blob: ee7fce2aeb019f889d0bd021426a01d70c42977f [file] [log] [blame]
Keir Mierle8ebb0732012-04-30 23:09:08 -07001// Ceres Solver - A fast non-linear least squares minimizer
2// Copyright 2010, 2011, 2012 Google Inc. All rights reserved.
3// http://code.google.com/p/ceres-solver/
4//
5// Redistribution and use in source and binary forms, with or without
6// modification, are permitted provided that the following conditions are met:
7//
8// * Redistributions of source code must retain the above copyright notice,
9// this list of conditions and the following disclaimer.
10// * Redistributions in binary form must reproduce the above copyright notice,
11// this list of conditions and the following disclaimer in the documentation
12// and/or other materials provided with the distribution.
13// * Neither the name of Google Inc. nor the names of its contributors may be
14// used to endorse or promote products derived from this software without
15// specific prior written permission.
16//
17// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
21// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27// POSSIBILITY OF SUCH DAMAGE.
28//
29// Author: sameeragarwal@google.com (Sameer Agarwal)
30//
31// Abstract interface for objects solving linear systems of various
32// kinds.
33
34#ifndef CERES_INTERNAL_LINEAR_SOLVER_H_
35#define CERES_INTERNAL_LINEAR_SOLVER_H_
36
37#include <cstddef>
Sameer Agarwal0c52f1e2012-09-17 11:30:14 -070038#include <vector>
Keir Mierle8ebb0732012-04-30 23:09:08 -070039
40#include <glog/logging.h>
41#include "ceres/block_sparse_matrix.h"
42#include "ceres/casts.h"
43#include "ceres/compressed_row_sparse_matrix.h"
44#include "ceres/dense_sparse_matrix.h"
45#include "ceres/triplet_sparse_matrix.h"
46#include "ceres/types.h"
47
48namespace ceres {
49namespace internal {
50
51class LinearOperator;
52
53// Abstract base class for objects that implement algorithms for
54// solving linear systems
55//
56// Ax = b
57//
58// It is expected that a single instance of a LinearSolver object
Sameer Agarwala9d8ef82012-05-14 02:28:05 -070059// maybe used multiple times for solving multiple linear systems with
60// the same sparsity structure. This allows them to cache and reuse
61// information across solves. This means that calling Solve on the
62// same LinearSolver instance with two different linear systems will
63// result in undefined behaviour.
Keir Mierle8ebb0732012-04-30 23:09:08 -070064//
65// Subclasses of LinearSolver use two structs to configure themselves.
66// The Options struct configures the LinearSolver object for its
67// lifetime. The PerSolveOptions struct is used to specify options for
68// a particular Solve call.
69class LinearSolver {
70 public:
71 struct Options {
72 Options()
73 : type(SPARSE_NORMAL_CHOLESKY),
74 preconditioner_type(JACOBI),
Sameer Agarwalb0518732012-05-29 00:27:57 -070075 sparse_linear_algebra_library(SUITE_SPARSE),
Sameer Agarwal7a3c43b2012-06-05 23:10:59 -070076 use_block_amd(true),
Keir Mierle8ebb0732012-04-30 23:09:08 -070077 min_num_iterations(1),
78 max_num_iterations(1),
79 num_threads(1),
Keir Mierle8ebb0732012-04-30 23:09:08 -070080 residual_reset_period(10),
81 row_block_size(Dynamic),
82 e_block_size(Dynamic),
83 f_block_size(Dynamic) {
84 }
85
86 LinearSolverType type;
87
88 PreconditionerType preconditioner_type;
89
Sameer Agarwalb0518732012-05-29 00:27:57 -070090 SparseLinearAlgebraLibraryType sparse_linear_algebra_library;
91
Sameer Agarwal7a3c43b2012-06-05 23:10:59 -070092 // See solver.h for explanation of this option.
93 bool use_block_amd;
94
Keir Mierle8ebb0732012-04-30 23:09:08 -070095 // Number of internal iterations that the solver uses. This
96 // parameter only makes sense for iterative solvers like CG.
97 int min_num_iterations;
98 int max_num_iterations;
99
100 // If possible, how many threads can the solver use.
101 int num_threads;
102
Sameer Agarwal0c52f1e2012-09-17 11:30:14 -0700103 // Hints about the order in which the parameter blocks should be
104 // eliminated by the linear solver.
105 //
106 // For example if elimination_groups is a vector of size k, then
107 // the linear solver is informed that it should eliminate the
108 // parameter blocks 0 - elimination_groups[0] - 1 first, and then
109 // elimination_groups[0] - elimination_groups[1] and so on. Within
110 // each elimination group, the linear solver is free to choose how
111 // the parameter blocks are ordered. Different linear solvers have
112 // differing requirements on elimination_groups.
113 //
114 // The most common use is for Schur type solvers, where there
115 // should be at least two elimination groups and the first
116 // elimination group must form an independent set in the normal
117 // equations. The first elimination group corresponds to the
118 // num_eliminate_blocks in the Schur type solvers.
119 vector<int> elimination_groups;
Keir Mierle8ebb0732012-04-30 23:09:08 -0700120
121 // Iterative solvers, e.g. Preconditioned Conjugate Gradients
122 // maintain a cheap estimate of the residual which may become
123 // inaccurate over time. Thus for non-zero values of this
124 // parameter, the solver can be told to recalculate the value of
125 // the residual using a |b - Ax| evaluation.
126 int residual_reset_period;
127
128 // If the block sizes in a BlockSparseMatrix are fixed, then in
129 // some cases the Schur complement based solvers can detect and
130 // specialize on them.
131 //
132 // It is expected that these parameters are set programmatically
133 // rather than manually.
134 //
Sameer Agarwala9d8ef82012-05-14 02:28:05 -0700135 // Please see schur_complement_solver.h and schur_eliminator.h for
136 // more details.
Keir Mierle8ebb0732012-04-30 23:09:08 -0700137 int row_block_size;
138 int e_block_size;
139 int f_block_size;
140 };
141
142 // Options for the Solve method.
143 struct PerSolveOptions {
144 PerSolveOptions()
145 : D(NULL),
146 preconditioner(NULL),
147 r_tolerance(0.0),
148 q_tolerance(0.0) {
149 }
150
151 // This option only makes sense for unsymmetric linear solvers
152 // that can solve rectangular linear systems.
153 //
154 // Given a matrix A, an optional diagonal matrix D as a vector,
155 // and a vector b, the linear solver will solve for
156 //
157 // | A | x = | b |
158 // | D | | 0 |
159 //
160 // If D is null, then it is treated as zero, and the solver returns
161 // the solution to
162 //
163 // A x = b
164 //
165 // In either case, x is the vector that solves the following
166 // optimization problem.
167 //
Keir Mierlef7898fb2012-05-05 20:55:08 -0700168 // arg min_x ||Ax - b||^2 + ||Dx||^2
Keir Mierle8ebb0732012-04-30 23:09:08 -0700169 //
170 // Here A is a matrix of size m x n, with full column rank. If A
171 // does not have full column rank, the results returned by the
172 // solver cannot be relied on. D, if it is not null is an array of
173 // size n. b is an array of size m and x is an array of size n.
174 double * D;
175
176 // This option only makes sense for iterative solvers.
177 //
178 // In general the performance of an iterative linear solver
179 // depends on the condition number of the matrix A. For example
180 // the convergence rate of the conjugate gradients algorithm
181 // is proportional to the square root of the condition number.
182 //
183 // One particularly useful technique for improving the
184 // conditioning of a linear system is to precondition it. In its
185 // simplest form a preconditioner is a matrix M such that instead
186 // of solving Ax = b, we solve the linear system AM^{-1} y = b
187 // instead, where M is such that the condition number k(AM^{-1})
188 // is smaller than the conditioner k(A). Given the solution to
189 // this system, x = M^{-1} y. The iterative solver takes care of
190 // the mechanics of solving the preconditioned system and
191 // returning the corrected solution x. The user only needs to
192 // supply a linear operator.
193 //
194 // A null preconditioner is equivalent to an identity matrix being
195 // used a preconditioner.
196 LinearOperator* preconditioner;
197
198
199 // The following tolerance related options only makes sense for
200 // iterative solvers. Direct solvers ignore them.
201
202 // Solver terminates when
203 //
204 // |Ax - b| <= r_tolerance * |b|.
205 //
206 // This is the most commonly used termination criterion for
207 // iterative solvers.
208 double r_tolerance;
209
210 // For PSD matrices A, let
211 //
212 // Q(x) = x'Ax - 2b'x
213 //
214 // be the cost of the quadratic function defined by A and b. Then,
215 // the solver terminates at iteration i if
216 //
217 // i * (Q(x_i) - Q(x_i-1)) / Q(x_i) < q_tolerance.
218 //
219 // This termination criterion is more useful when using CG to
220 // solve the Newton step. This particular convergence test comes
221 // from Stephen Nash's work on truncated Newton
222 // methods. References:
223 //
224 // 1. Stephen G. Nash & Ariela Sofer, Assessing A Search
225 // Direction Within A Truncated Newton Method, Operation
226 // Research Letters 9(1990) 219-221.
227 //
228 // 2. Stephen G. Nash, A Survey of Truncated Newton Methods,
229 // Journal of Computational and Applied Mathematics,
230 // 124(1-2), 45-59, 2000.
231 //
232 double q_tolerance;
233 };
234
235 // Summary of a call to the Solve method. We should move away from
236 // the true/false method for determining solver success. We should
237 // let the summary object do the talking.
238 struct Summary {
239 Summary()
240 : residual_norm(0.0),
241 num_iterations(-1),
242 termination_type(FAILURE) {
243 }
244
245 double residual_norm;
246 int num_iterations;
247 LinearSolverTerminationType termination_type;
248 };
249
250 virtual ~LinearSolver();
251
252 // Solve Ax = b.
253 virtual Summary Solve(LinearOperator* A,
254 const double* b,
255 const PerSolveOptions& per_solve_options,
256 double* x) = 0;
257
Sameer Agarwala9d8ef82012-05-14 02:28:05 -0700258 // Factory
Keir Mierle8ebb0732012-04-30 23:09:08 -0700259 static LinearSolver* Create(const Options& options);
260};
261
262// This templated subclass of LinearSolver serves as a base class for
263// other linear solvers that depend on the particular matrix layout of
264// the underlying linear operator. For example some linear solvers
265// need low level access to the TripletSparseMatrix implementing the
266// LinearOperator interface. This class hides those implementation
267// details behind a private virtual method, and has the Solve method
268// perform the necessary upcasting.
269template <typename MatrixType>
270class TypedLinearSolver : public LinearSolver {
271 public:
272 virtual ~TypedLinearSolver() {}
273 virtual LinearSolver::Summary Solve(
274 LinearOperator* A,
275 const double* b,
276 const LinearSolver::PerSolveOptions& per_solve_options,
277 double* x) {
278 CHECK_NOTNULL(A);
279 CHECK_NOTNULL(b);
280 CHECK_NOTNULL(x);
281 return SolveImpl(down_cast<MatrixType*>(A), b, per_solve_options, x);
282 }
283
284 private:
285 virtual LinearSolver::Summary SolveImpl(
286 MatrixType* A,
287 const double* b,
288 const LinearSolver::PerSolveOptions& per_solve_options,
289 double* x) = 0;
290};
291
292// Linear solvers that depend on acccess to the low level structure of
293// a SparseMatrix.
294typedef TypedLinearSolver<BlockSparseMatrix> BlockSparseMatrixSolver; // NOLINT
295typedef TypedLinearSolver<BlockSparseMatrixBase> BlockSparseMatrixBaseSolver; // NOLINT
296typedef TypedLinearSolver<CompressedRowSparseMatrix> CompressedRowSparseMatrixSolver; // NOLINT
297typedef TypedLinearSolver<DenseSparseMatrix> DenseSparseMatrixSolver; // NOLINT
298typedef TypedLinearSolver<TripletSparseMatrix> TripletSparseMatrixSolver; // NOLINT
299
300} // namespace internal
301} // namespace ceres
302
303#endif // CERES_INTERNAL_LINEAR_SOLVER_H_