Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/rdft/scalar/r2cf/r2cfII_32.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 08:06:44 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_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cfII_32 -dft-II -include rdft/scalar/r2cfII.h */ | |
29 | |
30 /* | |
31 * This function contains 174 FP additions, 128 FP multiplications, | |
32 * (or, 46 additions, 0 multiplications, 128 fused multiply/add), | |
33 * 62 stack variables, 15 constants, and 64 memory accesses | |
34 */ | |
35 #include "rdft/scalar/r2cfII.h" | |
36 | |
37 static void r2cfII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
40 DK(KP820678790, +0.820678790828660330972281985331011598767386482); | |
41 DK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
42 DK(KP303346683, +0.303346683607342391675883946941299872384187453); | |
43 DK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
44 DK(KP098491403, +0.098491403357164253077197521291327432293052451); | |
45 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
46 DK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
47 DK(KP534511135, +0.534511135950791641089685961295362908582039528); | |
48 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
49 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
50 DK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
51 DK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
52 DK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
54 { | |
55 INT i; | |
56 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) { | |
57 E T5, T2B, T1z, T2n, Tc, T2C, T1C, T2o, Tm, T1l, T1J, T27, Tv, T1k, T1G; | |
58 E T26, T15, T1r, T1Y, T2e, T1c, T1s, T1V, T2d, TK, T1o, T1R, T2b, TR, T1p; | |
59 E T1O, T2a; | |
60 { | |
61 E T1, T2l, T4, T2m, T2, T3; | |
62 T1 = R0[0]; | |
63 T2l = R0[WS(rs, 8)]; | |
64 T2 = R0[WS(rs, 4)]; | |
65 T3 = R0[WS(rs, 12)]; | |
66 T4 = T2 - T3; | |
67 T2m = T2 + T3; | |
68 T5 = FNMS(KP707106781, T4, T1); | |
69 T2B = FNMS(KP707106781, T2m, T2l); | |
70 T1z = FMA(KP707106781, T4, T1); | |
71 T2n = FMA(KP707106781, T2m, T2l); | |
72 } | |
73 { | |
74 E T8, T1A, Tb, T1B; | |
75 { | |
76 E T6, T7, T9, Ta; | |
77 T6 = R0[WS(rs, 10)]; | |
78 T7 = R0[WS(rs, 2)]; | |
79 T8 = FMA(KP414213562, T7, T6); | |
80 T1A = FNMS(KP414213562, T6, T7); | |
81 T9 = R0[WS(rs, 6)]; | |
82 Ta = R0[WS(rs, 14)]; | |
83 Tb = FMA(KP414213562, Ta, T9); | |
84 T1B = FMS(KP414213562, T9, Ta); | |
85 } | |
86 Tc = T8 - Tb; | |
87 T2C = T1B - T1A; | |
88 T1C = T1A + T1B; | |
89 T2o = T8 + Tb; | |
90 } | |
91 { | |
92 E Te, Tj, Th, Tk, Tf, Tg; | |
93 Te = R0[WS(rs, 7)]; | |
94 Tj = R0[WS(rs, 15)]; | |
95 Tf = R0[WS(rs, 3)]; | |
96 Tg = R0[WS(rs, 11)]; | |
97 Th = Tf + Tg; | |
98 Tk = Tg - Tf; | |
99 { | |
100 E Ti, Tl, T1H, T1I; | |
101 Ti = FNMS(KP707106781, Th, Te); | |
102 Tl = FNMS(KP707106781, Tk, Tj); | |
103 Tm = FNMS(KP668178637, Tl, Ti); | |
104 T1l = FMA(KP668178637, Ti, Tl); | |
105 T1H = FMA(KP707106781, Th, Te); | |
106 T1I = FMA(KP707106781, Tk, Tj); | |
107 T1J = FMA(KP198912367, T1I, T1H); | |
108 T27 = FNMS(KP198912367, T1H, T1I); | |
109 } | |
110 } | |
111 { | |
112 E Tn, Ts, Tq, Tt, To, Tp; | |
113 Tn = R0[WS(rs, 9)]; | |
114 Ts = R0[WS(rs, 1)]; | |
115 To = R0[WS(rs, 5)]; | |
116 Tp = R0[WS(rs, 13)]; | |
117 Tq = To + Tp; | |
118 Tt = To - Tp; | |
119 { | |
120 E Tr, Tu, T1E, T1F; | |
121 Tr = FNMS(KP707106781, Tq, Tn); | |
122 Tu = FNMS(KP707106781, Tt, Ts); | |
123 Tv = FNMS(KP668178637, Tu, Tr); | |
124 T1k = FMA(KP668178637, Tr, Tu); | |
125 T1E = FMA(KP707106781, Tq, Tn); | |
126 T1F = FMA(KP707106781, Tt, Ts); | |
127 T1G = FMA(KP198912367, T1F, T1E); | |
128 T26 = FNMS(KP198912367, T1E, T1F); | |
129 } | |
130 } | |
131 { | |
132 E TT, T16, TW, T17, T10, T1a, T13, T19, TU, TV; | |
133 TT = R1[WS(rs, 15)]; | |
134 T16 = R1[WS(rs, 7)]; | |
135 TU = R1[WS(rs, 3)]; | |
136 TV = R1[WS(rs, 11)]; | |
137 TW = TU - TV; | |
138 T17 = TU + TV; | |
139 { | |
140 E TY, TZ, T11, T12; | |
141 TY = R1[WS(rs, 9)]; | |
142 TZ = R1[WS(rs, 1)]; | |
143 T10 = FMA(KP414213562, TZ, TY); | |
144 T1a = FNMS(KP414213562, TY, TZ); | |
145 T11 = R1[WS(rs, 5)]; | |
146 T12 = R1[WS(rs, 13)]; | |
147 T13 = FMA(KP414213562, T12, T11); | |
148 T19 = FMS(KP414213562, T11, T12); | |
149 } | |
150 { | |
151 E TX, T14, T1W, T1X; | |
152 TX = FMA(KP707106781, TW, TT); | |
153 T14 = T10 - T13; | |
154 T15 = FMA(KP923879532, T14, TX); | |
155 T1r = FNMS(KP923879532, T14, TX); | |
156 T1W = FMA(KP707106781, T17, T16); | |
157 T1X = T10 + T13; | |
158 T1Y = FNMS(KP923879532, T1X, T1W); | |
159 T2e = FMA(KP923879532, T1X, T1W); | |
160 } | |
161 { | |
162 E T18, T1b, T1T, T1U; | |
163 T18 = FNMS(KP707106781, T17, T16); | |
164 T1b = T19 - T1a; | |
165 T1c = FNMS(KP923879532, T1b, T18); | |
166 T1s = FMA(KP923879532, T1b, T18); | |
167 T1T = FMS(KP707106781, TW, TT); | |
168 T1U = T1a + T19; | |
169 T1V = FNMS(KP923879532, T1U, T1T); | |
170 T2d = FMA(KP923879532, T1U, T1T); | |
171 } | |
172 } | |
173 { | |
174 E Ty, TL, TB, TM, TF, TP, TI, TO, Tz, TA; | |
175 Ty = R1[0]; | |
176 TL = R1[WS(rs, 8)]; | |
177 Tz = R1[WS(rs, 4)]; | |
178 TA = R1[WS(rs, 12)]; | |
179 TB = Tz - TA; | |
180 TM = Tz + TA; | |
181 { | |
182 E TD, TE, TG, TH; | |
183 TD = R1[WS(rs, 10)]; | |
184 TE = R1[WS(rs, 2)]; | |
185 TF = FMA(KP414213562, TE, TD); | |
186 TP = FNMS(KP414213562, TD, TE); | |
187 TG = R1[WS(rs, 6)]; | |
188 TH = R1[WS(rs, 14)]; | |
189 TI = FMA(KP414213562, TH, TG); | |
190 TO = FMS(KP414213562, TG, TH); | |
191 } | |
192 { | |
193 E TC, TJ, T1P, T1Q; | |
194 TC = FNMS(KP707106781, TB, Ty); | |
195 TJ = TF - TI; | |
196 TK = FNMS(KP923879532, TJ, TC); | |
197 T1o = FMA(KP923879532, TJ, TC); | |
198 T1P = FMA(KP707106781, TM, TL); | |
199 T1Q = TF + TI; | |
200 T1R = FNMS(KP923879532, T1Q, T1P); | |
201 T2b = FMA(KP923879532, T1Q, T1P); | |
202 } | |
203 { | |
204 E TN, TQ, T1M, T1N; | |
205 TN = FNMS(KP707106781, TM, TL); | |
206 TQ = TO - TP; | |
207 TR = FNMS(KP923879532, TQ, TN); | |
208 T1p = FMA(KP923879532, TQ, TN); | |
209 T1M = FMA(KP707106781, TB, Ty); | |
210 T1N = TP + TO; | |
211 T1O = FNMS(KP923879532, T1N, T1M); | |
212 T2a = FMA(KP923879532, T1N, T1M); | |
213 } | |
214 } | |
215 { | |
216 E Tx, T1f, T2L, T2N, T1e, T2O, T1i, T2M; | |
217 { | |
218 E Td, Tw, T2J, T2K; | |
219 Td = FNMS(KP923879532, Tc, T5); | |
220 Tw = Tm - Tv; | |
221 Tx = FMA(KP831469612, Tw, Td); | |
222 T1f = FNMS(KP831469612, Tw, Td); | |
223 T2J = FNMS(KP923879532, T2C, T2B); | |
224 T2K = T1k + T1l; | |
225 T2L = FMA(KP831469612, T2K, T2J); | |
226 T2N = FNMS(KP831469612, T2K, T2J); | |
227 } | |
228 { | |
229 E TS, T1d, T1g, T1h; | |
230 TS = FNMS(KP534511135, TR, TK); | |
231 T1d = FNMS(KP534511135, T1c, T15); | |
232 T1e = TS - T1d; | |
233 T2O = TS + T1d; | |
234 T1g = FMA(KP534511135, TK, TR); | |
235 T1h = FMA(KP534511135, T15, T1c); | |
236 T1i = T1g - T1h; | |
237 T2M = T1g + T1h; | |
238 } | |
239 Cr[WS(csr, 13)] = FNMS(KP881921264, T1e, Tx); | |
240 Ci[WS(csi, 13)] = FNMS(KP881921264, T2M, T2L); | |
241 Cr[WS(csr, 2)] = FMA(KP881921264, T1e, Tx); | |
242 Ci[WS(csi, 2)] = -(FMA(KP881921264, T2M, T2L)); | |
243 Cr[WS(csr, 10)] = FNMS(KP881921264, T1i, T1f); | |
244 Ci[WS(csi, 10)] = -(FMA(KP881921264, T2O, T2N)); | |
245 Cr[WS(csr, 5)] = FMA(KP881921264, T1i, T1f); | |
246 Ci[WS(csi, 5)] = FNMS(KP881921264, T2O, T2N); | |
247 } | |
248 { | |
249 E T29, T2h, T2r, T2t, T2g, T2u, T2k, T2s; | |
250 { | |
251 E T25, T28, T2p, T2q; | |
252 T25 = FMA(KP923879532, T1C, T1z); | |
253 T28 = T26 - T27; | |
254 T29 = FMA(KP980785280, T28, T25); | |
255 T2h = FNMS(KP980785280, T28, T25); | |
256 T2p = FMA(KP923879532, T2o, T2n); | |
257 T2q = T1G + T1J; | |
258 T2r = FMA(KP980785280, T2q, T2p); | |
259 T2t = FNMS(KP980785280, T2q, T2p); | |
260 } | |
261 { | |
262 E T2c, T2f, T2i, T2j; | |
263 T2c = FNMS(KP098491403, T2b, T2a); | |
264 T2f = FMA(KP098491403, T2e, T2d); | |
265 T2g = T2c + T2f; | |
266 T2u = T2f - T2c; | |
267 T2i = FMA(KP098491403, T2a, T2b); | |
268 T2j = FNMS(KP098491403, T2d, T2e); | |
269 T2k = T2i - T2j; | |
270 T2s = T2i + T2j; | |
271 } | |
272 Cr[WS(csr, 15)] = FNMS(KP995184726, T2g, T29); | |
273 Ci[WS(csi, 15)] = FNMS(KP995184726, T2s, T2r); | |
274 Cr[0] = FMA(KP995184726, T2g, T29); | |
275 Ci[0] = -(FMA(KP995184726, T2s, T2r)); | |
276 Cr[WS(csr, 8)] = FNMS(KP995184726, T2k, T2h); | |
277 Ci[WS(csi, 8)] = FMS(KP995184726, T2u, T2t); | |
278 Cr[WS(csr, 7)] = FMA(KP995184726, T2k, T2h); | |
279 Ci[WS(csi, 7)] = FMA(KP995184726, T2u, T2t); | |
280 } | |
281 { | |
282 E T1n, T1v, T2F, T2H, T1u, T2I, T1y, T2G; | |
283 { | |
284 E T1j, T1m, T2D, T2E; | |
285 T1j = FMA(KP923879532, Tc, T5); | |
286 T1m = T1k - T1l; | |
287 T1n = FMA(KP831469612, T1m, T1j); | |
288 T1v = FNMS(KP831469612, T1m, T1j); | |
289 T2D = FMA(KP923879532, T2C, T2B); | |
290 T2E = Tv + Tm; | |
291 T2F = FMA(KP831469612, T2E, T2D); | |
292 T2H = FNMS(KP831469612, T2E, T2D); | |
293 } | |
294 { | |
295 E T1q, T1t, T1w, T1x; | |
296 T1q = FMA(KP303346683, T1p, T1o); | |
297 T1t = FMA(KP303346683, T1s, T1r); | |
298 T1u = T1q - T1t; | |
299 T2I = T1q + T1t; | |
300 T1w = FNMS(KP303346683, T1r, T1s); | |
301 T1x = FNMS(KP303346683, T1o, T1p); | |
302 T1y = T1w - T1x; | |
303 T2G = T1x + T1w; | |
304 } | |
305 Cr[WS(csr, 14)] = FNMS(KP956940335, T1u, T1n); | |
306 Ci[WS(csi, 14)] = FMS(KP956940335, T2G, T2F); | |
307 Cr[WS(csr, 1)] = FMA(KP956940335, T1u, T1n); | |
308 Ci[WS(csi, 1)] = FMA(KP956940335, T2G, T2F); | |
309 Cr[WS(csr, 9)] = FNMS(KP956940335, T1y, T1v); | |
310 Ci[WS(csi, 9)] = FNMS(KP956940335, T2I, T2H); | |
311 Cr[WS(csr, 6)] = FMA(KP956940335, T1y, T1v); | |
312 Ci[WS(csi, 6)] = -(FMA(KP956940335, T2I, T2H)); | |
313 } | |
314 { | |
315 E T1L, T21, T2x, T2z, T20, T2A, T24, T2y; | |
316 { | |
317 E T1D, T1K, T2v, T2w; | |
318 T1D = FNMS(KP923879532, T1C, T1z); | |
319 T1K = T1G - T1J; | |
320 T1L = FMA(KP980785280, T1K, T1D); | |
321 T21 = FNMS(KP980785280, T1K, T1D); | |
322 T2v = FNMS(KP923879532, T2o, T2n); | |
323 T2w = T26 + T27; | |
324 T2x = FNMS(KP980785280, T2w, T2v); | |
325 T2z = FMA(KP980785280, T2w, T2v); | |
326 } | |
327 { | |
328 E T1S, T1Z, T22, T23; | |
329 T1S = FMA(KP820678790, T1R, T1O); | |
330 T1Z = FNMS(KP820678790, T1Y, T1V); | |
331 T20 = T1S + T1Z; | |
332 T2A = T1Z - T1S; | |
333 T22 = FMA(KP820678790, T1V, T1Y); | |
334 T23 = FNMS(KP820678790, T1O, T1R); | |
335 T24 = T22 - T23; | |
336 T2y = T23 + T22; | |
337 } | |
338 Cr[WS(csr, 12)] = FNMS(KP773010453, T20, T1L); | |
339 Ci[WS(csi, 12)] = FMS(KP773010453, T2y, T2x); | |
340 Cr[WS(csr, 3)] = FMA(KP773010453, T20, T1L); | |
341 Ci[WS(csi, 3)] = FMA(KP773010453, T2y, T2x); | |
342 Cr[WS(csr, 11)] = FNMS(KP773010453, T24, T21); | |
343 Ci[WS(csi, 11)] = FMA(KP773010453, T2A, T2z); | |
344 Cr[WS(csr, 4)] = FMA(KP773010453, T24, T21); | |
345 Ci[WS(csi, 4)] = FMS(KP773010453, T2A, T2z); | |
346 } | |
347 } | |
348 } | |
349 } | |
350 | |
351 static const kr2c_desc desc = { 32, "r2cfII_32", {46, 0, 128, 0}, &GENUS }; | |
352 | |
353 void X(codelet_r2cfII_32) (planner *p) { | |
354 X(kr2c_register) (p, r2cfII_32, &desc); | |
355 } | |
356 | |
357 #else | |
358 | |
359 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cfII_32 -dft-II -include rdft/scalar/r2cfII.h */ | |
360 | |
361 /* | |
362 * This function contains 174 FP additions, 82 FP multiplications, | |
363 * (or, 138 additions, 46 multiplications, 36 fused multiply/add), | |
364 * 62 stack variables, 15 constants, and 64 memory accesses | |
365 */ | |
366 #include "rdft/scalar/r2cfII.h" | |
367 | |
368 static void r2cfII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
369 { | |
370 DK(KP471396736, +0.471396736825997648556387625905254377657460319); | |
371 DK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
372 DK(KP634393284, +0.634393284163645498215171613225493370675687095); | |
373 DK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
374 DK(KP290284677, +0.290284677254462367636192375817395274691476278); | |
375 DK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
376 DK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
377 DK(KP098017140, +0.098017140329560601994195563888641845861136673); | |
378 DK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
379 DK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
380 DK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
381 DK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
382 DK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
383 DK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
384 DK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
385 { | |
386 INT i; | |
387 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) { | |
388 E T5, T2D, T1z, T2q, Tc, T2C, T1C, T2n, Tm, T1k, T1J, T26, Tv, T1l, T1G; | |
389 E T27, T15, T1r, T1Y, T2e, T1c, T1s, T1V, T2d, TK, T1o, T1R, T2b, TR, T1p; | |
390 E T1O, T2a; | |
391 { | |
392 E T1, T2p, T4, T2o, T2, T3; | |
393 T1 = R0[0]; | |
394 T2p = R0[WS(rs, 8)]; | |
395 T2 = R0[WS(rs, 4)]; | |
396 T3 = R0[WS(rs, 12)]; | |
397 T4 = KP707106781 * (T2 - T3); | |
398 T2o = KP707106781 * (T2 + T3); | |
399 T5 = T1 + T4; | |
400 T2D = T2p - T2o; | |
401 T1z = T1 - T4; | |
402 T2q = T2o + T2p; | |
403 } | |
404 { | |
405 E T8, T1A, Tb, T1B; | |
406 { | |
407 E T6, T7, T9, Ta; | |
408 T6 = R0[WS(rs, 2)]; | |
409 T7 = R0[WS(rs, 10)]; | |
410 T8 = FNMS(KP382683432, T7, KP923879532 * T6); | |
411 T1A = FMA(KP382683432, T6, KP923879532 * T7); | |
412 T9 = R0[WS(rs, 6)]; | |
413 Ta = R0[WS(rs, 14)]; | |
414 Tb = FNMS(KP923879532, Ta, KP382683432 * T9); | |
415 T1B = FMA(KP923879532, T9, KP382683432 * Ta); | |
416 } | |
417 Tc = T8 + Tb; | |
418 T2C = Tb - T8; | |
419 T1C = T1A - T1B; | |
420 T2n = T1A + T1B; | |
421 } | |
422 { | |
423 E Te, Tk, Th, Tj, Tf, Tg; | |
424 Te = R0[WS(rs, 1)]; | |
425 Tk = R0[WS(rs, 9)]; | |
426 Tf = R0[WS(rs, 5)]; | |
427 Tg = R0[WS(rs, 13)]; | |
428 Th = KP707106781 * (Tf - Tg); | |
429 Tj = KP707106781 * (Tf + Tg); | |
430 { | |
431 E Ti, Tl, T1H, T1I; | |
432 Ti = Te + Th; | |
433 Tl = Tj + Tk; | |
434 Tm = FNMS(KP195090322, Tl, KP980785280 * Ti); | |
435 T1k = FMA(KP195090322, Ti, KP980785280 * Tl); | |
436 T1H = Tk - Tj; | |
437 T1I = Te - Th; | |
438 T1J = FNMS(KP555570233, T1I, KP831469612 * T1H); | |
439 T26 = FMA(KP831469612, T1I, KP555570233 * T1H); | |
440 } | |
441 } | |
442 { | |
443 E Tq, Tt, Tp, Ts, Tn, To; | |
444 Tq = R0[WS(rs, 15)]; | |
445 Tt = R0[WS(rs, 7)]; | |
446 Tn = R0[WS(rs, 3)]; | |
447 To = R0[WS(rs, 11)]; | |
448 Tp = KP707106781 * (Tn - To); | |
449 Ts = KP707106781 * (Tn + To); | |
450 { | |
451 E Tr, Tu, T1E, T1F; | |
452 Tr = Tp - Tq; | |
453 Tu = Ts + Tt; | |
454 Tv = FMA(KP980785280, Tr, KP195090322 * Tu); | |
455 T1l = FNMS(KP980785280, Tu, KP195090322 * Tr); | |
456 T1E = Tt - Ts; | |
457 T1F = Tp + Tq; | |
458 T1G = FNMS(KP555570233, T1F, KP831469612 * T1E); | |
459 T27 = FMA(KP831469612, T1F, KP555570233 * T1E); | |
460 } | |
461 } | |
462 { | |
463 E TW, T1a, TV, T19, T10, T16, T13, T17, TT, TU; | |
464 TW = R1[WS(rs, 15)]; | |
465 T1a = R1[WS(rs, 7)]; | |
466 TT = R1[WS(rs, 3)]; | |
467 TU = R1[WS(rs, 11)]; | |
468 TV = KP707106781 * (TT - TU); | |
469 T19 = KP707106781 * (TT + TU); | |
470 { | |
471 E TY, TZ, T11, T12; | |
472 TY = R1[WS(rs, 1)]; | |
473 TZ = R1[WS(rs, 9)]; | |
474 T10 = FNMS(KP382683432, TZ, KP923879532 * TY); | |
475 T16 = FMA(KP382683432, TY, KP923879532 * TZ); | |
476 T11 = R1[WS(rs, 5)]; | |
477 T12 = R1[WS(rs, 13)]; | |
478 T13 = FNMS(KP923879532, T12, KP382683432 * T11); | |
479 T17 = FMA(KP923879532, T11, KP382683432 * T12); | |
480 } | |
481 { | |
482 E TX, T14, T1W, T1X; | |
483 TX = TV - TW; | |
484 T14 = T10 + T13; | |
485 T15 = TX + T14; | |
486 T1r = TX - T14; | |
487 T1W = T13 - T10; | |
488 T1X = T1a - T19; | |
489 T1Y = T1W - T1X; | |
490 T2e = T1W + T1X; | |
491 } | |
492 { | |
493 E T18, T1b, T1T, T1U; | |
494 T18 = T16 + T17; | |
495 T1b = T19 + T1a; | |
496 T1c = T18 + T1b; | |
497 T1s = T1b - T18; | |
498 T1T = TV + TW; | |
499 T1U = T16 - T17; | |
500 T1V = T1T + T1U; | |
501 T2d = T1U - T1T; | |
502 } | |
503 } | |
504 { | |
505 E Ty, TP, TB, TO, TF, TL, TI, TM, Tz, TA; | |
506 Ty = R1[0]; | |
507 TP = R1[WS(rs, 8)]; | |
508 Tz = R1[WS(rs, 4)]; | |
509 TA = R1[WS(rs, 12)]; | |
510 TB = KP707106781 * (Tz - TA); | |
511 TO = KP707106781 * (Tz + TA); | |
512 { | |
513 E TD, TE, TG, TH; | |
514 TD = R1[WS(rs, 2)]; | |
515 TE = R1[WS(rs, 10)]; | |
516 TF = FNMS(KP382683432, TE, KP923879532 * TD); | |
517 TL = FMA(KP382683432, TD, KP923879532 * TE); | |
518 TG = R1[WS(rs, 6)]; | |
519 TH = R1[WS(rs, 14)]; | |
520 TI = FNMS(KP923879532, TH, KP382683432 * TG); | |
521 TM = FMA(KP923879532, TG, KP382683432 * TH); | |
522 } | |
523 { | |
524 E TC, TJ, T1P, T1Q; | |
525 TC = Ty + TB; | |
526 TJ = TF + TI; | |
527 TK = TC + TJ; | |
528 T1o = TC - TJ; | |
529 T1P = TI - TF; | |
530 T1Q = TP - TO; | |
531 T1R = T1P - T1Q; | |
532 T2b = T1P + T1Q; | |
533 } | |
534 { | |
535 E TN, TQ, T1M, T1N; | |
536 TN = TL + TM; | |
537 TQ = TO + TP; | |
538 TR = TN + TQ; | |
539 T1p = TQ - TN; | |
540 T1M = Ty - TB; | |
541 T1N = TL - TM; | |
542 T1O = T1M - T1N; | |
543 T2a = T1M + T1N; | |
544 } | |
545 } | |
546 { | |
547 E Tx, T1f, T2s, T2u, T1e, T2l, T1i, T2t; | |
548 { | |
549 E Td, Tw, T2m, T2r; | |
550 Td = T5 + Tc; | |
551 Tw = Tm + Tv; | |
552 Tx = Td - Tw; | |
553 T1f = Td + Tw; | |
554 T2m = T1l - T1k; | |
555 T2r = T2n + T2q; | |
556 T2s = T2m - T2r; | |
557 T2u = T2m + T2r; | |
558 } | |
559 { | |
560 E TS, T1d, T1g, T1h; | |
561 TS = FMA(KP098017140, TK, KP995184726 * TR); | |
562 T1d = FNMS(KP995184726, T1c, KP098017140 * T15); | |
563 T1e = TS + T1d; | |
564 T2l = T1d - TS; | |
565 T1g = FNMS(KP098017140, TR, KP995184726 * TK); | |
566 T1h = FMA(KP995184726, T15, KP098017140 * T1c); | |
567 T1i = T1g + T1h; | |
568 T2t = T1h - T1g; | |
569 } | |
570 Cr[WS(csr, 8)] = Tx - T1e; | |
571 Ci[WS(csi, 8)] = T2t - T2u; | |
572 Cr[WS(csr, 7)] = Tx + T1e; | |
573 Ci[WS(csi, 7)] = T2t + T2u; | |
574 Cr[WS(csr, 15)] = T1f - T1i; | |
575 Ci[WS(csi, 15)] = T2l - T2s; | |
576 Cr[0] = T1f + T1i; | |
577 Ci[0] = T2l + T2s; | |
578 } | |
579 { | |
580 E T29, T2h, T2M, T2O, T2g, T2J, T2k, T2N; | |
581 { | |
582 E T25, T28, T2K, T2L; | |
583 T25 = T1z + T1C; | |
584 T28 = T26 - T27; | |
585 T29 = T25 + T28; | |
586 T2h = T25 - T28; | |
587 T2K = T1J + T1G; | |
588 T2L = T2C + T2D; | |
589 T2M = T2K - T2L; | |
590 T2O = T2K + T2L; | |
591 } | |
592 { | |
593 E T2c, T2f, T2i, T2j; | |
594 T2c = FMA(KP956940335, T2a, KP290284677 * T2b); | |
595 T2f = FNMS(KP290284677, T2e, KP956940335 * T2d); | |
596 T2g = T2c + T2f; | |
597 T2J = T2f - T2c; | |
598 T2i = FMA(KP290284677, T2d, KP956940335 * T2e); | |
599 T2j = FNMS(KP290284677, T2a, KP956940335 * T2b); | |
600 T2k = T2i - T2j; | |
601 T2N = T2j + T2i; | |
602 } | |
603 Cr[WS(csr, 14)] = T29 - T2g; | |
604 Ci[WS(csi, 14)] = T2N - T2O; | |
605 Cr[WS(csr, 1)] = T29 + T2g; | |
606 Ci[WS(csi, 1)] = T2N + T2O; | |
607 Cr[WS(csr, 9)] = T2h - T2k; | |
608 Ci[WS(csi, 9)] = T2J - T2M; | |
609 Cr[WS(csr, 6)] = T2h + T2k; | |
610 Ci[WS(csi, 6)] = T2J + T2M; | |
611 } | |
612 { | |
613 E T1n, T1v, T2y, T2A, T1u, T2v, T1y, T2z; | |
614 { | |
615 E T1j, T1m, T2w, T2x; | |
616 T1j = T5 - Tc; | |
617 T1m = T1k + T1l; | |
618 T1n = T1j + T1m; | |
619 T1v = T1j - T1m; | |
620 T2w = Tv - Tm; | |
621 T2x = T2q - T2n; | |
622 T2y = T2w - T2x; | |
623 T2A = T2w + T2x; | |
624 } | |
625 { | |
626 E T1q, T1t, T1w, T1x; | |
627 T1q = FMA(KP773010453, T1o, KP634393284 * T1p); | |
628 T1t = FNMS(KP634393284, T1s, KP773010453 * T1r); | |
629 T1u = T1q + T1t; | |
630 T2v = T1t - T1q; | |
631 T1w = FMA(KP634393284, T1r, KP773010453 * T1s); | |
632 T1x = FNMS(KP634393284, T1o, KP773010453 * T1p); | |
633 T1y = T1w - T1x; | |
634 T2z = T1x + T1w; | |
635 } | |
636 Cr[WS(csr, 12)] = T1n - T1u; | |
637 Ci[WS(csi, 12)] = T2z - T2A; | |
638 Cr[WS(csr, 3)] = T1n + T1u; | |
639 Ci[WS(csi, 3)] = T2z + T2A; | |
640 Cr[WS(csr, 11)] = T1v - T1y; | |
641 Ci[WS(csi, 11)] = T2v - T2y; | |
642 Cr[WS(csr, 4)] = T1v + T1y; | |
643 Ci[WS(csi, 4)] = T2v + T2y; | |
644 } | |
645 { | |
646 E T1L, T21, T2G, T2I, T20, T2H, T24, T2B; | |
647 { | |
648 E T1D, T1K, T2E, T2F; | |
649 T1D = T1z - T1C; | |
650 T1K = T1G - T1J; | |
651 T1L = T1D + T1K; | |
652 T21 = T1D - T1K; | |
653 T2E = T2C - T2D; | |
654 T2F = T26 + T27; | |
655 T2G = T2E - T2F; | |
656 T2I = T2F + T2E; | |
657 } | |
658 { | |
659 E T1S, T1Z, T22, T23; | |
660 T1S = FMA(KP881921264, T1O, KP471396736 * T1R); | |
661 T1Z = FMA(KP881921264, T1V, KP471396736 * T1Y); | |
662 T20 = T1S - T1Z; | |
663 T2H = T1S + T1Z; | |
664 T22 = FNMS(KP471396736, T1V, KP881921264 * T1Y); | |
665 T23 = FNMS(KP471396736, T1O, KP881921264 * T1R); | |
666 T24 = T22 - T23; | |
667 T2B = T23 + T22; | |
668 } | |
669 Cr[WS(csr, 13)] = T1L - T20; | |
670 Ci[WS(csi, 13)] = T2B - T2G; | |
671 Cr[WS(csr, 2)] = T1L + T20; | |
672 Ci[WS(csi, 2)] = T2B + T2G; | |
673 Cr[WS(csr, 10)] = T21 - T24; | |
674 Ci[WS(csi, 10)] = T2I - T2H; | |
675 Cr[WS(csr, 5)] = T21 + T24; | |
676 Ci[WS(csi, 5)] = -(T2H + T2I); | |
677 } | |
678 } | |
679 } | |
680 } | |
681 | |
682 static const kr2c_desc desc = { 32, "r2cfII_32", {138, 46, 36, 0}, &GENUS }; | |
683 | |
684 void X(codelet_r2cfII_32) (planner *p) { | |
685 X(kr2c_register) (p, r2cfII_32, &desc); | |
686 } | |
687 | |
688 #endif |