1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
|
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_STABLENORM_H
#define EIGEN_STABLENORM_H
namespace Eigen {
namespace internal {
template<typename ExpressionType, typename Scalar>
inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
{
Scalar maxCoeff = bl.cwiseAbs().maxCoeff();
if(maxCoeff>scale)
{
ssq = ssq * numext::abs2(scale/maxCoeff);
Scalar tmp = Scalar(1)/maxCoeff;
if(tmp > NumTraits<Scalar>::highest())
{
invScale = NumTraits<Scalar>::highest();
scale = Scalar(1)/invScale;
}
else if(maxCoeff>NumTraits<Scalar>::highest()) // we got a INF
{
invScale = Scalar(1);
scale = maxCoeff;
}
else
{
scale = maxCoeff;
invScale = tmp;
}
}
else if(maxCoeff!=maxCoeff) // we got a NaN
{
scale = maxCoeff;
}
// TODO if the maxCoeff is much much smaller than the current scale,
// then we can neglect this sub vector
if(scale>Scalar(0)) // if scale==0, then bl is 0
ssq += (bl*invScale).squaredNorm();
}
template<typename VectorType, typename RealScalar>
void stable_norm_impl_inner_step(const VectorType &vec, RealScalar& ssq, RealScalar& scale, RealScalar& invScale)
{
typedef typename VectorType::Scalar Scalar;
const Index blockSize = 4096;
typedef typename internal::nested_eval<VectorType,2>::type VectorTypeCopy;
typedef typename internal::remove_all<VectorTypeCopy>::type VectorTypeCopyClean;
const VectorTypeCopy copy(vec);
enum {
CanAlign = ( (int(VectorTypeCopyClean::Flags)&DirectAccessBit)
|| (int(internal::evaluator<VectorTypeCopyClean>::Alignment)>0) // FIXME Alignment)>0 might not be enough
) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT)
&& (EIGEN_MAX_STATIC_ALIGN_BYTES>0) // if we cannot allocate on the stack, then let's not bother about this optimization
};
typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<VectorTypeCopyClean>::Alignment>,
typename VectorTypeCopyClean::ConstSegmentReturnType>::type SegmentWrapper;
Index n = vec.size();
Index bi = internal::first_default_aligned(copy);
if (bi>0)
internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale);
for (; bi<n; bi+=blockSize)
internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale);
}
template<typename VectorType>
typename VectorType::RealScalar
stable_norm_impl(const VectorType &vec, typename enable_if<VectorType::IsVectorAtCompileTime>::type* = 0 )
{
using std::sqrt;
using std::abs;
Index n = vec.size();
if(n==1)
return abs(vec.coeff(0));
typedef typename VectorType::RealScalar RealScalar;
RealScalar scale(0);
RealScalar invScale(1);
RealScalar ssq(0); // sum of squares
stable_norm_impl_inner_step(vec, ssq, scale, invScale);
return scale * sqrt(ssq);
}
template<typename MatrixType>
typename MatrixType::RealScalar
stable_norm_impl(const MatrixType &mat, typename enable_if<!MatrixType::IsVectorAtCompileTime>::type* = 0 )
{
using std::sqrt;
typedef typename MatrixType::RealScalar RealScalar;
RealScalar scale(0);
RealScalar invScale(1);
RealScalar ssq(0); // sum of squares
for(Index j=0; j<mat.outerSize(); ++j)
stable_norm_impl_inner_step(mat.innerVector(j), ssq, scale, invScale);
return scale * sqrt(ssq);
}
template<typename Derived>
inline typename NumTraits<typename traits<Derived>::Scalar>::Real
blueNorm_impl(const EigenBase<Derived>& _vec)
{
typedef typename Derived::RealScalar RealScalar;
using std::pow;
using std::sqrt;
using std::abs;
// This program calculates the machine-dependent constants
// bl, b2, slm, s2m, relerr overfl
// from the "basic" machine-dependent numbers
// nbig, ibeta, it, iemin, iemax, rbig.
// The following define the basic machine-dependent constants.
// For portability, the PORT subprograms "ilmaeh" and "rlmach"
// are used. For any specific computer, each of the assignment
// statements can be replaced
static const int ibeta = std::numeric_limits<RealScalar>::radix; // base for floating-point numbers
static const int it = NumTraits<RealScalar>::digits(); // number of base-beta digits in mantissa
static const int iemin = NumTraits<RealScalar>::min_exponent(); // minimum exponent
static const int iemax = NumTraits<RealScalar>::max_exponent(); // maximum exponent
static const RealScalar rbig = NumTraits<RealScalar>::highest(); // largest floating-point number
static const RealScalar b1 = RealScalar(pow(RealScalar(ibeta),RealScalar(-((1-iemin)/2)))); // lower boundary of midrange
static const RealScalar b2 = RealScalar(pow(RealScalar(ibeta),RealScalar((iemax + 1 - it)/2))); // upper boundary of midrange
static const RealScalar s1m = RealScalar(pow(RealScalar(ibeta),RealScalar((2-iemin)/2))); // scaling factor for lower range
static const RealScalar s2m = RealScalar(pow(RealScalar(ibeta),RealScalar(- ((iemax+it)/2)))); // scaling factor for upper range
static const RealScalar eps = RealScalar(pow(double(ibeta), 1-it));
static const RealScalar relerr = sqrt(eps); // tolerance for neglecting asml
const Derived& vec(_vec.derived());
Index n = vec.size();
RealScalar ab2 = b2 / RealScalar(n);
RealScalar asml = RealScalar(0);
RealScalar amed = RealScalar(0);
RealScalar abig = RealScalar(0);
for(Index j=0; j<vec.outerSize(); ++j)
{
for(typename Derived::InnerIterator iter(vec, j); iter; ++iter)
{
RealScalar ax = abs(iter.value());
if(ax > ab2) abig += numext::abs2(ax*s2m);
else if(ax < b1) asml += numext::abs2(ax*s1m);
else amed += numext::abs2(ax);
}
}
if(amed!=amed)
return amed; // we got a NaN
if(abig > RealScalar(0))
{
abig = sqrt(abig);
if(abig > rbig) // overflow, or *this contains INF values
return abig; // return INF
if(amed > RealScalar(0))
{
abig = abig/s2m;
amed = sqrt(amed);
}
else
return abig/s2m;
}
else if(asml > RealScalar(0))
{
if (amed > RealScalar(0))
{
abig = sqrt(amed);
amed = sqrt(asml) / s1m;
}
else
return sqrt(asml)/s1m;
}
else
return sqrt(amed);
asml = numext::mini(abig, amed);
abig = numext::maxi(abig, amed);
if(asml <= abig*relerr)
return abig;
else
return abig * sqrt(RealScalar(1) + numext::abs2(asml/abig));
}
} // end namespace internal
/** \returns the \em l2 norm of \c *this avoiding underflow and overflow.
* This version use a blockwise two passes algorithm:
* 1 - find the absolute largest coefficient \c s
* 2 - compute \f$ s \Vert \frac{*this}{s} \Vert \f$ in a standard way
*
* For architecture/scalar types supporting vectorization, this version
* is faster than blueNorm(). Otherwise the blueNorm() is much faster.
*
* \sa norm(), blueNorm(), hypotNorm()
*/
template<typename Derived>
inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
MatrixBase<Derived>::stableNorm() const
{
return internal::stable_norm_impl(derived());
}
/** \returns the \em l2 norm of \c *this using the Blue's algorithm.
* A Portable Fortran Program to Find the Euclidean Norm of a Vector,
* ACM TOMS, Vol 4, Issue 1, 1978.
*
* For architecture/scalar types without vectorization, this version
* is much faster than stableNorm(). Otherwise the stableNorm() is faster.
*
* \sa norm(), stableNorm(), hypotNorm()
*/
template<typename Derived>
inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
MatrixBase<Derived>::blueNorm() const
{
return internal::blueNorm_impl(*this);
}
/** \returns the \em l2 norm of \c *this avoiding undeflow and overflow.
* This version use a concatenation of hypot() calls, and it is very slow.
*
* \sa norm(), stableNorm()
*/
template<typename Derived>
inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
MatrixBase<Derived>::hypotNorm() const
{
if(size()==1)
return numext::abs(coeff(0,0));
else
return this->cwiseAbs().redux(internal::scalar_hypot_op<RealScalar>());
}
} // end namespace Eigen
#endif // EIGEN_STABLENORM_H
|
