changed reading hint
[gromacs/adressmacs.git] / src / fftw / fhb_9.c
blobdff25257184c75d5915cedd819de0cecc36cfce7
1 /*
2 * Copyright (c) 1997-1999 Massachusetts Institute of Technology
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 /* This file was automatically generated --- DO NOT EDIT */
21 /* Generated on Tue May 18 13:56:05 EDT 1999 */
23 #include <fftw-int.h>
24 #include <fftw.h>
26 /* Generated by: ./genfft -magic-alignment-check -magic-twiddle-load-all -magic-variables 4 -magic-loopi -hc2hc-backward 9 */
29 * This function contains 181 FP additions, 120 FP multiplications,
30 * (or, 129 additions, 68 multiplications, 52 fused multiply/add),
31 * 43 stack variables, and 72 memory accesses
33 static const fftw_real K663413948 = FFTW_KONST(+0.663413948168938396205421319635891297216863310);
34 static const fftw_real K556670399 = FFTW_KONST(+0.556670399226419366452912952047023132968291906);
35 static const fftw_real K852868531 = FFTW_KONST(+0.852868531952443209628250963940074071936020296);
36 static const fftw_real K296198132 = FFTW_KONST(+0.296198132726023843175338011893050938967728390);
37 static const fftw_real K150383733 = FFTW_KONST(+0.150383733180435296639271897612501926072238258);
38 static const fftw_real K813797681 = FFTW_KONST(+0.813797681349373692844693217248393223289101568);
39 static const fftw_real K939692620 = FFTW_KONST(+0.939692620785908384054109277324731469936208134);
40 static const fftw_real K342020143 = FFTW_KONST(+0.342020143325668733044099614682259580763083368);
41 static const fftw_real K984807753 = FFTW_KONST(+0.984807753012208059366743024589523013670643252);
42 static const fftw_real K173648177 = FFTW_KONST(+0.173648177666930348851716626769314796000375677);
43 static const fftw_real K300767466 = FFTW_KONST(+0.300767466360870593278543795225003852144476517);
44 static const fftw_real K1_705737063 = FFTW_KONST(+1.705737063904886419256501927880148143872040591);
45 static const fftw_real K642787609 = FFTW_KONST(+0.642787609686539326322643409907263432907559884);
46 static const fftw_real K766044443 = FFTW_KONST(+0.766044443118978035202392650555416673935832457);
47 static const fftw_real K1_326827896 = FFTW_KONST(+1.326827896337876792410842639271782594433726619);
48 static const fftw_real K1_113340798 = FFTW_KONST(+1.113340798452838732905825904094046265936583811);
49 static const fftw_real K500000000 = FFTW_KONST(+0.500000000000000000000000000000000000000000000);
50 static const fftw_real K866025403 = FFTW_KONST(+0.866025403784438646763723170752936183471402627);
51 static const fftw_real K2_000000000 = FFTW_KONST(+2.000000000000000000000000000000000000000000000);
52 static const fftw_real K1_732050807 = FFTW_KONST(+1.732050807568877293527446341505872366942805254);
55 * Generator Id's :
56 * $Id$
57 * $Id$
58 * $Id$
61 void fftw_hc2hc_backward_9(fftw_real *A, const fftw_complex *W, int iostride, int m, int dist)
63 int i;
64 fftw_real *X;
65 fftw_real *Y;
66 X = A;
67 Y = A + (9 * iostride);
69 fftw_real tmp155;
70 fftw_real tmp183;
71 fftw_real tmp177;
72 fftw_real tmp162;
73 fftw_real tmp165;
74 fftw_real tmp160;
75 fftw_real tmp169;
76 fftw_real tmp180;
77 fftw_real tmp172;
78 fftw_real tmp181;
79 fftw_real tmp161;
80 fftw_real tmp166;
81 ASSERT_ALIGNED_DOUBLE();
83 fftw_real tmp176;
84 fftw_real tmp153;
85 fftw_real tmp154;
86 fftw_real tmp174;
87 fftw_real tmp175;
88 ASSERT_ALIGNED_DOUBLE();
89 tmp175 = Y[-3 * iostride];
90 tmp176 = K1_732050807 * tmp175;
91 tmp153 = X[0];
92 tmp154 = X[3 * iostride];
93 tmp174 = tmp154 - tmp153;
94 tmp155 = tmp153 + (K2_000000000 * tmp154);
95 tmp183 = tmp176 - tmp174;
96 tmp177 = tmp174 + tmp176;
99 fftw_real tmp156;
100 fftw_real tmp159;
101 fftw_real tmp170;
102 fftw_real tmp168;
103 fftw_real tmp167;
104 fftw_real tmp171;
105 ASSERT_ALIGNED_DOUBLE();
106 tmp156 = X[iostride];
107 tmp162 = Y[-iostride];
109 fftw_real tmp157;
110 fftw_real tmp158;
111 fftw_real tmp163;
112 fftw_real tmp164;
113 ASSERT_ALIGNED_DOUBLE();
114 tmp157 = X[4 * iostride];
115 tmp158 = X[2 * iostride];
116 tmp159 = tmp157 + tmp158;
117 tmp170 = K866025403 * (tmp157 - tmp158);
118 tmp163 = Y[-2 * iostride];
119 tmp164 = Y[-4 * iostride];
120 tmp165 = tmp163 - tmp164;
121 tmp168 = K866025403 * (tmp164 + tmp163);
123 tmp160 = tmp156 + tmp159;
124 tmp167 = tmp156 - (K500000000 * tmp159);
125 tmp169 = tmp167 - tmp168;
126 tmp180 = tmp167 + tmp168;
127 tmp171 = (K500000000 * tmp165) + tmp162;
128 tmp172 = tmp170 + tmp171;
129 tmp181 = tmp171 - tmp170;
131 X[0] = tmp155 + (K2_000000000 * tmp160);
132 tmp161 = tmp155 - tmp160;
133 tmp166 = K1_732050807 * (tmp162 - tmp165);
134 X[6 * iostride] = tmp161 + tmp166;
135 X[3 * iostride] = tmp161 - tmp166;
137 fftw_real tmp178;
138 fftw_real tmp173;
139 fftw_real tmp179;
140 fftw_real tmp185;
141 fftw_real tmp182;
142 fftw_real tmp184;
143 ASSERT_ALIGNED_DOUBLE();
144 tmp178 = (K1_113340798 * tmp169) + (K1_326827896 * tmp172);
145 tmp173 = (K766044443 * tmp169) - (K642787609 * tmp172);
146 tmp179 = tmp173 + tmp177;
147 X[iostride] = (K2_000000000 * tmp173) - tmp177;
148 X[7 * iostride] = tmp178 - tmp179;
149 X[4 * iostride] = -(tmp178 + tmp179);
150 tmp185 = (K1_705737063 * tmp180) + (K300767466 * tmp181);
151 tmp182 = (K173648177 * tmp180) - (K984807753 * tmp181);
152 tmp184 = tmp183 - tmp182;
153 X[2 * iostride] = (K2_000000000 * tmp182) + tmp183;
154 X[8 * iostride] = tmp185 + tmp184;
155 X[5 * iostride] = tmp184 - tmp185;
158 X = X + dist;
159 Y = Y - dist;
160 for (i = 2; i < m; i = i + 2, X = X + dist, Y = Y - dist, W = W + 8) {
161 fftw_real tmp45;
162 fftw_real tmp88;
163 fftw_real tmp136;
164 fftw_real tmp61;
165 fftw_real tmp108;
166 fftw_real tmp126;
167 fftw_real tmp50;
168 fftw_real tmp55;
169 fftw_real tmp56;
170 fftw_real tmp102;
171 fftw_real tmp110;
172 fftw_real tmp132;
173 fftw_real tmp138;
174 fftw_real tmp129;
175 fftw_real tmp137;
176 fftw_real tmp95;
177 fftw_real tmp109;
178 fftw_real tmp66;
179 fftw_real tmp71;
180 fftw_real tmp72;
181 ASSERT_ALIGNED_DOUBLE();
183 fftw_real tmp41;
184 fftw_real tmp57;
185 fftw_real tmp44;
186 fftw_real tmp106;
187 fftw_real tmp60;
188 fftw_real tmp87;
189 fftw_real tmp86;
190 fftw_real tmp107;
191 ASSERT_ALIGNED_DOUBLE();
192 tmp41 = X[0];
193 tmp57 = Y[0];
195 fftw_real tmp42;
196 fftw_real tmp43;
197 fftw_real tmp58;
198 fftw_real tmp59;
199 ASSERT_ALIGNED_DOUBLE();
200 tmp42 = X[3 * iostride];
201 tmp43 = Y[-6 * iostride];
202 tmp44 = tmp42 + tmp43;
203 tmp106 = K866025403 * (tmp42 - tmp43);
204 tmp58 = Y[-3 * iostride];
205 tmp59 = X[6 * iostride];
206 tmp60 = tmp58 - tmp59;
207 tmp87 = K866025403 * (tmp58 + tmp59);
209 tmp45 = tmp41 + tmp44;
210 tmp86 = tmp41 - (K500000000 * tmp44);
211 tmp88 = tmp86 - tmp87;
212 tmp136 = tmp86 + tmp87;
213 tmp61 = tmp57 + tmp60;
214 tmp107 = tmp57 - (K500000000 * tmp60);
215 tmp108 = tmp106 + tmp107;
216 tmp126 = tmp107 - tmp106;
219 fftw_real tmp46;
220 fftw_real tmp49;
221 fftw_real tmp89;
222 fftw_real tmp92;
223 fftw_real tmp62;
224 fftw_real tmp65;
225 fftw_real tmp90;
226 fftw_real tmp93;
227 fftw_real tmp51;
228 fftw_real tmp54;
229 fftw_real tmp96;
230 fftw_real tmp99;
231 fftw_real tmp67;
232 fftw_real tmp70;
233 fftw_real tmp97;
234 fftw_real tmp100;
235 ASSERT_ALIGNED_DOUBLE();
237 fftw_real tmp47;
238 fftw_real tmp48;
239 fftw_real tmp63;
240 fftw_real tmp64;
241 ASSERT_ALIGNED_DOUBLE();
242 tmp46 = X[iostride];
243 tmp47 = X[4 * iostride];
244 tmp48 = Y[-7 * iostride];
245 tmp49 = tmp47 + tmp48;
246 tmp89 = tmp46 - (K500000000 * tmp49);
247 tmp92 = K866025403 * (tmp47 - tmp48);
248 tmp62 = Y[-iostride];
249 tmp63 = Y[-4 * iostride];
250 tmp64 = X[7 * iostride];
251 tmp65 = tmp63 - tmp64;
252 tmp90 = K866025403 * (tmp63 + tmp64);
253 tmp93 = tmp62 - (K500000000 * tmp65);
256 fftw_real tmp52;
257 fftw_real tmp53;
258 fftw_real tmp68;
259 fftw_real tmp69;
260 ASSERT_ALIGNED_DOUBLE();
261 tmp51 = X[2 * iostride];
262 tmp52 = Y[-5 * iostride];
263 tmp53 = Y[-8 * iostride];
264 tmp54 = tmp52 + tmp53;
265 tmp96 = tmp51 - (K500000000 * tmp54);
266 tmp99 = K866025403 * (tmp52 - tmp53);
267 tmp67 = Y[-2 * iostride];
268 tmp68 = X[5 * iostride];
269 tmp69 = X[8 * iostride];
270 tmp70 = tmp68 + tmp69;
271 tmp97 = K866025403 * (tmp68 - tmp69);
272 tmp100 = tmp67 + (K500000000 * tmp70);
274 tmp50 = tmp46 + tmp49;
275 tmp55 = tmp51 + tmp54;
276 tmp56 = tmp50 + tmp55;
278 fftw_real tmp98;
279 fftw_real tmp101;
280 fftw_real tmp130;
281 fftw_real tmp131;
282 ASSERT_ALIGNED_DOUBLE();
283 tmp98 = tmp96 + tmp97;
284 tmp101 = tmp99 + tmp100;
285 tmp102 = (K173648177 * tmp98) - (K984807753 * tmp101);
286 tmp110 = (K984807753 * tmp98) + (K173648177 * tmp101);
287 tmp130 = tmp96 - tmp97;
288 tmp131 = tmp100 - tmp99;
289 tmp132 = (K342020143 * tmp130) - (K939692620 * tmp131);
290 tmp138 = (K939692620 * tmp130) + (K342020143 * tmp131);
293 fftw_real tmp127;
294 fftw_real tmp128;
295 fftw_real tmp91;
296 fftw_real tmp94;
297 ASSERT_ALIGNED_DOUBLE();
298 tmp127 = tmp93 - tmp92;
299 tmp128 = tmp89 + tmp90;
300 tmp129 = (K173648177 * tmp127) + (K984807753 * tmp128);
301 tmp137 = (K173648177 * tmp128) - (K984807753 * tmp127);
302 tmp91 = tmp89 - tmp90;
303 tmp94 = tmp92 + tmp93;
304 tmp95 = (K766044443 * tmp91) - (K642787609 * tmp94);
305 tmp109 = (K766044443 * tmp94) + (K642787609 * tmp91);
306 tmp66 = tmp62 + tmp65;
307 tmp71 = tmp67 - tmp70;
308 tmp72 = tmp66 + tmp71;
311 X[0] = tmp45 + tmp56;
313 fftw_real tmp76;
314 fftw_real tmp82;
315 fftw_real tmp80;
316 fftw_real tmp84;
317 ASSERT_ALIGNED_DOUBLE();
319 fftw_real tmp74;
320 fftw_real tmp75;
321 fftw_real tmp78;
322 fftw_real tmp79;
323 ASSERT_ALIGNED_DOUBLE();
324 tmp74 = tmp61 - (K500000000 * tmp72);
325 tmp75 = K866025403 * (tmp50 - tmp55);
326 tmp76 = tmp74 - tmp75;
327 tmp82 = tmp75 + tmp74;
328 tmp78 = tmp45 - (K500000000 * tmp56);
329 tmp79 = K866025403 * (tmp71 - tmp66);
330 tmp80 = tmp78 - tmp79;
331 tmp84 = tmp78 + tmp79;
334 fftw_real tmp73;
335 fftw_real tmp77;
336 fftw_real tmp81;
337 fftw_real tmp83;
338 ASSERT_ALIGNED_DOUBLE();
339 tmp73 = c_re(W[5]);
340 tmp77 = c_im(W[5]);
341 Y[-2 * iostride] = (tmp73 * tmp76) - (tmp77 * tmp80);
342 X[6 * iostride] = (tmp77 * tmp76) + (tmp73 * tmp80);
343 tmp81 = c_re(W[2]);
344 tmp83 = c_im(W[2]);
345 Y[-5 * iostride] = (tmp81 * tmp82) - (tmp83 * tmp84);
346 X[3 * iostride] = (tmp83 * tmp82) + (tmp81 * tmp84);
349 Y[-8 * iostride] = tmp61 + tmp72;
351 fftw_real tmp115;
352 fftw_real tmp104;
353 fftw_real tmp118;
354 fftw_real tmp119;
355 fftw_real tmp112;
356 fftw_real tmp114;
357 fftw_real tmp103;
358 fftw_real tmp111;
359 fftw_real tmp85;
360 fftw_real tmp105;
361 ASSERT_ALIGNED_DOUBLE();
362 tmp115 = K866025403 * (tmp95 - tmp102);
363 tmp103 = tmp95 + tmp102;
364 tmp104 = tmp88 + tmp103;
365 tmp118 = tmp88 - (K500000000 * tmp103);
366 tmp119 = K866025403 * (tmp110 - tmp109);
367 tmp111 = tmp109 + tmp110;
368 tmp112 = tmp108 + tmp111;
369 tmp114 = tmp108 - (K500000000 * tmp111);
370 tmp85 = c_re(W[0]);
371 tmp105 = c_im(W[0]);
372 X[iostride] = (tmp85 * tmp104) + (tmp105 * tmp112);
373 Y[-7 * iostride] = (tmp85 * tmp112) - (tmp105 * tmp104);
375 fftw_real tmp122;
376 fftw_real tmp124;
377 fftw_real tmp121;
378 fftw_real tmp123;
379 ASSERT_ALIGNED_DOUBLE();
380 tmp122 = tmp115 + tmp114;
381 tmp124 = tmp118 + tmp119;
382 tmp121 = c_re(W[3]);
383 tmp123 = c_im(W[3]);
384 Y[-4 * iostride] = (tmp121 * tmp122) - (tmp123 * tmp124);
385 X[4 * iostride] = (tmp123 * tmp122) + (tmp121 * tmp124);
388 fftw_real tmp116;
389 fftw_real tmp120;
390 fftw_real tmp113;
391 fftw_real tmp117;
392 ASSERT_ALIGNED_DOUBLE();
393 tmp116 = tmp114 - tmp115;
394 tmp120 = tmp118 - tmp119;
395 tmp113 = c_re(W[6]);
396 tmp117 = c_im(W[6]);
397 Y[-iostride] = (tmp113 * tmp116) - (tmp117 * tmp120);
398 X[7 * iostride] = (tmp117 * tmp116) + (tmp113 * tmp120);
402 fftw_real tmp143;
403 fftw_real tmp134;
404 fftw_real tmp146;
405 fftw_real tmp147;
406 fftw_real tmp140;
407 fftw_real tmp142;
408 fftw_real tmp133;
409 fftw_real tmp139;
410 fftw_real tmp125;
411 fftw_real tmp135;
412 ASSERT_ALIGNED_DOUBLE();
413 tmp143 = K866025403 * (tmp132 - tmp129);
414 tmp133 = tmp129 + tmp132;
415 tmp134 = tmp126 + tmp133;
416 tmp146 = tmp126 - (K500000000 * tmp133);
417 tmp147 = K866025403 * (tmp137 + tmp138);
418 tmp139 = tmp137 - tmp138;
419 tmp140 = tmp136 + tmp139;
420 tmp142 = tmp136 - (K500000000 * tmp139);
421 tmp125 = c_re(W[1]);
422 tmp135 = c_im(W[1]);
423 Y[-6 * iostride] = (tmp125 * tmp134) - (tmp135 * tmp140);
424 X[2 * iostride] = (tmp135 * tmp134) + (tmp125 * tmp140);
426 fftw_real tmp150;
427 fftw_real tmp152;
428 fftw_real tmp149;
429 fftw_real tmp151;
430 ASSERT_ALIGNED_DOUBLE();
431 tmp150 = tmp143 + tmp142;
432 tmp152 = tmp146 + tmp147;
433 tmp149 = c_re(W[4]);
434 tmp151 = c_im(W[4]);
435 X[5 * iostride] = (tmp149 * tmp150) + (tmp151 * tmp152);
436 Y[-3 * iostride] = (tmp149 * tmp152) - (tmp151 * tmp150);
439 fftw_real tmp144;
440 fftw_real tmp148;
441 fftw_real tmp141;
442 fftw_real tmp145;
443 ASSERT_ALIGNED_DOUBLE();
444 tmp144 = tmp142 - tmp143;
445 tmp148 = tmp146 - tmp147;
446 tmp141 = c_re(W[7]);
447 tmp145 = c_im(W[7]);
448 X[8 * iostride] = (tmp141 * tmp144) + (tmp145 * tmp148);
449 Y[0] = (tmp141 * tmp148) - (tmp145 * tmp144);
453 if (i == m) {
454 fftw_real tmp3;
455 fftw_real tmp33;
456 fftw_real tmp17;
457 fftw_real tmp12;
458 fftw_real tmp11;
459 fftw_real tmp20;
460 fftw_real tmp34;
461 fftw_real tmp7;
462 fftw_real tmp36;
463 fftw_real tmp35;
464 fftw_real tmp23;
465 fftw_real tmp29;
466 fftw_real tmp26;
467 fftw_real tmp37;
468 fftw_real tmp8;
469 fftw_real tmp13;
470 ASSERT_ALIGNED_DOUBLE();
472 fftw_real tmp16;
473 fftw_real tmp1;
474 fftw_real tmp2;
475 fftw_real tmp14;
476 fftw_real tmp15;
477 ASSERT_ALIGNED_DOUBLE();
478 tmp15 = Y[-iostride];
479 tmp16 = K1_732050807 * tmp15;
480 tmp1 = X[4 * iostride];
481 tmp2 = X[iostride];
482 tmp14 = tmp2 - tmp1;
483 tmp3 = tmp1 + (K2_000000000 * tmp2);
484 tmp33 = tmp14 - tmp16;
485 tmp17 = tmp14 + tmp16;
488 fftw_real tmp4;
489 fftw_real tmp6;
490 fftw_real tmp5;
491 fftw_real tmp24;
492 fftw_real tmp22;
493 fftw_real tmp27;
494 fftw_real tmp18;
495 fftw_real tmp10;
496 fftw_real tmp21;
497 fftw_real tmp19;
498 fftw_real tmp28;
499 fftw_real tmp9;
500 fftw_real tmp25;
501 ASSERT_ALIGNED_DOUBLE();
502 tmp4 = X[3 * iostride];
503 tmp6 = X[2 * iostride];
504 tmp5 = X[0];
505 tmp24 = K866025403 * (tmp4 - tmp5);
506 tmp22 = (K500000000 * (tmp4 + tmp5)) - tmp6;
507 tmp27 = K866025403 * (tmp6 - tmp5);
508 tmp18 = (K500000000 * (tmp5 + tmp6)) - tmp4;
509 tmp9 = Y[0];
510 tmp12 = Y[-2 * iostride];
511 tmp10 = Y[-3 * iostride];
512 tmp11 = tmp9 + tmp10;
513 tmp21 = K866025403 * (tmp9 - tmp10);
514 tmp19 = K866025403 * (tmp9 + tmp12);
515 tmp28 = K500000000 * (tmp12 - tmp9);
516 tmp20 = tmp18 + tmp19;
517 tmp34 = tmp18 - tmp19;
518 tmp7 = tmp4 + tmp5 + tmp6;
519 tmp36 = tmp27 + tmp28 + tmp10;
520 tmp35 = tmp22 + tmp21;
521 tmp23 = tmp21 - tmp22;
522 tmp29 = tmp27 - (tmp28 + tmp10);
523 tmp25 = K500000000 * tmp11;
524 tmp26 = tmp24 - tmp25 - tmp12;
525 tmp37 = (tmp24 + tmp12) + tmp25;
527 X[0] = tmp3 + (K2_000000000 * tmp7);
528 tmp8 = tmp7 - tmp3;
529 tmp13 = K1_732050807 * (tmp11 - tmp12);
530 X[3 * iostride] = tmp8 - tmp13;
531 X[6 * iostride] = -(tmp8 + tmp13);
533 fftw_real tmp32;
534 fftw_real tmp30;
535 fftw_real tmp31;
536 fftw_real tmp40;
537 fftw_real tmp38;
538 fftw_real tmp39;
539 ASSERT_ALIGNED_DOUBLE();
540 tmp32 = (K813797681 * tmp26) + (K150383733 * tmp29) - (K296198132 * tmp23) - (K852868531 * tmp20);
541 tmp30 = (K173648177 * tmp20) + (K939692620 * tmp23) + (K342020143 * tmp26) + (K984807753 * tmp29);
542 tmp31 = tmp17 - (K500000000 * tmp30);
543 X[2 * iostride] = -(tmp17 + tmp30);
544 X[8 * iostride] = tmp32 - tmp31;
545 X[5 * iostride] = tmp31 + tmp32;
546 tmp40 = (K556670399 * tmp34) + (K663413948 * tmp36) - (K150383733 * tmp37) - (K852868531 * tmp35);
547 tmp38 = (K766044443 * tmp34) + (K173648177 * tmp35) - (K642787609 * tmp36) - (K984807753 * tmp37);
548 tmp39 = (K500000000 * tmp38) - tmp33;
549 X[iostride] = tmp33 + tmp38;
550 X[7 * iostride] = tmp40 - tmp39;
551 X[4 * iostride] = tmp39 + tmp40;
556 static const int twiddle_order[] =
557 {1, 2, 3, 4, 5, 6, 7, 8};
558 fftw_codelet_desc fftw_hc2hc_backward_9_desc =
560 "fftw_hc2hc_backward_9",
561 (void (*)()) fftw_hc2hc_backward_9,
563 FFTW_BACKWARD,
564 FFTW_HC2HC,
565 212,
567 twiddle_order,