comparison src/fftw-3.3.8/rdft/scalar/r2cb/r2cbIII_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 08:07:45 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_r2cb.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include rdft/scalar/r2cbIII.h */
29
30 /*
31 * This function contains 174 FP additions, 100 FP multiplications,
32 * (or, 106 additions, 32 multiplications, 68 fused multiply/add),
33 * 65 stack variables, 18 constants, and 64 memory accesses
34 */
35 #include "rdft/scalar/r2cbIII.h"
36
37 static void r2cbIII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
38 {
39 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
40 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
41 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
42 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
43 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
44 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
45 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
46 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
47 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
48 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
49 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
50 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
51 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
52 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
54 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
55 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
56 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
57 {
58 INT i;
59 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
60 E T7, T2i, T2E, Tz, T1e, T1I, T1Z, T1x, Te, T22, T2F, T2j, T1h, T1y, TK;
61 E T1J, Tm, T2B, TW, T1k, T1C, T1M, T28, T2m, Tt, T2A, T17, T1j, T1F, T1L;
62 E T2d, T2l;
63 {
64 E T3, Tv, T1d, T2g, T6, T1a, Ty, T2h;
65 {
66 E T1, T2, T1b, T1c;
67 T1 = Cr[0];
68 T2 = Cr[WS(csr, 15)];
69 T3 = T1 + T2;
70 Tv = T1 - T2;
71 T1b = Ci[0];
72 T1c = Ci[WS(csi, 15)];
73 T1d = T1b + T1c;
74 T2g = T1c - T1b;
75 }
76 {
77 E T4, T5, Tw, Tx;
78 T4 = Cr[WS(csr, 8)];
79 T5 = Cr[WS(csr, 7)];
80 T6 = T4 + T5;
81 T1a = T4 - T5;
82 Tw = Ci[WS(csi, 8)];
83 Tx = Ci[WS(csi, 7)];
84 Ty = Tw + Tx;
85 T2h = Tx - Tw;
86 }
87 T7 = T3 + T6;
88 T2i = T2g - T2h;
89 T2E = T2h + T2g;
90 Tz = Tv - Ty;
91 T1e = T1a + T1d;
92 T1I = T1a - T1d;
93 T1Z = T3 - T6;
94 T1x = Tv + Ty;
95 }
96 {
97 E Ta, TA, TD, T20, Td, TF, TI, T21;
98 {
99 E T8, T9, TB, TC;
100 T8 = Cr[WS(csr, 4)];
101 T9 = Cr[WS(csr, 11)];
102 Ta = T8 + T9;
103 TA = T8 - T9;
104 TB = Ci[WS(csi, 4)];
105 TC = Ci[WS(csi, 11)];
106 TD = TB + TC;
107 T20 = TC - TB;
108 }
109 {
110 E Tb, Tc, TG, TH;
111 Tb = Cr[WS(csr, 3)];
112 Tc = Cr[WS(csr, 12)];
113 Td = Tb + Tc;
114 TF = Tb - Tc;
115 TG = Ci[WS(csi, 3)];
116 TH = Ci[WS(csi, 12)];
117 TI = TG + TH;
118 T21 = TG - TH;
119 }
120 Te = Ta + Td;
121 T22 = T20 - T21;
122 T2F = T20 + T21;
123 T2j = Ta - Td;
124 {
125 E T1f, T1g, TE, TJ;
126 T1f = TF + TI;
127 T1g = TA + TD;
128 T1h = T1f - T1g;
129 T1y = T1g + T1f;
130 TE = TA - TD;
131 TJ = TF - TI;
132 TK = TE + TJ;
133 T1J = TE - TJ;
134 }
135 }
136 {
137 E Ti, TM, TU, T25, Tl, TR, TP, T26, TQ, TV;
138 {
139 E Tg, Th, TS, TT;
140 Tg = Cr[WS(csr, 2)];
141 Th = Cr[WS(csr, 13)];
142 Ti = Tg + Th;
143 TM = Tg - Th;
144 TS = Ci[WS(csi, 2)];
145 TT = Ci[WS(csi, 13)];
146 TU = TS + TT;
147 T25 = TS - TT;
148 }
149 {
150 E Tj, Tk, TN, TO;
151 Tj = Cr[WS(csr, 10)];
152 Tk = Cr[WS(csr, 5)];
153 Tl = Tj + Tk;
154 TR = Tj - Tk;
155 TN = Ci[WS(csi, 10)];
156 TO = Ci[WS(csi, 5)];
157 TP = TN + TO;
158 T26 = TN - TO;
159 }
160 Tm = Ti + Tl;
161 T2B = T26 + T25;
162 TQ = TM - TP;
163 TV = TR + TU;
164 TW = FNMS(KP414213562, TV, TQ);
165 T1k = FMA(KP414213562, TQ, TV);
166 {
167 E T1A, T1B, T24, T27;
168 T1A = TR - TU;
169 T1B = TM + TP;
170 T1C = FMA(KP414213562, T1B, T1A);
171 T1M = FNMS(KP414213562, T1A, T1B);
172 T24 = Ti - Tl;
173 T27 = T25 - T26;
174 T28 = T24 - T27;
175 T2m = T24 + T27;
176 }
177 }
178 {
179 E Tp, TX, T14, T2a, Ts, T15, T10, T2b, T11, T16;
180 {
181 E Tn, To, T12, T13;
182 Tn = Cr[WS(csr, 1)];
183 To = Cr[WS(csr, 14)];
184 Tp = Tn + To;
185 TX = Tn - To;
186 T12 = Ci[WS(csi, 1)];
187 T13 = Ci[WS(csi, 14)];
188 T14 = T12 + T13;
189 T2a = T13 - T12;
190 }
191 {
192 E Tq, Tr, TY, TZ;
193 Tq = Cr[WS(csr, 6)];
194 Tr = Cr[WS(csr, 9)];
195 Ts = Tq + Tr;
196 T15 = Tq - Tr;
197 TY = Ci[WS(csi, 6)];
198 TZ = Ci[WS(csi, 9)];
199 T10 = TY + TZ;
200 T2b = TY - TZ;
201 }
202 Tt = Tp + Ts;
203 T2A = T2b + T2a;
204 T11 = TX - T10;
205 T16 = T14 - T15;
206 T17 = FNMS(KP414213562, T16, T11);
207 T1j = FMA(KP414213562, T11, T16);
208 {
209 E T1D, T1E, T29, T2c;
210 T1D = T15 + T14;
211 T1E = TX + T10;
212 T1F = FNMS(KP414213562, T1E, T1D);
213 T1L = FMA(KP414213562, T1D, T1E);
214 T29 = Tp - Ts;
215 T2c = T2a - T2b;
216 T2d = T29 + T2c;
217 T2l = T29 - T2c;
218 }
219 }
220 {
221 E Tf, Tu, T2L, T2M, T2N, T2O;
222 Tf = T7 + Te;
223 Tu = Tm + Tt;
224 T2L = Tf - Tu;
225 T2M = T2B + T2A;
226 T2N = T2F + T2E;
227 T2O = T2M + T2N;
228 R0[0] = KP2_000000000 * (Tf + Tu);
229 R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
230 R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
231 R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
232 }
233 {
234 E T2t, T2y, T2w, T2x;
235 {
236 E T2r, T2s, T2u, T2v;
237 T2r = T1Z - T22;
238 T2s = T2m + T2l;
239 T2t = FNMS(KP707106781, T2s, T2r);
240 T2y = FMA(KP707106781, T2s, T2r);
241 T2u = T2j + T2i;
242 T2v = T28 - T2d;
243 T2w = FNMS(KP707106781, T2v, T2u);
244 T2x = FMA(KP707106781, T2v, T2u);
245 }
246 R0[WS(rs, 3)] = KP1_662939224 * (FMA(KP668178637, T2w, T2t));
247 R0[WS(rs, 15)] = -(KP1_961570560 * (FNMS(KP198912367, T2x, T2y)));
248 R0[WS(rs, 11)] = KP1_662939224 * (FNMS(KP668178637, T2t, T2w));
249 R0[WS(rs, 7)] = KP1_961570560 * (FMA(KP198912367, T2y, T2x));
250 }
251 {
252 E T2D, T2K, T2I, T2J;
253 {
254 E T2z, T2C, T2G, T2H;
255 T2z = T7 - Te;
256 T2C = T2A - T2B;
257 T2D = T2z + T2C;
258 T2K = T2z - T2C;
259 T2G = T2E - T2F;
260 T2H = Tm - Tt;
261 T2I = T2G - T2H;
262 T2J = T2H + T2G;
263 }
264 R0[WS(rs, 2)] = KP1_847759065 * (FMA(KP414213562, T2I, T2D));
265 R0[WS(rs, 14)] = -(KP1_847759065 * (FNMS(KP414213562, T2J, T2K)));
266 R0[WS(rs, 10)] = KP1_847759065 * (FNMS(KP414213562, T2D, T2I));
267 R0[WS(rs, 6)] = KP1_847759065 * (FMA(KP414213562, T2K, T2J));
268 }
269 {
270 E T19, T1o, T1m, T1n;
271 {
272 E TL, T18, T1i, T1l;
273 TL = FMA(KP707106781, TK, Tz);
274 T18 = TW + T17;
275 T19 = FMA(KP923879532, T18, TL);
276 T1o = FNMS(KP923879532, T18, TL);
277 T1i = FNMS(KP707106781, T1h, T1e);
278 T1l = T1j - T1k;
279 T1m = FNMS(KP923879532, T1l, T1i);
280 T1n = FMA(KP923879532, T1l, T1i);
281 }
282 R1[0] = KP1_990369453 * (FNMS(KP098491403, T1m, T19));
283 R1[WS(rs, 12)] = -(KP1_546020906 * (FMA(KP820678790, T1n, T1o)));
284 R1[WS(rs, 8)] = -(KP1_990369453 * (FMA(KP098491403, T19, T1m)));
285 R1[WS(rs, 4)] = -(KP1_546020906 * (FNMS(KP820678790, T1o, T1n)));
286 }
287 {
288 E T1r, T1w, T1u, T1v;
289 {
290 E T1p, T1q, T1s, T1t;
291 T1p = FNMS(KP707106781, TK, Tz);
292 T1q = T1k + T1j;
293 T1r = FNMS(KP923879532, T1q, T1p);
294 T1w = FMA(KP923879532, T1q, T1p);
295 T1s = FMA(KP707106781, T1h, T1e);
296 T1t = TW - T17;
297 T1u = FMA(KP923879532, T1t, T1s);
298 T1v = FNMS(KP923879532, T1t, T1s);
299 }
300 R1[WS(rs, 2)] = KP1_763842528 * (FNMS(KP534511135, T1u, T1r));
301 R1[WS(rs, 14)] = -(KP1_913880671 * (FMA(KP303346683, T1v, T1w)));
302 R1[WS(rs, 10)] = -(KP1_763842528 * (FMA(KP534511135, T1r, T1u)));
303 R1[WS(rs, 6)] = -(KP1_913880671 * (FNMS(KP303346683, T1w, T1v)));
304 }
305 {
306 E T1T, T1Y, T1W, T1X;
307 {
308 E T1R, T1S, T1U, T1V;
309 T1R = FMA(KP707106781, T1y, T1x);
310 T1S = T1M + T1L;
311 T1T = FNMS(KP923879532, T1S, T1R);
312 T1Y = FMA(KP923879532, T1S, T1R);
313 T1U = FMA(KP707106781, T1J, T1I);
314 T1V = T1C + T1F;
315 T1W = FNMS(KP923879532, T1V, T1U);
316 T1X = FMA(KP923879532, T1V, T1U);
317 }
318 R1[WS(rs, 3)] = KP1_546020906 * (FMA(KP820678790, T1W, T1T));
319 R1[WS(rs, 15)] = -(KP1_990369453 * (FNMS(KP098491403, T1X, T1Y)));
320 R1[WS(rs, 11)] = KP1_546020906 * (FNMS(KP820678790, T1T, T1W));
321 R1[WS(rs, 7)] = KP1_990369453 * (FMA(KP098491403, T1Y, T1X));
322 }
323 {
324 E T2f, T2q, T2o, T2p;
325 {
326 E T23, T2e, T2k, T2n;
327 T23 = T1Z + T22;
328 T2e = T28 + T2d;
329 T2f = FMA(KP707106781, T2e, T23);
330 T2q = FNMS(KP707106781, T2e, T23);
331 T2k = T2i - T2j;
332 T2n = T2l - T2m;
333 T2o = FMA(KP707106781, T2n, T2k);
334 T2p = FNMS(KP707106781, T2n, T2k);
335 }
336 R0[WS(rs, 1)] = KP1_961570560 * (FMA(KP198912367, T2o, T2f));
337 R0[WS(rs, 13)] = -(KP1_662939224 * (FNMS(KP668178637, T2p, T2q)));
338 R0[WS(rs, 9)] = KP1_961570560 * (FNMS(KP198912367, T2f, T2o));
339 R0[WS(rs, 5)] = KP1_662939224 * (FMA(KP668178637, T2q, T2p));
340 }
341 {
342 E T1H, T1Q, T1O, T1P;
343 {
344 E T1z, T1G, T1K, T1N;
345 T1z = FNMS(KP707106781, T1y, T1x);
346 T1G = T1C - T1F;
347 T1H = FMA(KP923879532, T1G, T1z);
348 T1Q = FNMS(KP923879532, T1G, T1z);
349 T1K = FNMS(KP707106781, T1J, T1I);
350 T1N = T1L - T1M;
351 T1O = FMA(KP923879532, T1N, T1K);
352 T1P = FNMS(KP923879532, T1N, T1K);
353 }
354 R1[WS(rs, 1)] = KP1_913880671 * (FMA(KP303346683, T1O, T1H));
355 R1[WS(rs, 13)] = -(KP1_763842528 * (FNMS(KP534511135, T1P, T1Q)));
356 R1[WS(rs, 9)] = KP1_913880671 * (FNMS(KP303346683, T1H, T1O));
357 R1[WS(rs, 5)] = KP1_763842528 * (FMA(KP534511135, T1Q, T1P));
358 }
359 }
360 }
361 }
362
363 static const kr2c_desc desc = { 32, "r2cbIII_32", {106, 32, 68, 0}, &GENUS };
364
365 void X(codelet_r2cbIII_32) (planner *p) {
366 X(kr2c_register) (p, r2cbIII_32, &desc);
367 }
368
369 #else
370
371 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include rdft/scalar/r2cbIII.h */
372
373 /*
374 * This function contains 174 FP additions, 84 FP multiplications,
375 * (or, 138 additions, 48 multiplications, 36 fused multiply/add),
376 * 66 stack variables, 19 constants, and 64 memory accesses
377 */
378 #include "rdft/scalar/r2cbIII.h"
379
380 static void r2cbIII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
381 {
382 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
383 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
384 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
385 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
386 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
387 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
388 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
389 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
390 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
391 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
392 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
393 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
394 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
395 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
396 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
397 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
398 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
399 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
400 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
401 {
402 INT i;
403 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
404 E T7, T2i, T2F, Tz, T1k, T1I, T1Z, T1x, Te, T22, T2E, T2j, T1f, T1y, TK;
405 E T1J, Tm, T2B, TW, T1a, T1C, T1L, T28, T2l, Tt, T2A, T17, T1b, T1F, T1M;
406 E T2d, T2m;
407 {
408 E T3, Tv, T1j, T2h, T6, T1g, Ty, T2g;
409 {
410 E T1, T2, T1h, T1i;
411 T1 = Cr[0];
412 T2 = Cr[WS(csr, 15)];
413 T3 = T1 + T2;
414 Tv = T1 - T2;
415 T1h = Ci[0];
416 T1i = Ci[WS(csi, 15)];
417 T1j = T1h + T1i;
418 T2h = T1i - T1h;
419 }
420 {
421 E T4, T5, Tw, Tx;
422 T4 = Cr[WS(csr, 8)];
423 T5 = Cr[WS(csr, 7)];
424 T6 = T4 + T5;
425 T1g = T4 - T5;
426 Tw = Ci[WS(csi, 8)];
427 Tx = Ci[WS(csi, 7)];
428 Ty = Tw + Tx;
429 T2g = Tw - Tx;
430 }
431 T7 = T3 + T6;
432 T2i = T2g + T2h;
433 T2F = T2h - T2g;
434 Tz = Tv - Ty;
435 T1k = T1g + T1j;
436 T1I = T1g - T1j;
437 T1Z = T3 - T6;
438 T1x = Tv + Ty;
439 }
440 {
441 E Ta, TA, TD, T21, Td, TF, TI, T20;
442 {
443 E T8, T9, TB, TC;
444 T8 = Cr[WS(csr, 4)];
445 T9 = Cr[WS(csr, 11)];
446 Ta = T8 + T9;
447 TA = T8 - T9;
448 TB = Ci[WS(csi, 4)];
449 TC = Ci[WS(csi, 11)];
450 TD = TB + TC;
451 T21 = TB - TC;
452 }
453 {
454 E Tb, Tc, TG, TH;
455 Tb = Cr[WS(csr, 3)];
456 Tc = Cr[WS(csr, 12)];
457 Td = Tb + Tc;
458 TF = Tb - Tc;
459 TG = Ci[WS(csi, 3)];
460 TH = Ci[WS(csi, 12)];
461 TI = TG + TH;
462 T20 = TH - TG;
463 }
464 Te = Ta + Td;
465 T22 = T20 - T21;
466 T2E = T21 + T20;
467 T2j = Ta - Td;
468 {
469 E T1d, T1e, TE, TJ;
470 T1d = TA + TD;
471 T1e = TF + TI;
472 T1f = KP707106781 * (T1d - T1e);
473 T1y = KP707106781 * (T1d + T1e);
474 TE = TA - TD;
475 TJ = TF - TI;
476 TK = KP707106781 * (TE + TJ);
477 T1J = KP707106781 * (TE - TJ);
478 }
479 }
480 {
481 E Ti, TM, TU, T25, Tl, TR, TP, T26, TQ, TV;
482 {
483 E Tg, Th, TS, TT;
484 Tg = Cr[WS(csr, 2)];
485 Th = Cr[WS(csr, 13)];
486 Ti = Tg + Th;
487 TM = Tg - Th;
488 TS = Ci[WS(csi, 2)];
489 TT = Ci[WS(csi, 13)];
490 TU = TS + TT;
491 T25 = TS - TT;
492 }
493 {
494 E Tj, Tk, TN, TO;
495 Tj = Cr[WS(csr, 10)];
496 Tk = Cr[WS(csr, 5)];
497 Tl = Tj + Tk;
498 TR = Tj - Tk;
499 TN = Ci[WS(csi, 10)];
500 TO = Ci[WS(csi, 5)];
501 TP = TN + TO;
502 T26 = TN - TO;
503 }
504 Tm = Ti + Tl;
505 T2B = T26 + T25;
506 TQ = TM - TP;
507 TV = TR + TU;
508 TW = FNMS(KP382683432, TV, KP923879532 * TQ);
509 T1a = FMA(KP382683432, TQ, KP923879532 * TV);
510 {
511 E T1A, T1B, T24, T27;
512 T1A = TM + TP;
513 T1B = TU - TR;
514 T1C = FNMS(KP923879532, T1B, KP382683432 * T1A);
515 T1L = FMA(KP923879532, T1A, KP382683432 * T1B);
516 T24 = Ti - Tl;
517 T27 = T25 - T26;
518 T28 = T24 - T27;
519 T2l = T24 + T27;
520 }
521 }
522 {
523 E Tp, TX, T15, T2a, Ts, T12, T10, T2b, T11, T16;
524 {
525 E Tn, To, T13, T14;
526 Tn = Cr[WS(csr, 1)];
527 To = Cr[WS(csr, 14)];
528 Tp = Tn + To;
529 TX = Tn - To;
530 T13 = Ci[WS(csi, 1)];
531 T14 = Ci[WS(csi, 14)];
532 T15 = T13 + T14;
533 T2a = T14 - T13;
534 }
535 {
536 E Tq, Tr, TY, TZ;
537 Tq = Cr[WS(csr, 6)];
538 Tr = Cr[WS(csr, 9)];
539 Ts = Tq + Tr;
540 T12 = Tq - Tr;
541 TY = Ci[WS(csi, 6)];
542 TZ = Ci[WS(csi, 9)];
543 T10 = TY + TZ;
544 T2b = TY - TZ;
545 }
546 Tt = Tp + Ts;
547 T2A = T2b + T2a;
548 T11 = TX - T10;
549 T16 = T12 - T15;
550 T17 = FMA(KP923879532, T11, KP382683432 * T16);
551 T1b = FNMS(KP382683432, T11, KP923879532 * T16);
552 {
553 E T1D, T1E, T29, T2c;
554 T1D = TX + T10;
555 T1E = T12 + T15;
556 T1F = FNMS(KP923879532, T1E, KP382683432 * T1D);
557 T1M = FMA(KP923879532, T1D, KP382683432 * T1E);
558 T29 = Tp - Ts;
559 T2c = T2a - T2b;
560 T2d = T29 + T2c;
561 T2m = T2c - T29;
562 }
563 }
564 {
565 E Tf, Tu, T2L, T2M, T2N, T2O;
566 Tf = T7 + Te;
567 Tu = Tm + Tt;
568 T2L = Tf - Tu;
569 T2M = T2B + T2A;
570 T2N = T2F - T2E;
571 T2O = T2M + T2N;
572 R0[0] = KP2_000000000 * (Tf + Tu);
573 R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
574 R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
575 R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
576 }
577 {
578 E T2t, T2x, T2w, T2y;
579 {
580 E T2r, T2s, T2u, T2v;
581 T2r = T1Z - T22;
582 T2s = KP707106781 * (T2m - T2l);
583 T2t = T2r + T2s;
584 T2x = T2r - T2s;
585 T2u = T2j + T2i;
586 T2v = KP707106781 * (T28 - T2d);
587 T2w = T2u - T2v;
588 T2y = T2v + T2u;
589 }
590 R0[WS(rs, 3)] = FMA(KP1_662939224, T2t, KP1_111140466 * T2w);
591 R0[WS(rs, 15)] = FNMS(KP1_961570560, T2x, KP390180644 * T2y);
592 R0[WS(rs, 11)] = FNMS(KP1_111140466, T2t, KP1_662939224 * T2w);
593 R0[WS(rs, 7)] = FMA(KP390180644, T2x, KP1_961570560 * T2y);
594 }
595 {
596 E T2D, T2J, T2I, T2K;
597 {
598 E T2z, T2C, T2G, T2H;
599 T2z = T7 - Te;
600 T2C = T2A - T2B;
601 T2D = T2z + T2C;
602 T2J = T2z - T2C;
603 T2G = T2E + T2F;
604 T2H = Tm - Tt;
605 T2I = T2G - T2H;
606 T2K = T2H + T2G;
607 }
608 R0[WS(rs, 2)] = FMA(KP1_847759065, T2D, KP765366864 * T2I);
609 R0[WS(rs, 14)] = FNMS(KP1_847759065, T2J, KP765366864 * T2K);
610 R0[WS(rs, 10)] = FNMS(KP765366864, T2D, KP1_847759065 * T2I);
611 R0[WS(rs, 6)] = FMA(KP765366864, T2J, KP1_847759065 * T2K);
612 }
613 {
614 E T19, T1n, T1m, T1o;
615 {
616 E TL, T18, T1c, T1l;
617 TL = Tz + TK;
618 T18 = TW + T17;
619 T19 = TL + T18;
620 T1n = TL - T18;
621 T1c = T1a + T1b;
622 T1l = T1f + T1k;
623 T1m = T1c + T1l;
624 T1o = T1c - T1l;
625 }
626 R1[0] = FNMS(KP196034280, T1m, KP1_990369453 * T19);
627 R1[WS(rs, 12)] = FNMS(KP1_546020906, T1n, KP1_268786568 * T1o);
628 R1[WS(rs, 8)] = -(FMA(KP196034280, T19, KP1_990369453 * T1m));
629 R1[WS(rs, 4)] = FMA(KP1_268786568, T1n, KP1_546020906 * T1o);
630 }
631 {
632 E T1r, T1v, T1u, T1w;
633 {
634 E T1p, T1q, T1s, T1t;
635 T1p = Tz - TK;
636 T1q = T1b - T1a;
637 T1r = T1p + T1q;
638 T1v = T1p - T1q;
639 T1s = T1f - T1k;
640 T1t = TW - T17;
641 T1u = T1s - T1t;
642 T1w = T1t + T1s;
643 }
644 R1[WS(rs, 2)] = FMA(KP1_763842528, T1r, KP942793473 * T1u);
645 R1[WS(rs, 14)] = FNMS(KP1_913880671, T1v, KP580569354 * T1w);
646 R1[WS(rs, 10)] = FNMS(KP942793473, T1r, KP1_763842528 * T1u);
647 R1[WS(rs, 6)] = FMA(KP580569354, T1v, KP1_913880671 * T1w);
648 }
649 {
650 E T1T, T1X, T1W, T1Y;
651 {
652 E T1R, T1S, T1U, T1V;
653 T1R = T1x + T1y;
654 T1S = T1L + T1M;
655 T1T = T1R - T1S;
656 T1X = T1R + T1S;
657 T1U = T1J + T1I;
658 T1V = T1C - T1F;
659 T1W = T1U - T1V;
660 T1Y = T1V + T1U;
661 }
662 R1[WS(rs, 3)] = FMA(KP1_546020906, T1T, KP1_268786568 * T1W);
663 R1[WS(rs, 15)] = FNMS(KP1_990369453, T1X, KP196034280 * T1Y);
664 R1[WS(rs, 11)] = FNMS(KP1_268786568, T1T, KP1_546020906 * T1W);
665 R1[WS(rs, 7)] = FMA(KP196034280, T1X, KP1_990369453 * T1Y);
666 }
667 {
668 E T2f, T2p, T2o, T2q;
669 {
670 E T23, T2e, T2k, T2n;
671 T23 = T1Z + T22;
672 T2e = KP707106781 * (T28 + T2d);
673 T2f = T23 + T2e;
674 T2p = T23 - T2e;
675 T2k = T2i - T2j;
676 T2n = KP707106781 * (T2l + T2m);
677 T2o = T2k - T2n;
678 T2q = T2n + T2k;
679 }
680 R0[WS(rs, 1)] = FMA(KP1_961570560, T2f, KP390180644 * T2o);
681 R0[WS(rs, 13)] = FNMS(KP1_662939224, T2p, KP1_111140466 * T2q);
682 R0[WS(rs, 9)] = FNMS(KP390180644, T2f, KP1_961570560 * T2o);
683 R0[WS(rs, 5)] = FMA(KP1_111140466, T2p, KP1_662939224 * T2q);
684 }
685 {
686 E T1H, T1P, T1O, T1Q;
687 {
688 E T1z, T1G, T1K, T1N;
689 T1z = T1x - T1y;
690 T1G = T1C + T1F;
691 T1H = T1z + T1G;
692 T1P = T1z - T1G;
693 T1K = T1I - T1J;
694 T1N = T1L - T1M;
695 T1O = T1K - T1N;
696 T1Q = T1N + T1K;
697 }
698 R1[WS(rs, 1)] = FMA(KP1_913880671, T1H, KP580569354 * T1O);
699 R1[WS(rs, 13)] = FNMS(KP1_763842528, T1P, KP942793473 * T1Q);
700 R1[WS(rs, 9)] = FNMS(KP580569354, T1H, KP1_913880671 * T1O);
701 R1[WS(rs, 5)] = FMA(KP942793473, T1P, KP1_763842528 * T1Q);
702 }
703 }
704 }
705 }
706
707 static const kr2c_desc desc = { 32, "r2cbIII_32", {138, 48, 36, 0}, &GENUS };
708
709 void X(codelet_r2cbIII_32) (planner *p) {
710 X(kr2c_register) (p, r2cbIII_32, &desc);
711 }
712
713 #endif