Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/rdft/scalar/r2cf/hc2cfdft_16.c @ 167:bd3cc4d1df30
Add FFTW 3.3.8 source, and a Linux build
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Tue, 19 Nov 2019 14:52:55 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
166:cbd6d7e562c7 | 167:bd3cc4d1df30 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-14 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Thu May 24 08:07:12 EDT 2018 */ | |
23 | |
24 #include "rdft/codelet-rdft.h" | |
25 | |
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA) | |
27 | |
28 /* Generated by: ../../../genfft/gen_hc2cdft.native -fma -compact -variables 4 -pipeline-latency 4 -n 16 -dit -name hc2cfdft_16 -include rdft/scalar/hc2cf.h */ | |
29 | |
30 /* | |
31 * This function contains 206 FP additions, 132 FP multiplications, | |
32 * (or, 136 additions, 62 multiplications, 70 fused multiply/add), | |
33 * 67 stack variables, 4 constants, and 64 memory accesses | |
34 */ | |
35 #include "rdft/scalar/hc2cf.h" | |
36 | |
37 static void hc2cfdft_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
40 DK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
41 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
42 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
43 { | |
44 INT m; | |
45 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 30, MAKE_VOLATILE_STRIDE(64, rs)) { | |
46 E T1f, T2e, T1c, T2g, T1K, T3D, T2W, T3H, TR, T2j, T2R, T3E, T11, T2l, T1v; | |
47 E T3G, Ta, T2p, Tk, T2r, T3o, T3p, T1Y, T3z, T2G, T3w, Tv, T2u, TF, T2w; | |
48 E T3r, T3s, T2b, T3A, T2L, T3x; | |
49 { | |
50 E T1d, T1e, T1I, T16, T1A, T1D, T1E, T1C, T1G, T1H, T2U, T1b, T1z, T2S, T1w; | |
51 E T1y, T14, T15; | |
52 T1d = Ip[0]; | |
53 T1e = Im[0]; | |
54 T1I = T1d + T1e; | |
55 T14 = Ip[WS(rs, 4)]; | |
56 T15 = Im[WS(rs, 4)]; | |
57 T16 = T14 - T15; | |
58 T1A = T14 + T15; | |
59 { | |
60 E T1F, T19, T1a, T1x; | |
61 T1D = Rm[0]; | |
62 T1E = Rp[0]; | |
63 T1F = T1D - T1E; | |
64 T1C = W[0]; | |
65 T1G = T1C * T1F; | |
66 T1H = W[1]; | |
67 T2U = T1H * T1F; | |
68 T19 = Rp[WS(rs, 4)]; | |
69 T1a = Rm[WS(rs, 4)]; | |
70 T1x = T1a - T19; | |
71 T1b = T19 + T1a; | |
72 T1z = W[17]; | |
73 T2S = T1z * T1x; | |
74 T1w = W[16]; | |
75 T1y = T1w * T1x; | |
76 } | |
77 T1f = T1d - T1e; | |
78 T2e = T1E + T1D; | |
79 { | |
80 E T17, T2f, T13, T18; | |
81 T13 = W[14]; | |
82 T17 = T13 * T16; | |
83 T2f = T13 * T1b; | |
84 T18 = W[15]; | |
85 T1c = FNMS(T18, T1b, T17); | |
86 T2g = FMA(T18, T16, T2f); | |
87 } | |
88 { | |
89 E T1B, T1J, T2T, T2V; | |
90 T1B = FNMS(T1z, T1A, T1y); | |
91 T1J = FNMS(T1H, T1I, T1G); | |
92 T1K = T1B + T1J; | |
93 T3D = T1J - T1B; | |
94 T2T = FMA(T1w, T1A, T2S); | |
95 T2V = FMA(T1C, T1I, T2U); | |
96 T2W = T2T + T2V; | |
97 T3H = T2V - T2T; | |
98 } | |
99 } | |
100 { | |
101 E TL, T1n, TQ, T1m, T2N, T1j, T1l, TV, T1t, T10, T1s, T2P, T1p, T1r; | |
102 { | |
103 E TJ, TK, TO, TP, T1k; | |
104 TJ = Ip[WS(rs, 2)]; | |
105 TK = Im[WS(rs, 2)]; | |
106 TL = TJ - TK; | |
107 T1n = TJ + TK; | |
108 TO = Rp[WS(rs, 2)]; | |
109 TP = Rm[WS(rs, 2)]; | |
110 T1k = TP - TO; | |
111 TQ = TO + TP; | |
112 T1m = W[9]; | |
113 T2N = T1m * T1k; | |
114 T1j = W[8]; | |
115 T1l = T1j * T1k; | |
116 } | |
117 { | |
118 E TT, TU, TY, TZ, T1q; | |
119 TT = Ip[WS(rs, 6)]; | |
120 TU = Im[WS(rs, 6)]; | |
121 TV = TT - TU; | |
122 T1t = TT + TU; | |
123 TY = Rp[WS(rs, 6)]; | |
124 TZ = Rm[WS(rs, 6)]; | |
125 T1q = TZ - TY; | |
126 T10 = TY + TZ; | |
127 T1s = W[25]; | |
128 T2P = T1s * T1q; | |
129 T1p = W[24]; | |
130 T1r = T1p * T1q; | |
131 } | |
132 { | |
133 E T2O, T2Q, T1o, T1u; | |
134 { | |
135 E TM, T2i, TI, TN; | |
136 TI = W[6]; | |
137 TM = TI * TL; | |
138 T2i = TI * TQ; | |
139 TN = W[7]; | |
140 TR = FNMS(TN, TQ, TM); | |
141 T2j = FMA(TN, TL, T2i); | |
142 } | |
143 T2O = FMA(T1j, T1n, T2N); | |
144 T2Q = FMA(T1p, T1t, T2P); | |
145 T2R = T2O + T2Q; | |
146 T3E = T2O - T2Q; | |
147 { | |
148 E TW, T2k, TS, TX; | |
149 TS = W[22]; | |
150 TW = TS * TV; | |
151 T2k = TS * T10; | |
152 TX = W[23]; | |
153 T11 = FNMS(TX, T10, TW); | |
154 T2l = FMA(TX, TV, T2k); | |
155 } | |
156 T1o = FNMS(T1m, T1n, T1l); | |
157 T1u = FNMS(T1s, T1t, T1r); | |
158 T1v = T1o + T1u; | |
159 T3G = T1o - T1u; | |
160 } | |
161 } | |
162 { | |
163 E T4, T1Q, T9, T1N, T5, T2o, T1O, T2C, Te, T1W, Tj, T1T, Tf, T2q, T1U; | |
164 E T2E, T6, Tg; | |
165 { | |
166 E T1, T1M, Tb, T1S; | |
167 { | |
168 E T2, T3, T7, T8; | |
169 T2 = Ip[WS(rs, 1)]; | |
170 T3 = Im[WS(rs, 1)]; | |
171 T4 = T2 - T3; | |
172 T1Q = T2 + T3; | |
173 T7 = Rp[WS(rs, 1)]; | |
174 T8 = Rm[WS(rs, 1)]; | |
175 T9 = T7 + T8; | |
176 T1N = T7 - T8; | |
177 } | |
178 T1 = W[2]; | |
179 T5 = T1 * T4; | |
180 T2o = T1 * T9; | |
181 T1M = W[4]; | |
182 T1O = T1M * T1N; | |
183 T2C = T1M * T1Q; | |
184 { | |
185 E Tc, Td, Th, Ti; | |
186 Tc = Ip[WS(rs, 5)]; | |
187 Td = Im[WS(rs, 5)]; | |
188 Te = Tc - Td; | |
189 T1W = Tc + Td; | |
190 Th = Rp[WS(rs, 5)]; | |
191 Ti = Rm[WS(rs, 5)]; | |
192 Tj = Th + Ti; | |
193 T1T = Th - Ti; | |
194 } | |
195 Tb = W[18]; | |
196 Tf = Tb * Te; | |
197 T2q = Tb * Tj; | |
198 T1S = W[20]; | |
199 T1U = T1S * T1T; | |
200 T2E = T1S * T1W; | |
201 } | |
202 T6 = W[3]; | |
203 Ta = FNMS(T6, T9, T5); | |
204 T2p = FMA(T6, T4, T2o); | |
205 Tg = W[19]; | |
206 Tk = FNMS(Tg, Tj, Tf); | |
207 T2r = FMA(Tg, Te, T2q); | |
208 T3o = Ta - Tk; | |
209 T3p = T2p - T2r; | |
210 { | |
211 E T1R, T2D, T1X, T2F, T1P, T1V; | |
212 T1P = W[5]; | |
213 T1R = FMA(T1P, T1Q, T1O); | |
214 T2D = FNMS(T1P, T1N, T2C); | |
215 T1V = W[21]; | |
216 T1X = FMA(T1V, T1W, T1U); | |
217 T2F = FNMS(T1V, T1T, T2E); | |
218 T1Y = T1R + T1X; | |
219 T3z = T1X - T1R; | |
220 T2G = T2D + T2F; | |
221 T3w = T2F - T2D; | |
222 } | |
223 } | |
224 { | |
225 E Tp, T23, Tu, T20, Tq, T2t, T21, T2H, Tz, T29, TE, T26, TA, T2v, T27; | |
226 E T2J, Tr, TB; | |
227 { | |
228 E Tm, T1Z, Tw, T25; | |
229 { | |
230 E Tn, To, Ts, Tt; | |
231 Tn = Ip[WS(rs, 7)]; | |
232 To = Im[WS(rs, 7)]; | |
233 Tp = Tn - To; | |
234 T23 = Tn + To; | |
235 Ts = Rp[WS(rs, 7)]; | |
236 Tt = Rm[WS(rs, 7)]; | |
237 Tu = Ts + Tt; | |
238 T20 = Ts - Tt; | |
239 } | |
240 Tm = W[26]; | |
241 Tq = Tm * Tp; | |
242 T2t = Tm * Tu; | |
243 T1Z = W[28]; | |
244 T21 = T1Z * T20; | |
245 T2H = T1Z * T23; | |
246 { | |
247 E Tx, Ty, TC, TD; | |
248 Tx = Ip[WS(rs, 3)]; | |
249 Ty = Im[WS(rs, 3)]; | |
250 Tz = Tx - Ty; | |
251 T29 = Tx + Ty; | |
252 TC = Rp[WS(rs, 3)]; | |
253 TD = Rm[WS(rs, 3)]; | |
254 TE = TC + TD; | |
255 T26 = TC - TD; | |
256 } | |
257 Tw = W[10]; | |
258 TA = Tw * Tz; | |
259 T2v = Tw * TE; | |
260 T25 = W[12]; | |
261 T27 = T25 * T26; | |
262 T2J = T25 * T29; | |
263 } | |
264 Tr = W[27]; | |
265 Tv = FNMS(Tr, Tu, Tq); | |
266 T2u = FMA(Tr, Tp, T2t); | |
267 TB = W[11]; | |
268 TF = FNMS(TB, TE, TA); | |
269 T2w = FMA(TB, Tz, T2v); | |
270 T3r = T2u - T2w; | |
271 T3s = Tv - TF; | |
272 { | |
273 E T24, T2I, T2a, T2K, T22, T28; | |
274 T22 = W[29]; | |
275 T24 = FMA(T22, T23, T21); | |
276 T2I = FNMS(T22, T20, T2H); | |
277 T28 = W[13]; | |
278 T2a = FMA(T28, T29, T27); | |
279 T2K = FNMS(T28, T26, T2J); | |
280 T2b = T24 + T2a; | |
281 T3A = T2I - T2K; | |
282 T2L = T2I + T2K; | |
283 T3x = T2a - T24; | |
284 } | |
285 } | |
286 { | |
287 E TH, T3c, T36, T3g, T39, T3h, T1h, T32, T2d, T2A, T2y, T31, T2Y, T30, T2n; | |
288 E T3b; | |
289 { | |
290 E Tl, TG, T34, T35; | |
291 Tl = Ta + Tk; | |
292 TG = Tv + TF; | |
293 TH = Tl + TG; | |
294 T3c = Tl - TG; | |
295 T34 = T2L - T2G; | |
296 T35 = T1Y - T2b; | |
297 T36 = T34 + T35; | |
298 T3g = T34 - T35; | |
299 } | |
300 { | |
301 E T37, T38, T12, T1g; | |
302 T37 = T1K - T1v; | |
303 T38 = T2W - T2R; | |
304 T39 = T37 - T38; | |
305 T3h = T37 + T38; | |
306 T12 = TR + T11; | |
307 T1g = T1c + T1f; | |
308 T1h = T12 + T1g; | |
309 T32 = T1g - T12; | |
310 } | |
311 { | |
312 E T1L, T2c, T2s, T2x; | |
313 T1L = T1v + T1K; | |
314 T2c = T1Y + T2b; | |
315 T2d = T1L - T2c; | |
316 T2A = T2c + T1L; | |
317 T2s = T2p + T2r; | |
318 T2x = T2u + T2w; | |
319 T2y = T2s + T2x; | |
320 T31 = T2x - T2s; | |
321 } | |
322 { | |
323 E T2M, T2X, T2h, T2m; | |
324 T2M = T2G + T2L; | |
325 T2X = T2R + T2W; | |
326 T2Y = T2M - T2X; | |
327 T30 = T2M + T2X; | |
328 T2h = T2e + T2g; | |
329 T2m = T2j + T2l; | |
330 T2n = T2h + T2m; | |
331 T3b = T2h - T2m; | |
332 } | |
333 { | |
334 E T1i, T2Z, T2z, T2B; | |
335 T1i = TH + T1h; | |
336 Ip[0] = KP500000000 * (T1i + T2d); | |
337 Im[WS(rs, 7)] = KP500000000 * (T2d - T1i); | |
338 T2Z = T2n + T2y; | |
339 Rm[WS(rs, 7)] = KP500000000 * (T2Z - T30); | |
340 Rp[0] = KP500000000 * (T2Z + T30); | |
341 T2z = T2n - T2y; | |
342 Rm[WS(rs, 3)] = KP500000000 * (T2z - T2A); | |
343 Rp[WS(rs, 4)] = KP500000000 * (T2z + T2A); | |
344 T2B = T1h - TH; | |
345 Ip[WS(rs, 4)] = KP500000000 * (T2B + T2Y); | |
346 Im[WS(rs, 3)] = KP500000000 * (T2Y - T2B); | |
347 } | |
348 { | |
349 E T33, T3a, T3j, T3k; | |
350 T33 = T31 + T32; | |
351 T3a = T36 + T39; | |
352 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP707106781, T3a, T33)); | |
353 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP707106781, T3a, T33))); | |
354 T3j = T3b + T3c; | |
355 T3k = T3g + T3h; | |
356 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP707106781, T3k, T3j)); | |
357 Rp[WS(rs, 2)] = KP500000000 * (FMA(KP707106781, T3k, T3j)); | |
358 } | |
359 { | |
360 E T3d, T3e, T3f, T3i; | |
361 T3d = T3b - T3c; | |
362 T3e = T39 - T36; | |
363 Rm[WS(rs, 1)] = KP500000000 * (FNMS(KP707106781, T3e, T3d)); | |
364 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP707106781, T3e, T3d)); | |
365 T3f = T32 - T31; | |
366 T3i = T3g - T3h; | |
367 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP707106781, T3i, T3f)); | |
368 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP707106781, T3i, T3f))); | |
369 } | |
370 } | |
371 { | |
372 E T3n, T3Z, T44, T4e, T47, T4f, T3u, T4a, T3C, T3U, T3N, T49, T3Q, T40, T3J; | |
373 E T3V; | |
374 { | |
375 E T3l, T3m, T42, T43; | |
376 T3l = T1f - T1c; | |
377 T3m = T2j - T2l; | |
378 T3n = T3l - T3m; | |
379 T3Z = T3m + T3l; | |
380 T42 = T3w - T3x; | |
381 T43 = T3A - T3z; | |
382 T44 = FMA(KP414213562, T43, T42); | |
383 T4e = FNMS(KP414213562, T42, T43); | |
384 } | |
385 { | |
386 E T45, T46, T3q, T3t; | |
387 T45 = T3E + T3D; | |
388 T46 = T3H - T3G; | |
389 T47 = FMA(KP414213562, T46, T45); | |
390 T4f = FNMS(KP414213562, T45, T46); | |
391 T3q = T3o - T3p; | |
392 T3t = T3r + T3s; | |
393 T3u = T3q + T3t; | |
394 T4a = T3q - T3t; | |
395 } | |
396 { | |
397 E T3y, T3B, T3L, T3M; | |
398 T3y = T3w + T3x; | |
399 T3B = T3z + T3A; | |
400 T3C = FMA(KP414213562, T3B, T3y); | |
401 T3U = FNMS(KP414213562, T3y, T3B); | |
402 T3L = T2e - T2g; | |
403 T3M = TR - T11; | |
404 T3N = T3L + T3M; | |
405 T49 = T3L - T3M; | |
406 } | |
407 { | |
408 E T3O, T3P, T3F, T3I; | |
409 T3O = T3p + T3o; | |
410 T3P = T3r - T3s; | |
411 T3Q = T3O + T3P; | |
412 T40 = T3P - T3O; | |
413 T3F = T3D - T3E; | |
414 T3I = T3G + T3H; | |
415 T3J = FNMS(KP414213562, T3I, T3F); | |
416 T3V = FMA(KP414213562, T3F, T3I); | |
417 } | |
418 { | |
419 E T3v, T3K, T3X, T3Y; | |
420 T3v = FMA(KP707106781, T3u, T3n); | |
421 T3K = T3C + T3J; | |
422 Ip[WS(rs, 1)] = KP500000000 * (FMA(KP923879532, T3K, T3v)); | |
423 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP923879532, T3K, T3v))); | |
424 T3X = FMA(KP707106781, T3Q, T3N); | |
425 T3Y = T3U + T3V; | |
426 Rm[WS(rs, 6)] = KP500000000 * (FNMS(KP923879532, T3Y, T3X)); | |
427 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP923879532, T3Y, T3X)); | |
428 } | |
429 { | |
430 E T3R, T3S, T3T, T3W; | |
431 T3R = FNMS(KP707106781, T3Q, T3N); | |
432 T3S = T3J - T3C; | |
433 Rm[WS(rs, 2)] = KP500000000 * (FNMS(KP923879532, T3S, T3R)); | |
434 Rp[WS(rs, 5)] = KP500000000 * (FMA(KP923879532, T3S, T3R)); | |
435 T3T = FNMS(KP707106781, T3u, T3n); | |
436 T3W = T3U - T3V; | |
437 Ip[WS(rs, 5)] = KP500000000 * (FMA(KP923879532, T3W, T3T)); | |
438 Im[WS(rs, 2)] = -(KP500000000 * (FNMS(KP923879532, T3W, T3T))); | |
439 } | |
440 { | |
441 E T41, T48, T4h, T4i; | |
442 T41 = FNMS(KP707106781, T40, T3Z); | |
443 T48 = T44 - T47; | |
444 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP923879532, T48, T41)); | |
445 Im[0] = -(KP500000000 * (FNMS(KP923879532, T48, T41))); | |
446 T4h = FNMS(KP707106781, T4a, T49); | |
447 T4i = T4e + T4f; | |
448 Rp[WS(rs, 7)] = KP500000000 * (FNMS(KP923879532, T4i, T4h)); | |
449 Rm[0] = KP500000000 * (FMA(KP923879532, T4i, T4h)); | |
450 } | |
451 { | |
452 E T4b, T4c, T4d, T4g; | |
453 T4b = FMA(KP707106781, T4a, T49); | |
454 T4c = T44 + T47; | |
455 Rm[WS(rs, 4)] = KP500000000 * (FNMS(KP923879532, T4c, T4b)); | |
456 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP923879532, T4c, T4b)); | |
457 T4d = FMA(KP707106781, T40, T3Z); | |
458 T4g = T4e - T4f; | |
459 Ip[WS(rs, 3)] = KP500000000 * (FMA(KP923879532, T4g, T4d)); | |
460 Im[WS(rs, 4)] = -(KP500000000 * (FNMS(KP923879532, T4g, T4d))); | |
461 } | |
462 } | |
463 } | |
464 } | |
465 } | |
466 | |
467 static const tw_instr twinstr[] = { | |
468 {TW_FULL, 1, 16}, | |
469 {TW_NEXT, 1, 0} | |
470 }; | |
471 | |
472 static const hc2c_desc desc = { 16, "hc2cfdft_16", twinstr, &GENUS, {136, 62, 70, 0} }; | |
473 | |
474 void X(codelet_hc2cfdft_16) (planner *p) { | |
475 X(khc2c_register) (p, hc2cfdft_16, &desc, HC2C_VIA_DFT); | |
476 } | |
477 #else | |
478 | |
479 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -n 16 -dit -name hc2cfdft_16 -include rdft/scalar/hc2cf.h */ | |
480 | |
481 /* | |
482 * This function contains 206 FP additions, 100 FP multiplications, | |
483 * (or, 168 additions, 62 multiplications, 38 fused multiply/add), | |
484 * 61 stack variables, 4 constants, and 64 memory accesses | |
485 */ | |
486 #include "rdft/scalar/hc2cf.h" | |
487 | |
488 static void hc2cfdft_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms) | |
489 { | |
490 DK(KP461939766, +0.461939766255643378064091594698394143411208313); | |
491 DK(KP191341716, +0.191341716182544885864229992015199433380672281); | |
492 DK(KP353553390, +0.353553390593273762200422181052424519642417969); | |
493 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
494 { | |
495 INT m; | |
496 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 30, MAKE_VOLATILE_STRIDE(64, rs)) { | |
497 E T19, T3h, T21, T2Y, T1o, T3d, T2s, T39, TW, T3i, T24, T2Z, T1z, T3c, T2p; | |
498 E T3a, Tj, T2S, T28, T2R, T1L, T36, T2i, T32, TC, T2V, T2b, T2U, T1W, T35; | |
499 E T2l, T33; | |
500 { | |
501 E T10, T1m, T14, T1k, T18, T1h, T1f, T1Z; | |
502 { | |
503 E TY, TZ, T12, T13; | |
504 TY = Ip[WS(rs, 4)]; | |
505 TZ = Im[WS(rs, 4)]; | |
506 T10 = TY - TZ; | |
507 T1m = TY + TZ; | |
508 T12 = Rp[WS(rs, 4)]; | |
509 T13 = Rm[WS(rs, 4)]; | |
510 T14 = T12 + T13; | |
511 T1k = T12 - T13; | |
512 } | |
513 { | |
514 E T16, T17, T1d, T1e; | |
515 T16 = Ip[0]; | |
516 T17 = Im[0]; | |
517 T18 = T16 - T17; | |
518 T1h = T16 + T17; | |
519 T1d = Rm[0]; | |
520 T1e = Rp[0]; | |
521 T1f = T1d - T1e; | |
522 T1Z = T1e + T1d; | |
523 } | |
524 { | |
525 E T15, T20, TX, T11; | |
526 TX = W[14]; | |
527 T11 = W[15]; | |
528 T15 = FNMS(T11, T14, TX * T10); | |
529 T20 = FMA(TX, T14, T11 * T10); | |
530 T19 = T15 + T18; | |
531 T3h = T1Z - T20; | |
532 T21 = T1Z + T20; | |
533 T2Y = T18 - T15; | |
534 } | |
535 { | |
536 E T1i, T2r, T1n, T2q; | |
537 { | |
538 E T1c, T1g, T1j, T1l; | |
539 T1c = W[0]; | |
540 T1g = W[1]; | |
541 T1i = FNMS(T1g, T1h, T1c * T1f); | |
542 T2r = FMA(T1g, T1f, T1c * T1h); | |
543 T1j = W[16]; | |
544 T1l = W[17]; | |
545 T1n = FMA(T1j, T1k, T1l * T1m); | |
546 T2q = FNMS(T1l, T1k, T1j * T1m); | |
547 } | |
548 T1o = T1i - T1n; | |
549 T3d = T2r - T2q; | |
550 T2s = T2q + T2r; | |
551 T39 = T1n + T1i; | |
552 } | |
553 } | |
554 { | |
555 E TH, T1s, TL, T1q, TQ, T1x, TU, T1v; | |
556 { | |
557 E TF, TG, TJ, TK; | |
558 TF = Ip[WS(rs, 2)]; | |
559 TG = Im[WS(rs, 2)]; | |
560 TH = TF - TG; | |
561 T1s = TF + TG; | |
562 TJ = Rp[WS(rs, 2)]; | |
563 TK = Rm[WS(rs, 2)]; | |
564 TL = TJ + TK; | |
565 T1q = TJ - TK; | |
566 } | |
567 { | |
568 E TO, TP, TS, TT; | |
569 TO = Ip[WS(rs, 6)]; | |
570 TP = Im[WS(rs, 6)]; | |
571 TQ = TO - TP; | |
572 T1x = TO + TP; | |
573 TS = Rp[WS(rs, 6)]; | |
574 TT = Rm[WS(rs, 6)]; | |
575 TU = TS + TT; | |
576 T1v = TS - TT; | |
577 } | |
578 { | |
579 E TM, T22, TV, T23; | |
580 { | |
581 E TE, TI, TN, TR; | |
582 TE = W[6]; | |
583 TI = W[7]; | |
584 TM = FNMS(TI, TL, TE * TH); | |
585 T22 = FMA(TE, TL, TI * TH); | |
586 TN = W[22]; | |
587 TR = W[23]; | |
588 TV = FNMS(TR, TU, TN * TQ); | |
589 T23 = FMA(TN, TU, TR * TQ); | |
590 } | |
591 TW = TM + TV; | |
592 T3i = TM - TV; | |
593 T24 = T22 + T23; | |
594 T2Z = T22 - T23; | |
595 } | |
596 { | |
597 E T1t, T2n, T1y, T2o; | |
598 { | |
599 E T1p, T1r, T1u, T1w; | |
600 T1p = W[8]; | |
601 T1r = W[9]; | |
602 T1t = FMA(T1p, T1q, T1r * T1s); | |
603 T2n = FNMS(T1r, T1q, T1p * T1s); | |
604 T1u = W[24]; | |
605 T1w = W[25]; | |
606 T1y = FMA(T1u, T1v, T1w * T1x); | |
607 T2o = FNMS(T1w, T1v, T1u * T1x); | |
608 } | |
609 T1z = T1t + T1y; | |
610 T3c = T1y - T1t; | |
611 T2p = T2n + T2o; | |
612 T3a = T2n - T2o; | |
613 } | |
614 } | |
615 { | |
616 E T4, T1E, T8, T1C, Td, T1J, Th, T1H; | |
617 { | |
618 E T2, T3, T6, T7; | |
619 T2 = Ip[WS(rs, 1)]; | |
620 T3 = Im[WS(rs, 1)]; | |
621 T4 = T2 - T3; | |
622 T1E = T2 + T3; | |
623 T6 = Rp[WS(rs, 1)]; | |
624 T7 = Rm[WS(rs, 1)]; | |
625 T8 = T6 + T7; | |
626 T1C = T6 - T7; | |
627 } | |
628 { | |
629 E Tb, Tc, Tf, Tg; | |
630 Tb = Ip[WS(rs, 5)]; | |
631 Tc = Im[WS(rs, 5)]; | |
632 Td = Tb - Tc; | |
633 T1J = Tb + Tc; | |
634 Tf = Rp[WS(rs, 5)]; | |
635 Tg = Rm[WS(rs, 5)]; | |
636 Th = Tf + Tg; | |
637 T1H = Tf - Tg; | |
638 } | |
639 { | |
640 E T9, T26, Ti, T27; | |
641 { | |
642 E T1, T5, Ta, Te; | |
643 T1 = W[2]; | |
644 T5 = W[3]; | |
645 T9 = FNMS(T5, T8, T1 * T4); | |
646 T26 = FMA(T1, T8, T5 * T4); | |
647 Ta = W[18]; | |
648 Te = W[19]; | |
649 Ti = FNMS(Te, Th, Ta * Td); | |
650 T27 = FMA(Ta, Th, Te * Td); | |
651 } | |
652 Tj = T9 + Ti; | |
653 T2S = T26 - T27; | |
654 T28 = T26 + T27; | |
655 T2R = T9 - Ti; | |
656 } | |
657 { | |
658 E T1F, T2g, T1K, T2h; | |
659 { | |
660 E T1B, T1D, T1G, T1I; | |
661 T1B = W[4]; | |
662 T1D = W[5]; | |
663 T1F = FMA(T1B, T1C, T1D * T1E); | |
664 T2g = FNMS(T1D, T1C, T1B * T1E); | |
665 T1G = W[20]; | |
666 T1I = W[21]; | |
667 T1K = FMA(T1G, T1H, T1I * T1J); | |
668 T2h = FNMS(T1I, T1H, T1G * T1J); | |
669 } | |
670 T1L = T1F + T1K; | |
671 T36 = T2g - T2h; | |
672 T2i = T2g + T2h; | |
673 T32 = T1K - T1F; | |
674 } | |
675 } | |
676 { | |
677 E Tn, T1P, Tr, T1N, Tw, T1U, TA, T1S; | |
678 { | |
679 E Tl, Tm, Tp, Tq; | |
680 Tl = Ip[WS(rs, 7)]; | |
681 Tm = Im[WS(rs, 7)]; | |
682 Tn = Tl - Tm; | |
683 T1P = Tl + Tm; | |
684 Tp = Rp[WS(rs, 7)]; | |
685 Tq = Rm[WS(rs, 7)]; | |
686 Tr = Tp + Tq; | |
687 T1N = Tp - Tq; | |
688 } | |
689 { | |
690 E Tu, Tv, Ty, Tz; | |
691 Tu = Ip[WS(rs, 3)]; | |
692 Tv = Im[WS(rs, 3)]; | |
693 Tw = Tu - Tv; | |
694 T1U = Tu + Tv; | |
695 Ty = Rp[WS(rs, 3)]; | |
696 Tz = Rm[WS(rs, 3)]; | |
697 TA = Ty + Tz; | |
698 T1S = Ty - Tz; | |
699 } | |
700 { | |
701 E Ts, T29, TB, T2a; | |
702 { | |
703 E Tk, To, Tt, Tx; | |
704 Tk = W[26]; | |
705 To = W[27]; | |
706 Ts = FNMS(To, Tr, Tk * Tn); | |
707 T29 = FMA(Tk, Tr, To * Tn); | |
708 Tt = W[10]; | |
709 Tx = W[11]; | |
710 TB = FNMS(Tx, TA, Tt * Tw); | |
711 T2a = FMA(Tt, TA, Tx * Tw); | |
712 } | |
713 TC = Ts + TB; | |
714 T2V = Ts - TB; | |
715 T2b = T29 + T2a; | |
716 T2U = T29 - T2a; | |
717 } | |
718 { | |
719 E T1Q, T2j, T1V, T2k; | |
720 { | |
721 E T1M, T1O, T1R, T1T; | |
722 T1M = W[28]; | |
723 T1O = W[29]; | |
724 T1Q = FMA(T1M, T1N, T1O * T1P); | |
725 T2j = FNMS(T1O, T1N, T1M * T1P); | |
726 T1R = W[12]; | |
727 T1T = W[13]; | |
728 T1V = FMA(T1R, T1S, T1T * T1U); | |
729 T2k = FNMS(T1T, T1S, T1R * T1U); | |
730 } | |
731 T1W = T1Q + T1V; | |
732 T35 = T1V - T1Q; | |
733 T2l = T2j + T2k; | |
734 T33 = T2j - T2k; | |
735 } | |
736 } | |
737 { | |
738 E T1b, T2f, T2u, T2w, T1Y, T2e, T2d, T2v; | |
739 { | |
740 E TD, T1a, T2m, T2t; | |
741 TD = Tj + TC; | |
742 T1a = TW + T19; | |
743 T1b = TD + T1a; | |
744 T2f = T1a - TD; | |
745 T2m = T2i + T2l; | |
746 T2t = T2p + T2s; | |
747 T2u = T2m - T2t; | |
748 T2w = T2m + T2t; | |
749 } | |
750 { | |
751 E T1A, T1X, T25, T2c; | |
752 T1A = T1o - T1z; | |
753 T1X = T1L + T1W; | |
754 T1Y = T1A - T1X; | |
755 T2e = T1X + T1A; | |
756 T25 = T21 + T24; | |
757 T2c = T28 + T2b; | |
758 T2d = T25 - T2c; | |
759 T2v = T25 + T2c; | |
760 } | |
761 Ip[0] = KP500000000 * (T1b + T1Y); | |
762 Rp[0] = KP500000000 * (T2v + T2w); | |
763 Im[WS(rs, 7)] = KP500000000 * (T1Y - T1b); | |
764 Rm[WS(rs, 7)] = KP500000000 * (T2v - T2w); | |
765 Rm[WS(rs, 3)] = KP500000000 * (T2d - T2e); | |
766 Im[WS(rs, 3)] = KP500000000 * (T2u - T2f); | |
767 Rp[WS(rs, 4)] = KP500000000 * (T2d + T2e); | |
768 Ip[WS(rs, 4)] = KP500000000 * (T2f + T2u); | |
769 } | |
770 { | |
771 E T2z, T2L, T2J, T2P, T2C, T2M, T2F, T2N; | |
772 { | |
773 E T2x, T2y, T2H, T2I; | |
774 T2x = T2b - T28; | |
775 T2y = T19 - TW; | |
776 T2z = KP500000000 * (T2x + T2y); | |
777 T2L = KP500000000 * (T2y - T2x); | |
778 T2H = T21 - T24; | |
779 T2I = Tj - TC; | |
780 T2J = KP500000000 * (T2H - T2I); | |
781 T2P = KP500000000 * (T2H + T2I); | |
782 } | |
783 { | |
784 E T2A, T2B, T2D, T2E; | |
785 T2A = T2l - T2i; | |
786 T2B = T1L - T1W; | |
787 T2C = T2A + T2B; | |
788 T2M = T2A - T2B; | |
789 T2D = T1z + T1o; | |
790 T2E = T2s - T2p; | |
791 T2F = T2D - T2E; | |
792 T2N = T2D + T2E; | |
793 } | |
794 { | |
795 E T2G, T2Q, T2K, T2O; | |
796 T2G = KP353553390 * (T2C + T2F); | |
797 Ip[WS(rs, 2)] = T2z + T2G; | |
798 Im[WS(rs, 5)] = T2G - T2z; | |
799 T2Q = KP353553390 * (T2M + T2N); | |
800 Rm[WS(rs, 5)] = T2P - T2Q; | |
801 Rp[WS(rs, 2)] = T2P + T2Q; | |
802 T2K = KP353553390 * (T2F - T2C); | |
803 Rm[WS(rs, 1)] = T2J - T2K; | |
804 Rp[WS(rs, 6)] = T2J + T2K; | |
805 T2O = KP353553390 * (T2M - T2N); | |
806 Ip[WS(rs, 6)] = T2L + T2O; | |
807 Im[WS(rs, 1)] = T2O - T2L; | |
808 } | |
809 } | |
810 { | |
811 E T30, T3w, T3F, T3j, T2X, T3G, T3D, T3L, T3m, T3v, T38, T3q, T3A, T3K, T3f; | |
812 E T3r; | |
813 { | |
814 E T2T, T2W, T34, T37; | |
815 T30 = KP500000000 * (T2Y - T2Z); | |
816 T3w = KP500000000 * (T2Z + T2Y); | |
817 T3F = KP500000000 * (T3h - T3i); | |
818 T3j = KP500000000 * (T3h + T3i); | |
819 T2T = T2R - T2S; | |
820 T2W = T2U + T2V; | |
821 T2X = KP353553390 * (T2T + T2W); | |
822 T3G = KP353553390 * (T2T - T2W); | |
823 { | |
824 E T3B, T3C, T3k, T3l; | |
825 T3B = T3a + T39; | |
826 T3C = T3d - T3c; | |
827 T3D = FNMS(KP461939766, T3C, KP191341716 * T3B); | |
828 T3L = FMA(KP461939766, T3B, KP191341716 * T3C); | |
829 T3k = T2S + T2R; | |
830 T3l = T2U - T2V; | |
831 T3m = KP353553390 * (T3k + T3l); | |
832 T3v = KP353553390 * (T3l - T3k); | |
833 } | |
834 T34 = T32 + T33; | |
835 T37 = T35 - T36; | |
836 T38 = FMA(KP191341716, T34, KP461939766 * T37); | |
837 T3q = FNMS(KP191341716, T37, KP461939766 * T34); | |
838 { | |
839 E T3y, T3z, T3b, T3e; | |
840 T3y = T33 - T32; | |
841 T3z = T36 + T35; | |
842 T3A = FMA(KP461939766, T3y, KP191341716 * T3z); | |
843 T3K = FNMS(KP461939766, T3z, KP191341716 * T3y); | |
844 T3b = T39 - T3a; | |
845 T3e = T3c + T3d; | |
846 T3f = FNMS(KP191341716, T3e, KP461939766 * T3b); | |
847 T3r = FMA(KP191341716, T3b, KP461939766 * T3e); | |
848 } | |
849 } | |
850 { | |
851 E T31, T3g, T3t, T3u; | |
852 T31 = T2X + T30; | |
853 T3g = T38 + T3f; | |
854 Ip[WS(rs, 1)] = T31 + T3g; | |
855 Im[WS(rs, 6)] = T3g - T31; | |
856 T3t = T3j + T3m; | |
857 T3u = T3q + T3r; | |
858 Rm[WS(rs, 6)] = T3t - T3u; | |
859 Rp[WS(rs, 1)] = T3t + T3u; | |
860 } | |
861 { | |
862 E T3n, T3o, T3p, T3s; | |
863 T3n = T3j - T3m; | |
864 T3o = T3f - T38; | |
865 Rm[WS(rs, 2)] = T3n - T3o; | |
866 Rp[WS(rs, 5)] = T3n + T3o; | |
867 T3p = T30 - T2X; | |
868 T3s = T3q - T3r; | |
869 Ip[WS(rs, 5)] = T3p + T3s; | |
870 Im[WS(rs, 2)] = T3s - T3p; | |
871 } | |
872 { | |
873 E T3x, T3E, T3N, T3O; | |
874 T3x = T3v + T3w; | |
875 T3E = T3A + T3D; | |
876 Ip[WS(rs, 3)] = T3x + T3E; | |
877 Im[WS(rs, 4)] = T3E - T3x; | |
878 T3N = T3F + T3G; | |
879 T3O = T3K + T3L; | |
880 Rm[WS(rs, 4)] = T3N - T3O; | |
881 Rp[WS(rs, 3)] = T3N + T3O; | |
882 } | |
883 { | |
884 E T3H, T3I, T3J, T3M; | |
885 T3H = T3F - T3G; | |
886 T3I = T3D - T3A; | |
887 Rm[0] = T3H - T3I; | |
888 Rp[WS(rs, 7)] = T3H + T3I; | |
889 T3J = T3w - T3v; | |
890 T3M = T3K - T3L; | |
891 Ip[WS(rs, 7)] = T3J + T3M; | |
892 Im[0] = T3M - T3J; | |
893 } | |
894 } | |
895 } | |
896 } | |
897 } | |
898 | |
899 static const tw_instr twinstr[] = { | |
900 {TW_FULL, 1, 16}, | |
901 {TW_NEXT, 1, 0} | |
902 }; | |
903 | |
904 static const hc2c_desc desc = { 16, "hc2cfdft_16", twinstr, &GENUS, {168, 62, 38, 0} }; | |
905 | |
906 void X(codelet_hc2cfdft_16) (planner *p) { | |
907 X(khc2c_register) (p, hc2cfdft_16, &desc, HC2C_VIA_DFT); | |
908 } | |
909 #endif |