Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/rdft/scalar/r2cf/hc2cfdft_20.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 08:07:13 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 20 -dit -name hc2cfdft_20 -include rdft/scalar/hc2cf.h */ | |
29 | |
30 /* | |
31 * This function contains 286 FP additions, 188 FP multiplications, | |
32 * (or, 176 additions, 78 multiplications, 110 fused multiply/add), | |
33 * 153 stack variables, 5 constants, and 80 memory accesses | |
34 */ | |
35 #include "rdft/scalar/hc2cf.h" | |
36 | |
37 static void hc2cfdft_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
41 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
42 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
43 DK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
44 { | |
45 INT m; | |
46 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 38, MAKE_VOLATILE_STRIDE(80, rs)) { | |
47 E T2E, T4W, T3v, T4k, T2M, T4V, T3w, T4j, T2p, T2T, T5a, T5A, T3o, T3D, T4b; | |
48 E T4B, T1Y, T2S, T57, T5z, T3h, T3C, T44, T4A, TH, T2P, T50, T5x, T32, T3z; | |
49 E T3P, T4D, T1o, T2Q, T53, T5w, T39, T3A, T3W, T4E; | |
50 { | |
51 E T9, T1V, Tu, T2w, T1, T5, T6, T2Y, T1R, T1T, T1U, T40, T10, T2F, TE; | |
52 E T2C, T1y, T2m, T4g, TX, T33, TS, TW, Tw, TA, TB, T3L, T2y, T2A, T2B; | |
53 E T3t, T1q, T1u, T1v, T3d, T2i, T2k, T2l, T48, Tm, Tq, Tr, T3J, T2s, T2u; | |
54 E T2v, T3r, T1g, T20, T1l, T23, T1h, T3S, T21, T3k, TL, T2H, TQ, T2K, TM; | |
55 E T35, T2I, T4h, T1I, T1D, T2g, T2f, T46, T2c, T2e, T1E, T3b, T16, T1b, T29; | |
56 E T28, T3i, T25, T27, T17, T3Q, Tj, Te, T1P, T1O, T3Y, T1L, T1N, Tf, T2W; | |
57 E T2x, T2D; | |
58 { | |
59 E T7, T8, Ts, Tt; | |
60 T7 = Rp[WS(rs, 9)]; | |
61 T8 = Rm[WS(rs, 9)]; | |
62 T9 = T7 - T8; | |
63 T1V = T7 + T8; | |
64 Ts = Rp[WS(rs, 2)]; | |
65 Tt = Rm[WS(rs, 2)]; | |
66 Tu = Ts + Tt; | |
67 T2w = Ts - Tt; | |
68 { | |
69 E T4, T1S, T2, T3; | |
70 T2 = Ip[WS(rs, 9)]; | |
71 T3 = Im[WS(rs, 9)]; | |
72 T4 = T2 + T3; | |
73 T1S = T2 - T3; | |
74 T1 = W[36]; | |
75 T5 = T1 * T4; | |
76 T6 = W[37]; | |
77 T2Y = T6 * T4; | |
78 T1R = W[34]; | |
79 T1T = T1R * T1S; | |
80 T1U = W[35]; | |
81 T40 = T1U * T1S; | |
82 } | |
83 } | |
84 { | |
85 E TY, TZ, TC, TD; | |
86 TY = Ip[0]; | |
87 TZ = Im[0]; | |
88 T10 = TY + TZ; | |
89 T2F = TY - TZ; | |
90 TC = Rp[WS(rs, 7)]; | |
91 TD = Rm[WS(rs, 7)]; | |
92 TE = TC + TD; | |
93 T2C = TC - TD; | |
94 } | |
95 { | |
96 E T1w, T1x, TT, TU, TV; | |
97 T1w = Rp[WS(rs, 1)]; | |
98 T1x = Rm[WS(rs, 1)]; | |
99 T1y = T1w - T1x; | |
100 T2m = T1w + T1x; | |
101 TT = Rm[0]; | |
102 TU = Rp[0]; | |
103 TV = TT - TU; | |
104 T4g = TU + TT; | |
105 TX = W[0]; | |
106 T33 = TX * TV; | |
107 TS = W[1]; | |
108 TW = TS * TV; | |
109 } | |
110 { | |
111 E T1d, T1Z, TI, T2G; | |
112 { | |
113 E Tz, T2z, Tx, Ty; | |
114 Tx = Ip[WS(rs, 7)]; | |
115 Ty = Im[WS(rs, 7)]; | |
116 Tz = Tx - Ty; | |
117 T2z = Tx + Ty; | |
118 Tw = W[26]; | |
119 TA = Tw * Tz; | |
120 TB = W[27]; | |
121 T3L = TB * Tz; | |
122 T2y = W[28]; | |
123 T2A = T2y * T2z; | |
124 T2B = W[29]; | |
125 T3t = T2B * T2z; | |
126 } | |
127 { | |
128 E T1t, T2j, T1r, T1s; | |
129 T1r = Ip[WS(rs, 1)]; | |
130 T1s = Im[WS(rs, 1)]; | |
131 T1t = T1r + T1s; | |
132 T2j = T1r - T1s; | |
133 T1q = W[4]; | |
134 T1u = T1q * T1t; | |
135 T1v = W[5]; | |
136 T3d = T1v * T1t; | |
137 T2i = W[2]; | |
138 T2k = T2i * T2j; | |
139 T2l = W[3]; | |
140 T48 = T2l * T2j; | |
141 } | |
142 { | |
143 E Tp, T2t, Tn, To; | |
144 Tn = Ip[WS(rs, 2)]; | |
145 To = Im[WS(rs, 2)]; | |
146 Tp = Tn - To; | |
147 T2t = Tn + To; | |
148 Tm = W[6]; | |
149 Tq = Tm * Tp; | |
150 Tr = W[7]; | |
151 T3J = Tr * Tp; | |
152 T2s = W[8]; | |
153 T2u = T2s * T2t; | |
154 T2v = W[9]; | |
155 T3r = T2v * T2t; | |
156 } | |
157 { | |
158 E T1e, T1f, T1j, T1k; | |
159 T1e = Ip[WS(rs, 3)]; | |
160 T1f = Im[WS(rs, 3)]; | |
161 T1g = T1e - T1f; | |
162 T20 = T1e + T1f; | |
163 T1j = Rp[WS(rs, 3)]; | |
164 T1k = Rm[WS(rs, 3)]; | |
165 T1l = T1j + T1k; | |
166 T23 = T1j - T1k; | |
167 } | |
168 T1d = W[10]; | |
169 T1h = T1d * T1g; | |
170 T3S = T1d * T1l; | |
171 T1Z = W[12]; | |
172 T21 = T1Z * T20; | |
173 T3k = T1Z * T23; | |
174 { | |
175 E TJ, TK, TO, TP; | |
176 TJ = Ip[WS(rs, 5)]; | |
177 TK = Im[WS(rs, 5)]; | |
178 TL = TJ + TK; | |
179 T2H = TJ - TK; | |
180 TO = Rp[WS(rs, 5)]; | |
181 TP = Rm[WS(rs, 5)]; | |
182 TQ = TO - TP; | |
183 T2K = TO + TP; | |
184 } | |
185 TI = W[20]; | |
186 TM = TI * TL; | |
187 T35 = TI * TQ; | |
188 T2G = W[18]; | |
189 T2I = T2G * T2H; | |
190 T4h = T2G * T2K; | |
191 { | |
192 E T1G, T1H, T2d, T1B, T1C, T1A; | |
193 T1G = Rm[WS(rs, 6)]; | |
194 T1H = Rp[WS(rs, 6)]; | |
195 T1I = T1G - T1H; | |
196 T1B = Ip[WS(rs, 6)]; | |
197 T1C = Im[WS(rs, 6)]; | |
198 T1D = T1B + T1C; | |
199 T2d = T1B - T1C; | |
200 T2g = T1H + T1G; | |
201 T2f = W[23]; | |
202 T46 = T2f * T2d; | |
203 T2c = W[22]; | |
204 T2e = T2c * T2d; | |
205 T1A = W[24]; | |
206 T1E = T1A * T1D; | |
207 T3b = T1A * T1I; | |
208 } | |
209 { | |
210 E T14, T15, T26, T19, T1a, T13; | |
211 T14 = Ip[WS(rs, 8)]; | |
212 T15 = Im[WS(rs, 8)]; | |
213 T16 = T14 - T15; | |
214 T19 = Rp[WS(rs, 8)]; | |
215 T1a = Rm[WS(rs, 8)]; | |
216 T1b = T19 + T1a; | |
217 T26 = T1a - T19; | |
218 T29 = T14 + T15; | |
219 T28 = W[32]; | |
220 T3i = T28 * T26; | |
221 T25 = W[33]; | |
222 T27 = T25 * T26; | |
223 T13 = W[30]; | |
224 T17 = T13 * T16; | |
225 T3Q = T13 * T1b; | |
226 } | |
227 { | |
228 E Th, Ti, T1M, Tc, Td, Tb; | |
229 Th = Rm[WS(rs, 4)]; | |
230 Ti = Rp[WS(rs, 4)]; | |
231 Tj = Th - Ti; | |
232 Tc = Ip[WS(rs, 4)]; | |
233 Td = Im[WS(rs, 4)]; | |
234 Te = Tc + Td; | |
235 T1M = Tc - Td; | |
236 T1P = Ti + Th; | |
237 T1O = W[15]; | |
238 T3Y = T1O * T1M; | |
239 T1L = W[14]; | |
240 T1N = T1L * T1M; | |
241 Tb = W[16]; | |
242 Tf = Tb * Te; | |
243 T2W = Tb * Tj; | |
244 } | |
245 } | |
246 T2x = FNMS(T2v, T2w, T2u); | |
247 T2D = FNMS(T2B, T2C, T2A); | |
248 T2E = T2x - T2D; | |
249 T4W = T2x + T2D; | |
250 { | |
251 E T3s, T3u, T2L, T4i, T2J; | |
252 T3s = FMA(T2s, T2w, T3r); | |
253 T3u = FMA(T2y, T2C, T3t); | |
254 T3v = T3s + T3u; | |
255 T4k = T3u - T3s; | |
256 T2J = W[19]; | |
257 T2L = FNMS(T2J, T2K, T2I); | |
258 T4i = FMA(T2J, T2H, T4h); | |
259 T2M = T2F - T2L; | |
260 T4V = T4g + T4i; | |
261 T3w = T2L + T2F; | |
262 T4j = T4g - T4i; | |
263 } | |
264 { | |
265 E T2a, T3j, T24, T3l, T2o, T3n, T4a, T59, T22; | |
266 T2a = FMA(T28, T29, T27); | |
267 T3j = FNMS(T25, T29, T3i); | |
268 T22 = W[13]; | |
269 T24 = FNMS(T22, T23, T21); | |
270 T3l = FMA(T22, T20, T3k); | |
271 { | |
272 E T2h, T2n, T47, T49; | |
273 T2h = FNMS(T2f, T2g, T2e); | |
274 T2n = FNMS(T2l, T2m, T2k); | |
275 T2o = T2h - T2n; | |
276 T3n = T2h + T2n; | |
277 T47 = FMA(T2c, T2g, T46); | |
278 T49 = FMA(T2i, T2m, T48); | |
279 T4a = T47 - T49; | |
280 T59 = T47 + T49; | |
281 } | |
282 { | |
283 E T2b, T58, T3m, T45; | |
284 T2b = T24 - T2a; | |
285 T2p = T2b - T2o; | |
286 T2T = T2b + T2o; | |
287 T58 = T2a + T24; | |
288 T5a = T58 + T59; | |
289 T5A = T59 - T58; | |
290 T3m = T3j - T3l; | |
291 T3o = T3m - T3n; | |
292 T3D = T3m + T3n; | |
293 T45 = T3j + T3l; | |
294 T4b = T45 + T4a; | |
295 T4B = T4a - T45; | |
296 } | |
297 } | |
298 { | |
299 E T1z, T3e, T1J, T3c, T1X, T3g, T42, T55, T1F; | |
300 T1z = FNMS(T1v, T1y, T1u); | |
301 T3e = FMA(T1q, T1y, T3d); | |
302 T1F = W[25]; | |
303 T1J = FMA(T1F, T1I, T1E); | |
304 T3c = FNMS(T1F, T1D, T3b); | |
305 { | |
306 E T1Q, T1W, T3Z, T41; | |
307 T1Q = FNMS(T1O, T1P, T1N); | |
308 T1W = FNMS(T1U, T1V, T1T); | |
309 T1X = T1Q - T1W; | |
310 T3g = T1Q + T1W; | |
311 T3Z = FMA(T1L, T1P, T3Y); | |
312 T41 = FMA(T1R, T1V, T40); | |
313 T42 = T3Z - T41; | |
314 T55 = T3Z + T41; | |
315 } | |
316 { | |
317 E T1K, T56, T3f, T43; | |
318 T1K = T1z - T1J; | |
319 T1Y = T1K - T1X; | |
320 T2S = T1X + T1K; | |
321 T56 = T1J + T1z; | |
322 T57 = T55 + T56; | |
323 T5z = T55 - T56; | |
324 T3f = T3c - T3e; | |
325 T3h = T3f - T3g; | |
326 T3C = T3g + T3f; | |
327 T43 = T3c + T3e; | |
328 T44 = T42 + T43; | |
329 T4A = T42 - T43; | |
330 } | |
331 } | |
332 { | |
333 E Ta, T2Z, Tk, T2X, TG, T31, T3N, T4Y, Tg; | |
334 Ta = FNMS(T6, T9, T5); | |
335 T2Z = FMA(T1, T9, T2Y); | |
336 Tg = W[17]; | |
337 Tk = FMA(Tg, Tj, Tf); | |
338 T2X = FNMS(Tg, Te, T2W); | |
339 { | |
340 E Tv, TF, T3K, T3M; | |
341 Tv = FNMS(Tr, Tu, Tq); | |
342 TF = FNMS(TB, TE, TA); | |
343 TG = Tv - TF; | |
344 T31 = Tv + TF; | |
345 T3K = FMA(Tm, Tu, T3J); | |
346 T3M = FMA(Tw, TE, T3L); | |
347 T3N = T3K - T3M; | |
348 T4Y = T3K + T3M; | |
349 } | |
350 { | |
351 E Tl, T4Z, T30, T3O; | |
352 Tl = Ta - Tk; | |
353 TH = Tl - TG; | |
354 T2P = TG + Tl; | |
355 T4Z = Tk + Ta; | |
356 T50 = T4Y + T4Z; | |
357 T5x = T4Y - T4Z; | |
358 T30 = T2X - T2Z; | |
359 T32 = T30 - T31; | |
360 T3z = T31 + T30; | |
361 T3O = T2X + T2Z; | |
362 T3P = T3N + T3O; | |
363 T4D = T3N - T3O; | |
364 } | |
365 } | |
366 { | |
367 E T11, T34, TR, T36, T1c, T3R, T1m, T3T, TN, T18, T1i; | |
368 T11 = FMA(TX, T10, TW); | |
369 T34 = FNMS(TS, T10, T33); | |
370 TN = W[21]; | |
371 TR = FNMS(TN, TQ, TM); | |
372 T36 = FMA(TN, TL, T35); | |
373 T18 = W[31]; | |
374 T1c = FNMS(T18, T1b, T17); | |
375 T3R = FMA(T18, T16, T3Q); | |
376 T1i = W[11]; | |
377 T1m = FNMS(T1i, T1l, T1h); | |
378 T3T = FMA(T1i, T1g, T3S); | |
379 { | |
380 E T12, T1n, T51, T52; | |
381 T12 = TR - T11; | |
382 T1n = T1c - T1m; | |
383 T1o = T12 - T1n; | |
384 T2Q = T1n + T12; | |
385 T51 = T3R + T3T; | |
386 T52 = TR + T11; | |
387 T53 = T51 + T52; | |
388 T5w = T51 - T52; | |
389 } | |
390 { | |
391 E T37, T38, T3U, T3V; | |
392 T37 = T34 - T36; | |
393 T38 = T1c + T1m; | |
394 T39 = T37 - T38; | |
395 T3A = T38 + T37; | |
396 T3U = T3R - T3T; | |
397 T3V = T36 + T34; | |
398 T3W = T3U + T3V; | |
399 T4E = T3U - T3V; | |
400 } | |
401 } | |
402 } | |
403 { | |
404 E T4G, T4I, T2N, T2r, T4x, T4y, T4H, T4z; | |
405 { | |
406 E T4C, T4F, T1p, T2q; | |
407 T4C = T4A - T4B; | |
408 T4F = T4D - T4E; | |
409 T4G = FNMS(KP618033988, T4F, T4C); | |
410 T4I = FMA(KP618033988, T4C, T4F); | |
411 T2N = T2E + T2M; | |
412 T1p = TH + T1o; | |
413 T2q = T1Y + T2p; | |
414 T2r = T1p + T2q; | |
415 T4x = FMA(KP250000000, T2r, T2N); | |
416 T4y = T1p - T2q; | |
417 } | |
418 Im[WS(rs, 4)] = KP500000000 * (T2r - T2N); | |
419 T4H = FNMS(KP559016994, T4y, T4x); | |
420 Im[0] = -(KP500000000 * (FMA(KP951056516, T4I, T4H))); | |
421 Im[WS(rs, 8)] = -(KP500000000 * (FNMS(KP951056516, T4I, T4H))); | |
422 T4z = FMA(KP559016994, T4y, T4x); | |
423 Ip[WS(rs, 3)] = KP500000000 * (FNMS(KP951056516, T4G, T4z)); | |
424 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP951056516, T4G, T4z)); | |
425 } | |
426 { | |
427 E T4S, T4U, T4J, T4M, T4N, T4O, T4T, T4P; | |
428 { | |
429 E T4Q, T4R, T4K, T4L; | |
430 T4Q = T2p - T1Y; | |
431 T4R = T1o - TH; | |
432 T4S = FNMS(KP618033988, T4R, T4Q); | |
433 T4U = FMA(KP618033988, T4Q, T4R); | |
434 T4J = T4j - T4k; | |
435 T4K = T4D + T4E; | |
436 T4L = T4A + T4B; | |
437 T4M = T4K + T4L; | |
438 T4N = FNMS(KP250000000, T4M, T4J); | |
439 T4O = T4K - T4L; | |
440 } | |
441 Rm[WS(rs, 4)] = KP500000000 * (T4J + T4M); | |
442 T4T = FMA(KP559016994, T4O, T4N); | |
443 Rm[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T4U, T4T)); | |
444 Rm[0] = KP500000000 * (FNMS(KP951056516, T4U, T4T)); | |
445 T4P = FNMS(KP559016994, T4O, T4N); | |
446 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T4S, T4P)); | |
447 Rp[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T4S, T4P)); | |
448 } | |
449 { | |
450 E T4d, T4f, T2O, T2V, T3G, T3H, T4e, T3I; | |
451 { | |
452 E T3X, T4c, T2R, T2U; | |
453 T3X = T3P - T3W; | |
454 T4c = T44 - T4b; | |
455 T4d = FMA(KP618033988, T4c, T3X); | |
456 T4f = FNMS(KP618033988, T3X, T4c); | |
457 T2O = T2M - T2E; | |
458 T2R = T2P + T2Q; | |
459 T2U = T2S + T2T; | |
460 T2V = T2R + T2U; | |
461 T3G = FNMS(KP250000000, T2V, T2O); | |
462 T3H = T2R - T2U; | |
463 } | |
464 Ip[WS(rs, 5)] = KP500000000 * (T2O + T2V); | |
465 T4e = FNMS(KP559016994, T3H, T3G); | |
466 Im[WS(rs, 2)] = -(KP500000000 * (FMA(KP951056516, T4f, T4e))); | |
467 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP951056516, T4f, T4e))); | |
468 T3I = FMA(KP559016994, T3H, T3G); | |
469 Ip[WS(rs, 1)] = KP500000000 * (FNMS(KP951056516, T4d, T3I)); | |
470 Ip[WS(rs, 9)] = KP500000000 * (FMA(KP951056516, T4d, T3I)); | |
471 } | |
472 { | |
473 E T4u, T4w, T4l, T4o, T4p, T4q, T4v, T4r; | |
474 { | |
475 E T4s, T4t, T4m, T4n; | |
476 T4s = T2P - T2Q; | |
477 T4t = T2S - T2T; | |
478 T4u = FMA(KP618033988, T4t, T4s); | |
479 T4w = FNMS(KP618033988, T4s, T4t); | |
480 T4l = T4j + T4k; | |
481 T4m = T3P + T3W; | |
482 T4n = T44 + T4b; | |
483 T4o = T4m + T4n; | |
484 T4p = FNMS(KP250000000, T4o, T4l); | |
485 T4q = T4m - T4n; | |
486 } | |
487 Rp[WS(rs, 5)] = KP500000000 * (T4l + T4o); | |
488 T4v = FNMS(KP559016994, T4q, T4p); | |
489 Rm[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T4w, T4v)); | |
490 Rm[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T4w, T4v)); | |
491 T4r = FMA(KP559016994, T4q, T4p); | |
492 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T4u, T4r)); | |
493 Rp[WS(rs, 9)] = KP500000000 * (FNMS(KP951056516, T4u, T4r)); | |
494 } | |
495 { | |
496 E T5C, T5E, T3x, T3q, T5t, T5u, T5D, T5v; | |
497 { | |
498 E T5y, T5B, T3a, T3p; | |
499 T5y = T5w - T5x; | |
500 T5B = T5z - T5A; | |
501 T5C = FNMS(KP618033988, T5B, T5y); | |
502 T5E = FMA(KP618033988, T5y, T5B); | |
503 T3x = T3v + T3w; | |
504 T3a = T32 + T39; | |
505 T3p = T3h + T3o; | |
506 T3q = T3a + T3p; | |
507 T5t = FMA(KP250000000, T3q, T3x); | |
508 T5u = T3p - T3a; | |
509 } | |
510 Im[WS(rs, 9)] = KP500000000 * (T3q - T3x); | |
511 T5D = FNMS(KP559016994, T5u, T5t); | |
512 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP951056516, T5E, T5D)); | |
513 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP951056516, T5E, T5D))); | |
514 T5v = FMA(KP559016994, T5u, T5t); | |
515 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T5C, T5v)); | |
516 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP951056516, T5C, T5v))); | |
517 } | |
518 { | |
519 E T5O, T5Q, T5F, T5I, T5J, T5K, T5P, T5L; | |
520 { | |
521 E T5M, T5N, T5G, T5H; | |
522 T5M = T3o - T3h; | |
523 T5N = T39 - T32; | |
524 T5O = FNMS(KP618033988, T5N, T5M); | |
525 T5Q = FMA(KP618033988, T5M, T5N); | |
526 T5F = T4V - T4W; | |
527 T5G = T5x + T5w; | |
528 T5H = T5z + T5A; | |
529 T5I = T5G + T5H; | |
530 T5J = FNMS(KP250000000, T5I, T5F); | |
531 T5K = T5G - T5H; | |
532 } | |
533 Rm[WS(rs, 9)] = KP500000000 * (T5F + T5I); | |
534 T5P = FMA(KP559016994, T5K, T5J); | |
535 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T5Q, T5P)); | |
536 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP951056516, T5Q, T5P)); | |
537 T5L = FNMS(KP559016994, T5K, T5J); | |
538 Rp[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T5O, T5L)); | |
539 Rm[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T5O, T5L)); | |
540 } | |
541 { | |
542 E T5q, T5s, T3y, T3F, T5l, T5m, T5r, T5n; | |
543 { | |
544 E T5o, T5p, T3B, T3E; | |
545 T5o = T50 - T53; | |
546 T5p = T5a - T57; | |
547 T5q = FNMS(KP618033988, T5p, T5o); | |
548 T5s = FMA(KP618033988, T5o, T5p); | |
549 T3y = T3w - T3v; | |
550 T3B = T3z + T3A; | |
551 T3E = T3C + T3D; | |
552 T3F = T3B + T3E; | |
553 T5l = FNMS(KP250000000, T3F, T3y); | |
554 T5m = T3B - T3E; | |
555 } | |
556 Ip[0] = KP500000000 * (T3y + T3F); | |
557 T5r = FNMS(KP559016994, T5m, T5l); | |
558 Ip[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5s, T5r)); | |
559 Im[WS(rs, 7)] = -(KP500000000 * (FNMS(KP951056516, T5s, T5r))); | |
560 T5n = FMA(KP559016994, T5m, T5l); | |
561 Ip[WS(rs, 4)] = KP500000000 * (FMA(KP951056516, T5q, T5n)); | |
562 Im[WS(rs, 3)] = -(KP500000000 * (FNMS(KP951056516, T5q, T5n))); | |
563 } | |
564 { | |
565 E T5i, T5k, T4X, T5c, T5d, T5e, T5j, T5f; | |
566 { | |
567 E T5g, T5h, T54, T5b; | |
568 T5g = T3z - T3A; | |
569 T5h = T3C - T3D; | |
570 T5i = FMA(KP618033988, T5h, T5g); | |
571 T5k = FNMS(KP618033988, T5g, T5h); | |
572 T4X = T4V + T4W; | |
573 T54 = T50 + T53; | |
574 T5b = T57 + T5a; | |
575 T5c = T54 + T5b; | |
576 T5d = FNMS(KP250000000, T5c, T4X); | |
577 T5e = T54 - T5b; | |
578 } | |
579 Rp[0] = KP500000000 * (T4X + T5c); | |
580 T5j = FNMS(KP559016994, T5e, T5d); | |
581 Rp[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5k, T5j)); | |
582 Rm[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T5k, T5j)); | |
583 T5f = FMA(KP559016994, T5e, T5d); | |
584 Rp[WS(rs, 4)] = KP500000000 * (FNMS(KP951056516, T5i, T5f)); | |
585 Rm[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T5i, T5f)); | |
586 } | |
587 } | |
588 } | |
589 } | |
590 | |
591 static const tw_instr twinstr[] = { | |
592 {TW_FULL, 1, 20}, | |
593 {TW_NEXT, 1, 0} | |
594 }; | |
595 | |
596 static const hc2c_desc desc = { 20, "hc2cfdft_20", twinstr, &GENUS, {176, 78, 110, 0} }; | |
597 | |
598 void X(codelet_hc2cfdft_20) (planner *p) { | |
599 X(khc2c_register) (p, hc2cfdft_20, &desc, HC2C_VIA_DFT); | |
600 } | |
601 #else | |
602 | |
603 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hc2cfdft_20 -include rdft/scalar/hc2cf.h */ | |
604 | |
605 /* | |
606 * This function contains 286 FP additions, 140 FP multiplications, | |
607 * (or, 224 additions, 78 multiplications, 62 fused multiply/add), | |
608 * 98 stack variables, 5 constants, and 80 memory accesses | |
609 */ | |
610 #include "rdft/scalar/hc2cf.h" | |
611 | |
612 static void hc2cfdft_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms) | |
613 { | |
614 DK(KP125000000, +0.125000000000000000000000000000000000000000000); | |
615 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
616 DK(KP279508497, +0.279508497187473712051146708591409529430077295); | |
617 DK(KP293892626, +0.293892626146236564584352977319536384298826219); | |
618 DK(KP475528258, +0.475528258147576786058219666689691071702849317); | |
619 { | |
620 INT m; | |
621 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 38, MAKE_VOLATILE_STRIDE(80, rs)) { | |
622 E T12, T2w, T4o, T4V, T2H, T3a, T4y, T4Y, T1z, T2v, T25, T2y, T2s, T2z, T4v; | |
623 E T4X, T4r, T4U, T3A, T3Z, T2X, T37, T3k, T41, T2M, T39, T3v, T3Y, T2S, T36; | |
624 E T3p, T42, Td, T4G, T33, T3N, Tw, T4H, T32, T3O; | |
625 { | |
626 E T3, T3L, T1x, T2V, Th, Tl, TC, T3g, Tq, Tu, TH, T3h, T7, Tb, T1q; | |
627 E T2U, TR, T2P, T1F, T3r, T23, T2K, T2f, T3y, T1k, T3m, T2q, T2E, T10, T2Q; | |
628 E T1K, T3s, T1U, T2J, T2a, T3x, T1b, T3l, T2l, T2D; | |
629 { | |
630 E T1, T2, T1s, T1u, T1v, T1w, T1r, T1t; | |
631 T1 = Ip[0]; | |
632 T2 = Im[0]; | |
633 T1s = T1 + T2; | |
634 T1u = Rp[0]; | |
635 T1v = Rm[0]; | |
636 T1w = T1u - T1v; | |
637 T3 = T1 - T2; | |
638 T3L = T1u + T1v; | |
639 T1r = W[0]; | |
640 T1t = W[1]; | |
641 T1x = FNMS(T1t, T1w, T1r * T1s); | |
642 T2V = FMA(T1r, T1w, T1t * T1s); | |
643 } | |
644 { | |
645 E Tf, Tg, Tz, Tj, Tk, TB, Ty, TA; | |
646 Tf = Ip[WS(rs, 2)]; | |
647 Tg = Im[WS(rs, 2)]; | |
648 Tz = Tf - Tg; | |
649 Tj = Rp[WS(rs, 2)]; | |
650 Tk = Rm[WS(rs, 2)]; | |
651 TB = Tj + Tk; | |
652 Th = Tf + Tg; | |
653 Tl = Tj - Tk; | |
654 Ty = W[6]; | |
655 TA = W[7]; | |
656 TC = FNMS(TA, TB, Ty * Tz); | |
657 T3g = FMA(TA, Tz, Ty * TB); | |
658 } | |
659 { | |
660 E To, Tp, TE, Ts, Tt, TG, TD, TF; | |
661 To = Ip[WS(rs, 7)]; | |
662 Tp = Im[WS(rs, 7)]; | |
663 TE = To - Tp; | |
664 Ts = Rp[WS(rs, 7)]; | |
665 Tt = Rm[WS(rs, 7)]; | |
666 TG = Ts + Tt; | |
667 Tq = To + Tp; | |
668 Tu = Ts - Tt; | |
669 TD = W[26]; | |
670 TF = W[27]; | |
671 TH = FNMS(TF, TG, TD * TE); | |
672 T3h = FMA(TF, TE, TD * TG); | |
673 } | |
674 { | |
675 E T5, T6, T1n, T9, Ta, T1p, T1m, T1o; | |
676 T5 = Ip[WS(rs, 5)]; | |
677 T6 = Im[WS(rs, 5)]; | |
678 T1n = T5 + T6; | |
679 T9 = Rp[WS(rs, 5)]; | |
680 Ta = Rm[WS(rs, 5)]; | |
681 T1p = T9 - Ta; | |
682 T7 = T5 - T6; | |
683 Tb = T9 + Ta; | |
684 T1m = W[20]; | |
685 T1o = W[21]; | |
686 T1q = FNMS(T1o, T1p, T1m * T1n); | |
687 T2U = FMA(T1m, T1p, T1o * T1n); | |
688 } | |
689 { | |
690 E TM, T1C, TQ, T1E; | |
691 { | |
692 E TK, TL, TO, TP; | |
693 TK = Ip[WS(rs, 4)]; | |
694 TL = Im[WS(rs, 4)]; | |
695 TM = TK + TL; | |
696 T1C = TK - TL; | |
697 TO = Rp[WS(rs, 4)]; | |
698 TP = Rm[WS(rs, 4)]; | |
699 TQ = TO - TP; | |
700 T1E = TO + TP; | |
701 } | |
702 { | |
703 E TJ, TN, T1B, T1D; | |
704 TJ = W[16]; | |
705 TN = W[17]; | |
706 TR = FNMS(TN, TQ, TJ * TM); | |
707 T2P = FMA(TN, TM, TJ * TQ); | |
708 T1B = W[14]; | |
709 T1D = W[15]; | |
710 T1F = FNMS(T1D, T1E, T1B * T1C); | |
711 T3r = FMA(T1D, T1C, T1B * T1E); | |
712 } | |
713 } | |
714 { | |
715 E T1Y, T2c, T22, T2e; | |
716 { | |
717 E T1W, T1X, T20, T21; | |
718 T1W = Ip[WS(rs, 1)]; | |
719 T1X = Im[WS(rs, 1)]; | |
720 T1Y = T1W + T1X; | |
721 T2c = T1W - T1X; | |
722 T20 = Rp[WS(rs, 1)]; | |
723 T21 = Rm[WS(rs, 1)]; | |
724 T22 = T20 - T21; | |
725 T2e = T20 + T21; | |
726 } | |
727 { | |
728 E T1V, T1Z, T2b, T2d; | |
729 T1V = W[4]; | |
730 T1Z = W[5]; | |
731 T23 = FNMS(T1Z, T22, T1V * T1Y); | |
732 T2K = FMA(T1Z, T1Y, T1V * T22); | |
733 T2b = W[2]; | |
734 T2d = W[3]; | |
735 T2f = FNMS(T2d, T2e, T2b * T2c); | |
736 T3y = FMA(T2d, T2c, T2b * T2e); | |
737 } | |
738 } | |
739 { | |
740 E T1f, T2n, T1j, T2p; | |
741 { | |
742 E T1d, T1e, T1h, T1i; | |
743 T1d = Ip[WS(rs, 3)]; | |
744 T1e = Im[WS(rs, 3)]; | |
745 T1f = T1d - T1e; | |
746 T2n = T1d + T1e; | |
747 T1h = Rp[WS(rs, 3)]; | |
748 T1i = Rm[WS(rs, 3)]; | |
749 T1j = T1h + T1i; | |
750 T2p = T1h - T1i; | |
751 } | |
752 { | |
753 E T1c, T1g, T2m, T2o; | |
754 T1c = W[10]; | |
755 T1g = W[11]; | |
756 T1k = FNMS(T1g, T1j, T1c * T1f); | |
757 T3m = FMA(T1c, T1j, T1g * T1f); | |
758 T2m = W[12]; | |
759 T2o = W[13]; | |
760 T2q = FNMS(T2o, T2p, T2m * T2n); | |
761 T2E = FMA(T2m, T2p, T2o * T2n); | |
762 } | |
763 } | |
764 { | |
765 E TV, T1H, TZ, T1J; | |
766 { | |
767 E TT, TU, TX, TY; | |
768 TT = Ip[WS(rs, 9)]; | |
769 TU = Im[WS(rs, 9)]; | |
770 TV = TT + TU; | |
771 T1H = TT - TU; | |
772 TX = Rp[WS(rs, 9)]; | |
773 TY = Rm[WS(rs, 9)]; | |
774 TZ = TX - TY; | |
775 T1J = TX + TY; | |
776 } | |
777 { | |
778 E TS, TW, T1G, T1I; | |
779 TS = W[36]; | |
780 TW = W[37]; | |
781 T10 = FNMS(TW, TZ, TS * TV); | |
782 T2Q = FMA(TW, TV, TS * TZ); | |
783 T1G = W[34]; | |
784 T1I = W[35]; | |
785 T1K = FNMS(T1I, T1J, T1G * T1H); | |
786 T3s = FMA(T1I, T1H, T1G * T1J); | |
787 } | |
788 } | |
789 { | |
790 E T1P, T27, T1T, T29; | |
791 { | |
792 E T1N, T1O, T1R, T1S; | |
793 T1N = Ip[WS(rs, 6)]; | |
794 T1O = Im[WS(rs, 6)]; | |
795 T1P = T1N + T1O; | |
796 T27 = T1N - T1O; | |
797 T1R = Rp[WS(rs, 6)]; | |
798 T1S = Rm[WS(rs, 6)]; | |
799 T1T = T1R - T1S; | |
800 T29 = T1R + T1S; | |
801 } | |
802 { | |
803 E T1M, T1Q, T26, T28; | |
804 T1M = W[24]; | |
805 T1Q = W[25]; | |
806 T1U = FNMS(T1Q, T1T, T1M * T1P); | |
807 T2J = FMA(T1Q, T1P, T1M * T1T); | |
808 T26 = W[22]; | |
809 T28 = W[23]; | |
810 T2a = FNMS(T28, T29, T26 * T27); | |
811 T3x = FMA(T28, T27, T26 * T29); | |
812 } | |
813 } | |
814 { | |
815 E T16, T2k, T1a, T2i; | |
816 { | |
817 E T14, T15, T18, T19; | |
818 T14 = Ip[WS(rs, 8)]; | |
819 T15 = Im[WS(rs, 8)]; | |
820 T16 = T14 - T15; | |
821 T2k = T14 + T15; | |
822 T18 = Rp[WS(rs, 8)]; | |
823 T19 = Rm[WS(rs, 8)]; | |
824 T1a = T18 + T19; | |
825 T2i = T19 - T18; | |
826 } | |
827 { | |
828 E T13, T17, T2h, T2j; | |
829 T13 = W[30]; | |
830 T17 = W[31]; | |
831 T1b = FNMS(T17, T1a, T13 * T16); | |
832 T3l = FMA(T13, T1a, T17 * T16); | |
833 T2h = W[33]; | |
834 T2j = W[32]; | |
835 T2l = FMA(T2h, T2i, T2j * T2k); | |
836 T2D = FNMS(T2h, T2k, T2j * T2i); | |
837 } | |
838 } | |
839 { | |
840 E T2g, T2r, T3n, T3o; | |
841 { | |
842 E TI, T11, T4m, T4n; | |
843 TI = TC - TH; | |
844 T11 = TR - T10; | |
845 T12 = TI - T11; | |
846 T2w = TI + T11; | |
847 T4m = T3g + T3h; | |
848 T4n = TR + T10; | |
849 T4o = T4m + T4n; | |
850 T4V = T4m - T4n; | |
851 } | |
852 { | |
853 E T2F, T2G, T4w, T4x; | |
854 T2F = T2D - T2E; | |
855 T2G = T2a + T2f; | |
856 T2H = T2F - T2G; | |
857 T3a = T2F + T2G; | |
858 T4w = T2l + T2q; | |
859 T4x = T3x + T3y; | |
860 T4y = T4w + T4x; | |
861 T4Y = T4x - T4w; | |
862 } | |
863 { | |
864 E T1l, T1y, T1L, T24; | |
865 T1l = T1b - T1k; | |
866 T1y = T1q - T1x; | |
867 T1z = T1l + T1y; | |
868 T2v = T1y - T1l; | |
869 T1L = T1F - T1K; | |
870 T24 = T1U - T23; | |
871 T25 = T1L - T24; | |
872 T2y = T1L + T24; | |
873 } | |
874 T2g = T2a - T2f; | |
875 T2r = T2l - T2q; | |
876 T2s = T2g - T2r; | |
877 T2z = T2r + T2g; | |
878 { | |
879 E T4t, T4u, T4p, T4q; | |
880 T4t = T3r + T3s; | |
881 T4u = T1U + T23; | |
882 T4v = T4t + T4u; | |
883 T4X = T4t - T4u; | |
884 T4p = T3l + T3m; | |
885 T4q = T1q + T1x; | |
886 T4r = T4p + T4q; | |
887 T4U = T4p - T4q; | |
888 } | |
889 { | |
890 E T3w, T3z, T2T, T2W; | |
891 T3w = T2D + T2E; | |
892 T3z = T3x - T3y; | |
893 T3A = T3w + T3z; | |
894 T3Z = T3z - T3w; | |
895 T2T = T1b + T1k; | |
896 T2W = T2U + T2V; | |
897 T2X = T2T + T2W; | |
898 T37 = T2T - T2W; | |
899 } | |
900 { | |
901 E T3i, T3j, T2I, T2L; | |
902 T3i = T3g - T3h; | |
903 T3j = T2Q - T2P; | |
904 T3k = T3i + T3j; | |
905 T41 = T3i - T3j; | |
906 T2I = T1F + T1K; | |
907 T2L = T2J + T2K; | |
908 T2M = T2I + T2L; | |
909 T39 = T2I - T2L; | |
910 } | |
911 { | |
912 E T3t, T3u, T2O, T2R; | |
913 T3t = T3r - T3s; | |
914 T3u = T2K - T2J; | |
915 T3v = T3t + T3u; | |
916 T3Y = T3t - T3u; | |
917 T2O = TC + TH; | |
918 T2R = T2P + T2Q; | |
919 T2S = T2O + T2R; | |
920 T36 = T2O - T2R; | |
921 } | |
922 T3n = T3l - T3m; | |
923 T3o = T2U - T2V; | |
924 T3p = T3n + T3o; | |
925 T42 = T3n - T3o; | |
926 { | |
927 E Tc, T3M, T4, T8; | |
928 T4 = W[18]; | |
929 T8 = W[19]; | |
930 Tc = FNMS(T8, Tb, T4 * T7); | |
931 T3M = FMA(T4, Tb, T8 * T7); | |
932 Td = T3 - Tc; | |
933 T4G = T3L + T3M; | |
934 T33 = Tc + T3; | |
935 T3N = T3L - T3M; | |
936 } | |
937 { | |
938 E Tm, T30, Tv, T31; | |
939 { | |
940 E Te, Ti, Tn, Tr; | |
941 Te = W[8]; | |
942 Ti = W[9]; | |
943 Tm = FNMS(Ti, Tl, Te * Th); | |
944 T30 = FMA(Ti, Th, Te * Tl); | |
945 Tn = W[28]; | |
946 Tr = W[29]; | |
947 Tv = FNMS(Tr, Tu, Tn * Tq); | |
948 T31 = FMA(Tr, Tq, Tn * Tu); | |
949 } | |
950 Tw = Tm - Tv; | |
951 T4H = Tm + Tv; | |
952 T32 = T30 + T31; | |
953 T3O = T31 - T30; | |
954 } | |
955 } | |
956 } | |
957 { | |
958 E T3C, T3E, Tx, T2u, T3d, T3e, T3D, T3f; | |
959 { | |
960 E T3q, T3B, T1A, T2t; | |
961 T3q = T3k - T3p; | |
962 T3B = T3v - T3A; | |
963 T3C = FMA(KP475528258, T3q, KP293892626 * T3B); | |
964 T3E = FNMS(KP293892626, T3q, KP475528258 * T3B); | |
965 Tx = Td - Tw; | |
966 T1A = T12 + T1z; | |
967 T2t = T25 + T2s; | |
968 T2u = T1A + T2t; | |
969 T3d = KP279508497 * (T1A - T2t); | |
970 T3e = FNMS(KP125000000, T2u, KP500000000 * Tx); | |
971 } | |
972 Ip[WS(rs, 5)] = KP500000000 * (Tx + T2u); | |
973 T3D = T3d - T3e; | |
974 Im[WS(rs, 2)] = T3D - T3E; | |
975 Im[WS(rs, 6)] = T3D + T3E; | |
976 T3f = T3d + T3e; | |
977 Ip[WS(rs, 1)] = T3f - T3C; | |
978 Ip[WS(rs, 9)] = T3f + T3C; | |
979 } | |
980 { | |
981 E T3H, T3T, T3P, T3Q, T3K, T3R, T3U, T3S; | |
982 { | |
983 E T3F, T3G, T3I, T3J; | |
984 T3F = T12 - T1z; | |
985 T3G = T25 - T2s; | |
986 T3H = FMA(KP475528258, T3F, KP293892626 * T3G); | |
987 T3T = FNMS(KP293892626, T3F, KP475528258 * T3G); | |
988 T3P = T3N + T3O; | |
989 T3I = T3k + T3p; | |
990 T3J = T3v + T3A; | |
991 T3Q = T3I + T3J; | |
992 T3K = KP279508497 * (T3I - T3J); | |
993 T3R = FNMS(KP125000000, T3Q, KP500000000 * T3P); | |
994 } | |
995 Rp[WS(rs, 5)] = KP500000000 * (T3P + T3Q); | |
996 T3U = T3R - T3K; | |
997 Rm[WS(rs, 6)] = T3T + T3U; | |
998 Rm[WS(rs, 2)] = T3U - T3T; | |
999 T3S = T3K + T3R; | |
1000 Rp[WS(rs, 1)] = T3H + T3S; | |
1001 Rp[WS(rs, 9)] = T3S - T3H; | |
1002 } | |
1003 { | |
1004 E T44, T46, T2C, T2B, T3V, T3W, T45, T3X; | |
1005 { | |
1006 E T40, T43, T2x, T2A; | |
1007 T40 = T3Y - T3Z; | |
1008 T43 = T41 - T42; | |
1009 T44 = FNMS(KP293892626, T43, KP475528258 * T40); | |
1010 T46 = FMA(KP475528258, T43, KP293892626 * T40); | |
1011 T2C = Tw + Td; | |
1012 T2x = T2v - T2w; | |
1013 T2A = T2y + T2z; | |
1014 T2B = T2x - T2A; | |
1015 T3V = FMA(KP500000000, T2C, KP125000000 * T2B); | |
1016 T3W = KP279508497 * (T2x + T2A); | |
1017 } | |
1018 Im[WS(rs, 4)] = KP500000000 * (T2B - T2C); | |
1019 T45 = T3W - T3V; | |
1020 Im[0] = T45 - T46; | |
1021 Im[WS(rs, 8)] = T45 + T46; | |
1022 T3X = T3V + T3W; | |
1023 Ip[WS(rs, 3)] = T3X - T44; | |
1024 Ip[WS(rs, 7)] = T3X + T44; | |
1025 } | |
1026 { | |
1027 E T49, T4h, T4a, T4d, T4e, T4f, T4i, T4g; | |
1028 { | |
1029 E T47, T48, T4b, T4c; | |
1030 T47 = T2y - T2z; | |
1031 T48 = T2w + T2v; | |
1032 T49 = FNMS(KP293892626, T48, KP475528258 * T47); | |
1033 T4h = FMA(KP475528258, T48, KP293892626 * T47); | |
1034 T4a = T3N - T3O; | |
1035 T4b = T41 + T42; | |
1036 T4c = T3Y + T3Z; | |
1037 T4d = T4b + T4c; | |
1038 T4e = FNMS(KP125000000, T4d, KP500000000 * T4a); | |
1039 T4f = KP279508497 * (T4b - T4c); | |
1040 } | |
1041 Rm[WS(rs, 4)] = KP500000000 * (T4a + T4d); | |
1042 T4i = T4f + T4e; | |
1043 Rm[WS(rs, 8)] = T4h + T4i; | |
1044 Rm[0] = T4i - T4h; | |
1045 T4g = T4e - T4f; | |
1046 Rp[WS(rs, 3)] = T49 + T4g; | |
1047 Rp[WS(rs, 7)] = T4g - T49; | |
1048 } | |
1049 { | |
1050 E T50, T52, T34, T2Z, T4R, T4S, T51, T4T; | |
1051 { | |
1052 E T4W, T4Z, T2N, T2Y; | |
1053 T4W = T4U - T4V; | |
1054 T4Z = T4X - T4Y; | |
1055 T50 = FNMS(KP293892626, T4Z, KP475528258 * T4W); | |
1056 T52 = FMA(KP293892626, T4W, KP475528258 * T4Z); | |
1057 T34 = T32 + T33; | |
1058 T2N = T2H - T2M; | |
1059 T2Y = T2S + T2X; | |
1060 T2Z = T2N - T2Y; | |
1061 T4R = FMA(KP500000000, T34, KP125000000 * T2Z); | |
1062 T4S = KP279508497 * (T2Y + T2N); | |
1063 } | |
1064 Im[WS(rs, 9)] = KP500000000 * (T2Z - T34); | |
1065 T51 = T4R - T4S; | |
1066 Ip[WS(rs, 2)] = T51 + T52; | |
1067 Im[WS(rs, 1)] = T52 - T51; | |
1068 T4T = T4R + T4S; | |
1069 Ip[WS(rs, 6)] = T4T + T50; | |
1070 Im[WS(rs, 5)] = T50 - T4T; | |
1071 } | |
1072 { | |
1073 E T5c, T5d, T53, T56, T57, T58, T5e, T59; | |
1074 { | |
1075 E T5a, T5b, T54, T55; | |
1076 T5a = T2M + T2H; | |
1077 T5b = T2S - T2X; | |
1078 T5c = FNMS(KP293892626, T5b, KP475528258 * T5a); | |
1079 T5d = FMA(KP475528258, T5b, KP293892626 * T5a); | |
1080 T53 = T4G - T4H; | |
1081 T54 = T4V + T4U; | |
1082 T55 = T4X + T4Y; | |
1083 T56 = T54 + T55; | |
1084 T57 = FNMS(KP125000000, T56, KP500000000 * T53); | |
1085 T58 = KP279508497 * (T54 - T55); | |
1086 } | |
1087 Rm[WS(rs, 9)] = KP500000000 * (T53 + T56); | |
1088 T5e = T58 + T57; | |
1089 Rp[WS(rs, 6)] = T5d + T5e; | |
1090 Rm[WS(rs, 5)] = T5e - T5d; | |
1091 T59 = T57 - T58; | |
1092 Rp[WS(rs, 2)] = T59 - T5c; | |
1093 Rm[WS(rs, 1)] = T5c + T59; | |
1094 } | |
1095 { | |
1096 E T4A, T4C, T35, T3c, T4j, T4k, T4B, T4l; | |
1097 { | |
1098 E T4s, T4z, T38, T3b; | |
1099 T4s = T4o - T4r; | |
1100 T4z = T4v - T4y; | |
1101 T4A = FNMS(KP475528258, T4z, KP293892626 * T4s); | |
1102 T4C = FMA(KP475528258, T4s, KP293892626 * T4z); | |
1103 T35 = T33 - T32; | |
1104 T38 = T36 + T37; | |
1105 T3b = T39 + T3a; | |
1106 T3c = T38 + T3b; | |
1107 T4j = FNMS(KP125000000, T3c, KP500000000 * T35); | |
1108 T4k = KP279508497 * (T38 - T3b); | |
1109 } | |
1110 Ip[0] = KP500000000 * (T35 + T3c); | |
1111 T4B = T4k + T4j; | |
1112 Ip[WS(rs, 4)] = T4B + T4C; | |
1113 Im[WS(rs, 3)] = T4C - T4B; | |
1114 T4l = T4j - T4k; | |
1115 Ip[WS(rs, 8)] = T4l + T4A; | |
1116 Im[WS(rs, 7)] = T4A - T4l; | |
1117 } | |
1118 { | |
1119 E T4O, T4P, T4I, T4J, T4F, T4K, T4Q, T4L; | |
1120 { | |
1121 E T4M, T4N, T4D, T4E; | |
1122 T4M = T36 - T37; | |
1123 T4N = T39 - T3a; | |
1124 T4O = FMA(KP475528258, T4M, KP293892626 * T4N); | |
1125 T4P = FNMS(KP293892626, T4M, KP475528258 * T4N); | |
1126 T4I = T4G + T4H; | |
1127 T4D = T4o + T4r; | |
1128 T4E = T4v + T4y; | |
1129 T4J = T4D + T4E; | |
1130 T4F = KP279508497 * (T4D - T4E); | |
1131 T4K = FNMS(KP125000000, T4J, KP500000000 * T4I); | |
1132 } | |
1133 Rp[0] = KP500000000 * (T4I + T4J); | |
1134 T4Q = T4K - T4F; | |
1135 Rp[WS(rs, 8)] = T4P + T4Q; | |
1136 Rm[WS(rs, 7)] = T4Q - T4P; | |
1137 T4L = T4F + T4K; | |
1138 Rp[WS(rs, 4)] = T4L - T4O; | |
1139 Rm[WS(rs, 3)] = T4O + T4L; | |
1140 } | |
1141 } | |
1142 } | |
1143 } | |
1144 | |
1145 static const tw_instr twinstr[] = { | |
1146 {TW_FULL, 1, 20}, | |
1147 {TW_NEXT, 1, 0} | |
1148 }; | |
1149 | |
1150 static const hc2c_desc desc = { 20, "hc2cfdft_20", twinstr, &GENUS, {224, 78, 62, 0} }; | |
1151 | |
1152 void X(codelet_hc2cfdft_20) (planner *p) { | |
1153 X(khc2c_register) (p, hc2cfdft_20, &desc, HC2C_VIA_DFT); | |
1154 } | |
1155 #endif |