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