$darkmode
Eigen  5.0.1-dev
MathFunctions.h
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2007 Julien Pommier
5 // Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
6 // Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
7 //
8 // This Source Code Form is subject to the terms of the Mozilla
9 // Public License v. 2.0. If a copy of the MPL was not distributed
10 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
11 
12 /* The sin, cos, exp, and log functions of this file come from
13  * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
14  */
15 
16 #ifndef EIGEN_MATH_FUNCTIONS_ZVECTOR_H
17 #define EIGEN_MATH_FUNCTIONS_ZVECTOR_H
18 
19 // IWYU pragma: private
20 #include "../../InternalHeaderCheck.h"
21 
22 namespace Eigen {
23 
24 namespace internal {
25 
26 EIGEN_DOUBLE_PACKET_FUNCTION(atanh, Packet2d)
27 EIGEN_DOUBLE_PACKET_FUNCTION(log, Packet2d)
28 EIGEN_DOUBLE_PACKET_FUNCTION(log2, Packet2d)
29 EIGEN_DOUBLE_PACKET_FUNCTION(tanh, Packet2d)
30 
31 EIGEN_FLOAT_PACKET_FUNCTION(atanh, Packet4f)
32 EIGEN_FLOAT_PACKET_FUNCTION(log, Packet4f)
33 EIGEN_FLOAT_PACKET_FUNCTION(log2, Packet4f)
34 
35 EIGEN_GENERIC_PACKET_FUNCTION(atan, Packet2d)
36 EIGEN_GENERIC_PACKET_FUNCTION(atan, Packet4f)
37 EIGEN_GENERIC_PACKET_FUNCTION(exp2, Packet2d)
38 EIGEN_GENERIC_PACKET_FUNCTION(exp2, Packet4f)
39 
40 #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
41 static EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
42 static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
43 static EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
44 static EIGEN_DECLARE_CONST_Packet4i(23, 23);
45 
46 static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
47 
48 /* the smallest non denormalized float number */
49 static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos, 0x00800000);
50 static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf, 0xff800000); // -1.f/0.f
51 static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan, 0xffffffff);
52 
53 /* natural logarithm computed for 4 simultaneous float
54  return NaN for x <= 0
55 */
56 static EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
57 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
58 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, -1.1514610310E-1f);
59 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
60 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, -1.2420140846E-1f);
61 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, +1.4249322787E-1f);
62 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, -1.6668057665E-1f);
63 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, +2.0000714765E-1f);
64 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, -2.4999993993E-1f);
65 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, +3.3333331174E-1f);
66 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
67 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
68 
69 static EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f);
70 static EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
71 
72 static EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
73 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
74 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
75 
76 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
77 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
78 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
79 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
80 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
81 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
82 #endif
83 
84 static EIGEN_DECLARE_CONST_Packet2d(1, 1.0);
85 static EIGEN_DECLARE_CONST_Packet2d(2, 2.0);
86 static EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
87 
88 static EIGEN_DECLARE_CONST_Packet2d(exp_hi, 709.437);
89 static EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
90 
91 static EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
92 
93 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
94 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
95 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
96 
97 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
98 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
99 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
100 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
101 
102 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
103 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
104 
105 template <>
106 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet2d pexp<Packet2d>(const Packet2d& _x) {
107  Packet2d x = _x;
108 
109  Packet2d tmp, fx;
110  Packet2l emm0;
111 
112  // clamp x
113  x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
114  /* express exp(x) as exp(g + n*log(2)) */
115  fx = pmadd(p2d_cephes_LOG2EF, x, p2d_half);
116 
117  fx = vec_floor(fx);
118 
119  tmp = pmul(fx, p2d_cephes_exp_C1);
120  Packet2d z = pmul(fx, p2d_cephes_exp_C2);
121  x = psub(x, tmp);
122  x = psub(x, z);
123 
124  Packet2d x2 = pmul(x, x);
125 
126  Packet2d px = p2d_cephes_exp_p0;
127  px = pmadd(px, x2, p2d_cephes_exp_p1);
128  px = pmadd(px, x2, p2d_cephes_exp_p2);
129  px = pmul(px, x);
130 
131  Packet2d qx = p2d_cephes_exp_q0;
132  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
133  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
134  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
135 
136  x = pdiv(px, psub(qx, px));
137  x = pmadd(p2d_2, x, p2d_1);
138 
139  // build 2^n
140  emm0 = vec_ctsl(fx, 0);
141 
142  static const Packet2l p2l_1023 = {1023, 1023};
143  static const Packet2ul p2ul_52 = {52, 52};
144 
145  emm0 = emm0 + p2l_1023;
146  emm0 = emm0 << reinterpret_cast<Packet2l>(p2ul_52);
147 
148  // Altivec's max & min operators just drop silent NaNs. Check NaNs in
149  // inputs and return them unmodified.
150  Packet2ul isnumber_mask = reinterpret_cast<Packet2ul>(vec_cmpeq(_x, _x));
151  return vec_sel(_x, pmax(pmul(x, reinterpret_cast<Packet2d>(emm0)), _x), isnumber_mask);
152 }
153 
154 template <>
155 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f pexp<Packet4f>(const Packet4f& _x) {
156 #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
157  Packet4f x = _x;
158 
159  Packet4f tmp, fx;
160  Packet4i emm0;
161 
162  // clamp x
163  x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo);
164 
165  // express exp(x) as exp(g + n*log(2))
166  fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half);
167 
168  fx = pfloor(fx);
169 
170  tmp = pmul(fx, p4f_cephes_exp_C1);
171  Packet4f z = pmul(fx, p4f_cephes_exp_C2);
172  x = psub(x, tmp);
173  x = psub(x, z);
174 
175  z = pmul(x, x);
176 
177  Packet4f y = p4f_cephes_exp_p0;
178  y = pmadd(y, x, p4f_cephes_exp_p1);
179  y = pmadd(y, x, p4f_cephes_exp_p2);
180  y = pmadd(y, x, p4f_cephes_exp_p3);
181  y = pmadd(y, x, p4f_cephes_exp_p4);
182  y = pmadd(y, x, p4f_cephes_exp_p5);
183  y = pmadd(y, z, x);
184  y = padd(y, p4f_1);
185 
186  // build 2^n
187  emm0 = Packet4i{(int)fx[0], (int)fx[1], (int)fx[2], (int)fx[3]};
188  emm0 = emm0 + p4i_0x7f;
189  emm0 = emm0 << reinterpret_cast<Packet4i>(p4i_23);
190 
191  return pmax(pmul(y, reinterpret_cast<Packet4f>(emm0)), _x);
192 #else
193  Packet4f res;
194  res.v4f[0] = pexp<Packet2d>(_x.v4f[0]);
195  res.v4f[1] = pexp<Packet2d>(_x.v4f[1]);
196  return res;
197 #endif
198 }
199 
200 template <>
201 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet2d psqrt<Packet2d>(const Packet2d& x) {
202  return vec_sqrt(x);
203 }
204 
205 template <>
206 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f psqrt<Packet4f>(const Packet4f& x) {
207  Packet4f res;
208 #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
209  res = vec_sqrt(x);
210 #else
211  res.v4f[0] = psqrt<Packet2d>(x.v4f[0]);
212  res.v4f[1] = psqrt<Packet2d>(x.v4f[1]);
213 #endif
214  return res;
215 }
216 
217 template <>
218 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet2d prsqrt<Packet2d>(const Packet2d& x) {
219  return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x);
220 }
221 
222 template <>
223 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f prsqrt<Packet4f>(const Packet4f& x) {
224  Packet4f res;
225 #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
226  res = pset1<Packet4f>(1.0) / psqrt<Packet4f>(x);
227 #else
228  res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]);
229  res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]);
230 #endif
231  return res;
232 }
233 
234 // Hyperbolic Tangent function.
235 template <>
236 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f ptanh<Packet4f>(const Packet4f& x) {
237  return ptanh_float(x);
238 }
239 
240 } // end namespace internal
241 
242 } // end namespace Eigen
243 
244 #endif // EIGEN_MATH_FUNCTIONS_ZVECTOR_H
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_tanh_op< typename Derived::Scalar >, const Derived > tanh(const Eigen::ArrayBase< Derived > &x)
Namespace containing all symbols from the Eigen library.
Definition: B01_Experimental.dox:1
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_exp2_op< typename Derived::Scalar >, const Derived > exp2(const Eigen::ArrayBase< Derived > &x)
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_log_op< typename Derived::Scalar >, const Derived > log(const Eigen::ArrayBase< Derived > &x)
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_atanh_op< typename Derived::Scalar >, const Derived > atanh(const Eigen::ArrayBase< Derived > &x)
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_log2_op< typename Derived::Scalar >, const Derived > log2(const Eigen::ArrayBase< Derived > &x)
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_atan_op< typename Derived::Scalar >, const Derived > atan(const Eigen::ArrayBase< Derived > &x)