | |
320 T2e = Ci[WS(csi, 27)]; | |
321 TQ = Cr[WS(csr, 21)]; | |
322 T27 = TN - TO; | |
323 TP = TN + TO; | |
324 T4O = T2d - T2e; | |
325 T2f = T2d + T2e; | |
326 TR = Cr[WS(csr, 11)]; | |
327 T28 = Ci[WS(csi, 21)]; | |
328 T29 = Ci[WS(csi, 11)]; | |
329 } | |
330 { | |
331 E TU, TV, T2o, T2p; | |
332 TU = Cr[WS(csr, 3)]; | |
333 T2c = TQ - TR; | |
334 TS = TQ + TR; | |
335 T4P = T28 - T29; | |
336 T2a = T28 + T29; | |
337 TV = Cr[WS(csr, 29)]; | |
338 T2o = Ci[WS(csi, 3)]; | |
339 T2p = Ci[WS(csi, 29)]; | |
340 TX = Cr[WS(csr, 13)]; | |
341 T2i = TU - TV; | |
342 TW = TU + TV; | |
343 T4T = T2p - T2o; | |
344 T2q = T2o + T2p; | |
345 TY = Cr[WS(csr, 19)]; | |
346 T2j = Ci[WS(csi, 13)]; | |
347 T2k = Ci[WS(csi, 19)]; | |
348 } | |
349 } | |
350 { | |
351 E T4N, T2n, T2l, T4Q, T2b, T2g, TT, TZ, T4U; | |
352 T4N = TP - TS; | |
353 TT = TP + TS; | |
354 T2n = TX - TY; | |
355 TZ = TX + TY; | |
356 T4U = T2j - T2k; | |
357 T2l = T2j + T2k; | |
358 { | |
359 E T5S, T10, T4S, T4V, T5R; | |
360 T5S = T4P + T4O; | |
361 T4Q = T4O - T4P; | |
362 T10 = TW + TZ; | |
363 T4S = TW - TZ; | |
364 T4V = T4T - T4U; | |
365 T5R = T4U + T4T; | |
366 T3A = T27 + T2a; | |
367 T2b = T27 - T2a; | |
368 T5V = TT - T10; | |
369 T11 = TT + T10; | |
370 T6g = T5S + T5R; | |
371 T5T = T5R - T5S; | |
372 T55 = T4V - T4S; | |
373 T4W = T4S + T4V; | |
374 T2g = T2c + T2f; | |
375 T3z = T2f - T2c; | |
376 } | |
377 { | |
378 E T3D, T3C, T2m, T2r; | |
379 T3D = T2i + T2l; | |
380 T2m = T2i - T2l; | |
381 T2r = T2n - T2q; | |
382 T3C = T2n + T2q; | |
383 T2E = FMA(KP414213562, T2b, T2g); | |
384 T2h = FNMS(KP414213562, T2g, T2b); | |
385 T2F = FNMS(KP414213562, T2m, T2r); | |
386 T2s = FMA(KP414213562, T2r, T2m); | |
387 T3L = FMA(KP414213562, T3C, T3D); | |
388 T3E = FNMS(KP414213562, T3D, T3C); | |
389 T54 = T4N + T4Q; | |
390 T4R = T4N - T4Q; | |
391 } | |
392 } | |
393 } | |
394 } | |
395 } | |
396 } | |
397 { | |
398 E T3K, T3B, T5u, T5r, T5d, T5g; | |
399 { | |
400 E T6e, T6h, T6b, T5J, T5O, T5Z, T66, T69, T65, T67, T5U, T12, T6m, Th; | |
401 T6e = TM - T11; | |
402 T12 = TM + T11; | |
403 T6m = T6g + T6f; | |
404 T6h = T6f - T6g; | |
405 T6b = FNMS(KP2_000000000, Tg, T9); | |
406 Th = FMA(KP2_000000000, Tg, T9); | |
407 T3K = FMA(KP414213562, T3z, T3A); | |
408 T3B = FNMS(KP414213562, T3A, T3z); | |
409 { | |
410 E T63, T64, T6l, Tx; | |
411 T5J = FNMS(KP2_000000000, T5I, T5H); | |
412 T63 = FMA(KP2_000000000, T5I, T5H); | |
413 T64 = T5K + T5N; | |
414 T5O = T5K - T5N; | |
415 T5Z = T5V + T5Y; | |
416 T66 = T5Y - T5V; | |
417 T6l = FNMS(KP2_000000000, Tw, Th); | |
418 Tx = FMA(KP2_000000000, Tw, Th); | |
419 T69 = FMA(KP1_414213562, T64, T63); | |
420 T65 = FNMS(KP1_414213562, T64, T63); | |
421 R0[WS(rs, 8)] = FNMS(KP2_000000000, T6m, T6l); | |
422 R0[WS(rs, 24)] = FMA(KP2_000000000, T6m, T6l); | |
423 R0[0] = FMA(KP2_000000000, T12, Tx); | |
424 R0[WS(rs, 16)] = FNMS(KP2_000000000, T12, Tx); | |
425 T67 = T5Q - T5T; | |
426 T5U = T5Q + T5T; | |
427 } | |
428 { | |
429 E T6j, T6d, T6a, T68; | |
430 T6a = FMA(KP414213562, T66, T67); | |
431 T68 = FNMS(KP414213562, T67, T66); | |
432 T6j = FMA(KP2_000000000, T6c, T6b); | |
433 T6d = FNMS(KP2_000000000, T6c, T6b); | |
434 R0[WS(rs, 14)] = FNMS(KP1_847759065, T6a, T69); | |
435 R0[WS(rs, 30)] = FMA(KP1_847759065, T6a, T69); | |
436 R0[WS(rs, 22)] = FMA(KP1_847759065, T68, T65); | |
437 R0[WS(rs, 6)] = FNMS(KP1_847759065, T68, T65); | |
438 { | |
439 E T61, T5P, T6k, T6i; | |
440 T6k = T6e + T6h; | |
441 T6i = T6e - T6h; | |
442 T61 = FNMS(KP1_414213562, T5O, T5J); | |
443 T5P = FMA(KP1_414213562, T5O, T5J); | |
444 R0[WS(rs, 12)] = FNMS(KP1_414213562, T6k, T6j); | |
445 R0[WS(rs, 28)] = FMA(KP1_414213562, T6k, T6j); | |
446 R0[WS(rs, 4)] = FMA(KP1_414213562, T6i, T6d); | |
447 R0[WS(rs, 20)] = FNMS(KP1_414213562, T6i, T6d); | |
448 { | |
449 E T5b, T4v, T5f, T4Y, T5e, T57, T4G, T5c; | |
450 { | |
451 E T4X, T56, T62, T60; | |
452 T5u = T4W - T4R; | |
453 T4X = T4R + T4W; | |
454 T56 = T54 + T55; | |
455 T5r = T54 - T55; | |
456 T5b = FNMS(KP1_414213562, T4u, T4p); | |
457 T4v = FMA(KP1_414213562, T4u, T4p); | |
458 T62 = FMA(KP414213562, T5U, T5Z); | |
459 T60 = FNMS(KP414213562, T5Z, T5U); | |
460 T5f = FNMS(KP707106781, T4X, T4M); | |
461 T4Y = FMA(KP707106781, T4X, T4M); | |
462 T5e = FNMS(KP707106781, T56, T53); | |
463 T57 = FMA(KP707106781, T56, T53); | |
464 R0[WS(rs, 10)] = FNMS(KP1_847759065, T62, T61); | |
465 R0[WS(rs, 26)] = FMA(KP1_847759065, T62, T61); | |
466 R0[WS(rs, 2)] = FMA(KP1_847759065, T60, T5P); | |
467 R0[WS(rs, 18)] = FNMS(KP1_847759065, T60, T5P); | |
468 T4G = FNMS(KP414213562, T4F, T4A); | |
469 T5c = FMA(KP414213562, T4A, T4F); | |
470 } | |
471 { | |
472 E T5a, T59, T5h, T5i, T58, T4H; | |
473 T5a = FMA(KP198912367, T4Y, T57); | |
474 T58 = FNMS(KP198912367, T57, T4Y); | |
475 T59 = FNMS(KP1_847759065, T4G, T4v); | |
476 T4H = FMA(KP1_847759065, T4G, T4v); | |
477 T5h = FMA(KP1_847759065, T5c, T5b); | |
478 T5d = FNMS(KP1_847759065, T5c, T5b); | |
479 T5i = FMA(KP668178637, T5e, T5f); | |
480 T5g = FNMS(KP668178637, T5f, T5e); | |
481 R0[WS(rs, 1)] = FMA(KP1_961570560, T58, T4H); | |
482 R0[WS(rs, 17)] = FNMS(KP1_961570560, T58, T4H); | |
483 R0[WS(rs, 29)] = FMA(KP1_662939224, T5i, T5h); | |
484 R0[WS(rs, 13)] = FNMS(KP1_662939224, T5i, T5h); | |
485 R0[WS(rs, 25)] = FMA(KP1_961570560, T5a, T59); | |
486 R0[WS(rs, 9)] = FNMS(KP1_961570560, T5a, T59); | |
487 } | |
488 } | |
489 } | |
490 } | |
491 } | |
492 { | |
493 E T43, T42, T46, T4a, T49, T3V, T3G, T47, T3P, T3v, T3X, T3T, T3U, T3N, T5B; | |
494 E T5E; | |
495 { | |
496 E T5s, T5D, T5z, T5l, T5C, T5v, T5o, T5A; | |
497 R0[WS(rs, 21)] = FMA(KP1_662939224, T5g, T5d); | |
498 R0[WS(rs, 5)] = FNMS(KP1_662939224, T5g, T5d); | |
499 T5s = FNMS(KP707106781, T5r, T5q); | |
500 T5D = FMA(KP707106781, T5r, T5q); | |
501 T5z = FMA(KP1_414213562, T5k, T5j); | |
502 T5l = FNMS(KP1_414213562, T5k, T5j); | |
503 T5C = FMA(KP707106781, T5u, T5t); | |
504 T5v = FNMS(KP707106781, T5u, T5t); | |
505 T5o = FNMS(KP414213562, T5n, T5m); | |
506 T5A = FMA(KP414213562, T5m, T5n); | |
507 { | |
508 E T5y, T5x, T5F, T5G, T5w, T5p; | |
509 T5y = FMA(KP668178637, T5s, T5v); | |
510 T5w = FNMS(KP668178637, T5v, T5s); | |
511 T5x = FMA(KP1_847759065, T5o, T5l); | |
512 T5p = FNMS(KP1_847759065, T5o, T5l); | |
513 T5F = FMA(KP1_847759065, T5A, T5z); | |
514 T5B = FNMS(KP1_847759065, T5A, T5z); | |
515 T5G = FMA(KP198912367, T5C, T5D); | |
516 T5E = FNMS(KP198912367, T5D, T5C); | |
517 R0[WS(rs, 3)] = FMA(KP1_662939224, T5w, T5p); | |
518 R0[WS(rs, 19)] = FNMS(KP1_662939224, T5w, T5p); | |
519 R0[WS(rs, 31)] = FMA(KP1_961570560, T5G, T5F); | |
520 R0[WS(rs, 15)] = FNMS(KP1_961570560, T5G, T5F); | |
521 R0[WS(rs, 27)] = FMA(KP1_662939224, T5y, T5x); | |
522 R0[WS(rs, 11)] = FNMS(KP1_662939224, T5y, T5x); | |
523 } | |
524 } | |
525 { | |
526 E T3R, T3n, T3J, T3S, T3u, T3M; | |
527 T3R = FMA(KP1_847759065, T3m, T3j); | |
528 T3n = FNMS(KP1_847759065, T3m, T3j); | |
529 R0[WS(rs, 23)] = FMA(KP1_961570560, T5E, T5B); | |
530 R0[WS(rs, 7)] = FNMS(KP1_961570560, T5E, T5B); | |
531 { | |
532 E T3q, T3t, T3y, T3F; | |
533 T43 = FMA(KP707106781, T3p, T3o); | |
534 T3q = FNMS(KP707106781, T3p, T3o); | |
535 T3t = FNMS(KP707106781, T3s, T3r); | |
536 T42 = FMA(KP707106781, T3s, T3r); | |
537 T46 = FMA(KP707106781, T3x, T3w); | |
538 T3y = FNMS(KP707106781, T3x, T3w); | |
539 T3F = T3B + T3E; | |
540 T4a = T3B - T3E; | |
541 T49 = FMA(KP707106781, T3I, T3H); | |
542 T3J = FNMS(KP707106781, T3I, T3H); | |
543 T3S = FMA(KP668178637, T3q, T3t); | |
544 T3u = FNMS(KP668178637, T3t, T3q); | |
545 T3V = FMA(KP923879532, T3F, T3y); | |
546 T3G = FNMS(KP923879532, T3F, T3y); | |
547 T3M = T3K - T3L; | |
548 T47 = T3K + T3L; | |
549 } | |
550 T3P = FNMS(KP1_662939224, T3u, T3n); | |
551 T3v = FMA(KP1_662939224, T3u, T3n); | |
552 T3X = FMA(KP1_662939224, T3S, T3R); | |
553 T3T = FNMS(KP1_662939224, T3S, T3R); | |
554 T3U = FNMS(KP923879532, T3M, T3J); | |
555 T3N = FMA(KP923879532, T3M, T3J); | |
556 } | |
557 { | |
558 E T2X, T2W, T30, T34, T33, T2P, T2u, T31, T2J, T1P, T2R, T2N, T2O, T2H; | |
559 { | |
560 E T2L, T1n, T2D, T2M, T1O, T2G; | |
561 T2L = FNMS(KP1_847759065, T1m, T1b); | |
562 T1n = FMA(KP1_847759065, T1m, T1b); | |
563 { | |
564 E T3W, T3Y, T3Q, T3O; | |
565 T3W = FNMS(KP534511135, T3V, T3U); | |
566 T3Y = FMA(KP534511135, T3U, T3V); | |
567 T3Q = FMA(KP303346683, T3G, T3N); | |
568 T3O = FNMS(KP303346683, T3N, T3G); | |
569 R1[WS(rs, 21)] = FMA(KP1_763842528, T3W, T3T); | |
570 R1[WS(rs, 5)] = FNMS(KP1_763842528, T3W, T3T); | |
571 R1[WS(rs, 29)] = FMA(KP1_763842528, T3Y, T3X); | |
572 R1[WS(rs, 13)] = FNMS(KP1_763842528, T3Y, T3X); | |
573 R1[WS(rs, 25)] = FMA(KP1_913880671, T3Q, T3P); | |
574 R1[WS(rs, 9)] = FNMS(KP1_913880671, T3Q, T3P); | |
575 R1[WS(rs, 1)] = FMA(KP1_913880671, T3O, T3v); | |
576 R1[WS(rs, 17)] = FNMS(KP1_913880671, T3O, T3v); | |
577 } | |
578 { | |
579 E T1E, T1N, T26, T2t; | |
580 T2X = FNMS(KP707106781, T1D, T1s); | |
581 T1E = FMA(KP707106781, T1D, T1s); | |
582 T1N = FMA(KP707106781, T1M, T1J); | |
583 T2W = FNMS(KP707106781, T1M, T1J); | |
584 T30 = FNMS(KP707106781, T25, T1U); | |
585 T26 = FMA(KP707106781, T25, T1U); | |
586 T2t = T2h + T2s; | |
587 T34 = T2s - T2h; | |
588 T33 = FNMS(KP707106781, T2C, T2z); | |
589 T2D = FMA(KP707106781, T2C, T2z); | |
590 T2M = FMA(KP198912367, T1E, T1N); | |
591 T1O = FNMS(KP198912367, T1N, T1E); | |
592 T2P = FNMS(KP923879532, T2t, T26); | |
593 T2u = FMA(KP923879532, T2t, T26); | |
594 T2G = T2E + T2F; | |
595 T31 = T2E - T2F; | |
596 } | |
597 T2J = FNMS(KP1_961570560, T1O, T1n); | |
598 T1P = FMA(KP1_961570560, T1O, T1n); | |
599 T2R = FMA(KP1_961570560, T2M, T2L); | |
600 T2N = FNMS(KP1_961570560, T2M, T2L); | |
601 T2O = FNMS(KP923879532, T2G, T2D); | |
602 T2H = FMA(KP923879532, T2G, T2D); | |
603 } | |
604 { | |
605 E T4j, T48, T4d, T45, T4l, T4h, T4i, T4b; | |
606 { | |
607 E T4f, T41, T4g, T44; | |
608 T4f = FMA(KP1_847759065, T40, T3Z); | |
609 T41 = FNMS(KP1_847759065, T40, T3Z); | |
610 { | |
611 E T2Q, T2S, T2K, T2I; | |
612 T2Q = FNMS(KP820678790, T2P, T2O); | |
613 T2S = FMA(KP820678790, T2O, T2P); | |
614 T2K = FMA(KP098491403, T2u, T2H); | |
615 T2I = FNMS(KP098491403, T2H, T2u); | |
616 R1[WS(rs, 20)] = FMA(KP1_546020906, T2Q, T2N); | |
617 R1[WS(rs, 4)] = FNMS(KP1_546020906, T2Q, T2N); | |
618 R1[WS(rs, 28)] = FMA(KP1_546020906, T2S, T2R); | |
619 R1[WS(rs, 12)] = FNMS(KP1_546020906, T2S, T2R); | |
620 R1[WS(rs, 24)] = FMA(KP1_990369453, T2K, T2J); | |
621 R1[WS(rs, 8)] = FNMS(KP1_990369453, T2K, T2J); | |
622 R1[0] = FMA(KP1_990369453, T2I, T1P); | |
623 R1[WS(rs, 16)] = FNMS(KP1_990369453, T2I, T1P); | |
624 } | |
625 T4g = FMA(KP198912367, T42, T43); | |
626 T44 = FNMS(KP198912367, T43, T42); | |
627 T4j = FMA(KP923879532, T47, T46); | |
628 T48 = FNMS(KP923879532, T47, T46); | |
629 T4d = FMA(KP1_961570560, T44, T41); | |
630 T45 = FNMS(KP1_961570560, T44, T41); | |
631 T4l = FMA(KP1_961570560, T4g, T4f); | |
632 T4h = FNMS(KP1_961570560, T4g, T4f); | |
633 T4i = FMA(KP923879532, T4a, T49); | |
634 T4b = FNMS(KP923879532, T4a, T49); | |
635 } | |
636 { | |
637 E T39, T2V, T3a, T2Y; | |
638 T39 = FMA(KP1_847759065, T2U, T2T); | |
639 T2V = FNMS(KP1_847759065, T2U, T2T); | |
640 { | |
641 E T4k, T4m, T4e, T4c; | |
642 T4k = FNMS(KP098491403, T4j, T4i); | |
643 T4m = FMA(KP098491403, T4i, T4j); | |
644 T4e = FMA(KP820678790, T48, T4b); | |
645 T4c = FNMS(KP820678790, T4b, T48); | |
646 R1[WS(rs, 23)] = FMA(KP1_990369453, T4k, T4h); | |
647 R1[WS(rs, 7)] = FNMS(KP1_990369453, T4k, T4h); | |
648 R1[WS(rs, 31)] = FMA(KP1_990369453, T4m, T4l); | |
649 R1[WS(rs, 15)] = FNMS(KP1_990369453, T4m, T4l); | |
650 R1[WS(rs, 27)] = FMA(KP1_546020906, T4e, T4d); | |
651 R1[WS(rs, 11)] = FNMS(KP1_546020906, T4e, T4d); | |
652 R1[WS(rs, 3)] = FMA(KP1_546020906, T4c, T45); | |
653 R1[WS(rs, 19)] = FNMS(KP1_546020906, T4c, T45); | |
654 } | |
655 T3a = FMA(KP668178637, T2W, T2X); | |
656 T2Y = FNMS(KP668178637, T2X, T2W); | |
657 T3d = FMA(KP923879532, T31, T30); | |
658 T32 = FNMS(KP923879532, T31, T30); | |
659 T37 = FMA(KP1_662939224, T2Y, T2V); | |
660 T2Z = FNMS(KP1_662939224, T2Y, T2V); | |
661 T3f = FMA(KP1_662939224, T3a, T39); | |
662 T3b = FNMS(KP1_662939224, T3a, T39); | |
663 T3c = FMA(KP923879532, T34, T33); | |
664 T35 = FNMS(KP923879532, T34, T33); | |
665 } | |
666 } | |
667 } | |
668 } | |
669 } | |
670 } | |
671 { | |
672 E T3g, T3e, T36, T38; | |
673 T3g = FMA(KP303346683, T3c, T3d); | |
674 T3e = FNMS(KP303346683, T3d, T3c); | |
675 T36 = FNMS(KP534511135, T35, T32); | |
676 T38 = FMA(KP534511135, T32, T35); | |
677 R1[WS(rs, 22)] = FMA(KP1_913880671, T3e, T3b); | |
678 R1[WS(rs, 6)] = FNMS(KP1_913880671, T3e, T3b); | |
679 R1[WS(rs, 30)] = FMA(KP1_913880671, T3g, T3f); | |
680 R1[WS(rs, 14)] = FNMS(KP1_913880671, T3g, T3f); | |
681 R1[WS(rs, 26)] = FMA(KP1_763842528, T38, T37); | |
682 R1[WS(rs, 10)] = FNMS(KP1_763842528, T38, T37); | |
683 R1[WS(rs, 2)] = FMA(KP1_763842528, T36, T2Z); | |
684 R1[WS(rs, 18)] = FNMS(KP1_763842528, T36, T2Z); | |
685 } | |
686 } | |
687 } | |
688 } | |
689 | |
690 static const kr2c_desc desc = { 64, "r2cb_64", {178, 0, 216, 0}, &GENUS }; | |
691 | |
692 void X(codelet_r2cb_64) (planner *p) { | |
693 X(kr2c_register) (p, r2cb_64, &desc); | |
694 } | |
695 | |
696 #else /* HAVE_FMA */ | |
697 | |
698 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cb_64 -include r2cb.h */ | |
699 | |
700 /* | |
701 * This function contains 394 FP additions, 134 FP multiplications, | |
702 * (or, 342 additions, 82 multiplications, 52 fused multiply/add), | |
703 * 110 stack variables, 19 constants, and 128 memory accesses | |
704 */ | |
705 #include "r2cb.h" | |
706 | |
707 static void r2cb_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
708 { | |
709 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190); | |
710 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586); | |
711 DK(KP196034280, +0.196034280659121203988391127777283691722273346); | |
712 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737); | |
713 DK(KP942793473, +0.942793473651995297112775251810508755314920638); | |
714 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241); | |
715 DK(KP580569354, +0.580569354508924735272384751634790549382952557); | |
716 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411); | |
717 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382); | |
718 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624); | |
719 DK(KP390180644, +0.390180644032256535696569736954044481855383236); | |
720 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462); | |
721 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
722 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
723 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
724 DK(KP765366864, +0.765366864730179543456919968060797733522689125); | |
725 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252); | |
726 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875); | |
727 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
728 { | |
729 INT i; | |
730 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) { | |
731 E Ta, T2S, T18, T2u, T3F, T4V, T5l, T61, Th, T2T, T1h, T2v, T3M, T4W, T5o; | |
732 E T62, T3Q, T5q, T5u, T44, Tp, Tw, T2V, T2W, T2X, T2Y, T3X, T5t, T1r, T2x; | |
733 E T41, T5r, T1A, T2y, T4a, T5y, T5N, T4H, TN, T31, T4E, T5z, T39, T3q, T1L; | |
734 E T2B, T4h, T5M, T2h, T2F, T12, T36, T5D, T5J, T5G, T5K, T1U, T26, T23, T27; | |
735 E T4p, T4z, T4w, T4A, T34, T3r; | |
736 { | |
737 E T5, T3A, T3, T3y, T9, T3C, T17, T3D, T6, T14; | |
738 { | |
739 E T4, T3z, T1, T2; | |
740 T4 = Cr[WS(csr, 16)]; | |
741 T5 = KP2_000000000 * T4; | |
742 T3z = Ci[WS(csi, 16)]; | |
743 T3A = KP2_000000000 * T3z; | |
744 T1 = Cr[0]; | |
745 T2 = Cr[WS(csr, 32)]; | |
746 T3 = T1 + T2; | |
747 T3y = T1 - T2; | |
748 { | |
749 E T7, T8, T15, T16; | |
750 T7 = Cr[WS(csr, 8)]; | |
751 T8 = Cr[WS(csr, 24)]; | |
752 T9 = KP2_000000000 * (T7 + T8); | |
753 T3C = T7 - T8; | |
754 T15 = Ci[WS(csi, 8)]; | |
755 T16 = Ci[WS(csi, 24)]; | |
756 T17 = KP2_000000000 * (T15 - T16); | |
757 T3D = T15 + T16; | |
758 } | |
759 } | |
760 T6 = T3 + T5; | |
761 Ta = T6 + T9; | |
762 T2S = T6 - T9; | |
763 T14 = T3 - T5; | |
764 T18 = T14 - T17; | |
765 T2u = T14 + T17; | |
766 { | |
767 E T3B, T3E, T5j, T5k; | |
768 T3B = T3y - T3A; | |
769 T3E = KP1_414213562 * (T3C - T3D); | |
770 T3F = T3B + T3E; | |
771 T4V = T3B - T3E; | |
772 T5j = T3y + T3A; | |
773 T5k = KP1_414213562 * (T3C + T3D); | |
774 T5l = T5j - T5k; | |
775 T61 = T5j + T5k; | |
776 } | |
777 } | |
778 { | |
779 E Td, T3G, T1c, T3K, Tg, T3J, T1f, T3H, T19, T1g; | |
780 { | |
781 E Tb, Tc, T1a, T1b; | |
782 Tb = Cr[WS(csr, 4)]; | |
783 Tc = Cr[WS(csr, 28)]; | |
784 Td = Tb + Tc; | |
785 T3G = Tb - Tc; | |
786 T1a = Ci[WS(csi, 4)]; | |
787 T1b = Ci[WS(csi, 28)]; | |
788 T1c = T1a - T1b; | |
789 T3K = T1a + T1b; | |
790 } | |
791 { | |
792 E Te, Tf, T1d, T1e; | |
793 Te = Cr[WS(csr, 20)]; | |
794 Tf = Cr[WS(csr, 12)]; | |
795 Tg = Te + Tf; | |
796 T3J = Te - Tf; | |
797 T1d = Ci[WS(csi, 20)]; | |
798 T1e = Ci[WS(csi, 12)]; | |
799 T1f = T1d - T1e; | |
800 T3H = T1d + T1e; | |
801 } | |
802 Th = KP2_000000000 * (Td + Tg); | |
803 T2T = KP2_000000000 * (T1f + T1c); | |
804 T19 = Td - Tg; | |
805 T1g = T1c - T1f; | |
806 T1h = KP1_414213562 * (T19 - T1g); | |
807 T2v = KP1_414213562 * (T19 + T1g); | |
808 { | |
809 E T3I, T3L, T5m, T5n; | |
810 T3I = T3G - T3H; | |
811 T3L = T3J + T3K; | |
812 T3M = FNMS(KP765366864, T3L, KP1_847759065 * T3I); | |
813 T4W = FMA(KP765366864, T3I, KP1_847759065 * T3L); | |
814 T5m = T3G + T3H; | |
815 T5n = T3K - T3J; | |
816 T5o = FNMS(KP1_847759065, T5n, KP765366864 * T5m); | |
817 T62 = FMA(KP1_847759065, T5m, KP765366864 * T5n); | |
818 } | |
819 } | |
820 { | |
821 E Tl, T3O, T1v, T43, To, T42, T1y, T3P, Ts, T3R, T1p, T3S, Tv, T3U, T1m; | |
822 E T3V; | |
823 { | |
824 E Tj, Tk, T1t, T1u; | |
825 Tj = Cr[WS(csr, 2)]; | |
826 Tk = Cr[WS(csr, 30)]; | |
827 Tl = Tj + Tk; | |
828 T3O = Tj - Tk; | |
829 T1t = Ci[WS(csi, 2)]; | |
830 T1u = Ci[WS(csi, 30)]; | |
831 T1v = T1t - T1u; | |
832 T43 = T1t + T1u; | |
833 } | |
834 { | |
835 E Tm, Tn, T1w, T1x; | |
836 Tm = Cr[WS(csr, 18)]; | |
837 Tn = Cr[WS(csr, 14)]; | |
838 To = Tm + Tn; | |
839 T42 = Tm - Tn; | |
840 T1w = Ci[WS(csi, 18)]; | |
841 T1x = Ci[WS(csi, 14)]; | |
842 T1y = T1w - T1x; | |
843 T3P = T1w + T1x; | |
844 } | |
845 { | |
846 E Tq, Tr, T1n, T1o; | |
847 Tq = Cr[WS(csr, 10)]; | |
848 Tr = Cr[WS(csr, 22)]; | |
849 Ts = Tq + Tr; | |
850 T3R = Tq - Tr; | |
851 T1n = Ci[WS(csi, 10)]; | |
852 T1o = Ci[WS(csi, 22)]; | |
853 T1p = T1n - T1o; | |
854 T3S = T1n + T1o; | |
855 } | |
856 { | |
857 E Tt, Tu, T1k, T1l; | |
858 Tt = Cr[WS(csr, 6)]; | |
859 Tu = Cr[WS(csr, 26)]; | |
860 Tv = Tt + Tu; | |
861 T3U = Tt - Tu; | |
862 T1k = Ci[WS(csi, 26)]; | |
863 T1l = Ci[WS(csi, 6)]; | |
864 T1m = T1k - T1l; | |
865 T3V = T1l + T1k; | |
866 } | |
867 T3Q = T3O - T3P; | |
868 T5q = T3O + T3P; | |
869 T5u = T43 - T42; | |
870 T44 = T42 + T43; | |
871 Tp = Tl + To; | |
872 Tw = Ts + Tv; | |
873 T2V = Tp - Tw; | |
874 { | |
875 E T3T, T3W, T1j, T1q; | |
876 T2W = T1y + T1v; | |
877 T2X = T1p + T1m; | |
878 T2Y = T2W - T2X; | |
879 T3T = T3R - T3S; | |
880 T3W = T3U - T3V; | |
881 T3X = KP707106781 * (T3T + T3W); | |
882 T5t = KP707106781 * (T3T - T3W); | |
883 T1j = Tl - To; | |
884 T1q = T1m - T1p; | |
885 T1r = T1j + T1q; | |
886 T2x = T1j - T1q; | |
887 { | |
888 E T3Z, T40, T1s, T1z; | |
889 T3Z = T3R + T3S; | |
890 T40 = T3U + T3V; | |
891 T41 = KP707106781 * (T3Z - T40); | |
892 T5r = KP707106781 * (T3Z + T40); | |
893 T1s = Ts - Tv; | |
894 T1z = T1v - T1y; | |
895 T1A = T1s + T1z; | |
896 T2y = T1z - T1s; | |
897 } | |
898 } | |
899 } | |
900 { | |
901 E TB, T48, T2c, T4G, TE, T4F, T2f, T49, TI, T4b, T1J, T4c, TL, T4e, T1G; | |
902 E T4f; | |
903 { | |
904 E Tz, TA, T2a, T2b; | |
905 Tz = Cr[WS(csr, 1)]; | |
906 TA = Cr[WS(csr, 31)]; | |
907 TB = Tz + TA; | |
908 T48 = Tz - TA; | |
909 T2a = Ci[WS(csi, 1)]; | |
910 T2b = Ci[WS(csi, 31)]; | |
911 T2c = T2a - T2b; | |
912 T4G = T2a + T2b; | |
913 } | |
914 { | |
915 E TC, TD, T2d, T2e; | |
916 TC = Cr[WS(csr, 17)]; | |
917 TD = Cr[WS(csr, 15)]; | |
918 TE = TC + TD; | |
919 T4F = TC - TD; | |
920 T2d = Ci[WS(csi, 17)]; | |
921 T2e = Ci[WS(csi, 15)]; | |
922 T2f = T2d - T2e; | |
923 T49 = T2d + T2e; | |
924 } | |
925 { | |
926 E TG, TH, T1H, T1I; | |
927 TG = Cr[WS(csr, 9)]; | |
928 TH = Cr[WS(csr, 23)]; | |
929 TI = TG + TH; | |
930 T4b = TG - TH; | |
931 T1H = Ci[WS(csi, 9)]; | |
932 T1I = Ci[WS(csi, 23)]; | |
933 T1J = T1H - T1I; | |
934 T4c = T1H + T1I; | |
935 } | |
936 { | |
937 E TJ, TK, T1E, T1F; | |
938 TJ = Cr[WS(csr, 7)]; | |
939 TK = Cr[WS(csr, 25)]; | |
940 TL = TJ + TK; | |
941 T4e = TJ - TK; | |
942 T1E = Ci[WS(csi, 25)]; | |
943 T1F = Ci[WS(csi, 7)]; | |
944 T1G = T1E - T1F; | |
945 T4f = T1F + T1E; | |
946 } | |
947 { | |
948 E TF, TM, T1D, T1K; | |
949 T4a = T48 - T49; | |
950 T5y = T48 + T49; | |
951 T5N = T4G - T4F; | |
952 T4H = T4F + T4G; | |
953 TF = TB + TE; | |
954 TM = TI + TL; | |
955 TN = TF + TM; | |
956 T31 = TF - TM; | |
957 { | |
958 E T4C, T4D, T37, T38; | |
959 T4C = T4b + T4c; | |
960 T4D = T4e + T4f; | |
961 T4E = KP707106781 * (T4C - T4D); | |
962 T5z = KP707106781 * (T4C + T4D); | |
963 T37 = T2f + T2c; | |
964 T38 = T1J + T1G; | |
965 T39 = T37 - T38; | |
966 T3q = T38 + T37; | |
967 } | |
968 T1D = TB - TE; | |
969 T1K = T1G - T1J; | |
970 T1L = T1D + T1K; | |
971 T2B = T1D - T1K; | |
972 { | |
973 E T4d, T4g, T29, T2g; | |
974 T4d = T4b - T4c; | |
975 T4g = T4e - T4f; | |
976 T4h = KP707106781 * (T4d + T4g); | |
977 T5M = KP707106781 * (T4d - T4g); | |
978 T29 = TI - TL; | |
979 T2g = T2c - T2f; | |
980 T2h = T29 + T2g; | |
981 T2F = T2g - T29; | |
982 } | |
983 } | |
984 } | |
985 { | |
986 E TQ, T4j, T1P, T4n, TT, T4m, T1S, T4k, TX, T4q, T1Y, T4u, T10, T4t, T21; | |
987 E T4r; | |
988 { | |
989 E TO, TP, T1N, T1O; | |
990 TO = Cr[WS(csr, 5)]; | |
991 TP = Cr[WS(csr, 27)]; | |
992 TQ = TO + TP; | |
993 T4j = TO - TP; | |
994 T1N = Ci[WS(csi, 5)]; | |
995 T1O = Ci[WS(csi, 27)]; | |
996 T1P = T1N - T1O; | |
997 T4n = T1N + T1O; | |
998 } | |
999 { | |
1000 E TR, TS, T1Q, T1R; | |
1001 TR = Cr[WS(csr, 21)]; | |
1002 TS = Cr[WS(csr, 11)]; | |
1003 TT = TR + TS; | |
1004 T4m = TR - TS; | |
1005 T1Q = Ci[WS(csi, 21)]; | |
1006 T1R = Ci[WS(csi, 11)]; | |
1007 T1S = T1Q - T1R; | |
1008 T4k = T1Q + T1R; | |
1009 } | |
1010 { | |
1011 E TV, TW, T1W, T1X; | |
1012 TV = Cr[WS(csr, 3)]; | |
1013 TW = Cr[WS(csr, 29)]; | |
1014 TX = TV + TW; | |
1015 T4q = TV - TW; | |
1016 T1W = Ci[WS(csi, 29)]; | |
1017 T1X = Ci[WS(csi, 3)]; | |
1018 T1Y = T1W - T1X; | |
1019 T4u = T1X + T1W; | |
1020 } | |
1021 { | |
1022 E TY, TZ, T1Z, T20; | |
1023 TY = Cr[WS(csr, 13)]; | |
1024 TZ = Cr[WS(csr, 19)]; | |
1025 T10 = TY + TZ; | |
1026 T4t = TY - TZ; | |
1027 T1Z = Ci[WS(csi, 13)]; | |
1028 T20 = Ci[WS(csi, 19)]; | |
1029 T21 = T1Z - T20; | |
1030 T4r = T1Z + T20; | |
1031 } | |
1032 { | |
1033 E TU, T11, T5B, T5C; | |
1034 TU = TQ + TT; | |
1035 T11 = TX + T10; | |
1036 T12 = TU + T11; | |
1037 T36 = TU - T11; | |
1038 T5B = T4j + T4k; | |
1039 T5C = T4n - T4m; | |
1040 T5D = FNMS(KP923879532, T5C, KP382683432 * T5B); | |
1041 T5J = FMA(KP923879532, T5B, KP382683432 * T5C); | |
1042 } | |
1043 { | |
1044 E T5E, T5F, T1M, T1T; | |
1045 T5E = T4q + T4r; | |
1046 T5F = T4t + T4u; | |
1047 T5G = FNMS(KP923879532, T5F, KP382683432 * T5E); | |
1048 T5K = FMA(KP923879532, T5E, KP382683432 * T5F); | |
1049 T1M = TQ - TT; | |
1050 T1T = T1P - T1S; | |
1051 T1U = T1M - T1T; | |
1052 T26 = T1M + T1T; | |
1053 } | |
1054 { | |
1055 E T1V, T22, T4l, T4o; | |
1056 T1V = TX - T10; | |
1057 T22 = T1Y - T21; | |
1058 T23 = T1V + T22; | |
1059 T27 = T22 - T1V; | |
1060 T4l = T4j - T4k; | |
1061 T4o = T4m + T4n; | |
1062 T4p = FNMS(KP382683432, T4o, KP923879532 * T4l); | |
1063 T4z = FMA(KP382683432, T4l, KP923879532 * T4o); | |
1064 } | |
1065 { | |
1066 E T4s, T4v, T32, T33; | |
1067 T4s = T4q - T4r; | |
1068 T4v = T4t - T4u; | |
1069 T4w = FMA(KP923879532, T4s, KP382683432 * T4v); | |
1070 T4A = FNMS(KP382683432, T4s, KP923879532 * T4v); | |
1071 T32 = T21 + T1Y; | |
1072 T33 = T1S + T1P; | |
1073 T34 = T32 - T33; | |
1074 T3r = T33 + T32; | |
1075 } | |
1076 } | |
1077 { | |
1078 E T13, T3x, Ty, T3w, Ti, Tx; | |
1079 T13 = KP2_000000000 * (TN + T12); | |
1080 T3x = KP2_000000000 * (T3r + T3q); | |
1081 Ti = Ta + Th; | |
1082 Tx = KP2_000000000 * (Tp + Tw); | |
1083 Ty = Ti + Tx; | |
1084 T3w = Ti - Tx; | |
1085 R0[WS(rs, 16)] = Ty - T13; | |
1086 R0[WS(rs, 24)] = T3w + T3x; | |
1087 R0[0] = Ty + T13; | |
1088 R0[WS(rs, 8)] = T3w - T3x; | |
1089 } | |
1090 { | |
1091 E T3g, T3k, T3j, T3l; | |
1092 { | |
1093 E T3e, T3f, T3h, T3i; | |
1094 T3e = T2S + T2T; | |
1095 T3f = KP1_414213562 * (T2V + T2Y); | |
1096 T3g = T3e - T3f; | |
1097 T3k = T3e + T3f; | |
1098 T3h = T31 - T34; | |
1099 T3i = T39 - T36; | |
1100 T3j = FNMS(KP1_847759065, T3i, KP765366864 * T3h); | |
1101 T3l = FMA(KP1_847759065, T3h, KP765366864 * T3i); | |
1102 } | |
1103 R0[WS(rs, 22)] = T3g - T3j; | |
1104 R0[WS(rs, 30)] = T3k + T3l; | |
1105 R0[WS(rs, 6)] = T3g + T3j; | |
1106 R0[WS(rs, 14)] = T3k - T3l; | |
1107 } | |
1108 { | |
1109 E T3o, T3u, T3t, T3v; | |
1110 { | |
1111 E T3m, T3n, T3p, T3s; | |
1112 T3m = Ta - Th; | |
1113 T3n = KP2_000000000 * (T2X + T2W); | |
1114 T3o = T3m - T3n; | |
1115 T3u = T3m + T3n; | |
1116 T3p = TN - T12; | |
1117 T3s = T3q - T3r; | |
1118 T3t = KP1_414213562 * (T3p - T3s); | |
1119 T3v = KP1_414213562 * (T3p + T3s); | |
1120 } | |
1121 R0[WS(rs, 20)] = T3o - T3t; | |
1122 R0[WS(rs, 28)] = T3u + T3v; | |
1123 R0[WS(rs, 4)] = T3o + T3t; | |
1124 R0[WS(rs, 12)] = T3u - T3v; | |
1125 } | |
1126 { | |
1127 E T30, T3c, T3b, T3d; | |
1128 { | |
1129 E T2U, T2Z, T35, T3a; | |
1130 T2U = T2S - T2T; | |
1131 T2Z = KP1_414213562 * (T2V - T2Y); | |
1132 T30 = T2U + T2Z; | |
1133 T3c = T2U - T2Z; | |
1134 T35 = T31 + T34; | |
1135 T3a = T36 + T39; | |
1136 T3b = FNMS(KP765366864, T3a, KP1_847759065 * T35); | |
1137 T3d = FMA(KP765366864, T35, KP1_847759065 * T3a); | |
1138 } | |
1139 R0[WS(rs, 18)] = T30 - T3b; | |
1140 R0[WS(rs, 26)] = T3c + T3d; | |
1141 R0[WS(rs, 2)] = T30 + T3b; | |
1142 R0[WS(rs, 10)] = T3c - T3d; | |
1143 } | |
1144 { | |
1145 E T25, T2p, T2i, T2q, T1C, T2k, T2o, T2s, T24, T28; | |
1146 T24 = KP707106781 * (T1U + T23); | |
1147 T25 = T1L + T24; | |
1148 T2p = T1L - T24; | |
1149 T28 = KP707106781 * (T26 + T27); | |
1150 T2i = T28 + T2h; | |
1151 T2q = T2h - T28; | |
1152 { | |
1153 E T1i, T1B, T2m, T2n; | |
1154 T1i = T18 + T1h; | |
1155 T1B = FNMS(KP765366864, T1A, KP1_847759065 * T1r); | |
1156 T1C = T1i + T1B; | |
1157 T2k = T1i - T1B; | |
1158 T2m = T18 - T1h; | |
1159 T2n = FMA(KP765366864, T1r, KP1_847759065 * T1A); | |
1160 T2o = T2m - T2n; | |
1161 T2s = T2m + T2n; | |
1162 } | |
1163 { | |
1164 E T2j, T2t, T2l, T2r; | |
1165 T2j = FNMS(KP390180644, T2i, KP1_961570560 * T25); | |
1166 R0[WS(rs, 17)] = T1C - T2j; | |
1167 R0[WS(rs, 1)] = T1C + T2j; | |
1168 T2t = FMA(KP1_662939224, T2p, KP1_111140466 * T2q); | |
1169 R0[WS(rs, 13)] = T2s - T2t; | |
1170 R0[WS(rs, 29)] = T2s + T2t; | |
1171 T2l = FMA(KP390180644, T25, KP1_961570560 * T2i); | |
1172 R0[WS(rs, 9)] = T2k - T2l; | |
1173 R0[WS(rs, 25)] = T2k + T2l; | |
1174 T2r = FNMS(KP1_662939224, T2q, KP1_111140466 * T2p); | |
1175 R0[WS(rs, 21)] = T2o - T2r; | |
1176 R0[WS(rs, 5)] = T2o + T2r; | |
1177 } | |
1178 } | |
1179 { | |
1180 E T2D, T2N, T2G, T2O, T2A, T2I, T2M, T2Q, T2C, T2E; | |
1181 T2C = KP707106781 * (T27 - T26); | |
1182 T2D = T2B + T2C; | |
1183 T2N = T2B - T2C; | |
1184 T2E = KP707106781 * (T1U - T23); | |
1185 T2G = T2E + T2F; | |
1186 T2O = T2F - T2E; | |
1187 { | |
1188 E T2w, T2z, T2K, T2L; | |
1189 T2w = T2u - T2v; | |
1190 T2z = FNMS(KP1_847759065, T2y, KP765366864 * T2x); | |
1191 T2A = T2w + T2z; | |
1192 T2I = T2w - T2z; | |
1193 T2K = T2u + T2v; | |
1194 T2L = FMA(KP1_847759065, T2x, KP765366864 * T2y); | |
1195 T2M = T2K - T2L; | |
1196 T2Q = T2K + T2L; | |
1197 } | |
1198 { | |
1199 E T2H, T2R, T2J, T2P; | |
1200 T2H = FNMS(KP1_111140466, T2G, KP1_662939224 * T2D); | |
1201 R0[WS(rs, 19)] = T2A - T2H; | |
1202 R0[WS(rs, 3)] = T2A + T2H; | |
1203 T2R = FMA(KP1_961570560, T2N, KP390180644 * T2O); | |
1204 R0[WS(rs, 15)] = T2Q - T2R; | |
1205 R0[WS(rs, 31)] = T2Q + T2R; | |
1206 T2J = FMA(KP1_111140466, T2D, KP1_662939224 * T2G); | |
1207 R0[WS(rs, 11)] = T2I - T2J; | |
1208 R0[WS(rs, 27)] = T2I + T2J; | |
1209 T2P = FNMS(KP1_961570560, T2O, KP390180644 * T2N); | |
1210 R0[WS(rs, 23)] = T2M - T2P; | |
1211 R0[WS(rs, 7)] = T2M + T2P; | |
1212 } | |
1213 } | |
1214 { | |
1215 E T5p, T5T, T5w, T5U, T5I, T5W, T5P, T5X, T5s, T5v; | |
1216 T5p = T5l + T5o; | |
1217 T5T = T5l - T5o; | |
1218 T5s = T5q - T5r; | |
1219 T5v = T5t + T5u; | |
1220 T5w = FNMS(KP1_111140466, T5v, KP1_662939224 * T5s); | |
1221 T5U = FMA(KP1_111140466, T5s, KP1_662939224 * T5v); | |
1222 { | |
1223 E T5A, T5H, T5L, T5O; | |
1224 T5A = T5y - T5z; | |
1225 T5H = T5D + T5G; | |
1226 T5I = T5A + T5H; | |
1227 T5W = T5A - T5H; | |
1228 T5L = T5J - T5K; | |
1229 T5O = T5M + T5N; | |
1230 T5P = T5L + T5O; | |
1231 T5X = T5O - T5L; | |
1232 } | |
1233 { | |
1234 E T5x, T5Q, T5Z, T60; | |
1235 T5x = T5p + T5w; | |
1236 T5Q = FNMS(KP580569354, T5P, KP1_913880671 * T5I); | |
1237 R1[WS(rs, 17)] = T5x - T5Q; | |
1238 R1[WS(rs, 1)] = T5x + T5Q; | |
1239 T5Z = T5T + T5U; | |
1240 T60 = FMA(KP1_763842528, T5W, KP942793473 * T5X); | |
1241 R1[WS(rs, 13)] = T5Z - T60; | |
1242 R1[WS(rs, 29)] = T5Z + T60; | |
1243 } | |
1244 { | |
1245 E T5R, T5S, T5V, T5Y; | |
1246 T5R = T5p - T5w; | |
1247 T5S = FMA(KP580569354, T5I, KP1_913880671 * T5P); | |
1248 R1[WS(rs, 9)] = T5R - T5S; | |
1249 R1[WS(rs, 25)] = T5R + T5S; | |
1250 T5V = T5T - T5U; | |
1251 T5Y = FNMS(KP1_763842528, T5X, KP942793473 * T5W); | |
1252 R1[WS(rs, 21)] = T5V - T5Y; | |
1253 R1[WS(rs, 5)] = T5V + T5Y; | |
1254 } | |
1255 } | |
1256 { | |
1257 E T3N, T4N, T46, T4O, T4y, T4Q, T4J, T4R, T3Y, T45; | |
1258 T3N = T3F + T3M; | |
1259 T4N = T3F - T3M; | |
1260 T3Y = T3Q + T3X; | |
1261 T45 = T41 + T44; | |
1262 T46 = FNMS(KP390180644, T45, KP1_961570560 * T3Y); | |
1263 T4O = FMA(KP390180644, T3Y, KP1_961570560 * T45); | |
1264 { | |
1265 E T4i, T4x, T4B, T4I; | |
1266 T4i = T4a + T4h; | |
1267 T4x = T4p + T4w; | |
1268 T4y = T4i + T4x; | |
1269 T4Q = T4i - T4x; | |
1270 T4B = T4z + T4A; | |
1271 T4I = T4E + T4H; | |
1272 T4J = T4B + T4I; | |
1273 T4R = T4I - T4B; | |
1274 } | |
1275 { | |
1276 E T47, T4K, T4T, T4U; | |
1277 T47 = T3N + T46; | |
1278 T4K = FNMS(KP196034280, T4J, KP1_990369453 * T4y); | |
1279 R1[WS(rs, 16)] = T47 - T4K; | |
1280 R1[0] = T47 + T4K; | |
1281 T4T = T4N + T4O; | |
1282 T4U = FMA(KP1_546020906, T4Q, KP1_268786568 * T4R); | |
1283 R1[WS(rs, 12)] = T4T - T4U; | |
1284 R1[WS(rs, 28)] = T4T + T4U; | |
1285 } | |
1286 { | |
1287 E T4L, T4M, T4P, T4S; | |
1288 T4L = T3N - T46; | |
1289 T4M = FMA(KP196034280, T4y, KP1_990369453 * T4J); | |
1290 R1[WS(rs, 8)] = T4L - T4M; | |
1291 R1[WS(rs, 24)] = T4L + T4M; | |
1292 T4P = T4N - T4O; | |
1293 T4S = FNMS(KP1_546020906, T4R, KP1_268786568 * T4Q); | |
1294 R1[WS(rs, 20)] = T4P - T4S; | |
1295 R1[WS(rs, 4)] = T4P + T4S; | |
1296 } | |
1297 } | |
1298 { | |
1299 E T63, T6h, T66, T6i, T6a, T6k, T6d, T6l, T64, T65; | |
1300 T63 = T61 - T62; | |
1301 T6h = T61 + T62; | |
1302 T64 = T5q + T5r; | |
1303 T65 = T5u - T5t; | |
1304 T66 = FNMS(KP1_961570560, T65, KP390180644 * T64); | |
1305 T6i = FMA(KP1_961570560, T64, KP390180644 * T65); | |
1306 { | |
1307 E T68, T69, T6b, T6c; | |
1308 T68 = T5y + T5z; | |
1309 T69 = T5J + T5K; | |
1310 T6a = T68 - T69; | |
1311 T6k = T68 + T69; | |
1312 T6b = T5D - T5G; | |
1313 T6c = T5N - T5M; | |
1314 T6d = T6b + T6c; | |
1315 T6l = T6c - T6b; | |
1316 } | |
1317 { | |
1318 E T67, T6e, T6n, T6o; | |
1319 T67 = T63 + T66; | |
1320 T6e = FNMS(KP1_268786568, T6d, KP1_546020906 * T6a); | |
1321 R1[WS(rs, 19)] = T67 - T6e; | |
1322 R1[WS(rs, 3)] = T67 + T6e; | |
1323 T6n = T6h + T6i; | |
1324 T6o = FMA(KP1_990369453, T6k, KP196034280 * T6l); | |
1325 R1[WS(rs, 15)] = T6n - T6o; | |
1326 R1[WS(rs, 31)] = T6n + T6o; | |
1327 } | |
1328 { | |
1329 E T6f, T6g, T6j, T6m; | |
1330 T6f = T63 - T66; | |
1331 T6g = FMA(KP1_268786568, T6a, KP1_546020906 * T6d); | |
1332 R1[WS(rs, 11)] = T6f - T6g; | |
1333 R1[WS(rs, 27)] = T6f + T6g; | |
1334 T6j = T6h - T6i; | |
1335 T6m = FNMS(KP1_990369453, T6l, KP196034280 * T6k); | |
1336 R1[WS(rs, 23)] = T6j - T6m; | |
1337 R1[WS(rs, 7)] = T6j + T6m; | |
1338 } | |
1339 } | |
1340 { | |
1341 E T4X, T5b, T50, T5c, T54, T5e, T57, T5f, T4Y, T4Z; | |
1342 T4X = T4V - T4W; | |
1343 T5b = T4V + T4W; | |
1344 T4Y = T3Q - T3X; | |
1345 T4Z = T44 - T41; | |
1346 T50 = FNMS(KP1_662939224, T4Z, KP1_111140466 * T4Y); | |
1347 T5c = FMA(KP1_662939224, T4Y, KP1_111140466 * T4Z); | |
1348 { | |
1349 E T52, T53, T55, T56; | |
1350 T52 = T4a - T4h; | |
1351 T53 = T4A - T4z; | |
1352 T54 = T52 + T53; | |
1353 T5e = T52 - T53; | |
1354 T55 = T4p - T4w; | |
1355 T56 = T4H - T4E; | |
1356 T57 = T55 + T56; | |
1357 T5f = T56 - T55; | |
1358 } | |
1359 { | |
1360 E T51, T58, T5h, T5i; | |
1361 T51 = T4X + T50; | |
1362 T58 = FNMS(KP942793473, T57, KP1_763842528 * T54); | |
1363 R1[WS(rs, 18)] = T51 - T58; | |
1364 R1[WS(rs, 2)] = T51 + T58; | |
1365 T5h = T5b + T5c; | |
1366 T5i = FMA(KP1_913880671, T5e, KP580569354 * T5f); | |
1367 R1[WS(rs, 14)] = T5h - T5i; | |
1368 R1[WS(rs, 30)] = T5h + T5i; | |
1369 } | |
1370 { | |
1371 E T59, T5a, T5d, T5g; | |
1372 T59 = T4X - T50; | |
1373 T5a = FMA(KP942793473, T54, KP1_763842528 * T57); | |
1374 R1[WS(rs, 10)] = T59 - T5a; | |
1375 R1[WS(rs, 26)] = T59 + T5a; | |
1376 T5d = T5b - T5c; | |
1377 T5g = FNMS(KP1_913880671, T5f, KP580569354 * T5e); | |
1378 R1[WS(rs, 22)] = T5d - T5g; | |
1379 R1[WS(rs, 6)] = T5d + T5g; | |
1380 } | |
1381 } | |
1382 } | |
1383 } | |
1384 } | |
1385 | |
1386 static const kr2c_desc desc = { 64, "r2cb_64", {342, 82, 52, 0}, &GENUS }; | |
1387 | |
1388 void X(codelet_r2cb_64) (planner *p) { | |
1389 X(kr2c_register) (p, r2cb_64, &desc); | |
1390 } | |
1391 | |
1392 #endif /* HAVE_FMA */ |