• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
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_STABLENORM_H
11 #define EIGEN_STABLENORM_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 template<typename ExpressionType, typename Scalar>
stable_norm_kernel(const ExpressionType & bl,Scalar & ssq,Scalar & scale,Scalar & invScale)17 inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
18 {
19   Scalar max = bl.cwiseAbs().maxCoeff();
20   if (max>scale)
21   {
22     ssq = ssq * abs2(scale/max);
23     scale = max;
24     invScale = Scalar(1)/scale;
25   }
26   // TODO if the max is much much smaller than the current scale,
27   // then we can neglect this sub vector
28   ssq += (bl*invScale).squaredNorm();
29 }
30 }
31 
32 /** \returns the \em l2 norm of \c *this avoiding underflow and overflow.
33   * This version use a blockwise two passes algorithm:
34   *  1 - find the absolute largest coefficient \c s
35   *  2 - compute \f$ s \Vert \frac{*this}{s} \Vert \f$ in a standard way
36   *
37   * For architecture/scalar types supporting vectorization, this version
38   * is faster than blueNorm(). Otherwise the blueNorm() is much faster.
39   *
40   * \sa norm(), blueNorm(), hypotNorm()
41   */
42 template<typename Derived>
43 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
stableNorm()44 MatrixBase<Derived>::stableNorm() const
45 {
46   using std::min;
47   const Index blockSize = 4096;
48   RealScalar scale(0);
49   RealScalar invScale(1);
50   RealScalar ssq(0); // sum of square
51   enum {
52     Alignment = (int(Flags)&DirectAccessBit) || (int(Flags)&AlignedBit) ? 1 : 0
53   };
54   Index n = size();
55   Index bi = internal::first_aligned(derived());
56   if (bi>0)
57     internal::stable_norm_kernel(this->head(bi), ssq, scale, invScale);
58   for (; bi<n; bi+=blockSize)
59     internal::stable_norm_kernel(this->segment(bi,(min)(blockSize, n - bi)).template forceAlignedAccessIf<Alignment>(), ssq, scale, invScale);
60   return scale * internal::sqrt(ssq);
61 }
62 
63 /** \returns the \em l2 norm of \c *this using the Blue's algorithm.
64   * A Portable Fortran Program to Find the Euclidean Norm of a Vector,
65   * ACM TOMS, Vol 4, Issue 1, 1978.
66   *
67   * For architecture/scalar types without vectorization, this version
68   * is much faster than stableNorm(). Otherwise the stableNorm() is faster.
69   *
70   * \sa norm(), stableNorm(), hypotNorm()
71   */
72 template<typename Derived>
73 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
blueNorm()74 MatrixBase<Derived>::blueNorm() const
75 {
76   using std::pow;
77   using std::min;
78   using std::max;
79   static Index nmax = -1;
80   static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr;
81   if(nmax <= 0)
82   {
83     int nbig, ibeta, it, iemin, iemax, iexp;
84     RealScalar abig, eps;
85     // This program calculates the machine-dependent constants
86     // bl, b2, slm, s2m, relerr overfl, nmax
87     // from the "basic" machine-dependent numbers
88     // nbig, ibeta, it, iemin, iemax, rbig.
89     // The following define the basic machine-dependent constants.
90     // For portability, the PORT subprograms "ilmaeh" and "rlmach"
91     // are used. For any specific computer, each of the assignment
92     // statements can be replaced
93     nbig  = (std::numeric_limits<Index>::max)();            // largest integer
94     ibeta = std::numeric_limits<RealScalar>::radix;         // base for floating-point numbers
95     it    = std::numeric_limits<RealScalar>::digits;        // number of base-beta digits in mantissa
96     iemin = std::numeric_limits<RealScalar>::min_exponent;  // minimum exponent
97     iemax = std::numeric_limits<RealScalar>::max_exponent;  // maximum exponent
98     rbig  = (std::numeric_limits<RealScalar>::max)();         // largest floating-point number
99 
100     iexp  = -((1-iemin)/2);
101     b1    = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));  // lower boundary of midrange
102     iexp  = (iemax + 1 - it)/2;
103     b2    = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));   // upper boundary of midrange
104 
105     iexp  = (2-iemin)/2;
106     s1m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));   // scaling factor for lower range
107     iexp  = - ((iemax+it)/2);
108     s2m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));   // scaling factor for upper range
109 
110     overfl  = rbig*s2m;             // overflow boundary for abig
111     eps     = RealScalar(pow(double(ibeta), 1-it));
112     relerr  = internal::sqrt(eps);         // tolerance for neglecting asml
113     abig    = RealScalar(1.0/eps - 1.0);
114     if (RealScalar(nbig)>abig)  nmax = int(abig);  // largest safe n
115     else                        nmax = nbig;
116   }
117   Index n = size();
118   RealScalar ab2 = b2 / RealScalar(n);
119   RealScalar asml = RealScalar(0);
120   RealScalar amed = RealScalar(0);
121   RealScalar abig = RealScalar(0);
122   for(Index j=0; j<n; ++j)
123   {
124     RealScalar ax = internal::abs(coeff(j));
125     if(ax > ab2)     abig += internal::abs2(ax*s2m);
126     else if(ax < b1) asml += internal::abs2(ax*s1m);
127     else             amed += internal::abs2(ax);
128   }
129   if(abig > RealScalar(0))
130   {
131     abig = internal::sqrt(abig);
132     if(abig > overfl)
133     {
134       eigen_assert(false && "overflow");
135       return rbig;
136     }
137     if(amed > RealScalar(0))
138     {
139       abig = abig/s2m;
140       amed = internal::sqrt(amed);
141     }
142     else
143       return abig/s2m;
144   }
145   else if(asml > RealScalar(0))
146   {
147     if (amed > RealScalar(0))
148     {
149       abig = internal::sqrt(amed);
150       amed = internal::sqrt(asml) / s1m;
151     }
152     else
153       return internal::sqrt(asml)/s1m;
154   }
155   else
156     return internal::sqrt(amed);
157   asml = (min)(abig, amed);
158   abig = (max)(abig, amed);
159   if(asml <= abig*relerr)
160     return abig;
161   else
162     return abig * internal::sqrt(RealScalar(1) + internal::abs2(asml/abig));
163 }
164 
165 /** \returns the \em l2 norm of \c *this avoiding undeflow and overflow.
166   * This version use a concatenation of hypot() calls, and it is very slow.
167   *
168   * \sa norm(), stableNorm()
169   */
170 template<typename Derived>
171 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
hypotNorm()172 MatrixBase<Derived>::hypotNorm() const
173 {
174   return this->cwiseAbs().redux(internal::scalar_hypot_op<RealScalar>());
175 }
176 
177 } // end namespace Eigen
178 
179 #endif // EIGEN_STABLENORM_H
180