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 #include "ceres/compressed_row_sparse_matrix.h"
32 
33 #include <algorithm>
34 #include <numeric>
35 #include <vector>
36 #include "ceres/crs_matrix.h"
37 #include "ceres/internal/port.h"
38 #include "ceres/triplet_sparse_matrix.h"
39 #include "glog/logging.h"
40 
41 namespace ceres {
42 namespace internal {
43 namespace {
44 
45 // Helper functor used by the constructor for reordering the contents
46 // of a TripletSparseMatrix. This comparator assumes thay there are no
47 // duplicates in the pair of arrays rows and cols, i.e., there is no
48 // indices i and j (not equal to each other) s.t.
49 //
50 //  rows[i] == rows[j] && cols[i] == cols[j]
51 //
52 // If this is the case, this functor will not be a StrictWeakOrdering.
53 struct RowColLessThan {
RowColLessThanceres::internal::__anon2c5187480111::RowColLessThan54   RowColLessThan(const int* rows, const int* cols)
55       : rows(rows), cols(cols) {
56   }
57 
operator ()ceres::internal::__anon2c5187480111::RowColLessThan58   bool operator()(const int x, const int y) const {
59     if (rows[x] == rows[y]) {
60       return (cols[x] < cols[y]);
61     }
62     return (rows[x] < rows[y]);
63   }
64 
65   const int* rows;
66   const int* cols;
67 };
68 
69 }  // namespace
70 
71 // This constructor gives you a semi-initialized CompressedRowSparseMatrix.
CompressedRowSparseMatrix(int num_rows,int num_cols,int max_num_nonzeros)72 CompressedRowSparseMatrix::CompressedRowSparseMatrix(int num_rows,
73                                                      int num_cols,
74                                                      int max_num_nonzeros) {
75   num_rows_ = num_rows;
76   num_cols_ = num_cols;
77   rows_.resize(num_rows + 1, 0);
78   cols_.resize(max_num_nonzeros, 0);
79   values_.resize(max_num_nonzeros, 0.0);
80 
81 
82   VLOG(1) << "# of rows: " << num_rows_
83           << " # of columns: " << num_cols_
84           << " max_num_nonzeros: " << cols_.size()
85           << ". Allocating " << (num_rows_ + 1) * sizeof(int) +  // NOLINT
86       cols_.size() * sizeof(int) +  // NOLINT
87       cols_.size() * sizeof(double);  // NOLINT
88 }
89 
CompressedRowSparseMatrix(const TripletSparseMatrix & m)90 CompressedRowSparseMatrix::CompressedRowSparseMatrix(
91     const TripletSparseMatrix& m) {
92   num_rows_ = m.num_rows();
93   num_cols_ = m.num_cols();
94 
95   rows_.resize(num_rows_ + 1, 0);
96   cols_.resize(m.num_nonzeros(), 0);
97   values_.resize(m.max_num_nonzeros(), 0.0);
98 
99   // index is the list of indices into the TripletSparseMatrix m.
100   vector<int> index(m.num_nonzeros(), 0);
101   for (int i = 0; i < m.num_nonzeros(); ++i) {
102     index[i] = i;
103   }
104 
105   // Sort index such that the entries of m are ordered by row and ties
106   // are broken by column.
107   sort(index.begin(), index.end(), RowColLessThan(m.rows(), m.cols()));
108 
109   VLOG(1) << "# of rows: " << num_rows_
110           << " # of columns: " << num_cols_
111           << " max_num_nonzeros: " << cols_.size()
112           << ". Allocating "
113           << ((num_rows_ + 1) * sizeof(int) +  // NOLINT
114               cols_.size() * sizeof(int) +     // NOLINT
115               cols_.size() * sizeof(double));  // NOLINT
116 
117   // Copy the contents of the cols and values array in the order given
118   // by index and count the number of entries in each row.
119   for (int i = 0; i < m.num_nonzeros(); ++i) {
120     const int idx = index[i];
121     ++rows_[m.rows()[idx] + 1];
122     cols_[i] = m.cols()[idx];
123     values_[i] = m.values()[idx];
124   }
125 
126   // Find the cumulative sum of the row counts.
127   for (int i = 1; i < num_rows_ + 1; ++i) {
128     rows_[i] += rows_[i - 1];
129   }
130 
131   CHECK_EQ(num_nonzeros(), m.num_nonzeros());
132 }
133 
CompressedRowSparseMatrix(const double * diagonal,int num_rows)134 CompressedRowSparseMatrix::CompressedRowSparseMatrix(const double* diagonal,
135                                                      int num_rows) {
136   CHECK_NOTNULL(diagonal);
137 
138   num_rows_ = num_rows;
139   num_cols_ = num_rows;
140   rows_.resize(num_rows + 1);
141   cols_.resize(num_rows);
142   values_.resize(num_rows);
143 
144   rows_[0] = 0;
145   for (int i = 0; i < num_rows_; ++i) {
146     cols_[i] = i;
147     values_[i] = diagonal[i];
148     rows_[i + 1] = i + 1;
149   }
150 
151   CHECK_EQ(num_nonzeros(), num_rows);
152 }
153 
~CompressedRowSparseMatrix()154 CompressedRowSparseMatrix::~CompressedRowSparseMatrix() {
155 }
156 
SetZero()157 void CompressedRowSparseMatrix::SetZero() {
158   fill(values_.begin(), values_.end(), 0);
159 }
160 
RightMultiply(const double * x,double * y) const161 void CompressedRowSparseMatrix::RightMultiply(const double* x,
162                                               double* y) const {
163   CHECK_NOTNULL(x);
164   CHECK_NOTNULL(y);
165 
166   for (int r = 0; r < num_rows_; ++r) {
167     for (int idx = rows_[r]; idx < rows_[r + 1]; ++idx) {
168       y[r] += values_[idx] * x[cols_[idx]];
169     }
170   }
171 }
172 
LeftMultiply(const double * x,double * y) const173 void CompressedRowSparseMatrix::LeftMultiply(const double* x, double* y) const {
174   CHECK_NOTNULL(x);
175   CHECK_NOTNULL(y);
176 
177   for (int r = 0; r < num_rows_; ++r) {
178     for (int idx = rows_[r]; idx < rows_[r + 1]; ++idx) {
179       y[cols_[idx]] += values_[idx] * x[r];
180     }
181   }
182 }
183 
SquaredColumnNorm(double * x) const184 void CompressedRowSparseMatrix::SquaredColumnNorm(double* x) const {
185   CHECK_NOTNULL(x);
186 
187   fill(x, x + num_cols_, 0.0);
188   for (int idx = 0; idx < rows_[num_rows_]; ++idx) {
189     x[cols_[idx]] += values_[idx] * values_[idx];
190   }
191 }
192 
ScaleColumns(const double * scale)193 void CompressedRowSparseMatrix::ScaleColumns(const double* scale) {
194   CHECK_NOTNULL(scale);
195 
196   for (int idx = 0; idx < rows_[num_rows_]; ++idx) {
197     values_[idx] *= scale[cols_[idx]];
198   }
199 }
200 
ToDenseMatrix(Matrix * dense_matrix) const201 void CompressedRowSparseMatrix::ToDenseMatrix(Matrix* dense_matrix) const {
202   CHECK_NOTNULL(dense_matrix);
203   dense_matrix->resize(num_rows_, num_cols_);
204   dense_matrix->setZero();
205 
206   for (int r = 0; r < num_rows_; ++r) {
207     for (int idx = rows_[r]; idx < rows_[r + 1]; ++idx) {
208       (*dense_matrix)(r, cols_[idx]) = values_[idx];
209     }
210   }
211 }
212 
DeleteRows(int delta_rows)213 void CompressedRowSparseMatrix::DeleteRows(int delta_rows) {
214   CHECK_GE(delta_rows, 0);
215   CHECK_LE(delta_rows, num_rows_);
216 
217   num_rows_ -= delta_rows;
218   rows_.resize(num_rows_ + 1);
219 
220   // Walk the list of row blocks until we reach the new number of rows
221   // and the drop the rest of the row blocks.
222   int num_row_blocks = 0;
223   int num_rows = 0;
224   while (num_row_blocks < row_blocks_.size() && num_rows < num_rows_) {
225     num_rows += row_blocks_[num_row_blocks];
226     ++num_row_blocks;
227   }
228 
229   row_blocks_.resize(num_row_blocks);
230 }
231 
AppendRows(const CompressedRowSparseMatrix & m)232 void CompressedRowSparseMatrix::AppendRows(const CompressedRowSparseMatrix& m) {
233   CHECK_EQ(m.num_cols(), num_cols_);
234 
235   CHECK(row_blocks_.size() == 0 || m.row_blocks().size() !=0)
236       << "Cannot append a matrix with row blocks to one without and vice versa."
237       << "This matrix has : " << row_blocks_.size() << " row blocks."
238       << "The matrix being appended has: " << m.row_blocks().size()
239       << " row blocks.";
240 
241   if (cols_.size() < num_nonzeros() + m.num_nonzeros()) {
242     cols_.resize(num_nonzeros() + m.num_nonzeros());
243     values_.resize(num_nonzeros() + m.num_nonzeros());
244   }
245 
246   // Copy the contents of m into this matrix.
247   copy(m.cols(), m.cols() + m.num_nonzeros(), &cols_[num_nonzeros()]);
248   copy(m.values(), m.values() + m.num_nonzeros(), &values_[num_nonzeros()]);
249   rows_.resize(num_rows_ + m.num_rows() + 1);
250   // new_rows = [rows_, m.row() + rows_[num_rows_]]
251   fill(rows_.begin() + num_rows_,
252        rows_.begin() + num_rows_ + m.num_rows() + 1,
253        rows_[num_rows_]);
254 
255   for (int r = 0; r < m.num_rows() + 1; ++r) {
256     rows_[num_rows_ + r] += m.rows()[r];
257   }
258 
259   num_rows_ += m.num_rows();
260   row_blocks_.insert(row_blocks_.end(), m.row_blocks().begin(), m.row_blocks().end());
261 }
262 
ToTextFile(FILE * file) const263 void CompressedRowSparseMatrix::ToTextFile(FILE* file) const {
264   CHECK_NOTNULL(file);
265   for (int r = 0; r < num_rows_; ++r) {
266     for (int idx = rows_[r]; idx < rows_[r + 1]; ++idx) {
267       fprintf(file,
268               "% 10d % 10d %17f\n",
269               r,
270               cols_[idx],
271               values_[idx]);
272     }
273   }
274 }
275 
ToCRSMatrix(CRSMatrix * matrix) const276 void CompressedRowSparseMatrix::ToCRSMatrix(CRSMatrix* matrix) const {
277   matrix->num_rows = num_rows_;
278   matrix->num_cols = num_cols_;
279   matrix->rows = rows_;
280   matrix->cols = cols_;
281   matrix->values = values_;
282 
283   // Trim.
284   matrix->rows.resize(matrix->num_rows + 1);
285   matrix->cols.resize(matrix->rows[matrix->num_rows]);
286   matrix->values.resize(matrix->rows[matrix->num_rows]);
287 }
288 
SetMaxNumNonZeros(int num_nonzeros)289 void CompressedRowSparseMatrix::SetMaxNumNonZeros(int num_nonzeros) {
290   CHECK_GE(num_nonzeros, 0);
291 
292   cols_.resize(num_nonzeros);
293   values_.resize(num_nonzeros);
294 }
295 
SolveLowerTriangularInPlace(double * solution) const296 void CompressedRowSparseMatrix::SolveLowerTriangularInPlace(
297     double* solution) const {
298   for (int r = 0; r < num_rows_; ++r) {
299     for (int idx = rows_[r]; idx < rows_[r + 1] - 1; ++idx) {
300       solution[r] -= values_[idx] * solution[cols_[idx]];
301     }
302     solution[r] /=  values_[rows_[r + 1] - 1];
303   }
304 }
305 
SolveLowerTriangularTransposeInPlace(double * solution) const306 void CompressedRowSparseMatrix::SolveLowerTriangularTransposeInPlace(
307     double* solution) const {
308   for (int r = num_rows_ - 1; r >= 0; --r) {
309     solution[r] /= values_[rows_[r + 1] - 1];
310     for (int idx = rows_[r + 1] - 2; idx >= rows_[r]; --idx) {
311       solution[cols_[idx]] -= values_[idx] * solution[r];
312     }
313   }
314 }
315 
CreateBlockDiagonalMatrix(const double * diagonal,const vector<int> & blocks)316 CompressedRowSparseMatrix* CompressedRowSparseMatrix::CreateBlockDiagonalMatrix(
317     const double* diagonal,
318     const vector<int>& blocks) {
319   int num_rows = 0;
320   int num_nonzeros = 0;
321   for (int i = 0; i < blocks.size(); ++i) {
322     num_rows += blocks[i];
323     num_nonzeros += blocks[i] * blocks[i];
324   }
325 
326   CompressedRowSparseMatrix* matrix =
327       new CompressedRowSparseMatrix(num_rows, num_rows, num_nonzeros);
328 
329   int* rows = matrix->mutable_rows();
330   int* cols = matrix->mutable_cols();
331   double* values = matrix->mutable_values();
332   fill(values, values + num_nonzeros, 0.0);
333 
334   int idx_cursor = 0;
335   int col_cursor = 0;
336   for (int i = 0; i < blocks.size(); ++i) {
337     const int block_size = blocks[i];
338     for (int r = 0; r < block_size; ++r) {
339       *(rows++) = idx_cursor;
340       values[idx_cursor + r] = diagonal[col_cursor + r];
341       for (int c = 0; c < block_size; ++c, ++idx_cursor) {
342         *(cols++) = col_cursor + c;
343       }
344     }
345     col_cursor += block_size;
346   }
347   *rows = idx_cursor;
348 
349   *matrix->mutable_row_blocks() = blocks;
350   *matrix->mutable_col_blocks() = blocks;
351 
352   CHECK_EQ(idx_cursor, num_nonzeros);
353   CHECK_EQ(col_cursor, num_rows);
354   return matrix;
355 }
356 
Transpose() const357 CompressedRowSparseMatrix* CompressedRowSparseMatrix::Transpose() const {
358   CompressedRowSparseMatrix* transpose =
359       new CompressedRowSparseMatrix(num_cols_, num_rows_, num_nonzeros());
360 
361   int* transpose_rows = transpose->mutable_rows();
362   int* transpose_cols = transpose->mutable_cols();
363   double* transpose_values = transpose->mutable_values();
364 
365   for (int idx = 0; idx < num_nonzeros(); ++idx) {
366     ++transpose_rows[cols_[idx] + 1];
367   }
368 
369   for (int i = 1; i < transpose->num_rows() + 1; ++i) {
370     transpose_rows[i] += transpose_rows[i - 1];
371   }
372 
373   for (int r = 0; r < num_rows(); ++r) {
374     for (int idx = rows_[r]; idx < rows_[r + 1]; ++idx) {
375       const int c = cols_[idx];
376       const int transpose_idx = transpose_rows[c]++;
377       transpose_cols[transpose_idx] = r;
378       transpose_values[transpose_idx] = values_[idx];
379     }
380   }
381 
382   for (int i = transpose->num_rows() - 1; i > 0 ; --i) {
383     transpose_rows[i] = transpose_rows[i - 1];
384   }
385   transpose_rows[0] = 0;
386 
387   *(transpose->mutable_row_blocks()) = col_blocks_;
388   *(transpose->mutable_col_blocks()) = row_blocks_;
389 
390   return transpose;
391 }
392 
393 namespace {
394 // A ProductTerm is a term in the outer product of a matrix with
395 // itself.
396 struct ProductTerm {
ProductTermceres::internal::__anon2c5187480211::ProductTerm397   ProductTerm(const int row, const int col, const int index)
398       : row(row), col(col), index(index) {
399   }
400 
operator <ceres::internal::__anon2c5187480211::ProductTerm401   bool operator<(const ProductTerm& right) const {
402     if (row == right.row) {
403       if (col == right.col) {
404         return index < right.index;
405       }
406       return col < right.col;
407     }
408     return row < right.row;
409   }
410 
411   int row;
412   int col;
413   int index;
414 };
415 
416 CompressedRowSparseMatrix*
CompressAndFillProgram(const int num_rows,const int num_cols,const vector<ProductTerm> & product,vector<int> * program)417 CompressAndFillProgram(const int num_rows,
418                        const int num_cols,
419                        const vector<ProductTerm>& product,
420                        vector<int>* program) {
421   CHECK_GT(product.size(), 0);
422 
423   // Count the number of unique product term, which in turn is the
424   // number of non-zeros in the outer product.
425   int num_nonzeros = 1;
426   for (int i = 1; i < product.size(); ++i) {
427     if (product[i].row != product[i - 1].row ||
428         product[i].col != product[i - 1].col) {
429       ++num_nonzeros;
430     }
431   }
432 
433   CompressedRowSparseMatrix* matrix =
434       new CompressedRowSparseMatrix(num_rows, num_cols, num_nonzeros);
435 
436   int* crsm_rows = matrix->mutable_rows();
437   std::fill(crsm_rows, crsm_rows + num_rows + 1, 0);
438   int* crsm_cols = matrix->mutable_cols();
439   std::fill(crsm_cols, crsm_cols + num_nonzeros, 0);
440 
441   CHECK_NOTNULL(program)->clear();
442   program->resize(product.size());
443 
444   // Iterate over the sorted product terms. This means each row is
445   // filled one at a time, and we are able to assign a position in the
446   // values array to each term.
447   //
448   // If terms repeat, i.e., they contribute to the same entry in the
449   // result matrix), then they do not affect the sparsity structure of
450   // the result matrix.
451   int nnz = 0;
452   crsm_cols[0] = product[0].col;
453   crsm_rows[product[0].row + 1]++;
454   (*program)[product[0].index] = nnz;
455   for (int i = 1; i < product.size(); ++i) {
456     const ProductTerm& previous = product[i - 1];
457     const ProductTerm& current = product[i];
458 
459     // Sparsity structure is updated only if the term is not a repeat.
460     if (previous.row != current.row || previous.col != current.col) {
461       crsm_cols[++nnz] = current.col;
462       crsm_rows[current.row + 1]++;
463     }
464 
465     // All terms get assigned the position in the values array where
466     // their value is accumulated.
467     (*program)[current.index] = nnz;
468   }
469 
470   for (int i = 1; i < num_rows + 1; ++i) {
471     crsm_rows[i] += crsm_rows[i - 1];
472   }
473 
474   return matrix;
475 }
476 
477 }  // namespace
478 
479 CompressedRowSparseMatrix*
CreateOuterProductMatrixAndProgram(const CompressedRowSparseMatrix & m,vector<int> * program)480 CompressedRowSparseMatrix::CreateOuterProductMatrixAndProgram(
481       const CompressedRowSparseMatrix& m,
482       vector<int>* program) {
483   CHECK_NOTNULL(program)->clear();
484   CHECK_GT(m.num_nonzeros(), 0) << "Congratulations, "
485                                 << "you found a bug in Ceres. Please report it.";
486 
487   vector<ProductTerm> product;
488   const vector<int>& row_blocks = m.row_blocks();
489   int row_block_begin = 0;
490   // Iterate over row blocks
491   for (int row_block = 0; row_block < row_blocks.size(); ++row_block) {
492     const int row_block_end = row_block_begin + row_blocks[row_block];
493     // Compute the outer product terms for just one row per row block.
494     const int r = row_block_begin;
495     // Compute the lower triangular part of the product.
496     for (int idx1 = m.rows()[r]; idx1 < m.rows()[r + 1]; ++idx1) {
497       for (int idx2 = m.rows()[r]; idx2 <= idx1; ++idx2) {
498         product.push_back(ProductTerm(m.cols()[idx1], m.cols()[idx2], product.size()));
499       }
500     }
501     row_block_begin = row_block_end;
502   }
503   CHECK_EQ(row_block_begin, m.num_rows());
504   sort(product.begin(), product.end());
505   return CompressAndFillProgram(m.num_cols(), m.num_cols(), product, program);
506 }
507 
ComputeOuterProduct(const CompressedRowSparseMatrix & m,const vector<int> & program,CompressedRowSparseMatrix * result)508 void CompressedRowSparseMatrix::ComputeOuterProduct(
509     const CompressedRowSparseMatrix& m,
510     const vector<int>& program,
511     CompressedRowSparseMatrix* result) {
512   result->SetZero();
513   double* values = result->mutable_values();
514   const vector<int>& row_blocks = m.row_blocks();
515 
516   int cursor = 0;
517   int row_block_begin = 0;
518   const double* m_values = m.values();
519   const int* m_rows = m.rows();
520   // Iterate over row blocks.
521   for (int row_block = 0; row_block < row_blocks.size(); ++row_block) {
522     const int row_block_end = row_block_begin + row_blocks[row_block];
523     const int saved_cursor = cursor;
524     for (int r = row_block_begin; r < row_block_end; ++r) {
525       // Reuse the program segment for each row in this row block.
526       cursor = saved_cursor;
527       const int row_begin = m_rows[r];
528       const int row_end = m_rows[r + 1];
529       for (int idx1 = row_begin; idx1 < row_end; ++idx1) {
530         const double v1 =  m_values[idx1];
531         for (int idx2 = row_begin; idx2 <= idx1; ++idx2, ++cursor) {
532           values[program[cursor]] += v1 * m_values[idx2];
533         }
534       }
535     }
536     row_block_begin = row_block_end;
537   }
538 
539   CHECK_EQ(row_block_begin, m.num_rows());
540   CHECK_EQ(cursor, program.size());
541 }
542 
543 }  // namespace internal
544 }  // namespace ceres
545