1// 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: keir@google.com (Keir Mierle)
30
31#include "ceres/compressed_row_jacobian_writer.h"
32
33#include "ceres/casts.h"
34#include "ceres/compressed_row_sparse_matrix.h"
35#include "ceres/parameter_block.h"
36#include "ceres/program.h"
37#include "ceres/residual_block.h"
38#include "ceres/scratch_evaluate_preparer.h"
39
40namespace ceres {
41namespace internal {
42
43void CompressedRowJacobianWriter::PopulateJacobianRowAndColumnBlockVectors(
44    const Program* program, CompressedRowSparseMatrix* jacobian) {
45  const vector<ParameterBlock*>& parameter_blocks =
46      program->parameter_blocks();
47  vector<int>& col_blocks = *(jacobian->mutable_col_blocks());
48  col_blocks.resize(parameter_blocks.size());
49  for (int i = 0; i < parameter_blocks.size(); ++i) {
50    col_blocks[i] = parameter_blocks[i]->LocalSize();
51  }
52
53  const vector<ResidualBlock*>& residual_blocks =
54      program->residual_blocks();
55  vector<int>& row_blocks = *(jacobian->mutable_row_blocks());
56  row_blocks.resize(residual_blocks.size());
57  for (int i = 0; i < residual_blocks.size(); ++i) {
58    row_blocks[i] = residual_blocks[i]->NumResiduals();
59  }
60}
61
62void CompressedRowJacobianWriter::GetOrderedParameterBlocks(
63      const Program* program,
64      int residual_id,
65      vector<pair<int, int> >* evaluated_jacobian_blocks) {
66  const ResidualBlock* residual_block =
67      program->residual_blocks()[residual_id];
68  const int num_parameter_blocks = residual_block->NumParameterBlocks();
69
70  for (int j = 0; j < num_parameter_blocks; ++j) {
71    const ParameterBlock* parameter_block =
72        residual_block->parameter_blocks()[j];
73    if (!parameter_block->IsConstant()) {
74      evaluated_jacobian_blocks->push_back(
75          make_pair(parameter_block->index(), j));
76    }
77  }
78  sort(evaluated_jacobian_blocks->begin(), evaluated_jacobian_blocks->end());
79}
80
81SparseMatrix* CompressedRowJacobianWriter::CreateJacobian() const {
82  const vector<ResidualBlock*>& residual_blocks =
83      program_->residual_blocks();
84
85  int total_num_residuals = program_->NumResiduals();
86  int total_num_effective_parameters = program_->NumEffectiveParameters();
87
88  // Count the number of jacobian nonzeros.
89  int num_jacobian_nonzeros = 0;
90  for (int i = 0; i < residual_blocks.size(); ++i) {
91    ResidualBlock* residual_block = residual_blocks[i];
92    const int num_residuals = residual_block->NumResiduals();
93    const int num_parameter_blocks = residual_block->NumParameterBlocks();
94    for (int j = 0; j < num_parameter_blocks; ++j) {
95      ParameterBlock* parameter_block = residual_block->parameter_blocks()[j];
96      if (!parameter_block->IsConstant()) {
97        num_jacobian_nonzeros += num_residuals * parameter_block->LocalSize();
98      }
99    }
100  }
101
102  // Allocate storage for the jacobian with some extra space at the end.
103  // Allocate more space than needed to store the jacobian so that when the LM
104  // algorithm adds the diagonal, no reallocation is necessary. This reduces
105  // peak memory usage significantly.
106  CompressedRowSparseMatrix* jacobian =
107      new CompressedRowSparseMatrix(
108          total_num_residuals,
109          total_num_effective_parameters,
110          num_jacobian_nonzeros + total_num_effective_parameters);
111
112  // At this stage, the CompressedRowSparseMatrix is an invalid state. But this
113  // seems to be the only way to construct it without doing a memory copy.
114  int* rows = jacobian->mutable_rows();
115  int* cols = jacobian->mutable_cols();
116  int row_pos = 0;
117  rows[0] = 0;
118  for (int i = 0; i < residual_blocks.size(); ++i) {
119    const ResidualBlock* residual_block = residual_blocks[i];
120    const int num_parameter_blocks = residual_block->NumParameterBlocks();
121
122    // Count the number of derivatives for a row of this residual block and
123    // build a list of active parameter block indices.
124    int num_derivatives = 0;
125    vector<int> parameter_indices;
126    for (int j = 0; j < num_parameter_blocks; ++j) {
127      ParameterBlock* parameter_block = residual_block->parameter_blocks()[j];
128      if (!parameter_block->IsConstant()) {
129        parameter_indices.push_back(parameter_block->index());
130        num_derivatives += parameter_block->LocalSize();
131      }
132    }
133
134    // Sort the parameters by their position in the state vector.
135    sort(parameter_indices.begin(), parameter_indices.end());
136    CHECK(unique(parameter_indices.begin(), parameter_indices.end()) ==
137          parameter_indices.end())
138          << "Ceres internal error:  "
139          << "Duplicate parameter blocks detected in a cost function. "
140          << "This should never happen. Please report this to "
141          << "the Ceres developers.";
142
143    // Update the row indices.
144    const int num_residuals = residual_block->NumResiduals();
145    for (int j = 0; j < num_residuals; ++j) {
146      rows[row_pos + j + 1] = rows[row_pos + j] + num_derivatives;
147    }
148
149    // Iterate over parameter blocks in the order which they occur in the
150    // parameter vector. This code mirrors that in Write(), where jacobian
151    // values are updated.
152    int col_pos = 0;
153    for (int j = 0; j < parameter_indices.size(); ++j) {
154      ParameterBlock* parameter_block =
155          program_->parameter_blocks()[parameter_indices[j]];
156      const int parameter_block_size = parameter_block->LocalSize();
157
158      for (int r = 0; r < num_residuals; ++r) {
159        // This is the position in the values array of the jacobian where this
160        // row of the jacobian block should go.
161        const int column_block_begin = rows[row_pos + r] + col_pos;
162
163        for (int c = 0; c < parameter_block_size; ++c) {
164          cols[column_block_begin + c] = parameter_block->delta_offset() + c;
165        }
166      }
167      col_pos += parameter_block_size;
168    }
169    row_pos += num_residuals;
170  }
171  CHECK_EQ(num_jacobian_nonzeros, rows[total_num_residuals]);
172
173  PopulateJacobianRowAndColumnBlockVectors(program_, jacobian);
174
175  return jacobian;
176}
177
178void CompressedRowJacobianWriter::Write(int residual_id,
179                                        int residual_offset,
180                                        double **jacobians,
181                                        SparseMatrix* base_jacobian) {
182  CompressedRowSparseMatrix* jacobian =
183      down_cast<CompressedRowSparseMatrix*>(base_jacobian);
184
185  double* jacobian_values = jacobian->mutable_values();
186  const int* jacobian_rows = jacobian->rows();
187
188  const ResidualBlock* residual_block =
189      program_->residual_blocks()[residual_id];
190  const int num_residuals = residual_block->NumResiduals();
191
192  vector<pair<int, int> > evaluated_jacobian_blocks;
193  GetOrderedParameterBlocks(program_, residual_id, &evaluated_jacobian_blocks);
194
195  // Where in the current row does the jacobian for a parameter block begin.
196  int col_pos = 0;
197
198  // Iterate over the jacobian blocks in increasing order of their
199  // positions in the reduced parameter vector.
200  for (int i = 0; i < evaluated_jacobian_blocks.size(); ++i) {
201    const ParameterBlock* parameter_block =
202        program_->parameter_blocks()[evaluated_jacobian_blocks[i].first];
203    const int argument = evaluated_jacobian_blocks[i].second;
204    const int parameter_block_size = parameter_block->LocalSize();
205
206    // Copy one row of the jacobian block at a time.
207    for (int r = 0; r < num_residuals; ++r) {
208      // Position of the r^th row of the current jacobian block.
209      const double* block_row_begin =
210          jacobians[argument] + r * parameter_block_size;
211
212      // Position in the values array of the jacobian where this
213      // row of the jacobian block should go.
214      double* column_block_begin =
215          jacobian_values + jacobian_rows[residual_offset + r] + col_pos;
216
217      copy(block_row_begin,
218           block_row_begin + parameter_block_size,
219           column_block_begin);
220    }
221    col_pos += parameter_block_size;
222  }
223}
224
225}  // namespace internal
226}  // namespace ceres
227