Mercurial > hg > js-dsp-test
comparison fft/fftw/fftw-3.3.4/rdft/scalar/r2cf/hf2_16.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 13:49:12 EST 2014 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 16 -dit -name hf2_16 -include hf.h */ | |
29 | |
30 /* | |
31 * This function contains 196 FP additions, 134 FP multiplications, | |
32 * (or, 104 additions, 42 multiplications, 92 fused multiply/add), | |
33 * 106 stack variables, 3 constants, and 64 memory accesses | |
34 */ | |
35 #include "hf.h" | |
36 | |
37 static void hf2_16(R *cr, R *ci, 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 { | |
43 INT m; | |
44 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(32, rs)) { | |
45 E T35, T32; | |
46 { | |
47 E T2, Tf, TM, TO, T3, Tg, TN, TS, T4, Tp, T6, T5, Th; | |
48 T2 = W[0]; | |
49 Tf = W[2]; | |
50 TM = W[6]; | |
51 TO = W[7]; | |
52 T3 = W[4]; | |
53 Tg = T2 * Tf; | |
54 TN = T2 * TM; | |
55 TS = T2 * TO; | |
56 T4 = T2 * T3; | |
57 Tp = Tf * T3; | |
58 T6 = W[5]; | |
59 T5 = W[1]; | |
60 Th = W[3]; | |
61 { | |
62 E TZ, Te, T1U, T3A, T3M, T2w, T1G, T2I, T3h, T1R, T2D, T2B, T3i, Tx, T3L; | |
63 E T1Z, T3w, TL, T21, T26, T38, T1d, T2h, T2s, T3c, T1s, T2t, T2m, T3d, TX; | |
64 E T10, TV, T2a, TY, T2b; | |
65 { | |
66 E TF, TP, TT, Tq, TW, Tz, Tu, TI, TC, T1m, T1f, T1p, T1j, Tr, Ts; | |
67 E Tv, To, T1W; | |
68 { | |
69 E Ti, Tm, T1L, T1O, T1D, T1A, T1x, T2G, T1F, T2F; | |
70 { | |
71 E T1, T7, Tb, T3z, T8, T1z, T9, Tc; | |
72 { | |
73 E T1i, T1e, T1C, T1y, Tt, Ta, Tl; | |
74 T1 = cr[0]; | |
75 Tt = Tf * T6; | |
76 Ta = T2 * T6; | |
77 T7 = FMA(T5, T6, T4); | |
78 TF = FNMS(T5, T6, T4); | |
79 TP = FMA(T5, TO, TN); | |
80 TT = FNMS(T5, TM, TS); | |
81 Tq = FNMS(Th, T6, Tp); | |
82 TW = FMA(Th, T6, Tp); | |
83 Tz = FMA(T5, Th, Tg); | |
84 Ti = FNMS(T5, Th, Tg); | |
85 Tl = T2 * Th; | |
86 Tu = FMA(Th, T3, Tt); | |
87 TZ = FNMS(Th, T3, Tt); | |
88 TI = FMA(T5, T3, Ta); | |
89 Tb = FNMS(T5, T3, Ta); | |
90 T1i = Ti * T6; | |
91 T1e = Ti * T3; | |
92 T1C = Tz * T6; | |
93 T1y = Tz * T3; | |
94 Tm = FMA(T5, Tf, Tl); | |
95 TC = FNMS(T5, Tf, Tl); | |
96 T3z = ci[0]; | |
97 T8 = cr[WS(rs, 8)]; | |
98 T1m = FNMS(Tm, T6, T1e); | |
99 T1f = FMA(Tm, T6, T1e); | |
100 T1p = FMA(Tm, T3, T1i); | |
101 T1j = FNMS(Tm, T3, T1i); | |
102 T1L = FNMS(TC, T6, T1y); | |
103 T1z = FMA(TC, T6, T1y); | |
104 T1O = FMA(TC, T3, T1C); | |
105 T1D = FNMS(TC, T3, T1C); | |
106 T9 = T7 * T8; | |
107 Tc = ci[WS(rs, 8)]; | |
108 } | |
109 { | |
110 E T1u, T1w, T1v, T2E, T3y, T1B, T1E, Td, T3x; | |
111 T1u = cr[WS(rs, 15)]; | |
112 T1w = ci[WS(rs, 15)]; | |
113 T1A = cr[WS(rs, 7)]; | |
114 Td = FMA(Tb, Tc, T9); | |
115 T3x = T7 * Tc; | |
116 T1v = TM * T1u; | |
117 T2E = TM * T1w; | |
118 Te = T1 + Td; | |
119 T1U = T1 - Td; | |
120 T3y = FNMS(Tb, T8, T3x); | |
121 T1B = T1z * T1A; | |
122 T1E = ci[WS(rs, 7)]; | |
123 T1x = FMA(TO, T1w, T1v); | |
124 T3A = T3y + T3z; | |
125 T3M = T3z - T3y; | |
126 T2G = T1z * T1E; | |
127 T1F = FMA(T1D, T1E, T1B); | |
128 T2F = FNMS(TO, T1u, T2E); | |
129 } | |
130 } | |
131 { | |
132 E T1H, T1I, T1J, T1M, T1P, T2H; | |
133 T1H = cr[WS(rs, 3)]; | |
134 T2H = FNMS(T1D, T1A, T2G); | |
135 T2w = T1x - T1F; | |
136 T1G = T1x + T1F; | |
137 T1I = Tf * T1H; | |
138 T2I = T2F - T2H; | |
139 T3h = T2F + T2H; | |
140 T1J = ci[WS(rs, 3)]; | |
141 T1M = cr[WS(rs, 11)]; | |
142 T1P = ci[WS(rs, 11)]; | |
143 { | |
144 E Tj, Tk, Tn, T1V; | |
145 { | |
146 E T1K, T2y, T1Q, T2A, T2x, T1N, T2z; | |
147 Tj = cr[WS(rs, 4)]; | |
148 T1K = FMA(Th, T1J, T1I); | |
149 T2x = Tf * T1J; | |
150 T1N = T1L * T1M; | |
151 T2z = T1L * T1P; | |
152 Tk = Ti * Tj; | |
153 T2y = FNMS(Th, T1H, T2x); | |
154 T1Q = FMA(T1O, T1P, T1N); | |
155 T2A = FNMS(T1O, T1M, T2z); | |
156 Tn = ci[WS(rs, 4)]; | |
157 Tr = cr[WS(rs, 12)]; | |
158 T1R = T1K + T1Q; | |
159 T2D = T1Q - T1K; | |
160 T2B = T2y - T2A; | |
161 T3i = T2y + T2A; | |
162 T1V = Ti * Tn; | |
163 Ts = Tq * Tr; | |
164 Tv = ci[WS(rs, 12)]; | |
165 } | |
166 To = FMA(Tm, Tn, Tk); | |
167 T1W = FNMS(Tm, Tj, T1V); | |
168 } | |
169 } | |
170 } | |
171 { | |
172 E T19, T1b, T18, T2p, T1a, T2q; | |
173 { | |
174 E TE, T23, TK, T25; | |
175 { | |
176 E TA, TD, TB, T22, TG, TJ, TH, T24, T1Y, Tw, T1X; | |
177 TA = cr[WS(rs, 2)]; | |
178 Tw = FMA(Tu, Tv, Ts); | |
179 T1X = Tq * Tv; | |
180 TD = ci[WS(rs, 2)]; | |
181 TB = Tz * TA; | |
182 Tx = To + Tw; | |
183 T3L = To - Tw; | |
184 T1Y = FNMS(Tu, Tr, T1X); | |
185 T22 = Tz * TD; | |
186 TG = cr[WS(rs, 10)]; | |
187 TJ = ci[WS(rs, 10)]; | |
188 T1Z = T1W - T1Y; | |
189 T3w = T1W + T1Y; | |
190 TH = TF * TG; | |
191 T24 = TF * TJ; | |
192 TE = FMA(TC, TD, TB); | |
193 T23 = FNMS(TC, TA, T22); | |
194 TK = FMA(TI, TJ, TH); | |
195 T25 = FNMS(TI, TG, T24); | |
196 } | |
197 { | |
198 E T15, T17, T16, T2o; | |
199 T15 = cr[WS(rs, 1)]; | |
200 T17 = ci[WS(rs, 1)]; | |
201 TL = TE + TK; | |
202 T21 = TE - TK; | |
203 T26 = T23 - T25; | |
204 T38 = T23 + T25; | |
205 T16 = T2 * T15; | |
206 T2o = T2 * T17; | |
207 T19 = cr[WS(rs, 9)]; | |
208 T1b = ci[WS(rs, 9)]; | |
209 T18 = FMA(T5, T17, T16); | |
210 T2p = FNMS(T5, T15, T2o); | |
211 T1a = T3 * T19; | |
212 T2q = T3 * T1b; | |
213 } | |
214 } | |
215 { | |
216 E T1n, T1q, T1l, T2j, T1o, T2k; | |
217 { | |
218 E T1g, T1k, T1h, T2i, T1c, T2r; | |
219 T1g = cr[WS(rs, 5)]; | |
220 T1k = ci[WS(rs, 5)]; | |
221 T1c = FMA(T6, T1b, T1a); | |
222 T2r = FNMS(T6, T19, T2q); | |
223 T1h = T1f * T1g; | |
224 T2i = T1f * T1k; | |
225 T1d = T18 + T1c; | |
226 T2h = T18 - T1c; | |
227 T2s = T2p - T2r; | |
228 T3c = T2p + T2r; | |
229 T1n = cr[WS(rs, 13)]; | |
230 T1q = ci[WS(rs, 13)]; | |
231 T1l = FMA(T1j, T1k, T1h); | |
232 T2j = FNMS(T1j, T1g, T2i); | |
233 T1o = T1m * T1n; | |
234 T2k = T1m * T1q; | |
235 } | |
236 { | |
237 E TQ, TU, TR, T29, T1r, T2l; | |
238 TQ = cr[WS(rs, 14)]; | |
239 TU = ci[WS(rs, 14)]; | |
240 T1r = FMA(T1p, T1q, T1o); | |
241 T2l = FNMS(T1p, T1n, T2k); | |
242 TR = TP * TQ; | |
243 T29 = TP * TU; | |
244 T1s = T1l + T1r; | |
245 T2t = T1l - T1r; | |
246 T2m = T2j - T2l; | |
247 T3d = T2j + T2l; | |
248 TX = cr[WS(rs, 6)]; | |
249 T10 = ci[WS(rs, 6)]; | |
250 TV = FMA(TT, TU, TR); | |
251 T2a = FNMS(TT, TQ, T29); | |
252 TY = TW * TX; | |
253 T2b = TW * T10; | |
254 } | |
255 } | |
256 } | |
257 } | |
258 { | |
259 E T36, T3G, T3b, T3g, T28, T2d, T3F, T39, T3j, T3q, T3C, T3e, T3u, T3t; | |
260 { | |
261 E T3D, T1T, T3r, T14, T3E, T3s; | |
262 { | |
263 E Ty, T3B, T11, T2c, T13, T3v; | |
264 T36 = Te - Tx; | |
265 Ty = Te + Tx; | |
266 T3B = T3w + T3A; | |
267 T3G = T3A - T3w; | |
268 T11 = FMA(TZ, T10, TY); | |
269 T2c = FNMS(TZ, TX, T2b); | |
270 { | |
271 E T1t, T1S, T12, T37; | |
272 T3b = T1d - T1s; | |
273 T1t = T1d + T1s; | |
274 T1S = T1G + T1R; | |
275 T3g = T1G - T1R; | |
276 T12 = TV + T11; | |
277 T28 = TV - T11; | |
278 T2d = T2a - T2c; | |
279 T37 = T2a + T2c; | |
280 T3D = T1S - T1t; | |
281 T1T = T1t + T1S; | |
282 T13 = TL + T12; | |
283 T3F = TL - T12; | |
284 T39 = T37 - T38; | |
285 T3v = T38 + T37; | |
286 } | |
287 T3j = T3h - T3i; | |
288 T3r = T3h + T3i; | |
289 T3q = Ty - T13; | |
290 T14 = Ty + T13; | |
291 T3E = T3B - T3v; | |
292 T3C = T3v + T3B; | |
293 T3s = T3c + T3d; | |
294 T3e = T3c - T3d; | |
295 } | |
296 ci[WS(rs, 7)] = T14 - T1T; | |
297 cr[WS(rs, 12)] = T3D - T3E; | |
298 ci[WS(rs, 11)] = T3D + T3E; | |
299 T3u = T3s + T3r; | |
300 T3t = T3r - T3s; | |
301 cr[0] = T14 + T1T; | |
302 } | |
303 { | |
304 E T3m, T3a, T3J, T3H; | |
305 ci[WS(rs, 15)] = T3u + T3C; | |
306 cr[WS(rs, 8)] = T3u - T3C; | |
307 ci[WS(rs, 3)] = T3q + T3t; | |
308 cr[WS(rs, 4)] = T3q - T3t; | |
309 T3m = T36 + T39; | |
310 T3a = T36 - T39; | |
311 T3J = T3G - T3F; | |
312 T3H = T3F + T3G; | |
313 { | |
314 E T2Q, T20, T3N, T3T, T2C, T2J, T3U, T2f, T33, T30, T2V, T2W, T3O, T2T, T2N; | |
315 E T2v; | |
316 { | |
317 E T2R, T27, T2e, T2S; | |
318 { | |
319 E T3n, T3f, T3o, T3k; | |
320 T2Q = T1U + T1Z; | |
321 T20 = T1U - T1Z; | |
322 T3n = T3b - T3e; | |
323 T3f = T3b + T3e; | |
324 T3o = T3g + T3j; | |
325 T3k = T3g - T3j; | |
326 T3N = T3L + T3M; | |
327 T3T = T3M - T3L; | |
328 { | |
329 E T3p, T3K, T3I, T3l; | |
330 T3p = T3n + T3o; | |
331 T3K = T3o - T3n; | |
332 T3I = T3k - T3f; | |
333 T3l = T3f + T3k; | |
334 ci[WS(rs, 1)] = FMA(KP707106781, T3p, T3m); | |
335 cr[WS(rs, 6)] = FNMS(KP707106781, T3p, T3m); | |
336 ci[WS(rs, 13)] = FMA(KP707106781, T3K, T3J); | |
337 cr[WS(rs, 10)] = FMS(KP707106781, T3K, T3J); | |
338 ci[WS(rs, 9)] = FMA(KP707106781, T3I, T3H); | |
339 cr[WS(rs, 14)] = FMS(KP707106781, T3I, T3H); | |
340 cr[WS(rs, 2)] = FMA(KP707106781, T3l, T3a); | |
341 ci[WS(rs, 5)] = FNMS(KP707106781, T3l, T3a); | |
342 T2R = T21 + T26; | |
343 T27 = T21 - T26; | |
344 T2e = T28 + T2d; | |
345 T2S = T28 - T2d; | |
346 } | |
347 } | |
348 { | |
349 E T2Y, T2Z, T2n, T2u; | |
350 T2C = T2w - T2B; | |
351 T2Y = T2w + T2B; | |
352 T2Z = T2I + T2D; | |
353 T2J = T2D - T2I; | |
354 T3U = T2e - T27; | |
355 T2f = T27 + T2e; | |
356 T33 = FMA(KP414213562, T2Y, T2Z); | |
357 T30 = FNMS(KP414213562, T2Z, T2Y); | |
358 T2V = T2h + T2m; | |
359 T2n = T2h - T2m; | |
360 T2u = T2s + T2t; | |
361 T2W = T2s - T2t; | |
362 T3O = T2R - T2S; | |
363 T2T = T2R + T2S; | |
364 T2N = FMA(KP414213562, T2n, T2u); | |
365 T2v = FNMS(KP414213562, T2u, T2n); | |
366 } | |
367 } | |
368 { | |
369 E T2M, T3S, T31, T2P, T3Q, T3R, T3P, T2U; | |
370 { | |
371 E T2g, T2X, T2O, T2K, T3V, T3X, T3W, T34, T2L, T3Y; | |
372 T2M = FNMS(KP707106781, T2f, T20); | |
373 T2g = FMA(KP707106781, T2f, T20); | |
374 T34 = FNMS(KP414213562, T2V, T2W); | |
375 T2X = FMA(KP414213562, T2W, T2V); | |
376 T2O = FMA(KP414213562, T2C, T2J); | |
377 T2K = FNMS(KP414213562, T2J, T2C); | |
378 T3V = FMA(KP707106781, T3U, T3T); | |
379 T3X = FNMS(KP707106781, T3U, T3T); | |
380 T35 = T33 - T34; | |
381 T3W = T34 + T33; | |
382 T3S = T2K - T2v; | |
383 T2L = T2v + T2K; | |
384 T3Y = T30 - T2X; | |
385 T31 = T2X + T30; | |
386 ci[WS(rs, 14)] = FMA(KP923879532, T3W, T3V); | |
387 cr[WS(rs, 9)] = FMS(KP923879532, T3W, T3V); | |
388 ci[0] = FMA(KP923879532, T2L, T2g); | |
389 cr[WS(rs, 7)] = FNMS(KP923879532, T2L, T2g); | |
390 cr[WS(rs, 13)] = FMS(KP923879532, T3Y, T3X); | |
391 ci[WS(rs, 10)] = FMA(KP923879532, T3Y, T3X); | |
392 T2P = T2N + T2O; | |
393 T3Q = T2O - T2N; | |
394 } | |
395 T32 = FNMS(KP707106781, T2T, T2Q); | |
396 T2U = FMA(KP707106781, T2T, T2Q); | |
397 T3R = FNMS(KP707106781, T3O, T3N); | |
398 T3P = FMA(KP707106781, T3O, T3N); | |
399 cr[WS(rs, 3)] = FMA(KP923879532, T2P, T2M); | |
400 ci[WS(rs, 4)] = FNMS(KP923879532, T2P, T2M); | |
401 cr[WS(rs, 1)] = FMA(KP923879532, T31, T2U); | |
402 ci[WS(rs, 6)] = FNMS(KP923879532, T31, T2U); | |
403 ci[WS(rs, 8)] = FMA(KP923879532, T3Q, T3P); | |
404 cr[WS(rs, 15)] = FMS(KP923879532, T3Q, T3P); | |
405 ci[WS(rs, 12)] = FMA(KP923879532, T3S, T3R); | |
406 cr[WS(rs, 11)] = FMS(KP923879532, T3S, T3R); | |
407 } | |
408 } | |
409 } | |
410 } | |
411 } | |
412 } | |
413 ci[WS(rs, 2)] = FMA(KP923879532, T35, T32); | |
414 cr[WS(rs, 5)] = FNMS(KP923879532, T35, T32); | |
415 } | |
416 } | |
417 } | |
418 | |
419 static const tw_instr twinstr[] = { | |
420 {TW_CEXP, 1, 1}, | |
421 {TW_CEXP, 1, 3}, | |
422 {TW_CEXP, 1, 9}, | |
423 {TW_CEXP, 1, 15}, | |
424 {TW_NEXT, 1, 0} | |
425 }; | |
426 | |
427 static const hc2hc_desc desc = { 16, "hf2_16", twinstr, &GENUS, {104, 42, 92, 0} }; | |
428 | |
429 void X(codelet_hf2_16) (planner *p) { | |
430 X(khc2hc_register) (p, hf2_16, &desc); | |
431 } | |
432 #else /* HAVE_FMA */ | |
433 | |
434 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 16 -dit -name hf2_16 -include hf.h */ | |
435 | |
436 /* | |
437 * This function contains 196 FP additions, 108 FP multiplications, | |
438 * (or, 156 additions, 68 multiplications, 40 fused multiply/add), | |
439 * 82 stack variables, 3 constants, and 64 memory accesses | |
440 */ | |
441 #include "hf.h" | |
442 | |
443 static void hf2_16(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms) | |
444 { | |
445 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
446 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
447 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
448 { | |
449 INT m; | |
450 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(32, rs)) { | |
451 E T2, T5, Tg, Ti, Tk, To, TE, TC, T6, T3, T8, TW, TJ, Tt, TU; | |
452 E Tc, Tx, TH, TN, TO, TP, TR, T1f, T1k, T1b, T1i, T1y, T1H, T1u, T1F; | |
453 { | |
454 E T7, Tv, Ta, Ts, T4, Tw, Tb, Tr; | |
455 { | |
456 E Th, Tn, Tj, Tm; | |
457 T2 = W[0]; | |
458 T5 = W[1]; | |
459 Tg = W[2]; | |
460 Ti = W[3]; | |
461 Th = T2 * Tg; | |
462 Tn = T5 * Tg; | |
463 Tj = T5 * Ti; | |
464 Tm = T2 * Ti; | |
465 Tk = Th - Tj; | |
466 To = Tm + Tn; | |
467 TE = Tm - Tn; | |
468 TC = Th + Tj; | |
469 T6 = W[5]; | |
470 T7 = T5 * T6; | |
471 Tv = Tg * T6; | |
472 Ta = T2 * T6; | |
473 Ts = Ti * T6; | |
474 T3 = W[4]; | |
475 T4 = T2 * T3; | |
476 Tw = Ti * T3; | |
477 Tb = T5 * T3; | |
478 Tr = Tg * T3; | |
479 } | |
480 T8 = T4 + T7; | |
481 TW = Tv - Tw; | |
482 TJ = Ta + Tb; | |
483 Tt = Tr - Ts; | |
484 TU = Tr + Ts; | |
485 Tc = Ta - Tb; | |
486 Tx = Tv + Tw; | |
487 TH = T4 - T7; | |
488 TN = W[6]; | |
489 TO = W[7]; | |
490 TP = FMA(T2, TN, T5 * TO); | |
491 TR = FNMS(T5, TN, T2 * TO); | |
492 { | |
493 E T1d, T1e, T19, T1a; | |
494 T1d = Tk * T6; | |
495 T1e = To * T3; | |
496 T1f = T1d - T1e; | |
497 T1k = T1d + T1e; | |
498 T19 = Tk * T3; | |
499 T1a = To * T6; | |
500 T1b = T19 + T1a; | |
501 T1i = T19 - T1a; | |
502 } | |
503 { | |
504 E T1w, T1x, T1s, T1t; | |
505 T1w = TC * T6; | |
506 T1x = TE * T3; | |
507 T1y = T1w - T1x; | |
508 T1H = T1w + T1x; | |
509 T1s = TC * T3; | |
510 T1t = TE * T6; | |
511 T1u = T1s + T1t; | |
512 T1F = T1s - T1t; | |
513 } | |
514 } | |
515 { | |
516 E Tf, T3s, T1N, T3e, TA, T3r, T1Q, T3b, TM, T2N, T1W, T2w, TZ, T2M, T21; | |
517 E T2x, T1B, T1K, T2V, T2W, T2X, T2Y, T2j, T2E, T2o, T2D, T18, T1n, T2Q, T2R; | |
518 E T2S, T2T, T28, T2B, T2d, T2A; | |
519 { | |
520 E T1, T3d, Te, T3c, T9, Td; | |
521 T1 = cr[0]; | |
522 T3d = ci[0]; | |
523 T9 = cr[WS(rs, 8)]; | |
524 Td = ci[WS(rs, 8)]; | |
525 Te = FMA(T8, T9, Tc * Td); | |
526 T3c = FNMS(Tc, T9, T8 * Td); | |
527 Tf = T1 + Te; | |
528 T3s = T3d - T3c; | |
529 T1N = T1 - Te; | |
530 T3e = T3c + T3d; | |
531 } | |
532 { | |
533 E Tq, T1O, Tz, T1P; | |
534 { | |
535 E Tl, Tp, Tu, Ty; | |
536 Tl = cr[WS(rs, 4)]; | |
537 Tp = ci[WS(rs, 4)]; | |
538 Tq = FMA(Tk, Tl, To * Tp); | |
539 T1O = FNMS(To, Tl, Tk * Tp); | |
540 Tu = cr[WS(rs, 12)]; | |
541 Ty = ci[WS(rs, 12)]; | |
542 Tz = FMA(Tt, Tu, Tx * Ty); | |
543 T1P = FNMS(Tx, Tu, Tt * Ty); | |
544 } | |
545 TA = Tq + Tz; | |
546 T3r = Tq - Tz; | |
547 T1Q = T1O - T1P; | |
548 T3b = T1O + T1P; | |
549 } | |
550 { | |
551 E TG, T1T, TL, T1U, T1S, T1V; | |
552 { | |
553 E TD, TF, TI, TK; | |
554 TD = cr[WS(rs, 2)]; | |
555 TF = ci[WS(rs, 2)]; | |
556 TG = FMA(TC, TD, TE * TF); | |
557 T1T = FNMS(TE, TD, TC * TF); | |
558 TI = cr[WS(rs, 10)]; | |
559 TK = ci[WS(rs, 10)]; | |
560 TL = FMA(TH, TI, TJ * TK); | |
561 T1U = FNMS(TJ, TI, TH * TK); | |
562 } | |
563 TM = TG + TL; | |
564 T2N = T1T + T1U; | |
565 T1S = TG - TL; | |
566 T1V = T1T - T1U; | |
567 T1W = T1S - T1V; | |
568 T2w = T1S + T1V; | |
569 } | |
570 { | |
571 E TT, T1Y, TY, T1Z, T1X, T20; | |
572 { | |
573 E TQ, TS, TV, TX; | |
574 TQ = cr[WS(rs, 14)]; | |
575 TS = ci[WS(rs, 14)]; | |
576 TT = FMA(TP, TQ, TR * TS); | |
577 T1Y = FNMS(TR, TQ, TP * TS); | |
578 TV = cr[WS(rs, 6)]; | |
579 TX = ci[WS(rs, 6)]; | |
580 TY = FMA(TU, TV, TW * TX); | |
581 T1Z = FNMS(TW, TV, TU * TX); | |
582 } | |
583 TZ = TT + TY; | |
584 T2M = T1Y + T1Z; | |
585 T1X = TT - TY; | |
586 T20 = T1Y - T1Z; | |
587 T21 = T1X + T20; | |
588 T2x = T1X - T20; | |
589 } | |
590 { | |
591 E T1r, T2f, T1J, T2m, T1A, T2g, T1E, T2l; | |
592 { | |
593 E T1p, T1q, T1G, T1I; | |
594 T1p = cr[WS(rs, 15)]; | |
595 T1q = ci[WS(rs, 15)]; | |
596 T1r = FMA(TN, T1p, TO * T1q); | |
597 T2f = FNMS(TO, T1p, TN * T1q); | |
598 T1G = cr[WS(rs, 11)]; | |
599 T1I = ci[WS(rs, 11)]; | |
600 T1J = FMA(T1F, T1G, T1H * T1I); | |
601 T2m = FNMS(T1H, T1G, T1F * T1I); | |
602 } | |
603 { | |
604 E T1v, T1z, T1C, T1D; | |
605 T1v = cr[WS(rs, 7)]; | |
606 T1z = ci[WS(rs, 7)]; | |
607 T1A = FMA(T1u, T1v, T1y * T1z); | |
608 T2g = FNMS(T1y, T1v, T1u * T1z); | |
609 T1C = cr[WS(rs, 3)]; | |
610 T1D = ci[WS(rs, 3)]; | |
611 T1E = FMA(Tg, T1C, Ti * T1D); | |
612 T2l = FNMS(Ti, T1C, Tg * T1D); | |
613 } | |
614 T1B = T1r + T1A; | |
615 T1K = T1E + T1J; | |
616 T2V = T1B - T1K; | |
617 T2W = T2f + T2g; | |
618 T2X = T2l + T2m; | |
619 T2Y = T2W - T2X; | |
620 { | |
621 E T2h, T2i, T2k, T2n; | |
622 T2h = T2f - T2g; | |
623 T2i = T1E - T1J; | |
624 T2j = T2h + T2i; | |
625 T2E = T2h - T2i; | |
626 T2k = T1r - T1A; | |
627 T2n = T2l - T2m; | |
628 T2o = T2k - T2n; | |
629 T2D = T2k + T2n; | |
630 } | |
631 } | |
632 { | |
633 E T14, T29, T1m, T26, T17, T2a, T1h, T25; | |
634 { | |
635 E T12, T13, T1j, T1l; | |
636 T12 = cr[WS(rs, 1)]; | |
637 T13 = ci[WS(rs, 1)]; | |
638 T14 = FMA(T2, T12, T5 * T13); | |
639 T29 = FNMS(T5, T12, T2 * T13); | |
640 T1j = cr[WS(rs, 13)]; | |
641 T1l = ci[WS(rs, 13)]; | |
642 T1m = FMA(T1i, T1j, T1k * T1l); | |
643 T26 = FNMS(T1k, T1j, T1i * T1l); | |
644 } | |
645 { | |
646 E T15, T16, T1c, T1g; | |
647 T15 = cr[WS(rs, 9)]; | |
648 T16 = ci[WS(rs, 9)]; | |
649 T17 = FMA(T3, T15, T6 * T16); | |
650 T2a = FNMS(T6, T15, T3 * T16); | |
651 T1c = cr[WS(rs, 5)]; | |
652 T1g = ci[WS(rs, 5)]; | |
653 T1h = FMA(T1b, T1c, T1f * T1g); | |
654 T25 = FNMS(T1f, T1c, T1b * T1g); | |
655 } | |
656 T18 = T14 + T17; | |
657 T1n = T1h + T1m; | |
658 T2Q = T18 - T1n; | |
659 T2R = T29 + T2a; | |
660 T2S = T25 + T26; | |
661 T2T = T2R - T2S; | |
662 { | |
663 E T24, T27, T2b, T2c; | |
664 T24 = T14 - T17; | |
665 T27 = T25 - T26; | |
666 T28 = T24 - T27; | |
667 T2B = T24 + T27; | |
668 T2b = T29 - T2a; | |
669 T2c = T1h - T1m; | |
670 T2d = T2b + T2c; | |
671 T2A = T2b - T2c; | |
672 } | |
673 } | |
674 { | |
675 E T23, T2r, T3u, T3w, T2q, T3v, T2u, T3p; | |
676 { | |
677 E T1R, T22, T3q, T3t; | |
678 T1R = T1N - T1Q; | |
679 T22 = KP707106781 * (T1W + T21); | |
680 T23 = T1R + T22; | |
681 T2r = T1R - T22; | |
682 T3q = KP707106781 * (T2w - T2x); | |
683 T3t = T3r + T3s; | |
684 T3u = T3q + T3t; | |
685 T3w = T3t - T3q; | |
686 } | |
687 { | |
688 E T2e, T2p, T2s, T2t; | |
689 T2e = FNMS(KP382683432, T2d, KP923879532 * T28); | |
690 T2p = FMA(KP382683432, T2j, KP923879532 * T2o); | |
691 T2q = T2e + T2p; | |
692 T3v = T2p - T2e; | |
693 T2s = FMA(KP923879532, T2d, KP382683432 * T28); | |
694 T2t = FNMS(KP923879532, T2j, KP382683432 * T2o); | |
695 T2u = T2s + T2t; | |
696 T3p = T2t - T2s; | |
697 } | |
698 cr[WS(rs, 7)] = T23 - T2q; | |
699 cr[WS(rs, 11)] = T3v - T3w; | |
700 ci[WS(rs, 12)] = T3v + T3w; | |
701 ci[0] = T23 + T2q; | |
702 ci[WS(rs, 4)] = T2r - T2u; | |
703 cr[WS(rs, 15)] = T3p - T3u; | |
704 ci[WS(rs, 8)] = T3p + T3u; | |
705 cr[WS(rs, 3)] = T2r + T2u; | |
706 } | |
707 { | |
708 E T11, T35, T3g, T3i, T1M, T3h, T38, T39; | |
709 { | |
710 E TB, T10, T3a, T3f; | |
711 TB = Tf + TA; | |
712 T10 = TM + TZ; | |
713 T11 = TB + T10; | |
714 T35 = TB - T10; | |
715 T3a = T2N + T2M; | |
716 T3f = T3b + T3e; | |
717 T3g = T3a + T3f; | |
718 T3i = T3f - T3a; | |
719 } | |
720 { | |
721 E T1o, T1L, T36, T37; | |
722 T1o = T18 + T1n; | |
723 T1L = T1B + T1K; | |
724 T1M = T1o + T1L; | |
725 T3h = T1L - T1o; | |
726 T36 = T2W + T2X; | |
727 T37 = T2R + T2S; | |
728 T38 = T36 - T37; | |
729 T39 = T37 + T36; | |
730 } | |
731 ci[WS(rs, 7)] = T11 - T1M; | |
732 cr[WS(rs, 12)] = T3h - T3i; | |
733 ci[WS(rs, 11)] = T3h + T3i; | |
734 cr[0] = T11 + T1M; | |
735 cr[WS(rs, 4)] = T35 - T38; | |
736 cr[WS(rs, 8)] = T39 - T3g; | |
737 ci[WS(rs, 15)] = T39 + T3g; | |
738 ci[WS(rs, 3)] = T35 + T38; | |
739 } | |
740 { | |
741 E T2z, T2H, T3A, T3C, T2G, T3B, T2K, T3x; | |
742 { | |
743 E T2v, T2y, T3y, T3z; | |
744 T2v = T1N + T1Q; | |
745 T2y = KP707106781 * (T2w + T2x); | |
746 T2z = T2v + T2y; | |
747 T2H = T2v - T2y; | |
748 T3y = KP707106781 * (T21 - T1W); | |
749 T3z = T3s - T3r; | |
750 T3A = T3y + T3z; | |
751 T3C = T3z - T3y; | |
752 } | |
753 { | |
754 E T2C, T2F, T2I, T2J; | |
755 T2C = FMA(KP382683432, T2A, KP923879532 * T2B); | |
756 T2F = FNMS(KP382683432, T2E, KP923879532 * T2D); | |
757 T2G = T2C + T2F; | |
758 T3B = T2F - T2C; | |
759 T2I = FNMS(KP923879532, T2A, KP382683432 * T2B); | |
760 T2J = FMA(KP923879532, T2E, KP382683432 * T2D); | |
761 T2K = T2I + T2J; | |
762 T3x = T2J - T2I; | |
763 } | |
764 ci[WS(rs, 6)] = T2z - T2G; | |
765 cr[WS(rs, 13)] = T3B - T3C; | |
766 ci[WS(rs, 10)] = T3B + T3C; | |
767 cr[WS(rs, 1)] = T2z + T2G; | |
768 cr[WS(rs, 5)] = T2H - T2K; | |
769 cr[WS(rs, 9)] = T3x - T3A; | |
770 ci[WS(rs, 14)] = T3x + T3A; | |
771 ci[WS(rs, 2)] = T2H + T2K; | |
772 } | |
773 { | |
774 E T2P, T31, T3m, T3o, T30, T3j, T34, T3n; | |
775 { | |
776 E T2L, T2O, T3k, T3l; | |
777 T2L = Tf - TA; | |
778 T2O = T2M - T2N; | |
779 T2P = T2L - T2O; | |
780 T31 = T2L + T2O; | |
781 T3k = TM - TZ; | |
782 T3l = T3e - T3b; | |
783 T3m = T3k + T3l; | |
784 T3o = T3l - T3k; | |
785 } | |
786 { | |
787 E T2U, T2Z, T32, T33; | |
788 T2U = T2Q + T2T; | |
789 T2Z = T2V - T2Y; | |
790 T30 = KP707106781 * (T2U + T2Z); | |
791 T3j = KP707106781 * (T2Z - T2U); | |
792 T32 = T2Q - T2T; | |
793 T33 = T2V + T2Y; | |
794 T34 = KP707106781 * (T32 + T33); | |
795 T3n = KP707106781 * (T33 - T32); | |
796 } | |
797 ci[WS(rs, 5)] = T2P - T30; | |
798 cr[WS(rs, 10)] = T3n - T3o; | |
799 ci[WS(rs, 13)] = T3n + T3o; | |
800 cr[WS(rs, 2)] = T2P + T30; | |
801 cr[WS(rs, 6)] = T31 - T34; | |
802 cr[WS(rs, 14)] = T3j - T3m; | |
803 ci[WS(rs, 9)] = T3j + T3m; | |
804 ci[WS(rs, 1)] = T31 + T34; | |
805 } | |
806 } | |
807 } | |
808 } | |
809 } | |
810 | |
811 static const tw_instr twinstr[] = { | |
812 {TW_CEXP, 1, 1}, | |
813 {TW_CEXP, 1, 3}, | |
814 {TW_CEXP, 1, 9}, | |
815 {TW_CEXP, 1, 15}, | |
816 {TW_NEXT, 1, 0} | |
817 }; | |
818 | |
819 static const hc2hc_desc desc = { 16, "hf2_16", twinstr, &GENUS, {156, 68, 40, 0} }; | |
820 | |
821 void X(codelet_hf2_16) (planner *p) { | |
822 X(khc2hc_register) (p, hf2_16, &desc); | |
823 } | |
824 #endif /* HAVE_FMA */ |