xref: /aosp_15_r20/external/eigen/unsupported/Eigen/src/SpecialFunctions/BesselFunctionsImpl.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2015 Eugene Brevdo <[email protected]>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_BESSEL_FUNCTIONS_H
11 #define EIGEN_BESSEL_FUNCTIONS_H
12 
13 namespace Eigen {
14 namespace internal {
15 
16 //  Parts of this code are based on the Cephes Math Library.
17 //
18 //  Cephes Math Library Release 2.8:  June, 2000
19 //  Copyright 1984, 1987, 1992, 2000 by Stephen L. Moshier
20 //
21 //  Permission has been kindly provided by the original author
22 //  to incorporate the Cephes software into the Eigen codebase:
23 //
24 //    From: Stephen Moshier
25 //    To: Eugene Brevdo
26 //    Subject: Re: Permission to wrap several cephes functions in Eigen
27 //
28 //    Hello Eugene,
29 //
30 //    Thank you for writing.
31 //
32 //    If your licensing is similar to BSD, the formal way that has been
33 //    handled is simply to add a statement to the effect that you are incorporating
34 //    the Cephes software by permission of the author.
35 //
36 //    Good luck with your project,
37 //    Steve
38 
39 
40 /****************************************************************************
41  * Implementation of Bessel function, based on Cephes                       *
42  ****************************************************************************/
43 
44 template <typename Scalar>
45 struct bessel_i0e_retval {
46   typedef Scalar type;
47 };
48 
49 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
50 struct generic_i0e {
51   EIGEN_DEVICE_FUNC
rungeneric_i0e52   static EIGEN_STRONG_INLINE T run(const T&) {
53     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
54                         THIS_TYPE_IS_NOT_SUPPORTED);
55     return ScalarType(0);
56   }
57 };
58 
59 template <typename T>
60 struct generic_i0e<T, float> {
61   EIGEN_DEVICE_FUNC
62   static EIGEN_STRONG_INLINE T run(const T& x) {
63     /*  i0ef.c
64      *
65      *  Modified Bessel function of order zero,
66      *  exponentially scaled
67      *
68      *
69      *
70      * SYNOPSIS:
71      *
72      * float x, y, i0ef();
73      *
74      * y = i0ef( x );
75      *
76      *
77      *
78      * DESCRIPTION:
79      *
80      * Returns exponentially scaled modified Bessel function
81      * of order zero of the argument.
82      *
83      * The function is defined as i0e(x) = exp(-|x|) j0( ix ).
84      *
85      *
86      *
87      * ACCURACY:
88      *
89      *                      Relative error:
90      * arithmetic   domain     # trials      peak         rms
91      *    IEEE      0,30        100000      3.7e-7      7.0e-8
92      * See i0f().
93      *
94      */
95 
96     const float A[] = {-1.30002500998624804212E-8f, 6.04699502254191894932E-8f,
97                        -2.67079385394061173391E-7f, 1.11738753912010371815E-6f,
98                        -4.41673835845875056359E-6f, 1.64484480707288970893E-5f,
99                        -5.75419501008210370398E-5f, 1.88502885095841655729E-4f,
100                        -5.76375574538582365885E-4f, 1.63947561694133579842E-3f,
101                        -4.32430999505057594430E-3f, 1.05464603945949983183E-2f,
102                        -2.37374148058994688156E-2f, 4.93052842396707084878E-2f,
103                        -9.49010970480476444210E-2f, 1.71620901522208775349E-1f,
104                        -3.04682672343198398683E-1f, 6.76795274409476084995E-1f};
105 
106     const float B[] = {3.39623202570838634515E-9f, 2.26666899049817806459E-8f,
107                        2.04891858946906374183E-7f, 2.89137052083475648297E-6f,
108                        6.88975834691682398426E-5f, 3.36911647825569408990E-3f,
109                        8.04490411014108831608E-1f};
110     T y = pabs(x);
111     T y_le_eight = internal::pchebevl<T, 18>::run(
112         pmadd(pset1<T>(0.5f), y, pset1<T>(-2.0f)), A);
113     T y_gt_eight = pmul(
114         internal::pchebevl<T, 7>::run(
115             psub(pdiv(pset1<T>(32.0f), y), pset1<T>(2.0f)), B),
116         prsqrt(y));
117     // TODO: Perhaps instead check whether all packet elements are in
118     // [-8, 8] and evaluate a branch based off of that. It's possible
119     // in practice most elements are in this region.
120     return pselect(pcmp_le(y, pset1<T>(8.0f)), y_le_eight, y_gt_eight);
121   }
122 };
123 
124 template <typename T>
125 struct generic_i0e<T, double> {
126   EIGEN_DEVICE_FUNC
127   static EIGEN_STRONG_INLINE T run(const T& x) {
128     /*  i0e.c
129      *
130      *  Modified Bessel function of order zero,
131      *  exponentially scaled
132      *
133      *
134      *
135      * SYNOPSIS:
136      *
137      * double x, y, i0e();
138      *
139      * y = i0e( x );
140      *
141      *
142      *
143      * DESCRIPTION:
144      *
145      * Returns exponentially scaled modified Bessel function
146      * of order zero of the argument.
147      *
148      * The function is defined as i0e(x) = exp(-|x|) j0( ix ).
149      *
150      *
151      *
152      * ACCURACY:
153      *
154      *                      Relative error:
155      * arithmetic   domain     # trials      peak         rms
156      *    IEEE      0,30        30000       5.4e-16     1.2e-16
157      * See i0().
158      *
159      */
160 
161     const double A[] = {-4.41534164647933937950E-18, 3.33079451882223809783E-17,
162                         -2.43127984654795469359E-16, 1.71539128555513303061E-15,
163                         -1.16853328779934516808E-14, 7.67618549860493561688E-14,
164                         -4.85644678311192946090E-13, 2.95505266312963983461E-12,
165                         -1.72682629144155570723E-11, 9.67580903537323691224E-11,
166                         -5.18979560163526290666E-10, 2.65982372468238665035E-9,
167                         -1.30002500998624804212E-8,  6.04699502254191894932E-8,
168                         -2.67079385394061173391E-7,  1.11738753912010371815E-6,
169                         -4.41673835845875056359E-6,  1.64484480707288970893E-5,
170                         -5.75419501008210370398E-5,  1.88502885095841655729E-4,
171                         -5.76375574538582365885E-4,  1.63947561694133579842E-3,
172                         -4.32430999505057594430E-3,  1.05464603945949983183E-2,
173                         -2.37374148058994688156E-2,  4.93052842396707084878E-2,
174                         -9.49010970480476444210E-2,  1.71620901522208775349E-1,
175                         -3.04682672343198398683E-1,  6.76795274409476084995E-1};
176     const double B[] = {
177         -7.23318048787475395456E-18, -4.83050448594418207126E-18,
178         4.46562142029675999901E-17,  3.46122286769746109310E-17,
179         -2.82762398051658348494E-16, -3.42548561967721913462E-16,
180         1.77256013305652638360E-15,  3.81168066935262242075E-15,
181         -9.55484669882830764870E-15, -4.15056934728722208663E-14,
182         1.54008621752140982691E-14,  3.85277838274214270114E-13,
183         7.18012445138366623367E-13,  -1.79417853150680611778E-12,
184         -1.32158118404477131188E-11, -3.14991652796324136454E-11,
185         1.18891471078464383424E-11,  4.94060238822496958910E-10,
186         3.39623202570838634515E-9,   2.26666899049817806459E-8,
187         2.04891858946906374183E-7,   2.89137052083475648297E-6,
188         6.88975834691682398426E-5,   3.36911647825569408990E-3,
189         8.04490411014108831608E-1};
190     T y = pabs(x);
191     T y_le_eight = internal::pchebevl<T, 30>::run(
192         pmadd(pset1<T>(0.5), y, pset1<T>(-2.0)), A);
193     T y_gt_eight = pmul(
194         internal::pchebevl<T, 25>::run(
195             psub(pdiv(pset1<T>(32.0), y), pset1<T>(2.0)), B),
196         prsqrt(y));
197     // TODO: Perhaps instead check whether all packet elements are in
198     // [-8, 8] and evaluate a branch based off of that. It's possible
199     // in practice most elements are in this region.
200     return pselect(pcmp_le(y, pset1<T>(8.0)), y_le_eight, y_gt_eight);
201   }
202 };
203 
204 template <typename T>
205 struct bessel_i0e_impl {
206   EIGEN_DEVICE_FUNC
207   static EIGEN_STRONG_INLINE T run(const T x) {
208     return generic_i0e<T>::run(x);
209   }
210 };
211 
212 template <typename Scalar>
213 struct bessel_i0_retval {
214   typedef Scalar type;
215 };
216 
217 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
218 struct generic_i0 {
219   EIGEN_DEVICE_FUNC
220   static EIGEN_STRONG_INLINE T run(const T& x) {
221     return pmul(
222         pexp(pabs(x)),
223         generic_i0e<T, ScalarType>::run(x));
224   }
225 };
226 
227 template <typename T>
228 struct bessel_i0_impl {
229   EIGEN_DEVICE_FUNC
230   static EIGEN_STRONG_INLINE T run(const T x) {
231     return generic_i0<T>::run(x);
232   }
233 };
234 
235 template <typename Scalar>
236 struct bessel_i1e_retval {
237   typedef Scalar type;
238 };
239 
240 template <typename T, typename ScalarType = typename unpacket_traits<T>::type >
241 struct generic_i1e {
242   EIGEN_DEVICE_FUNC
243   static EIGEN_STRONG_INLINE T run(const T&) {
244     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
245                         THIS_TYPE_IS_NOT_SUPPORTED);
246     return ScalarType(0);
247   }
248 };
249 
250 template <typename T>
251 struct generic_i1e<T, float> {
252   EIGEN_DEVICE_FUNC
253   static EIGEN_STRONG_INLINE T run(const T& x) {
254     /* i1ef.c
255      *
256      *  Modified Bessel function of order one,
257      *  exponentially scaled
258      *
259      *
260      *
261      * SYNOPSIS:
262      *
263      * float x, y, i1ef();
264      *
265      * y = i1ef( x );
266      *
267      *
268      *
269      * DESCRIPTION:
270      *
271      * Returns exponentially scaled modified Bessel function
272      * of order one of the argument.
273      *
274      * The function is defined as i1(x) = -i exp(-|x|) j1( ix ).
275      *
276      *
277      *
278      * ACCURACY:
279      *
280      *                      Relative error:
281      * arithmetic   domain     # trials      peak         rms
282      *    IEEE      0, 30       30000       1.5e-6      1.5e-7
283      * See i1().
284      *
285      */
286     const float A[] = {9.38153738649577178388E-9f, -4.44505912879632808065E-8f,
287                        2.00329475355213526229E-7f, -8.56872026469545474066E-7f,
288                        3.47025130813767847674E-6f, -1.32731636560394358279E-5f,
289                        4.78156510755005422638E-5f, -1.61760815825896745588E-4f,
290                        5.12285956168575772895E-4f, -1.51357245063125314899E-3f,
291                        4.15642294431288815669E-3f, -1.05640848946261981558E-2f,
292                        2.47264490306265168283E-2f, -5.29459812080949914269E-2f,
293                        1.02643658689847095384E-1f, -1.76416518357834055153E-1f,
294                        2.52587186443633654823E-1f};
295 
296     const float B[] = {-3.83538038596423702205E-9f, -2.63146884688951950684E-8f,
297                        -2.51223623787020892529E-7f, -3.88256480887769039346E-6f,
298                        -1.10588938762623716291E-4f, -9.76109749136146840777E-3f,
299                        7.78576235018280120474E-1f};
300 
301 
302     T y = pabs(x);
303     T y_le_eight = pmul(y, internal::pchebevl<T, 17>::run(
304         pmadd(pset1<T>(0.5f), y, pset1<T>(-2.0f)), A));
305     T y_gt_eight = pmul(
306         internal::pchebevl<T, 7>::run(
307             psub(pdiv(pset1<T>(32.0f), y),
308                  pset1<T>(2.0f)), B),
309         prsqrt(y));
310     // TODO: Perhaps instead check whether all packet elements are in
311     // [-8, 8] and evaluate a branch based off of that. It's possible
312     // in practice most elements are in this region.
313     y = pselect(pcmp_le(y, pset1<T>(8.0f)), y_le_eight, y_gt_eight);
314     return pselect(pcmp_lt(x, pset1<T>(0.0f)), pnegate(y), y);
315   }
316 };
317 
318 template <typename T>
319 struct generic_i1e<T, double> {
320   EIGEN_DEVICE_FUNC
321   static EIGEN_STRONG_INLINE T run(const T& x) {
322     /*  i1e.c
323      *
324      *  Modified Bessel function of order one,
325      *  exponentially scaled
326      *
327      *
328      *
329      * SYNOPSIS:
330      *
331      * double x, y, i1e();
332      *
333      * y = i1e( x );
334      *
335      *
336      *
337      * DESCRIPTION:
338      *
339      * Returns exponentially scaled modified Bessel function
340      * of order one of the argument.
341      *
342      * The function is defined as i1(x) = -i exp(-|x|) j1( ix ).
343      *
344      *
345      *
346      * ACCURACY:
347      *
348      *                      Relative error:
349      * arithmetic   domain     # trials      peak         rms
350      *    IEEE      0, 30       30000       2.0e-15     2.0e-16
351      * See i1().
352      *
353      */
354     const double A[] = {2.77791411276104639959E-18, -2.11142121435816608115E-17,
355                         1.55363195773620046921E-16, -1.10559694773538630805E-15,
356                         7.60068429473540693410E-15, -5.04218550472791168711E-14,
357                         3.22379336594557470981E-13, -1.98397439776494371520E-12,
358                         1.17361862988909016308E-11, -6.66348972350202774223E-11,
359                         3.62559028155211703701E-10, -1.88724975172282928790E-9,
360                         9.38153738649577178388E-9,  -4.44505912879632808065E-8,
361                         2.00329475355213526229E-7,  -8.56872026469545474066E-7,
362                         3.47025130813767847674E-6,  -1.32731636560394358279E-5,
363                         4.78156510755005422638E-5,  -1.61760815825896745588E-4,
364                         5.12285956168575772895E-4,  -1.51357245063125314899E-3,
365                         4.15642294431288815669E-3,  -1.05640848946261981558E-2,
366                         2.47264490306265168283E-2,  -5.29459812080949914269E-2,
367                         1.02643658689847095384E-1,  -1.76416518357834055153E-1,
368                         2.52587186443633654823E-1};
369     const double B[] = {
370         7.51729631084210481353E-18,  4.41434832307170791151E-18,
371         -4.65030536848935832153E-17, -3.20952592199342395980E-17,
372         2.96262899764595013876E-16,  3.30820231092092828324E-16,
373         -1.88035477551078244854E-15, -3.81440307243700780478E-15,
374         1.04202769841288027642E-14,  4.27244001671195135429E-14,
375         -2.10154184277266431302E-14, -4.08355111109219731823E-13,
376         -7.19855177624590851209E-13, 2.03562854414708950722E-12,
377         1.41258074366137813316E-11,  3.25260358301548823856E-11,
378         -1.89749581235054123450E-11, -5.58974346219658380687E-10,
379         -3.83538038596423702205E-9,  -2.63146884688951950684E-8,
380         -2.51223623787020892529E-7,  -3.88256480887769039346E-6,
381         -1.10588938762623716291E-4,  -9.76109749136146840777E-3,
382         7.78576235018280120474E-1};
383     T y = pabs(x);
384     T y_le_eight = pmul(y, internal::pchebevl<T, 29>::run(
385         pmadd(pset1<T>(0.5), y, pset1<T>(-2.0)), A));
386     T y_gt_eight = pmul(
387         internal::pchebevl<T, 25>::run(
388             psub(pdiv(pset1<T>(32.0), y),
389                  pset1<T>(2.0)), B),
390         prsqrt(y));
391     // TODO: Perhaps instead check whether all packet elements are in
392     // [-8, 8] and evaluate a branch based off of that. It's possible
393     // in practice most elements are in this region.
394     y = pselect(pcmp_le(y, pset1<T>(8.0)), y_le_eight, y_gt_eight);
395     return pselect(pcmp_lt(x, pset1<T>(0.0)), pnegate(y), y);
396   }
397 };
398 
399 template <typename T>
400 struct bessel_i1e_impl {
401   EIGEN_DEVICE_FUNC
402   static EIGEN_STRONG_INLINE T run(const T x) {
403     return generic_i1e<T>::run(x);
404   }
405 };
406 
407 template <typename T>
408 struct bessel_i1_retval {
409   typedef T type;
410 };
411 
412 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
413 struct generic_i1 {
414   EIGEN_DEVICE_FUNC
415   static EIGEN_STRONG_INLINE T run(const T& x) {
416     return pmul(
417         pexp(pabs(x)),
418         generic_i1e<T, ScalarType>::run(x));
419   }
420 };
421 
422 template <typename T>
423 struct bessel_i1_impl {
424   EIGEN_DEVICE_FUNC
425   static EIGEN_STRONG_INLINE T run(const T x) {
426     return generic_i1<T>::run(x);
427   }
428 };
429 
430 template <typename T>
431 struct bessel_k0e_retval {
432   typedef T type;
433 };
434 
435 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
436 struct generic_k0e {
437   EIGEN_DEVICE_FUNC
438   static EIGEN_STRONG_INLINE T run(const T&) {
439     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
440                         THIS_TYPE_IS_NOT_SUPPORTED);
441     return ScalarType(0);
442   }
443 };
444 
445 template <typename T>
446 struct generic_k0e<T, float> {
447   EIGEN_DEVICE_FUNC
448   static EIGEN_STRONG_INLINE T run(const T& x) {
449     /*  k0ef.c
450      *	Modified Bessel function, third kind, order zero,
451      *	exponentially scaled
452      *
453      *
454      *
455      * SYNOPSIS:
456      *
457      * float x, y, k0ef();
458      *
459      * y = k0ef( x );
460      *
461      *
462      *
463      * DESCRIPTION:
464      *
465      * Returns exponentially scaled modified Bessel function
466      * of the third kind of order zero of the argument.
467      *
468      *
469      *
470      * ACCURACY:
471      *
472      *                      Relative error:
473      * arithmetic   domain     # trials      peak         rms
474      *    IEEE      0, 30       30000       8.1e-7      7.8e-8
475      * See k0().
476      *
477      */
478 
479     const float A[] = {1.90451637722020886025E-9f, 2.53479107902614945675E-7f,
480                        2.28621210311945178607E-5f, 1.26461541144692592338E-3f,
481                        3.59799365153615016266E-2f, 3.44289899924628486886E-1f,
482                        -5.35327393233902768720E-1f};
483 
484     const float B[] = {-1.69753450938905987466E-9f, 8.57403401741422608519E-9f,
485                        -4.66048989768794782956E-8f, 2.76681363944501510342E-7f,
486                        -1.83175552271911948767E-6f, 1.39498137188764993662E-5f,
487                        -1.28495495816278026384E-4f, 1.56988388573005337491E-3f,
488                        -3.14481013119645005427E-2f, 2.44030308206595545468E0f};
489     const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
490     const T two = pset1<T>(2.0);
491     T x_le_two = internal::pchebevl<T, 7>::run(
492         pmadd(x, x, pset1<T>(-2.0)), A);
493     x_le_two = pmadd(
494         generic_i0<T, float>::run(x), pnegate(
495             plog(pmul(pset1<T>(0.5), x))), x_le_two);
496     x_le_two = pmul(pexp(x), x_le_two);
497     T x_gt_two = pmul(
498             internal::pchebevl<T, 10>::run(
499                 psub(pdiv(pset1<T>(8.0), x), two), B),
500             prsqrt(x));
501     return pselect(
502         pcmp_le(x, pset1<T>(0.0)),
503         MAXNUM,
504         pselect(pcmp_le(x, two), x_le_two, x_gt_two));
505   }
506 };
507 
508 template <typename T>
509 struct generic_k0e<T, double> {
510   EIGEN_DEVICE_FUNC
511   static EIGEN_STRONG_INLINE T run(const T& x) {
512     /*  k0e.c
513      *	Modified Bessel function, third kind, order zero,
514      *	exponentially scaled
515      *
516      *
517      *
518      * SYNOPSIS:
519      *
520      * double x, y, k0e();
521      *
522      * y = k0e( x );
523      *
524      *
525      *
526      * DESCRIPTION:
527      *
528      * Returns exponentially scaled modified Bessel function
529      * of the third kind of order zero of the argument.
530      *
531      *
532      *
533      * ACCURACY:
534      *
535      *                      Relative error:
536      * arithmetic   domain     # trials      peak         rms
537      *    IEEE      0, 30       30000       1.4e-15     1.4e-16
538      * See k0().
539      *
540      */
541 
542     const double A[] = {
543       1.37446543561352307156E-16,
544       4.25981614279661018399E-14,
545       1.03496952576338420167E-11,
546       1.90451637722020886025E-9,
547       2.53479107902614945675E-7,
548       2.28621210311945178607E-5,
549       1.26461541144692592338E-3,
550       3.59799365153615016266E-2,
551       3.44289899924628486886E-1,
552       -5.35327393233902768720E-1};
553     const double B[] = {
554        5.30043377268626276149E-18, -1.64758043015242134646E-17,
555        5.21039150503902756861E-17, -1.67823109680541210385E-16,
556        5.51205597852431940784E-16, -1.84859337734377901440E-15,
557        6.34007647740507060557E-15, -2.22751332699166985548E-14,
558        8.03289077536357521100E-14, -2.98009692317273043925E-13,
559        1.14034058820847496303E-12, -4.51459788337394416547E-12,
560        1.85594911495471785253E-11, -7.95748924447710747776E-11,
561        3.57739728140030116597E-10, -1.69753450938905987466E-9,
562        8.57403401741422608519E-9, -4.66048989768794782956E-8,
563        2.76681363944501510342E-7, -1.83175552271911948767E-6,
564        1.39498137188764993662E-5, -1.28495495816278026384E-4,
565        1.56988388573005337491E-3, -3.14481013119645005427E-2,
566        2.44030308206595545468E0
567     };
568     const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
569     const T two = pset1<T>(2.0);
570     T x_le_two = internal::pchebevl<T, 10>::run(
571         pmadd(x, x, pset1<T>(-2.0)), A);
572     x_le_two = pmadd(
573         generic_i0<T, double>::run(x), pmul(
574             pset1<T>(-1.0), plog(pmul(pset1<T>(0.5), x))), x_le_two);
575     x_le_two = pmul(pexp(x), x_le_two);
576     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
577     T x_gt_two = pmul(
578             internal::pchebevl<T, 25>::run(
579                 psub(pdiv(pset1<T>(8.0), x), two), B),
580             prsqrt(x));
581     return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
582   }
583 };
584 
585 template <typename T>
586 struct bessel_k0e_impl {
587   EIGEN_DEVICE_FUNC
588   static EIGEN_STRONG_INLINE T run(const T x) {
589     return generic_k0e<T>::run(x);
590   }
591 };
592 
593 template <typename T>
594 struct bessel_k0_retval {
595   typedef T type;
596 };
597 
598 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
599 struct generic_k0 {
600   EIGEN_DEVICE_FUNC
601   static EIGEN_STRONG_INLINE T run(const T&) {
602     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
603                         THIS_TYPE_IS_NOT_SUPPORTED);
604     return ScalarType(0);
605   }
606 };
607 
608 template <typename T>
609 struct generic_k0<T, float> {
610   EIGEN_DEVICE_FUNC
611   static EIGEN_STRONG_INLINE T run(const T& x) {
612     /*  k0f.c
613      *	Modified Bessel function, third kind, order zero
614      *
615      *
616      *
617      * SYNOPSIS:
618      *
619      * float x, y, k0f();
620      *
621      * y = k0f( x );
622      *
623      *
624      *
625      * DESCRIPTION:
626      *
627      * Returns modified Bessel function of the third kind
628      * of order zero of the argument.
629      *
630      * The range is partitioned into the two intervals [0,8] and
631      * (8, infinity).  Chebyshev polynomial expansions are employed
632      * in each interval.
633      *
634      *
635      *
636      * ACCURACY:
637      *
638      * Tested at 2000 random points between 0 and 8.  Peak absolute
639      * error (relative when K0 > 1) was 1.46e-14; rms, 4.26e-15.
640      *                      Relative error:
641      * arithmetic   domain     # trials      peak         rms
642      *    IEEE      0, 30       30000       7.8e-7      8.5e-8
643      *
644      * ERROR MESSAGES:
645      *
646      *   message         condition      value returned
647      *  K0 domain          x <= 0          MAXNUM
648      *
649      */
650 
651     const float A[] = {1.90451637722020886025E-9f, 2.53479107902614945675E-7f,
652                        2.28621210311945178607E-5f, 1.26461541144692592338E-3f,
653                        3.59799365153615016266E-2f, 3.44289899924628486886E-1f,
654                        -5.35327393233902768720E-1f};
655 
656     const float B[] = {-1.69753450938905987466E-9f, 8.57403401741422608519E-9f,
657                        -4.66048989768794782956E-8f, 2.76681363944501510342E-7f,
658                        -1.83175552271911948767E-6f, 1.39498137188764993662E-5f,
659                        -1.28495495816278026384E-4f, 1.56988388573005337491E-3f,
660                        -3.14481013119645005427E-2f, 2.44030308206595545468E0f};
661     const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
662     const T two = pset1<T>(2.0);
663     T x_le_two = internal::pchebevl<T, 7>::run(
664         pmadd(x, x, pset1<T>(-2.0)), A);
665     x_le_two = pmadd(
666         generic_i0<T, float>::run(x), pnegate(
667             plog(pmul(pset1<T>(0.5), x))), x_le_two);
668     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
669     T x_gt_two = pmul(
670         pmul(
671             pexp(pnegate(x)),
672             internal::pchebevl<T, 10>::run(
673                 psub(pdiv(pset1<T>(8.0), x), two), B)),
674         prsqrt(x));
675     return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
676   }
677 };
678 
679 template <typename T>
680 struct generic_k0<T, double> {
681   EIGEN_DEVICE_FUNC
682   static EIGEN_STRONG_INLINE T run(const T& x) {
683     /*
684      *
685      *	Modified Bessel function, third kind, order zero,
686      *	exponentially scaled
687      *
688      *
689      *
690      * SYNOPSIS:
691      *
692      * double x, y, k0();
693      *
694      * y = k0( x );
695      *
696      *
697      *
698      * DESCRIPTION:
699      *
700      * Returns exponentially scaled modified Bessel function
701      * of the third kind of order zero of the argument.
702      *
703      *
704      *
705      * ACCURACY:
706      *
707      *                      Relative error:
708      * arithmetic   domain     # trials      peak         rms
709      *    IEEE      0, 30       30000       1.4e-15     1.4e-16
710      * See k0().
711      *
712      */
713     const double A[] = {
714       1.37446543561352307156E-16,
715       4.25981614279661018399E-14,
716       1.03496952576338420167E-11,
717       1.90451637722020886025E-9,
718       2.53479107902614945675E-7,
719       2.28621210311945178607E-5,
720       1.26461541144692592338E-3,
721       3.59799365153615016266E-2,
722       3.44289899924628486886E-1,
723       -5.35327393233902768720E-1};
724     const double B[] = {
725        5.30043377268626276149E-18, -1.64758043015242134646E-17,
726        5.21039150503902756861E-17, -1.67823109680541210385E-16,
727        5.51205597852431940784E-16, -1.84859337734377901440E-15,
728        6.34007647740507060557E-15, -2.22751332699166985548E-14,
729        8.03289077536357521100E-14, -2.98009692317273043925E-13,
730        1.14034058820847496303E-12, -4.51459788337394416547E-12,
731        1.85594911495471785253E-11, -7.95748924447710747776E-11,
732        3.57739728140030116597E-10, -1.69753450938905987466E-9,
733        8.57403401741422608519E-9, -4.66048989768794782956E-8,
734        2.76681363944501510342E-7, -1.83175552271911948767E-6,
735        1.39498137188764993662E-5, -1.28495495816278026384E-4,
736        1.56988388573005337491E-3, -3.14481013119645005427E-2,
737        2.44030308206595545468E0
738     };
739     const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
740     const T two = pset1<T>(2.0);
741     T x_le_two = internal::pchebevl<T, 10>::run(
742         pmadd(x, x, pset1<T>(-2.0)), A);
743     x_le_two = pmadd(
744         generic_i0<T, double>::run(x), pnegate(
745             plog(pmul(pset1<T>(0.5), x))), x_le_two);
746     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
747     T x_gt_two = pmul(
748         pmul(
749             pexp(-x),
750             internal::pchebevl<T, 25>::run(
751                 psub(pdiv(pset1<T>(8.0), x), two), B)),
752         prsqrt(x));
753     return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
754   }
755 };
756 
757 template <typename T>
758 struct bessel_k0_impl {
759   EIGEN_DEVICE_FUNC
760   static EIGEN_STRONG_INLINE T run(const T x) {
761     return generic_k0<T>::run(x);
762   }
763 };
764 
765 template <typename T>
766 struct bessel_k1e_retval {
767   typedef T type;
768 };
769 
770 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
771 struct generic_k1e {
772   EIGEN_DEVICE_FUNC
773   static EIGEN_STRONG_INLINE T run(const T&) {
774     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
775                         THIS_TYPE_IS_NOT_SUPPORTED);
776     return ScalarType(0);
777   }
778 };
779 
780 template <typename T>
781 struct generic_k1e<T, float> {
782   EIGEN_DEVICE_FUNC
783   static EIGEN_STRONG_INLINE T run(const T& x) {
784     /* k1ef.c
785      *
786      *	Modified Bessel function, third kind, order one,
787      *	exponentially scaled
788      *
789      *
790      *
791      * SYNOPSIS:
792      *
793      * float x, y, k1ef();
794      *
795      * y = k1ef( x );
796      *
797      *
798      *
799      * DESCRIPTION:
800      *
801      * Returns exponentially scaled modified Bessel function
802      * of the third kind of order one of the argument:
803      *
804      *      k1e(x) = exp(x) * k1(x).
805      *
806      *
807      *
808      * ACCURACY:
809      *
810      *                      Relative error:
811      * arithmetic   domain     # trials      peak         rms
812      *    IEEE      0, 30       30000       4.9e-7      6.7e-8
813      * See k1().
814      *
815      */
816 
817     const float A[] = {-2.21338763073472585583E-8f, -2.43340614156596823496E-6f,
818                         -1.73028895751305206302E-4f, -6.97572385963986435018E-3f,
819                         -1.22611180822657148235E-1f, -3.53155960776544875667E-1f,
820                         1.52530022733894777053E0f};
821     const float B[] = {2.01504975519703286596E-9f, -1.03457624656780970260E-8f,
822                        5.74108412545004946722E-8f, -3.50196060308781257119E-7f,
823                        2.40648494783721712015E-6f, -1.93619797416608296024E-5f,
824                        1.95215518471351631108E-4f, -2.85781685962277938680E-3f,
825                        1.03923736576817238437E-1f, 2.72062619048444266945E0f};
826     const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
827     const T two = pset1<T>(2.0);
828     T x_le_two = pdiv(internal::pchebevl<T, 7>::run(
829         pmadd(x, x, pset1<T>(-2.0)), A), x);
830     x_le_two = pmadd(
831         generic_i1<T, float>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
832     x_le_two = pmul(x_le_two, pexp(x));
833     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
834     T x_gt_two = pmul(
835         internal::pchebevl<T, 10>::run(
836             psub(pdiv(pset1<T>(8.0), x), two), B),
837         prsqrt(x));
838     return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
839   }
840 };
841 
842 template <typename T>
843 struct generic_k1e<T, double> {
844   EIGEN_DEVICE_FUNC
845   static EIGEN_STRONG_INLINE T run(const T& x) {
846     /*  k1e.c
847      *
848      *	Modified Bessel function, third kind, order one,
849      *	exponentially scaled
850      *
851      *
852      *
853      * SYNOPSIS:
854      *
855      * double x, y, k1e();
856      *
857      * y = k1e( x );
858      *
859      *
860      *
861      * DESCRIPTION:
862      *
863      * Returns exponentially scaled modified Bessel function
864      * of the third kind of order one of the argument:
865      *
866      *      k1e(x) = exp(x) * k1(x).
867      *
868      *
869      *
870      * ACCURACY:
871      *
872      *                      Relative error:
873      * arithmetic   domain     # trials      peak         rms
874      *    IEEE      0, 30       30000       7.8e-16     1.2e-16
875      * See k1().
876      *
877      */
878     const double A[] = {-7.02386347938628759343E-18, -2.42744985051936593393E-15,
879                         -6.66690169419932900609E-13, -1.41148839263352776110E-10,
880                         -2.21338763073472585583E-8, -2.43340614156596823496E-6,
881                         -1.73028895751305206302E-4, -6.97572385963986435018E-3,
882                         -1.22611180822657148235E-1, -3.53155960776544875667E-1,
883                         1.52530022733894777053E0};
884     const double B[] = {-5.75674448366501715755E-18, 1.79405087314755922667E-17,
885                         -5.68946255844285935196E-17, 1.83809354436663880070E-16,
886                         -6.05704724837331885336E-16, 2.03870316562433424052E-15,
887                         -7.01983709041831346144E-15, 2.47715442448130437068E-14,
888                         -8.97670518232499435011E-14, 3.34841966607842919884E-13,
889                         -1.28917396095102890680E-12, 5.13963967348173025100E-12,
890                         -2.12996783842756842877E-11, 9.21831518760500529508E-11,
891                         -4.19035475934189648750E-10, 2.01504975519703286596E-9,
892                         -1.03457624656780970260E-8, 5.74108412545004946722E-8,
893                         -3.50196060308781257119E-7, 2.40648494783721712015E-6,
894                         -1.93619797416608296024E-5, 1.95215518471351631108E-4,
895                         -2.85781685962277938680E-3, 1.03923736576817238437E-1,
896                         2.72062619048444266945E0};
897     const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
898     const T two = pset1<T>(2.0);
899     T x_le_two = pdiv(internal::pchebevl<T, 11>::run(
900         pmadd(x, x, pset1<T>(-2.0)), A), x);
901     x_le_two = pmadd(
902         generic_i1<T, double>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
903     x_le_two = pmul(x_le_two, pexp(x));
904     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
905     T x_gt_two = pmul(
906         internal::pchebevl<T, 25>::run(
907             psub(pdiv(pset1<T>(8.0), x), two), B),
908         prsqrt(x));
909     return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
910   }
911 };
912 
913 template <typename T>
914 struct bessel_k1e_impl {
915   EIGEN_DEVICE_FUNC
916   static EIGEN_STRONG_INLINE T run(const T x) {
917     return generic_k1e<T>::run(x);
918   }
919 };
920 
921 template <typename T>
922 struct bessel_k1_retval {
923   typedef T type;
924 };
925 
926 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
927 struct generic_k1 {
928   EIGEN_DEVICE_FUNC
929   static EIGEN_STRONG_INLINE T run(const T&) {
930     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
931                         THIS_TYPE_IS_NOT_SUPPORTED);
932     return ScalarType(0);
933   }
934 };
935 
936 template <typename T>
937 struct generic_k1<T, float> {
938   EIGEN_DEVICE_FUNC
939   static EIGEN_STRONG_INLINE T run(const T& x) {
940     /* k1f.c
941      *	Modified Bessel function, third kind, order one
942      *
943      *
944      *
945      * SYNOPSIS:
946      *
947      * float x, y, k1f();
948      *
949      * y = k1f( x );
950      *
951      *
952      *
953      * DESCRIPTION:
954      *
955      * Computes the modified Bessel function of the third kind
956      * of order one of the argument.
957      *
958      * The range is partitioned into the two intervals [0,2] and
959      * (2, infinity).  Chebyshev polynomial expansions are employed
960      * in each interval.
961      *
962      *
963      *
964      * ACCURACY:
965      *
966      *                      Relative error:
967      * arithmetic   domain     # trials      peak         rms
968      *    IEEE      0, 30       30000       4.6e-7      7.6e-8
969      *
970      * ERROR MESSAGES:
971      *
972      *   message         condition      value returned
973      * k1 domain          x <= 0          MAXNUM
974      *
975      */
976 
977     const float A[] = {-2.21338763073472585583E-8f, -2.43340614156596823496E-6f,
978                         -1.73028895751305206302E-4f, -6.97572385963986435018E-3f,
979                         -1.22611180822657148235E-1f, -3.53155960776544875667E-1f,
980                         1.52530022733894777053E0f};
981     const float B[] = {2.01504975519703286596E-9f, -1.03457624656780970260E-8f,
982                        5.74108412545004946722E-8f, -3.50196060308781257119E-7f,
983                        2.40648494783721712015E-6f, -1.93619797416608296024E-5f,
984                        1.95215518471351631108E-4f, -2.85781685962277938680E-3f,
985                        1.03923736576817238437E-1f, 2.72062619048444266945E0f};
986     const T MAXNUM = pset1<T>(NumTraits<float>::infinity());
987     const T two = pset1<T>(2.0);
988     T x_le_two = pdiv(internal::pchebevl<T, 7>::run(
989         pmadd(x, x, pset1<T>(-2.0)), A), x);
990     x_le_two = pmadd(
991         generic_i1<T, float>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
992     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
993     T x_gt_two = pmul(
994         pexp(pnegate(x)),
995         pmul(
996             internal::pchebevl<T, 10>::run(
997                 psub(pdiv(pset1<T>(8.0), x), two), B),
998             prsqrt(x)));
999     return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
1000   }
1001 };
1002 
1003 template <typename T>
1004 struct generic_k1<T, double> {
1005   EIGEN_DEVICE_FUNC
1006   static EIGEN_STRONG_INLINE T run(const T& x) {
1007     /*  k1.c
1008      *	Modified Bessel function, third kind, order one
1009      *
1010      *
1011      *
1012      * SYNOPSIS:
1013      *
1014      * float x, y, k1f();
1015      *
1016      * y = k1f( x );
1017      *
1018      *
1019      *
1020      * DESCRIPTION:
1021      *
1022      * Computes the modified Bessel function of the third kind
1023      * of order one of the argument.
1024      *
1025      * The range is partitioned into the two intervals [0,2] and
1026      * (2, infinity).  Chebyshev polynomial expansions are employed
1027      * in each interval.
1028      *
1029      *
1030      *
1031      * ACCURACY:
1032      *
1033      *                      Relative error:
1034      * arithmetic   domain     # trials      peak         rms
1035      *    IEEE      0, 30       30000       4.6e-7      7.6e-8
1036      *
1037      * ERROR MESSAGES:
1038      *
1039      *   message         condition      value returned
1040      * k1 domain          x <= 0          MAXNUM
1041      *
1042      */
1043     const double A[] = {-7.02386347938628759343E-18, -2.42744985051936593393E-15,
1044                         -6.66690169419932900609E-13, -1.41148839263352776110E-10,
1045                         -2.21338763073472585583E-8, -2.43340614156596823496E-6,
1046                         -1.73028895751305206302E-4, -6.97572385963986435018E-3,
1047                         -1.22611180822657148235E-1, -3.53155960776544875667E-1,
1048                         1.52530022733894777053E0};
1049     const double B[] = {-5.75674448366501715755E-18, 1.79405087314755922667E-17,
1050                         -5.68946255844285935196E-17, 1.83809354436663880070E-16,
1051                         -6.05704724837331885336E-16, 2.03870316562433424052E-15,
1052                         -7.01983709041831346144E-15, 2.47715442448130437068E-14,
1053                         -8.97670518232499435011E-14, 3.34841966607842919884E-13,
1054                         -1.28917396095102890680E-12, 5.13963967348173025100E-12,
1055                         -2.12996783842756842877E-11, 9.21831518760500529508E-11,
1056                         -4.19035475934189648750E-10, 2.01504975519703286596E-9,
1057                         -1.03457624656780970260E-8, 5.74108412545004946722E-8,
1058                         -3.50196060308781257119E-7, 2.40648494783721712015E-6,
1059                         -1.93619797416608296024E-5, 1.95215518471351631108E-4,
1060                         -2.85781685962277938680E-3, 1.03923736576817238437E-1,
1061                         2.72062619048444266945E0};
1062     const T MAXNUM = pset1<T>(NumTraits<double>::infinity());
1063     const T two = pset1<T>(2.0);
1064     T x_le_two = pdiv(internal::pchebevl<T, 11>::run(
1065         pmadd(x, x, pset1<T>(-2.0)), A), x);
1066     x_le_two = pmadd(
1067         generic_i1<T, double>::run(x), plog(pmul(pset1<T>(0.5), x)), x_le_two);
1068     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), MAXNUM, x_le_two);
1069     T x_gt_two = pmul(
1070         pexp(-x),
1071         pmul(
1072             internal::pchebevl<T, 25>::run(
1073                 psub(pdiv(pset1<T>(8.0), x), two), B),
1074             prsqrt(x)));
1075     return pselect(pcmp_le(x, two), x_le_two, x_gt_two);
1076   }
1077 };
1078 
1079 template <typename T>
1080 struct bessel_k1_impl {
1081   EIGEN_DEVICE_FUNC
1082   static EIGEN_STRONG_INLINE T run(const T x) {
1083     return generic_k1<T>::run(x);
1084   }
1085 };
1086 
1087 template <typename T>
1088 struct bessel_j0_retval {
1089   typedef T type;
1090 };
1091 
1092 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
1093 struct generic_j0 {
1094   EIGEN_DEVICE_FUNC
1095   static EIGEN_STRONG_INLINE T run(const T&) {
1096     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
1097                         THIS_TYPE_IS_NOT_SUPPORTED);
1098     return ScalarType(0);
1099   }
1100 };
1101 
1102 template <typename T>
1103 struct generic_j0<T, float> {
1104   EIGEN_DEVICE_FUNC
1105   static EIGEN_STRONG_INLINE T run(const T& x) {
1106     /* j0f.c
1107      *	Bessel function of order zero
1108      *
1109      *
1110      *
1111      * SYNOPSIS:
1112      *
1113      * float x, y, j0f();
1114      *
1115      * y = j0f( x );
1116      *
1117      *
1118      *
1119      * DESCRIPTION:
1120      *
1121      * Returns Bessel function of order zero of the argument.
1122      *
1123      * The domain is divided into the intervals [0, 2] and
1124      * (2, infinity). In the first interval the following polynomial
1125      * approximation is used:
1126      *
1127      *
1128      *        2         2         2
1129      * (w - r  ) (w - r  ) (w - r  ) P(w)
1130      *       1         2         3
1131      *
1132      *            2
1133      * where w = x  and the three r's are zeros of the function.
1134      *
1135      * In the second interval, the modulus and phase are approximated
1136      * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
1137      * and Phase(x) = x + 1/x R(1/x^2) - pi/4.  The function is
1138      *
1139      *   j0(x) = Modulus(x) cos( Phase(x) ).
1140      *
1141      *
1142      *
1143      * ACCURACY:
1144      *
1145      *                      Absolute error:
1146      * arithmetic   domain     # trials      peak         rms
1147      *    IEEE      0, 2        100000      1.3e-7      3.6e-8
1148      *    IEEE      2, 32       100000      1.9e-7      5.4e-8
1149      *
1150      */
1151 
1152     const float JP[] = {-6.068350350393235E-008f, 6.388945720783375E-006f,
1153                         -3.969646342510940E-004f, 1.332913422519003E-002f,
1154                         -1.729150680240724E-001f};
1155     const float MO[] = {-6.838999669318810E-002f, 1.864949361379502E-001f,
1156                         -2.145007480346739E-001f, 1.197549369473540E-001f,
1157                         -3.560281861530129E-003f, -4.969382655296620E-002f,
1158                         -3.355424622293709E-006f, 7.978845717621440E-001f};
1159     const float PH[] = {3.242077816988247E+001f, -3.630592630518434E+001f,
1160                         1.756221482109099E+001f, -4.974978466280903E+000f,
1161                         1.001973420681837E+000f, -1.939906941791308E-001f,
1162                         6.490598792654666E-002f, -1.249992184872738E-001f};
1163     const T DR1 =  pset1<T>(5.78318596294678452118f);
1164     const T NEG_PIO4F = pset1<T>(-0.7853981633974483096f); /* -pi / 4 */
1165     T y = pabs(x);
1166     T z = pmul(y, y);
1167     T y_le_two = pselect(
1168         pcmp_lt(y, pset1<T>(1.0e-3f)),
1169         pmadd(z, pset1<T>(-0.25f), pset1<T>(1.0f)),
1170         pmul(psub(z, DR1), internal::ppolevl<T, 4>::run(z, JP)));
1171     T q = pdiv(pset1<T>(1.0f), y);
1172     T w = prsqrt(y);
1173     T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO));
1174     w = pmul(q, q);
1175     T yn = pmadd(q, internal::ppolevl<T, 7>::run(w, PH), NEG_PIO4F);
1176     T y_gt_two = pmul(p, pcos(padd(yn, y)));
1177     return pselect(pcmp_le(y, pset1<T>(2.0)), y_le_two, y_gt_two);
1178   }
1179 };
1180 
1181 template <typename T>
1182 struct generic_j0<T, double> {
1183   EIGEN_DEVICE_FUNC
1184   static EIGEN_STRONG_INLINE T run(const T& x) {
1185     /*  j0.c
1186      *	Bessel function of order zero
1187      *
1188      *
1189      *
1190      * SYNOPSIS:
1191      *
1192      * double x, y, j0();
1193      *
1194      * y = j0( x );
1195      *
1196      *
1197      *
1198      * DESCRIPTION:
1199      *
1200      * Returns Bessel function of order zero of the argument.
1201      *
1202      * The domain is divided into the intervals [0, 5] and
1203      * (5, infinity). In the first interval the following rational
1204      * approximation is used:
1205      *
1206      *
1207      *        2         2
1208      * (w - r  ) (w - r  ) P (w) / Q (w)
1209      *       1         2    3       8
1210      *
1211      *            2
1212      * where w = x  and the two r's are zeros of the function.
1213      *
1214      * In the second interval, the Hankel asymptotic expansion
1215      * is employed with two rational functions of degree 6/6
1216      * and 7/7.
1217      *
1218      *
1219      *
1220      * ACCURACY:
1221      *
1222      *                      Absolute error:
1223      * arithmetic   domain     # trials      peak         rms
1224      *    DEC       0, 30       10000       4.4e-17     6.3e-18
1225      *    IEEE      0, 30       60000       4.2e-16     1.1e-16
1226      *
1227      */
1228     const double PP[] = {7.96936729297347051624E-4, 8.28352392107440799803E-2,
1229                         1.23953371646414299388E0, 5.44725003058768775090E0,
1230                         8.74716500199817011941E0, 5.30324038235394892183E0,
1231                         9.99999999999999997821E-1};
1232     const double PQ[] = {9.24408810558863637013E-4, 8.56288474354474431428E-2,
1233                          1.25352743901058953537E0, 5.47097740330417105182E0,
1234                          8.76190883237069594232E0, 5.30605288235394617618E0,
1235                          1.00000000000000000218E0};
1236     const double QP[] = {-1.13663838898469149931E-2, -1.28252718670509318512E0,
1237                          -1.95539544257735972385E1, -9.32060152123768231369E1,
1238                          -1.77681167980488050595E2, -1.47077505154951170175E2,
1239                          -5.14105326766599330220E1, -6.05014350600728481186E0};
1240     const double QQ[] = {1.00000000000000000000E0, 6.43178256118178023184E1,
1241                          8.56430025976980587198E2, 3.88240183605401609683E3,
1242                          7.24046774195652478189E3, 5.93072701187316984827E3,
1243                          2.06209331660327847417E3, 2.42005740240291393179E2};
1244     const double RP[] = {-4.79443220978201773821E9, 1.95617491946556577543E12,
1245                          -2.49248344360967716204E14, 9.70862251047306323952E15};
1246     const double RQ[] = {1.00000000000000000000E0, 4.99563147152651017219E2,
1247                          1.73785401676374683123E5, 4.84409658339962045305E7,
1248                          1.11855537045356834862E10, 2.11277520115489217587E12,
1249                          3.10518229857422583814E14, 3.18121955943204943306E16,
1250                          1.71086294081043136091E18};
1251     const T DR1 = pset1<T>(5.78318596294678452118E0);
1252     const T DR2 = pset1<T>(3.04712623436620863991E1);
1253     const T SQ2OPI = pset1<T>(7.9788456080286535587989E-1); /* sqrt(2 / pi) */
1254     const T NEG_PIO4 = pset1<T>(-0.7853981633974483096); /* pi / 4 */
1255 
1256     T y = pabs(x);
1257     T z = pmul(y, y);
1258     T y_le_five = pselect(
1259         pcmp_lt(y, pset1<T>(1.0e-5)),
1260         pmadd(z, pset1<T>(-0.25), pset1<T>(1.0)),
1261         pmul(pmul(psub(z, DR1), psub(z, DR2)),
1262              pdiv(internal::ppolevl<T, 3>::run(z, RP),
1263                   internal::ppolevl<T, 8>::run(z, RQ))));
1264     T s = pdiv(pset1<T>(25.0), z);
1265     T p = pdiv(
1266         internal::ppolevl<T, 6>::run(s, PP),
1267         internal::ppolevl<T, 6>::run(s, PQ));
1268     T q = pdiv(
1269         internal::ppolevl<T, 7>::run(s, QP),
1270         internal::ppolevl<T, 7>::run(s, QQ));
1271     T yn = padd(y, NEG_PIO4);
1272     T w = pdiv(pset1<T>(-5.0), y);
1273     p = pmadd(p, pcos(yn), pmul(w, pmul(q, psin(yn))));
1274     T y_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(y)));
1275     return pselect(pcmp_le(y, pset1<T>(5.0)), y_le_five, y_gt_five);
1276   }
1277 };
1278 
1279 template <typename T>
1280 struct bessel_j0_impl {
1281   EIGEN_DEVICE_FUNC
1282   static EIGEN_STRONG_INLINE T run(const T x) {
1283     return generic_j0<T>::run(x);
1284   }
1285 };
1286 
1287 template <typename T>
1288 struct bessel_y0_retval {
1289   typedef T type;
1290 };
1291 
1292 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
1293 struct generic_y0 {
1294   EIGEN_DEVICE_FUNC
1295   static EIGEN_STRONG_INLINE T run(const T&) {
1296     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
1297                         THIS_TYPE_IS_NOT_SUPPORTED);
1298     return ScalarType(0);
1299   }
1300 };
1301 
1302 template <typename T>
1303 struct generic_y0<T, float> {
1304   EIGEN_DEVICE_FUNC
1305   static EIGEN_STRONG_INLINE T run(const T& x) {
1306     /* j0f.c
1307      * 	Bessel function of the second kind, order zero
1308      *
1309      *
1310      *
1311      * SYNOPSIS:
1312      *
1313      * float x, y, y0f();
1314      *
1315      * y = y0f( x );
1316      *
1317      *
1318      *
1319      * DESCRIPTION:
1320      *
1321      * Returns Bessel function of the second kind, of order
1322      * zero, of the argument.
1323      *
1324      * The domain is divided into the intervals [0, 2] and
1325      * (2, infinity). In the first interval a rational approximation
1326      * R(x) is employed to compute
1327      *
1328      *                  2         2         2
1329      * y0(x)  =  (w - r  ) (w - r  ) (w - r  ) R(x)  +  2/pi ln(x) j0(x).
1330      *                 1         2         3
1331      *
1332      * Thus a call to j0() is required.  The three zeros are removed
1333      * from R(x) to improve its numerical stability.
1334      *
1335      * In the second interval, the modulus and phase are approximated
1336      * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
1337      * and Phase(x) = x + 1/x S(1/x^2) - pi/4.  Then the function is
1338      *
1339      *   y0(x) = Modulus(x) sin( Phase(x) ).
1340      *
1341      *
1342      *
1343      *
1344      * ACCURACY:
1345      *
1346      *  Absolute error, when y0(x) < 1; else relative error:
1347      *
1348      * arithmetic   domain     # trials      peak         rms
1349      *    IEEE      0,  2       100000      2.4e-7      3.4e-8
1350      *    IEEE      2, 32       100000      1.8e-7      5.3e-8
1351      *
1352      */
1353 
1354     const float YP[] = {9.454583683980369E-008f, -9.413212653797057E-006f,
1355                         5.344486707214273E-004f, -1.584289289821316E-002f,
1356                         1.707584643733568E-001f};
1357     const float MO[] = {-6.838999669318810E-002f, 1.864949361379502E-001f,
1358                         -2.145007480346739E-001f, 1.197549369473540E-001f,
1359                         -3.560281861530129E-003f, -4.969382655296620E-002f,
1360                         -3.355424622293709E-006f, 7.978845717621440E-001f};
1361     const float PH[] = {3.242077816988247E+001f, -3.630592630518434E+001f,
1362                         1.756221482109099E+001f, -4.974978466280903E+000f,
1363                         1.001973420681837E+000f, -1.939906941791308E-001f,
1364                         6.490598792654666E-002f, -1.249992184872738E-001f};
1365     const T YZ1 = pset1<T>(0.43221455686510834878f);
1366     const T TWOOPI =  pset1<T>(0.636619772367581343075535f); /* 2 / pi */
1367     const T NEG_PIO4F = pset1<T>(-0.7853981633974483096f); /* -pi / 4 */
1368     const T NEG_MAXNUM = pset1<T>(-NumTraits<float>::infinity());
1369     T z = pmul(x, x);
1370     T x_le_two = pmul(TWOOPI, pmul(plog(x), generic_j0<T, float>::run(x)));
1371     x_le_two = pmadd(
1372         psub(z, YZ1), internal::ppolevl<T, 4>::run(z, YP), x_le_two);
1373     x_le_two = pselect(pcmp_le(x, pset1<T>(0.0)), NEG_MAXNUM, x_le_two);
1374     T q = pdiv(pset1<T>(1.0), x);
1375     T w = prsqrt(x);
1376     T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO));
1377     T u = pmul(q, q);
1378     T xn = pmadd(q, internal::ppolevl<T, 7>::run(u, PH), NEG_PIO4F);
1379     T x_gt_two = pmul(p, psin(padd(xn, x)));
1380     return pselect(pcmp_le(x, pset1<T>(2.0)), x_le_two, x_gt_two);
1381   }
1382 };
1383 
1384 template <typename T>
1385 struct generic_y0<T, double> {
1386   EIGEN_DEVICE_FUNC
1387   static EIGEN_STRONG_INLINE T run(const T& x) {
1388     /*  j0.c
1389      *	Bessel function of the second kind, order zero
1390      *
1391      *
1392      *
1393      * SYNOPSIS:
1394      *
1395      * double x, y, y0();
1396      *
1397      * y = y0( x );
1398      *
1399      *
1400      *
1401      * DESCRIPTION:
1402      *
1403      * Returns Bessel function of the second kind, of order
1404      * zero, of the argument.
1405      *
1406      * The domain is divided into the intervals [0, 5] and
1407      * (5, infinity). In the first interval a rational approximation
1408      * R(x) is employed to compute
1409      *   y0(x)  = R(x)  +   2 * log(x) * j0(x) / PI.
1410      * Thus a call to j0() is required.
1411      *
1412      * In the second interval, the Hankel asymptotic expansion
1413      * is employed with two rational functions of degree 6/6
1414      * and 7/7.
1415      *
1416      *
1417      *
1418      * ACCURACY:
1419      *
1420      *  Absolute error, when y0(x) < 1; else relative error:
1421      *
1422      * arithmetic   domain     # trials      peak         rms
1423      *    DEC       0, 30        9400       7.0e-17     7.9e-18
1424      *    IEEE      0, 30       30000       1.3e-15     1.6e-16
1425      *
1426      */
1427     const double PP[] = {7.96936729297347051624E-4, 8.28352392107440799803E-2,
1428                         1.23953371646414299388E0, 5.44725003058768775090E0,
1429                         8.74716500199817011941E0, 5.30324038235394892183E0,
1430                         9.99999999999999997821E-1};
1431     const double PQ[] = {9.24408810558863637013E-4, 8.56288474354474431428E-2,
1432                          1.25352743901058953537E0, 5.47097740330417105182E0,
1433                          8.76190883237069594232E0, 5.30605288235394617618E0,
1434                          1.00000000000000000218E0};
1435     const double QP[] = {-1.13663838898469149931E-2, -1.28252718670509318512E0,
1436                          -1.95539544257735972385E1, -9.32060152123768231369E1,
1437                          -1.77681167980488050595E2, -1.47077505154951170175E2,
1438                          -5.14105326766599330220E1, -6.05014350600728481186E0};
1439     const double QQ[] = {1.00000000000000000000E0, 6.43178256118178023184E1,
1440                          8.56430025976980587198E2, 3.88240183605401609683E3,
1441                          7.24046774195652478189E3, 5.93072701187316984827E3,
1442                          2.06209331660327847417E3, 2.42005740240291393179E2};
1443     const double YP[] = {1.55924367855235737965E4, -1.46639295903971606143E7,
1444                          5.43526477051876500413E9, -9.82136065717911466409E11,
1445                          8.75906394395366999549E13, -3.46628303384729719441E15,
1446                          4.42733268572569800351E16, -1.84950800436986690637E16};
1447     const double YQ[] = {1.00000000000000000000E0,  1.04128353664259848412E3,
1448                          6.26107330137134956842E5, 2.68919633393814121987E8,
1449                          8.64002487103935000337E10, 2.02979612750105546709E13,
1450                          3.17157752842975028269E15, 2.50596256172653059228E17};
1451     const T SQ2OPI = pset1<T>(7.9788456080286535587989E-1); /* sqrt(2 / pi) */
1452     const T TWOOPI =  pset1<T>(0.636619772367581343075535); /* 2 / pi */
1453     const T NEG_PIO4 = pset1<T>(-0.7853981633974483096); /* -pi / 4 */
1454     const T NEG_MAXNUM = pset1<T>(-NumTraits<double>::infinity());
1455 
1456     T z = pmul(x, x);
1457     T x_le_five = pdiv(internal::ppolevl<T, 7>::run(z, YP),
1458                        internal::ppolevl<T, 7>::run(z, YQ));
1459     x_le_five = pmadd(
1460         pmul(TWOOPI, plog(x)), generic_j0<T, double>::run(x), x_le_five);
1461     x_le_five = pselect(pcmp_le(x, pset1<T>(0.0)), NEG_MAXNUM, x_le_five);
1462     T s = pdiv(pset1<T>(25.0), z);
1463     T p = pdiv(
1464         internal::ppolevl<T, 6>::run(s, PP),
1465         internal::ppolevl<T, 6>::run(s, PQ));
1466     T q = pdiv(
1467         internal::ppolevl<T, 7>::run(s, QP),
1468         internal::ppolevl<T, 7>::run(s, QQ));
1469     T xn = padd(x, NEG_PIO4);
1470     T w = pdiv(pset1<T>(5.0), x);
1471     p = pmadd(p, psin(xn), pmul(w, pmul(q, pcos(xn))));
1472     T x_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(x)));
1473     return pselect(pcmp_le(x, pset1<T>(5.0)), x_le_five, x_gt_five);
1474   }
1475 };
1476 
1477 template <typename T>
1478 struct bessel_y0_impl {
1479   EIGEN_DEVICE_FUNC
1480   static EIGEN_STRONG_INLINE T run(const T x) {
1481     return generic_y0<T>::run(x);
1482   }
1483 };
1484 
1485 template <typename T>
1486 struct bessel_j1_retval {
1487   typedef T type;
1488 };
1489 
1490 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
1491 struct generic_j1 {
1492   EIGEN_DEVICE_FUNC
1493   static EIGEN_STRONG_INLINE T run(const T&) {
1494     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
1495                         THIS_TYPE_IS_NOT_SUPPORTED);
1496     return ScalarType(0);
1497   }
1498 };
1499 
1500 template <typename T>
1501 struct generic_j1<T, float> {
1502   EIGEN_DEVICE_FUNC
1503   static EIGEN_STRONG_INLINE T run(const T& x) {
1504     /* j1f.c
1505      *	Bessel function of order one
1506      *
1507      *
1508      *
1509      * SYNOPSIS:
1510      *
1511      * float x, y, j1f();
1512      *
1513      * y = j1f( x );
1514      *
1515      *
1516      *
1517      * DESCRIPTION:
1518      *
1519      * Returns Bessel function of order one of the argument.
1520      *
1521      * The domain is divided into the intervals [0, 2] and
1522      * (2, infinity). In the first interval a polynomial approximation
1523      *        2
1524      * (w - r  ) x P(w)
1525      *       1
1526      *                     2
1527      * is used, where w = x  and r is the first zero of the function.
1528      *
1529      * In the second interval, the modulus and phase are approximated
1530      * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
1531      * and Phase(x) = x + 1/x R(1/x^2) - 3pi/4.  The function is
1532      *
1533      *   j0(x) = Modulus(x) cos( Phase(x) ).
1534      *
1535      *
1536      *
1537      * ACCURACY:
1538      *
1539      *                      Absolute error:
1540      * arithmetic   domain      # trials      peak       rms
1541      *    IEEE      0,  2       100000       1.2e-7     2.5e-8
1542      *    IEEE      2, 32       100000       2.0e-7     5.3e-8
1543      *
1544      *
1545      */
1546 
1547     const float JP[] = {-4.878788132172128E-009f, 6.009061827883699E-007f,
1548                         -4.541343896997497E-005f, 1.937383947804541E-003f,
1549                         -3.405537384615824E-002f};
1550     const float MO1[] = {6.913942741265801E-002f, -2.284801500053359E-001f,
1551                         3.138238455499697E-001f, -2.102302420403875E-001f,
1552                         5.435364690523026E-003f, 1.493389585089498E-001f,
1553                         4.976029650847191E-006f, 7.978845453073848E-001f};
1554     const float PH1[] = {-4.497014141919556E+001f, 5.073465654089319E+001f,
1555                         -2.485774108720340E+001f, 7.222973196770240E+000f,
1556                         -1.544842782180211E+000f, 3.503787691653334E-001f,
1557                         -1.637986776941202E-001f, 3.749989509080821E-001f};
1558     const T Z1 = pset1<T>(1.46819706421238932572E1f);
1559     const T NEG_THPIO4F = pset1<T>(-2.35619449019234492885f);    /* -3*pi/4 */
1560 
1561     T y = pabs(x);
1562     T z = pmul(y, y);
1563     T y_le_two = pmul(
1564         psub(z, Z1),
1565         pmul(x, internal::ppolevl<T, 4>::run(z, JP)));
1566     T q = pdiv(pset1<T>(1.0f), y);
1567     T w = prsqrt(y);
1568     T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO1));
1569     w = pmul(q, q);
1570     T yn = pmadd(q, internal::ppolevl<T, 7>::run(w, PH1), NEG_THPIO4F);
1571     T y_gt_two = pmul(p, pcos(padd(yn, y)));
1572     // j1 is an odd function. This implementation differs from cephes to
1573     // take this fact in to account. Cephes returns -j1(x) for y > 2 range.
1574     y_gt_two = pselect(
1575         pcmp_lt(x, pset1<T>(0.0f)), pnegate(y_gt_two), y_gt_two);
1576     return pselect(pcmp_le(y, pset1<T>(2.0f)), y_le_two, y_gt_two);
1577   }
1578 };
1579 
1580 template <typename T>
1581 struct generic_j1<T, double> {
1582   EIGEN_DEVICE_FUNC
1583   static EIGEN_STRONG_INLINE T run(const T& x) {
1584     /*  j1.c
1585      *	Bessel function of order one
1586      *
1587      *
1588      *
1589      * SYNOPSIS:
1590      *
1591      * double x, y, j1();
1592      *
1593      * y = j1( x );
1594      *
1595      *
1596      *
1597      * DESCRIPTION:
1598      *
1599      * Returns Bessel function of order one of the argument.
1600      *
1601      * The domain is divided into the intervals [0, 8] and
1602      * (8, infinity). In the first interval a 24 term Chebyshev
1603      * expansion is used. In the second, the asymptotic
1604      * trigonometric representation is employed using two
1605      * rational functions of degree 5/5.
1606      *
1607      *
1608      *
1609      * ACCURACY:
1610      *
1611      *                      Absolute error:
1612      * arithmetic   domain      # trials      peak         rms
1613      *    DEC       0, 30       10000       4.0e-17     1.1e-17
1614      *    IEEE      0, 30       30000       2.6e-16     1.1e-16
1615      *
1616      */
1617     const double PP[] = {7.62125616208173112003E-4, 7.31397056940917570436E-2,
1618                          1.12719608129684925192E0, 5.11207951146807644818E0,
1619                          8.42404590141772420927E0, 5.21451598682361504063E0,
1620                          1.00000000000000000254E0};
1621     const double PQ[] = {5.71323128072548699714E-4, 6.88455908754495404082E-2,
1622                          1.10514232634061696926E0, 5.07386386128601488557E0,
1623                          8.39985554327604159757E0, 5.20982848682361821619E0,
1624                          9.99999999999999997461E-1};
1625     const double QP[] = {5.10862594750176621635E-2, 4.98213872951233449420E0,
1626                          7.58238284132545283818E1, 3.66779609360150777800E2,
1627                          7.10856304998926107277E2, 5.97489612400613639965E2,
1628                          2.11688757100572135698E2, 2.52070205858023719784E1};
1629     const double QQ[] = {1.00000000000000000000E0, 7.42373277035675149943E1,
1630                          1.05644886038262816351E3, 4.98641058337653607651E3,
1631                          9.56231892404756170795E3, 7.99704160447350683650E3,
1632                          2.82619278517639096600E3, 3.36093607810698293419E2};
1633     const double RP[] = {-8.99971225705559398224E8, 4.52228297998194034323E11,
1634                          -7.27494245221818276015E13, 3.68295732863852883286E15};
1635     const double RQ[] = {1.00000000000000000000E0, 6.20836478118054335476E2,
1636                          2.56987256757748830383E5, 8.35146791431949253037E7,
1637                          2.21511595479792499675E10, 4.74914122079991414898E12,
1638                          7.84369607876235854894E14, 8.95222336184627338078E16,
1639                          5.32278620332680085395E18};
1640     const T Z1 = pset1<T>(1.46819706421238932572E1);
1641     const T Z2 = pset1<T>(4.92184563216946036703E1);
1642     const T NEG_THPIO4 = pset1<T>(-2.35619449019234492885);    /* -3*pi/4 */
1643     const T SQ2OPI = pset1<T>(7.9788456080286535587989E-1); /* sqrt(2 / pi) */
1644     T y = pabs(x);
1645     T z = pmul(y, y);
1646     T y_le_five = pdiv(internal::ppolevl<T, 3>::run(z, RP),
1647                        internal::ppolevl<T, 8>::run(z, RQ));
1648     y_le_five = pmul(pmul(pmul(y_le_five, x), psub(z, Z1)), psub(z, Z2));
1649     T s = pdiv(pset1<T>(25.0), z);
1650     T p = pdiv(
1651         internal::ppolevl<T, 6>::run(s, PP),
1652         internal::ppolevl<T, 6>::run(s, PQ));
1653     T q = pdiv(
1654         internal::ppolevl<T, 7>::run(s, QP),
1655         internal::ppolevl<T, 7>::run(s, QQ));
1656     T yn = padd(y, NEG_THPIO4);
1657     T w = pdiv(pset1<T>(-5.0), y);
1658     p = pmadd(p, pcos(yn), pmul(w, pmul(q, psin(yn))));
1659     T y_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(y)));
1660     // j1 is an odd function. This implementation differs from cephes to
1661     // take this fact in to account. Cephes returns -j1(x) for y > 5 range.
1662     y_gt_five = pselect(
1663         pcmp_lt(x, pset1<T>(0.0)), pnegate(y_gt_five), y_gt_five);
1664     return pselect(pcmp_le(y, pset1<T>(5.0)), y_le_five, y_gt_five);
1665   }
1666 };
1667 
1668 template <typename T>
1669 struct bessel_j1_impl {
1670   EIGEN_DEVICE_FUNC
1671   static EIGEN_STRONG_INLINE T run(const T x) {
1672     return generic_j1<T>::run(x);
1673   }
1674 };
1675 
1676 template <typename T>
1677 struct bessel_y1_retval {
1678   typedef T type;
1679 };
1680 
1681 template <typename T, typename ScalarType = typename unpacket_traits<T>::type>
1682 struct generic_y1 {
1683   EIGEN_DEVICE_FUNC
1684   static EIGEN_STRONG_INLINE T run(const T&) {
1685     EIGEN_STATIC_ASSERT((internal::is_same<T, T>::value == false),
1686                         THIS_TYPE_IS_NOT_SUPPORTED);
1687     return ScalarType(0);
1688   }
1689 };
1690 
1691 template <typename T>
1692 struct generic_y1<T, float> {
1693   EIGEN_DEVICE_FUNC
1694   static EIGEN_STRONG_INLINE T run(const T& x) {
1695     /* j1f.c
1696      *	Bessel function of second kind of order one
1697      *
1698      *
1699      *
1700      * SYNOPSIS:
1701      *
1702      * double x, y, y1();
1703      *
1704      * y = y1( x );
1705      *
1706      *
1707      *
1708      * DESCRIPTION:
1709      *
1710      * Returns Bessel function of the second kind of order one
1711      * of the argument.
1712      *
1713      * The domain is divided into the intervals [0, 2] and
1714      * (2, infinity). In the first interval a rational approximation
1715      * R(x) is employed to compute
1716      *
1717      *                  2
1718      * y0(x)  =  (w - r  ) x R(x^2)  +  2/pi (ln(x) j1(x) - 1/x) .
1719      *                 1
1720      *
1721      * Thus a call to j1() is required.
1722      *
1723      * In the second interval, the modulus and phase are approximated
1724      * by polynomials of the form Modulus(x) = sqrt(1/x) Q(1/x)
1725      * and Phase(x) = x + 1/x S(1/x^2) - 3pi/4.  Then the function is
1726      *
1727      *   y0(x) = Modulus(x) sin( Phase(x) ).
1728      *
1729      *
1730      *
1731      *
1732      * ACCURACY:
1733      *
1734      *                      Absolute error:
1735      * arithmetic   domain      # trials      peak         rms
1736      *    IEEE      0,  2       100000       2.2e-7     4.6e-8
1737      *    IEEE      2, 32       100000       1.9e-7     5.3e-8
1738      *
1739      * (error criterion relative when |y1| > 1).
1740      *
1741      */
1742 
1743     const float YP[] = {8.061978323326852E-009f, -9.496460629917016E-007f,
1744                         6.719543806674249E-005f, -2.641785726447862E-003f,
1745                         4.202369946500099E-002f};
1746     const float MO1[] = {6.913942741265801E-002f, -2.284801500053359E-001f,
1747                         3.138238455499697E-001f, -2.102302420403875E-001f,
1748                         5.435364690523026E-003f, 1.493389585089498E-001f,
1749                         4.976029650847191E-006f, 7.978845453073848E-001f};
1750     const float PH1[] = {-4.497014141919556E+001f, 5.073465654089319E+001f,
1751                         -2.485774108720340E+001f, 7.222973196770240E+000f,
1752                         -1.544842782180211E+000f, 3.503787691653334E-001f,
1753                         -1.637986776941202E-001f, 3.749989509080821E-001f};
1754     const T YO1 = pset1<T>(4.66539330185668857532f);
1755     const T NEG_THPIO4F = pset1<T>(-2.35619449019234492885f);    /* -3*pi/4 */
1756     const T TWOOPI = pset1<T>(0.636619772367581343075535f); /* 2/pi */
1757     const T NEG_MAXNUM = pset1<T>(-NumTraits<float>::infinity());
1758 
1759     T z = pmul(x, x);
1760     T x_le_two = pmul(psub(z, YO1), internal::ppolevl<T, 4>::run(z, YP));
1761     x_le_two = pmadd(
1762        x_le_two, x,
1763        pmul(TWOOPI, pmadd(
1764            generic_j1<T, float>::run(x), plog(x),
1765            pdiv(pset1<T>(-1.0f), x))));
1766     x_le_two = pselect(pcmp_lt(x, pset1<T>(0.0f)), NEG_MAXNUM, x_le_two);
1767 
1768     T q = pdiv(pset1<T>(1.0), x);
1769     T w = prsqrt(x);
1770     T p = pmul(w, internal::ppolevl<T, 7>::run(q, MO1));
1771     w = pmul(q, q);
1772     T xn = pmadd(q, internal::ppolevl<T, 7>::run(w, PH1), NEG_THPIO4F);
1773     T x_gt_two = pmul(p, psin(padd(xn, x)));
1774     return pselect(pcmp_le(x, pset1<T>(2.0)), x_le_two, x_gt_two);
1775   }
1776 };
1777 
1778 template <typename T>
1779 struct generic_y1<T, double> {
1780   EIGEN_DEVICE_FUNC
1781   static EIGEN_STRONG_INLINE T run(const T& x) {
1782     /*  j1.c
1783      *	Bessel function of second kind of order one
1784      *
1785      *
1786      *
1787      * SYNOPSIS:
1788      *
1789      * double x, y, y1();
1790      *
1791      * y = y1( x );
1792      *
1793      *
1794      *
1795      * DESCRIPTION:
1796      *
1797      * Returns Bessel function of the second kind of order one
1798      * of the argument.
1799      *
1800      * The domain is divided into the intervals [0, 8] and
1801      * (8, infinity). In the first interval a 25 term Chebyshev
1802      * expansion is used, and a call to j1() is required.
1803      * In the second, the asymptotic trigonometric representation
1804      * is employed using two rational functions of degree 5/5.
1805      *
1806      *
1807      *
1808      * ACCURACY:
1809      *
1810      *                      Absolute error:
1811      * arithmetic   domain      # trials      peak         rms
1812      *    DEC       0, 30       10000       8.6e-17     1.3e-17
1813      *    IEEE      0, 30       30000       1.0e-15     1.3e-16
1814      *
1815      * (error criterion relative when |y1| > 1).
1816      *
1817      */
1818     const double PP[] = {7.62125616208173112003E-4, 7.31397056940917570436E-2,
1819                          1.12719608129684925192E0, 5.11207951146807644818E0,
1820                          8.42404590141772420927E0, 5.21451598682361504063E0,
1821                          1.00000000000000000254E0};
1822     const double PQ[] = {5.71323128072548699714E-4, 6.88455908754495404082E-2,
1823                          1.10514232634061696926E0, 5.07386386128601488557E0,
1824                          8.39985554327604159757E0, 5.20982848682361821619E0,
1825                          9.99999999999999997461E-1};
1826     const double QP[] = {5.10862594750176621635E-2, 4.98213872951233449420E0,
1827                          7.58238284132545283818E1, 3.66779609360150777800E2,
1828                          7.10856304998926107277E2, 5.97489612400613639965E2,
1829                          2.11688757100572135698E2, 2.52070205858023719784E1};
1830     const double QQ[] = {1.00000000000000000000E0, 7.42373277035675149943E1,
1831                          1.05644886038262816351E3, 4.98641058337653607651E3,
1832                          9.56231892404756170795E3, 7.99704160447350683650E3,
1833                          2.82619278517639096600E3, 3.36093607810698293419E2};
1834     const double YP[] = {1.26320474790178026440E9, -6.47355876379160291031E11,
1835                          1.14509511541823727583E14, -8.12770255501325109621E15,
1836                          2.02439475713594898196E17, -7.78877196265950026825E17};
1837     const double YQ[] = {1.00000000000000000000E0, 5.94301592346128195359E2,
1838                          2.35564092943068577943E5, 7.34811944459721705660E7,
1839                          1.87601316108706159478E10, 3.88231277496238566008E12,
1840                          6.20557727146953693363E14, 6.87141087355300489866E16,
1841                          3.97270608116560655612E18};
1842     const T SQ2OPI = pset1<T>(.79788456080286535588);
1843     const T NEG_THPIO4 = pset1<T>(-2.35619449019234492885);    /* -3*pi/4 */
1844     const T TWOOPI = pset1<T>(0.636619772367581343075535); /* 2/pi */
1845     const T NEG_MAXNUM = pset1<T>(-NumTraits<double>::infinity());
1846 
1847     T z = pmul(x, x);
1848     T x_le_five = pdiv(internal::ppolevl<T, 5>::run(z, YP),
1849                    internal::ppolevl<T, 8>::run(z, YQ));
1850     x_le_five = pmadd(
1851         x_le_five, x, pmul(
1852             TWOOPI, pmadd(generic_j1<T, double>::run(x), plog(x),
1853                           pdiv(pset1<T>(-1.0), x))));
1854 
1855     x_le_five = pselect(pcmp_le(x, pset1<T>(0.0)), NEG_MAXNUM, x_le_five);
1856     T s = pdiv(pset1<T>(25.0), z);
1857     T p = pdiv(
1858         internal::ppolevl<T, 6>::run(s, PP),
1859         internal::ppolevl<T, 6>::run(s, PQ));
1860     T q = pdiv(
1861         internal::ppolevl<T, 7>::run(s, QP),
1862         internal::ppolevl<T, 7>::run(s, QQ));
1863     T xn = padd(x, NEG_THPIO4);
1864     T w = pdiv(pset1<T>(5.0), x);
1865     p = pmadd(p, psin(xn), pmul(w, pmul(q, pcos(xn))));
1866     T x_gt_five = pmul(p, pmul(SQ2OPI, prsqrt(x)));
1867     return pselect(pcmp_le(x, pset1<T>(5.0)), x_le_five, x_gt_five);
1868   }
1869 };
1870 
1871 template <typename T>
1872 struct bessel_y1_impl {
1873   EIGEN_DEVICE_FUNC
1874   static EIGEN_STRONG_INLINE T run(const T x) {
1875     return generic_y1<T>::run(x);
1876   }
1877 };
1878 
1879 }  // end namespace internal
1880 
1881 namespace numext {
1882 
1883 template <typename Scalar>
1884 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i0, Scalar)
1885     bessel_i0(const Scalar& x) {
1886   return EIGEN_MATHFUNC_IMPL(bessel_i0, Scalar)::run(x);
1887 }
1888 
1889 template <typename Scalar>
1890 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i0e, Scalar)
1891     bessel_i0e(const Scalar& x) {
1892   return EIGEN_MATHFUNC_IMPL(bessel_i0e, Scalar)::run(x);
1893 }
1894 
1895 template <typename Scalar>
1896 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i1, Scalar)
1897     bessel_i1(const Scalar& x) {
1898   return EIGEN_MATHFUNC_IMPL(bessel_i1, Scalar)::run(x);
1899 }
1900 
1901 template <typename Scalar>
1902 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_i1e, Scalar)
1903     bessel_i1e(const Scalar& x) {
1904   return EIGEN_MATHFUNC_IMPL(bessel_i1e, Scalar)::run(x);
1905 }
1906 
1907 template <typename Scalar>
1908 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k0, Scalar)
1909     bessel_k0(const Scalar& x) {
1910   return EIGEN_MATHFUNC_IMPL(bessel_k0, Scalar)::run(x);
1911 }
1912 
1913 template <typename Scalar>
1914 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k0e, Scalar)
1915     bessel_k0e(const Scalar& x) {
1916   return EIGEN_MATHFUNC_IMPL(bessel_k0e, Scalar)::run(x);
1917 }
1918 
1919 template <typename Scalar>
1920 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k1, Scalar)
1921     bessel_k1(const Scalar& x) {
1922   return EIGEN_MATHFUNC_IMPL(bessel_k1, Scalar)::run(x);
1923 }
1924 
1925 template <typename Scalar>
1926 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_k1e, Scalar)
1927     bessel_k1e(const Scalar& x) {
1928   return EIGEN_MATHFUNC_IMPL(bessel_k1e, Scalar)::run(x);
1929 }
1930 
1931 template <typename Scalar>
1932 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_j0, Scalar)
1933     bessel_j0(const Scalar& x) {
1934   return EIGEN_MATHFUNC_IMPL(bessel_j0, Scalar)::run(x);
1935 }
1936 
1937 template <typename Scalar>
1938 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_y0, Scalar)
1939     bessel_y0(const Scalar& x) {
1940   return EIGEN_MATHFUNC_IMPL(bessel_y0, Scalar)::run(x);
1941 }
1942 
1943 template <typename Scalar>
1944 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_j1, Scalar)
1945     bessel_j1(const Scalar& x) {
1946   return EIGEN_MATHFUNC_IMPL(bessel_j1, Scalar)::run(x);
1947 }
1948 
1949 template <typename Scalar>
1950 EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(bessel_y1, Scalar)
1951     bessel_y1(const Scalar& x) {
1952   return EIGEN_MATHFUNC_IMPL(bessel_y1, Scalar)::run(x);
1953 }
1954 
1955 }  // end namespace numext
1956 
1957 }  // end namespace Eigen
1958 
1959 #endif  // EIGEN_BESSEL_FUNCTIONS_H
1960