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: sameeragarwal@google.com (Sameer Agarwal)
30
31 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 10
32
33 #include "ceres/partitioned_matrix_view.h"
34
35 #include <algorithm>
36 #include <cstring>
37 #include <vector>
38 #include "ceres/block_sparse_matrix.h"
39 #include "ceres/block_structure.h"
40 #include "ceres/internal/eigen.h"
41 #include "ceres/small_blas.h"
42 #include "glog/logging.h"
43
44 namespace ceres {
45 namespace internal {
46
PartitionedMatrixView(const BlockSparseMatrix & matrix,int num_col_blocks_a)47 PartitionedMatrixView::PartitionedMatrixView(
48 const BlockSparseMatrix& matrix,
49 int num_col_blocks_a)
50 : matrix_(matrix),
51 num_col_blocks_e_(num_col_blocks_a) {
52 const CompressedRowBlockStructure* bs = matrix_.block_structure();
53 CHECK_NOTNULL(bs);
54
55 num_col_blocks_f_ = bs->cols.size() - num_col_blocks_a;
56
57 // Compute the number of row blocks in E. The number of row blocks
58 // in E maybe less than the number of row blocks in the input matrix
59 // as some of the row blocks at the bottom may not have any
60 // e_blocks. For a definition of what an e_block is, please see
61 // explicit_schur_complement_solver.h
62 num_row_blocks_e_ = 0;
63 for (int r = 0; r < bs->rows.size(); ++r) {
64 const vector<Cell>& cells = bs->rows[r].cells;
65 if (cells[0].block_id < num_col_blocks_a) {
66 ++num_row_blocks_e_;
67 }
68 }
69
70 // Compute the number of columns in E and F.
71 num_cols_e_ = 0;
72 num_cols_f_ = 0;
73
74 for (int c = 0; c < bs->cols.size(); ++c) {
75 const Block& block = bs->cols[c];
76 if (c < num_col_blocks_a) {
77 num_cols_e_ += block.size;
78 } else {
79 num_cols_f_ += block.size;
80 }
81 }
82
83 CHECK_EQ(num_cols_e_ + num_cols_f_, matrix_.num_cols());
84 }
85
~PartitionedMatrixView()86 PartitionedMatrixView::~PartitionedMatrixView() {
87 }
88
89 // The next four methods don't seem to be particularly cache
90 // friendly. This is an artifact of how the BlockStructure of the
91 // input matrix is constructed. These methods will benefit from
92 // multithreading as well as improved data layout.
93
RightMultiplyE(const double * x,double * y) const94 void PartitionedMatrixView::RightMultiplyE(const double* x, double* y) const {
95 const CompressedRowBlockStructure* bs = matrix_.block_structure();
96
97 // Iterate over the first num_row_blocks_e_ row blocks, and multiply
98 // by the first cell in each row block.
99 const double* values = matrix_.values();
100 for (int r = 0; r < num_row_blocks_e_; ++r) {
101 const Cell& cell = bs->rows[r].cells[0];
102 const int row_block_pos = bs->rows[r].block.position;
103 const int row_block_size = bs->rows[r].block.size;
104 const int col_block_id = cell.block_id;
105 const int col_block_pos = bs->cols[col_block_id].position;
106 const int col_block_size = bs->cols[col_block_id].size;
107 MatrixVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
108 values + cell.position, row_block_size, col_block_size,
109 x + col_block_pos,
110 y + row_block_pos);
111 }
112 }
113
RightMultiplyF(const double * x,double * y) const114 void PartitionedMatrixView::RightMultiplyF(const double* x, double* y) const {
115 const CompressedRowBlockStructure* bs = matrix_.block_structure();
116
117 // Iterate over row blocks, and if the row block is in E, then
118 // multiply by all the cells except the first one which is of type
119 // E. If the row block is not in E (i.e its in the bottom
120 // num_row_blocks - num_row_blocks_e row blocks), then all the cells
121 // are of type F and multiply by them all.
122 const double* values = matrix_.values();
123 for (int r = 0; r < bs->rows.size(); ++r) {
124 const int row_block_pos = bs->rows[r].block.position;
125 const int row_block_size = bs->rows[r].block.size;
126 const vector<Cell>& cells = bs->rows[r].cells;
127 for (int c = (r < num_row_blocks_e_) ? 1 : 0; c < cells.size(); ++c) {
128 const int col_block_id = cells[c].block_id;
129 const int col_block_pos = bs->cols[col_block_id].position;
130 const int col_block_size = bs->cols[col_block_id].size;
131 MatrixVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
132 values + cells[c].position, row_block_size, col_block_size,
133 x + col_block_pos - num_cols_e(),
134 y + row_block_pos);
135 }
136 }
137 }
138
LeftMultiplyE(const double * x,double * y) const139 void PartitionedMatrixView::LeftMultiplyE(const double* x, double* y) const {
140 const CompressedRowBlockStructure* bs = matrix_.block_structure();
141
142 // Iterate over the first num_row_blocks_e_ row blocks, and multiply
143 // by the first cell in each row block.
144 const double* values = matrix_.values();
145 for (int r = 0; r < num_row_blocks_e_; ++r) {
146 const Cell& cell = bs->rows[r].cells[0];
147 const int row_block_pos = bs->rows[r].block.position;
148 const int row_block_size = bs->rows[r].block.size;
149 const int col_block_id = cell.block_id;
150 const int col_block_pos = bs->cols[col_block_id].position;
151 const int col_block_size = bs->cols[col_block_id].size;
152 MatrixTransposeVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
153 values + cell.position, row_block_size, col_block_size,
154 x + row_block_pos,
155 y + col_block_pos);
156 }
157 }
158
LeftMultiplyF(const double * x,double * y) const159 void PartitionedMatrixView::LeftMultiplyF(const double* x, double* y) const {
160 const CompressedRowBlockStructure* bs = matrix_.block_structure();
161
162 // Iterate over row blocks, and if the row block is in E, then
163 // multiply by all the cells except the first one which is of type
164 // E. If the row block is not in E (i.e its in the bottom
165 // num_row_blocks - num_row_blocks_e row blocks), then all the cells
166 // are of type F and multiply by them all.
167 const double* values = matrix_.values();
168 for (int r = 0; r < bs->rows.size(); ++r) {
169 const int row_block_pos = bs->rows[r].block.position;
170 const int row_block_size = bs->rows[r].block.size;
171 const vector<Cell>& cells = bs->rows[r].cells;
172 for (int c = (r < num_row_blocks_e_) ? 1 : 0; c < cells.size(); ++c) {
173 const int col_block_id = cells[c].block_id;
174 const int col_block_pos = bs->cols[col_block_id].position;
175 const int col_block_size = bs->cols[col_block_id].size;
176 MatrixTransposeVectorMultiply<Eigen::Dynamic, Eigen::Dynamic, 1>(
177 values + cells[c].position, row_block_size, col_block_size,
178 x + row_block_pos,
179 y + col_block_pos - num_cols_e());
180 }
181 }
182 }
183
184 // Given a range of columns blocks of a matrix m, compute the block
185 // structure of the block diagonal of the matrix m(:,
186 // start_col_block:end_col_block)'m(:, start_col_block:end_col_block)
187 // and return a BlockSparseMatrix with the this block structure. The
188 // caller owns the result.
CreateBlockDiagonalMatrixLayout(int start_col_block,int end_col_block) const189 BlockSparseMatrix* PartitionedMatrixView::CreateBlockDiagonalMatrixLayout(
190 int start_col_block, int end_col_block) const {
191 const CompressedRowBlockStructure* bs = matrix_.block_structure();
192 CompressedRowBlockStructure* block_diagonal_structure =
193 new CompressedRowBlockStructure;
194
195 int block_position = 0;
196 int diagonal_cell_position = 0;
197
198 // Iterate over the column blocks, creating a new diagonal block for
199 // each column block.
200 for (int c = start_col_block; c < end_col_block; ++c) {
201 const Block& block = bs->cols[c];
202 block_diagonal_structure->cols.push_back(Block());
203 Block& diagonal_block = block_diagonal_structure->cols.back();
204 diagonal_block.size = block.size;
205 diagonal_block.position = block_position;
206
207 block_diagonal_structure->rows.push_back(CompressedRow());
208 CompressedRow& row = block_diagonal_structure->rows.back();
209 row.block = diagonal_block;
210
211 row.cells.push_back(Cell());
212 Cell& cell = row.cells.back();
213 cell.block_id = c - start_col_block;
214 cell.position = diagonal_cell_position;
215
216 block_position += block.size;
217 diagonal_cell_position += block.size * block.size;
218 }
219
220 // Build a BlockSparseMatrix with the just computed block
221 // structure.
222 return new BlockSparseMatrix(block_diagonal_structure);
223 }
224
CreateBlockDiagonalEtE() const225 BlockSparseMatrix* PartitionedMatrixView::CreateBlockDiagonalEtE() const {
226 BlockSparseMatrix* block_diagonal =
227 CreateBlockDiagonalMatrixLayout(0, num_col_blocks_e_);
228 UpdateBlockDiagonalEtE(block_diagonal);
229 return block_diagonal;
230 }
231
CreateBlockDiagonalFtF() const232 BlockSparseMatrix* PartitionedMatrixView::CreateBlockDiagonalFtF() const {
233 BlockSparseMatrix* block_diagonal =
234 CreateBlockDiagonalMatrixLayout(
235 num_col_blocks_e_, num_col_blocks_e_ + num_col_blocks_f_);
236 UpdateBlockDiagonalFtF(block_diagonal);
237 return block_diagonal;
238 }
239
240 // Similar to the code in RightMultiplyE, except instead of the matrix
241 // vector multiply its an outer product.
242 //
243 // block_diagonal = block_diagonal(E'E)
UpdateBlockDiagonalEtE(BlockSparseMatrix * block_diagonal) const244 void PartitionedMatrixView::UpdateBlockDiagonalEtE(
245 BlockSparseMatrix* block_diagonal) const {
246 const CompressedRowBlockStructure* bs = matrix_.block_structure();
247 const CompressedRowBlockStructure* block_diagonal_structure =
248 block_diagonal->block_structure();
249
250 block_diagonal->SetZero();
251 const double* values = matrix_.values();
252 for (int r = 0; r < num_row_blocks_e_ ; ++r) {
253 const Cell& cell = bs->rows[r].cells[0];
254 const int row_block_size = bs->rows[r].block.size;
255 const int block_id = cell.block_id;
256 const int col_block_size = bs->cols[block_id].size;
257 const int cell_position =
258 block_diagonal_structure->rows[block_id].cells[0].position;
259
260 MatrixTransposeMatrixMultiply
261 <Eigen::Dynamic, Eigen::Dynamic, Eigen::Dynamic, Eigen::Dynamic, 1>(
262 values + cell.position, row_block_size, col_block_size,
263 values + cell.position, row_block_size, col_block_size,
264 block_diagonal->mutable_values() + cell_position,
265 0, 0, col_block_size, col_block_size);
266 }
267 }
268
269 // Similar to the code in RightMultiplyF, except instead of the matrix
270 // vector multiply its an outer product.
271 //
272 // block_diagonal = block_diagonal(F'F)
273 //
UpdateBlockDiagonalFtF(BlockSparseMatrix * block_diagonal) const274 void PartitionedMatrixView::UpdateBlockDiagonalFtF(
275 BlockSparseMatrix* block_diagonal) const {
276 const CompressedRowBlockStructure* bs = matrix_.block_structure();
277 const CompressedRowBlockStructure* block_diagonal_structure =
278 block_diagonal->block_structure();
279
280 block_diagonal->SetZero();
281 const double* values = matrix_.values();
282 for (int r = 0; r < bs->rows.size(); ++r) {
283 const int row_block_size = bs->rows[r].block.size;
284 const vector<Cell>& cells = bs->rows[r].cells;
285 for (int c = (r < num_row_blocks_e_) ? 1 : 0; c < cells.size(); ++c) {
286 const int col_block_id = cells[c].block_id;
287 const int col_block_size = bs->cols[col_block_id].size;
288 const int diagonal_block_id = col_block_id - num_col_blocks_e_;
289 const int cell_position =
290 block_diagonal_structure->rows[diagonal_block_id].cells[0].position;
291
292 MatrixTransposeMatrixMultiply
293 <Eigen::Dynamic, Eigen::Dynamic, Eigen::Dynamic, Eigen::Dynamic, 1>(
294 values + cells[c].position, row_block_size, col_block_size,
295 values + cells[c].position, row_block_size, col_block_size,
296 block_diagonal->mutable_values() + cell_position,
297 0, 0, col_block_size, col_block_size);
298 }
299 }
300 }
301
302 } // namespace internal
303 } // namespace ceres
304