• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 // TODO(sameeragarwal): row_block_counter can perhaps be replaced by
32 // Chunk::start ?
33 
34 #ifndef CERES_INTERNAL_SCHUR_ELIMINATOR_IMPL_H_
35 #define CERES_INTERNAL_SCHUR_ELIMINATOR_IMPL_H_
36 
37 #ifdef CERES_USE_OPENMP
38 #include <omp.h>
39 #endif
40 
41 // Eigen has an internal threshold switching between different matrix
42 // multiplication algorithms. In particular for matrices larger than
43 // EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD it uses a cache friendly
44 // matrix matrix product algorithm that has a higher setup cost. For
45 // matrix sizes close to this threshold, especially when the matrices
46 // are thin and long, the default choice may not be optimal. This is
47 // the case for us, as the default choice causes a 30% performance
48 // regression when we moved from Eigen2 to Eigen3.
49 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 10
50 
51 #include <algorithm>
52 #include <map>
53 #include <glog/logging.h>
54 #include "Eigen/Dense"
55 #include "ceres/block_random_access_matrix.h"
56 #include "ceres/block_sparse_matrix.h"
57 #include "ceres/block_structure.h"
58 #include "ceres/map_util.h"
59 #include "ceres/schur_eliminator.h"
60 #include "ceres/stl_util.h"
61 #include "ceres/internal/eigen.h"
62 #include "ceres/internal/scoped_ptr.h"
63 
64 namespace ceres {
65 namespace internal {
66 
67 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
~SchurEliminator()68 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::~SchurEliminator() {
69   STLDeleteElements(&rhs_locks_);
70 }
71 
72 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
73 void
74 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
Init(int num_eliminate_blocks,const CompressedRowBlockStructure * bs)75 Init(int num_eliminate_blocks, const CompressedRowBlockStructure* bs) {
76   CHECK_GT(num_eliminate_blocks, 0)
77       << "SchurComplementSolver cannot be initialized with "
78       << "num_eliminate_blocks = 0.";
79 
80   num_eliminate_blocks_ = num_eliminate_blocks;
81 
82   const int num_col_blocks = bs->cols.size();
83   const int num_row_blocks = bs->rows.size();
84 
85   buffer_size_ = 1;
86   chunks_.clear();
87   lhs_row_layout_.clear();
88 
89   int lhs_num_rows = 0;
90   // Add a map object for each block in the reduced linear system
91   // and build the row/column block structure of the reduced linear
92   // system.
93   lhs_row_layout_.resize(num_col_blocks - num_eliminate_blocks_);
94   for (int i = num_eliminate_blocks_; i < num_col_blocks; ++i) {
95     lhs_row_layout_[i - num_eliminate_blocks_] = lhs_num_rows;
96     lhs_num_rows += bs->cols[i].size;
97   }
98 
99   int r = 0;
100   // Iterate over the row blocks of A, and detect the chunks. The
101   // matrix should already have been ordered so that all rows
102   // containing the same y block are vertically contiguous. Along
103   // the way also compute the amount of space each chunk will need
104   // to perform the elimination.
105   while (r < num_row_blocks) {
106     const int chunk_block_id = bs->rows[r].cells.front().block_id;
107     if (chunk_block_id >= num_eliminate_blocks_) {
108       break;
109     }
110 
111     chunks_.push_back(Chunk());
112     Chunk& chunk = chunks_.back();
113     chunk.size = 0;
114     chunk.start = r;
115     int buffer_size = 0;
116     const int e_block_size = bs->cols[chunk_block_id].size;
117 
118     // Add to the chunk until the first block in the row is
119     // different than the one in the first row for the chunk.
120     while (r + chunk.size < num_row_blocks) {
121       const CompressedRow& row = bs->rows[r + chunk.size];
122       if (row.cells.front().block_id != chunk_block_id) {
123         break;
124       }
125 
126       // Iterate over the blocks in the row, ignoring the first
127       // block since it is the one to be eliminated.
128       for (int c = 1; c < row.cells.size(); ++c) {
129         const Cell& cell = row.cells[c];
130         if (InsertIfNotPresent(
131                 &(chunk.buffer_layout), cell.block_id, buffer_size)) {
132           buffer_size += e_block_size * bs->cols[cell.block_id].size;
133         }
134       }
135 
136       buffer_size_ = max(buffer_size, buffer_size_);
137       ++chunk.size;
138     }
139 
140     CHECK_GT(chunk.size, 0);
141     r += chunk.size;
142   }
143   const Chunk& chunk = chunks_.back();
144 
145   uneliminated_row_begins_ = chunk.start + chunk.size;
146   if (num_threads_ > 1) {
147     random_shuffle(chunks_.begin(), chunks_.end());
148   }
149 
150   buffer_.reset(new double[buffer_size_ * num_threads_]);
151 
152   STLDeleteElements(&rhs_locks_);
153   rhs_locks_.resize(num_col_blocks - num_eliminate_blocks_);
154   for (int i = 0; i < num_col_blocks - num_eliminate_blocks_; ++i) {
155     rhs_locks_[i] = new Mutex;
156   }
157 
158   VLOG(1) << "Eliminator threads: " << num_threads_;
159 }
160 
161 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
162 void
163 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
Eliminate(const BlockSparseMatrixBase * A,const double * b,const double * D,BlockRandomAccessMatrix * lhs,double * rhs)164 Eliminate(const BlockSparseMatrixBase* A,
165           const double* b,
166           const double* D,
167           BlockRandomAccessMatrix* lhs,
168           double* rhs) {
169   if (lhs->num_rows() > 0) {
170     lhs->SetZero();
171     VectorRef(rhs, lhs->num_rows()).setZero();
172   }
173 
174   const CompressedRowBlockStructure* bs = A->block_structure();
175   const int num_col_blocks = bs->cols.size();
176 
177   // Add the diagonal to the schur complement.
178   if (D != NULL) {
179 #pragma omp parallel for num_threads(num_threads_) schedule(dynamic)
180     for (int i = num_eliminate_blocks_; i < num_col_blocks; ++i) {
181       const int block_id = i - num_eliminate_blocks_;
182       int r, c, row_stride, col_stride;
183       CellInfo* cell_info = lhs->GetCell(block_id, block_id,
184                                          &r, &c,
185                                          &row_stride, &col_stride);
186       if (cell_info != NULL) {
187         const int block_size = bs->cols[i].size;
188         typename EigenTypes<kFBlockSize>::ConstVectorRef
189             diag(D + bs->cols[i].position, block_size);
190 
191         CeresMutexLock l(&cell_info->m);
192         MatrixRef m(cell_info->values, row_stride, col_stride);
193         m.block(r, c, block_size, block_size).diagonal()
194             += diag.array().square().matrix();
195       }
196     }
197   }
198 
199   // Eliminate y blocks one chunk at a time.  For each chunk,x3
200   // compute the entries of the normal equations and the gradient
201   // vector block corresponding to the y block and then apply
202   // Gaussian elimination to them. The matrix ete stores the normal
203   // matrix corresponding to the block being eliminated and array
204   // buffer_ contains the non-zero blocks in the row corresponding
205   // to this y block in the normal equations. This computation is
206   // done in ChunkDiagonalBlockAndGradient. UpdateRhs then applies
207   // gaussian elimination to the rhs of the normal equations,
208   // updating the rhs of the reduced linear system by modifying rhs
209   // blocks for all the z blocks that share a row block/residual
210   // term with the y block. EliminateRowOuterProduct does the
211   // corresponding operation for the lhs of the reduced linear
212   // system.
213 #pragma omp parallel for num_threads(num_threads_) schedule(dynamic)
214   for (int i = 0; i < chunks_.size(); ++i) {
215 #ifdef CERES_USE_OPENMP
216     int thread_id = omp_get_thread_num();
217 #else
218     int thread_id = 0;
219 #endif
220     double* buffer = buffer_.get() + thread_id * buffer_size_;
221     const Chunk& chunk = chunks_[i];
222     const int e_block_id = bs->rows[chunk.start].cells.front().block_id;
223     const int e_block_size = bs->cols[e_block_id].size;
224 
225     VectorRef(buffer, buffer_size_).setZero();
226 
227     typename EigenTypes<kEBlockSize, kEBlockSize>::Matrix
228         ete(e_block_size, e_block_size);
229 
230     if (D != NULL) {
231       const typename EigenTypes<kEBlockSize>::ConstVectorRef
232           diag(D + bs->cols[e_block_id].position, e_block_size);
233       ete = diag.array().square().matrix().asDiagonal();
234     } else {
235       ete.setZero();
236     }
237 
238     typename EigenTypes<kEBlockSize>::Vector g(e_block_size);
239     g.setZero();
240 
241     // We are going to be computing
242     //
243     //   S += F'F - F'E(E'E)^{-1}E'F
244     //
245     // for each Chunk. The computation is broken down into a number of
246     // function calls as below.
247 
248     // Compute the outer product of the e_blocks with themselves (ete
249     // = E'E). Compute the product of the e_blocks with the
250     // corresonding f_blocks (buffer = E'F), the gradient of the terms
251     // in this chunk (g) and add the outer product of the f_blocks to
252     // Schur complement (S += F'F).
253     ChunkDiagonalBlockAndGradient(
254         chunk, A, b, chunk.start, &ete, &g, buffer, lhs);
255 
256     // Normally one wouldn't compute the inverse explicitly, but
257     // e_block_size will typically be a small number like 3, in
258     // which case its much faster to compute the inverse once and
259     // use it to multiply other matrices/vectors instead of doing a
260     // Solve call over and over again.
261     typename EigenTypes<kEBlockSize, kEBlockSize>::Matrix inverse_ete =
262         ete
263         .template selfadjointView<Eigen::Upper>()
264         .ldlt()
265         .solve(Matrix::Identity(e_block_size, e_block_size));
266 
267     // For the current chunk compute and update the rhs of the reduced
268     // linear system.
269     //
270     //   rhs = F'b - F'E(E'E)^(-1) E'b
271     UpdateRhs(chunk, A, b, chunk.start, inverse_ete * g, rhs);
272 
273     // S -= F'E(E'E)^{-1}E'F
274     ChunkOuterProduct(bs, inverse_ete, buffer, chunk.buffer_layout, lhs);
275   }
276 
277   // For rows with no e_blocks, the schur complement update reduces to
278   // S += F'F.
279   NoEBlockRowsUpdate(A, b,  uneliminated_row_begins_, lhs, rhs);
280 }
281 
282 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
283 void
284 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
BackSubstitute(const BlockSparseMatrixBase * A,const double * b,const double * D,const double * z,double * y)285 BackSubstitute(const BlockSparseMatrixBase* A,
286                const double* b,
287                const double* D,
288                const double* z,
289                double* y) {
290   const CompressedRowBlockStructure* bs = A->block_structure();
291 #pragma omp parallel for num_threads(num_threads_) schedule(dynamic)
292   for (int i = 0; i < chunks_.size(); ++i) {
293     const Chunk& chunk = chunks_[i];
294     const int e_block_id = bs->rows[chunk.start].cells.front().block_id;
295     const int e_block_size = bs->cols[e_block_id].size;
296 
297     typename EigenTypes<kEBlockSize>::VectorRef y_block(
298         y +  bs->cols[e_block_id].position, e_block_size);
299 
300     typename EigenTypes<kEBlockSize, kEBlockSize>::Matrix
301         ete(e_block_size, e_block_size);
302     if (D != NULL) {
303       const typename EigenTypes<kEBlockSize>::ConstVectorRef
304           diag(D + bs->cols[e_block_id].position, e_block_size);
305       ete = diag.array().square().matrix().asDiagonal();
306     } else {
307       ete.setZero();
308     }
309 
310     for (int j = 0; j < chunk.size; ++j) {
311       const CompressedRow& row = bs->rows[chunk.start + j];
312       const double* row_values = A->RowBlockValues(chunk.start + j);
313       const Cell& e_cell = row.cells.front();
314       DCHECK_EQ(e_block_id, e_cell.block_id);
315       const typename EigenTypes<kRowBlockSize, kEBlockSize>::ConstMatrixRef
316           e_block(row_values + e_cell.position,
317                   row.block.size,
318                   e_block_size);
319 
320       typename EigenTypes<kRowBlockSize>::Vector
321           sj =
322           typename EigenTypes<kRowBlockSize>::ConstVectorRef
323           (b + bs->rows[chunk.start + j].block.position,
324            row.block.size);
325 
326       for (int c = 1; c < row.cells.size(); ++c) {
327         const int f_block_id = row.cells[c].block_id;
328         const int f_block_size = bs->cols[f_block_id].size;
329         const typename EigenTypes<kRowBlockSize, kFBlockSize>::ConstMatrixRef
330             f_block(row_values + row.cells[c].position,
331                     row.block.size, f_block_size);
332         const int r_block = f_block_id - num_eliminate_blocks_;
333 
334         sj -= f_block *
335             typename EigenTypes<kFBlockSize>::ConstVectorRef
336             (z + lhs_row_layout_[r_block], f_block_size);
337       }
338 
339       y_block += e_block.transpose() * sj;
340       ete.template selfadjointView<Eigen::Upper>()
341           .rankUpdate(e_block.transpose(), 1.0);
342     }
343 
344     y_block =
345         ete
346         .template selfadjointView<Eigen::Upper>()
347         .ldlt()
348         .solve(y_block);
349   }
350 }
351 
352 // Update the rhs of the reduced linear system. Compute
353 //
354 //   F'b - F'E(E'E)^(-1) E'b
355 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
356 void
357 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
UpdateRhs(const Chunk & chunk,const BlockSparseMatrixBase * A,const double * b,int row_block_counter,const Vector & inverse_ete_g,double * rhs)358 UpdateRhs(const Chunk& chunk,
359           const BlockSparseMatrixBase* A,
360           const double* b,
361           int row_block_counter,
362           const Vector& inverse_ete_g,
363           double* rhs) {
364   const CompressedRowBlockStructure* bs = A->block_structure();
365   const int e_block_size = inverse_ete_g.rows();
366   int b_pos = bs->rows[row_block_counter].block.position;
367   for (int j = 0; j < chunk.size; ++j) {
368     const CompressedRow& row = bs->rows[row_block_counter + j];
369     const double *row_values = A->RowBlockValues(row_block_counter + j);
370     const Cell& e_cell = row.cells.front();
371 
372     const typename EigenTypes<kRowBlockSize, kEBlockSize>::ConstMatrixRef
373         e_block(row_values + e_cell.position,
374                 row.block.size,
375                 e_block_size);
376 
377     const typename EigenTypes<kRowBlockSize>::Vector
378         sj =
379         typename EigenTypes<kRowBlockSize>::ConstVectorRef
380          (b + b_pos, row.block.size) - e_block * (inverse_ete_g);
381 
382     for (int c = 1; c < row.cells.size(); ++c) {
383       const int block_id = row.cells[c].block_id;
384       const int block_size = bs->cols[block_id].size;
385       const typename EigenTypes<kRowBlockSize, kFBlockSize>::ConstMatrixRef
386           b(row_values + row.cells[c].position,
387             row.block.size, block_size);
388 
389       const int block = block_id - num_eliminate_blocks_;
390       CeresMutexLock l(rhs_locks_[block]);
391       typename EigenTypes<kFBlockSize>::VectorRef
392           (rhs + lhs_row_layout_[block], block_size).noalias()
393           += b.transpose() * sj;
394     }
395     b_pos += row.block.size;
396   }
397 }
398 
399 // Given a Chunk - set of rows with the same e_block, e.g. in the
400 // following Chunk with two rows.
401 //
402 //                E                   F
403 //      [ y11   0   0   0 |  z11     0    0   0    z51]
404 //      [ y12   0   0   0 |  z12   z22    0   0      0]
405 //
406 // this function computes twp matrices. The diagonal block matrix
407 //
408 //   ete = y11 * y11' + y12 * y12'
409 //
410 // and the off diagonal blocks in the Guass Newton Hessian.
411 //
412 //   buffer = [y11'(z11 + z12), y12' * z22, y11' * z51]
413 //
414 // which are zero compressed versions of the block sparse matrices E'E
415 // and E'F.
416 //
417 // and the gradient of the e_block, E'b.
418 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
419 void
420 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
ChunkDiagonalBlockAndGradient(const Chunk & chunk,const BlockSparseMatrixBase * A,const double * b,int row_block_counter,typename EigenTypes<kEBlockSize,kEBlockSize>::Matrix * ete,typename EigenTypes<kEBlockSize>::Vector * g,double * buffer,BlockRandomAccessMatrix * lhs)421 ChunkDiagonalBlockAndGradient(
422     const Chunk& chunk,
423     const BlockSparseMatrixBase* A,
424     const double* b,
425     int row_block_counter,
426     typename EigenTypes<kEBlockSize, kEBlockSize>::Matrix* ete,
427     typename EigenTypes<kEBlockSize>::Vector* g,
428     double* buffer,
429     BlockRandomAccessMatrix* lhs) {
430   const CompressedRowBlockStructure* bs = A->block_structure();
431 
432   int b_pos = bs->rows[row_block_counter].block.position;
433   const int e_block_size = ete->rows();
434 
435   // Iterate over the rows in this chunk, for each row, compute the
436   // contribution of its F blocks to the Schur complement, the
437   // contribution of its E block to the matrix EE' (ete), and the
438   // corresponding block in the gradient vector.
439   for (int j = 0; j < chunk.size; ++j) {
440     const CompressedRow& row = bs->rows[row_block_counter + j];
441     const double *row_values = A->RowBlockValues(row_block_counter + j);
442 
443     if (row.cells.size() > 1) {
444       EBlockRowOuterProduct(A, row_block_counter + j, lhs);
445     }
446 
447     // Extract the e_block, ETE += E_i' E_i
448     const Cell& e_cell = row.cells.front();
449     const typename EigenTypes<kRowBlockSize, kEBlockSize>::ConstMatrixRef
450         e_block(row_values + e_cell.position,
451                 row.block.size,
452                 e_block_size);
453 
454     ete->template selfadjointView<Eigen::Upper>()
455         .rankUpdate(e_block.transpose(), 1.0);
456 
457     // g += E_i' b_i
458     g->noalias() += e_block.transpose() *
459         typename EigenTypes<kRowBlockSize>::ConstVectorRef
460         (b + b_pos, row.block.size);
461 
462     // buffer = E'F. This computation is done by iterating over the
463     // f_blocks for each row in the chunk.
464     for (int c = 1; c < row.cells.size(); ++c) {
465       const int f_block_id = row.cells[c].block_id;
466       const int f_block_size = bs->cols[f_block_id].size;
467       const typename EigenTypes<kRowBlockSize, kFBlockSize>::ConstMatrixRef
468           f_block(row_values + row.cells[c].position,
469                   row.block.size, f_block_size);
470 
471       double* buffer_ptr =
472           buffer +  FindOrDie(chunk.buffer_layout, f_block_id);
473 
474       typename EigenTypes<kEBlockSize, kFBlockSize>::MatrixRef
475           (buffer_ptr,  e_block_size, f_block_size).noalias()
476           += e_block.transpose() * f_block;
477     }
478     b_pos += row.block.size;
479   }
480 }
481 
482 // Compute the outer product F'E(E'E)^{-1}E'F and subtract it from the
483 // Schur complement matrix, i.e
484 //
485 //  S -= F'E(E'E)^{-1}E'F.
486 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
487 void
488 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
ChunkOuterProduct(const CompressedRowBlockStructure * bs,const Matrix & inverse_ete,const double * buffer,const BufferLayoutType & buffer_layout,BlockRandomAccessMatrix * lhs)489 ChunkOuterProduct(const CompressedRowBlockStructure* bs,
490                   const Matrix& inverse_ete,
491                   const double* buffer,
492                   const BufferLayoutType& buffer_layout,
493                   BlockRandomAccessMatrix* lhs) {
494   // This is the most computationally expensive part of this
495   // code. Profiling experiments reveal that the bottleneck is not the
496   // computation of the right-hand matrix product, but memory
497   // references to the left hand side.
498   const int e_block_size = inverse_ete.rows();
499   BufferLayoutType::const_iterator it1 = buffer_layout.begin();
500   // S(i,j) -= bi' * ete^{-1} b_j
501   for (; it1 != buffer_layout.end(); ++it1) {
502     const int block1 = it1->first - num_eliminate_blocks_;
503     const int block1_size = bs->cols[it1->first].size;
504 
505     const typename EigenTypes<kEBlockSize, kFBlockSize>::ConstMatrixRef
506         b1(buffer + it1->second, e_block_size, block1_size);
507     const typename EigenTypes<kFBlockSize, kEBlockSize>::Matrix
508         b1_transpose_inverse_ete = b1.transpose() * inverse_ete;
509 
510     BufferLayoutType::const_iterator it2 = it1;
511     for (; it2 != buffer_layout.end(); ++it2) {
512       const int block2 = it2->first - num_eliminate_blocks_;
513 
514       int r, c, row_stride, col_stride;
515       CellInfo* cell_info = lhs->GetCell(block1, block2,
516                                          &r, &c,
517                                          &row_stride, &col_stride);
518       if (cell_info == NULL) {
519         continue;
520       }
521 
522       const int block2_size = bs->cols[it2->first].size;
523       const typename EigenTypes<kEBlockSize, kFBlockSize>::ConstMatrixRef
524           b2(buffer + it2->second, e_block_size, block2_size);
525 
526       CeresMutexLock l(&cell_info->m);
527       MatrixRef m(cell_info->values, row_stride, col_stride);
528 
529       // We explicitly construct a block object here instead of using
530       // m.block(), as m.block() variant of the constructor does not
531       // allow mixing of template sizing and runtime sizing parameters
532       // like the Matrix class does.
533       Eigen::Block<MatrixRef, kFBlockSize, kFBlockSize>
534           block(m, r, c,  block1_size, block2_size);
535 #ifdef CERES_WORK_AROUND_ANDROID_NDK_COMPILER_BUG
536       // Removing the ".noalias()" annotation on the following statement is
537       // necessary to produce a correct build with the Android NDK, including
538       // versions 6, 7, 8, and 8b, when built with STLPort and the
539       // non-standalone toolchain (i.e. ndk-build). This appears to be a
540       // compiler bug; if the workaround is not in place, the line
541       //
542       //   block.noalias() -= b1_transpose_inverse_ete * b2;
543       //
544       // gets compiled to
545       //
546       //   block.noalias() += b1_transpose_inverse_ete * b2;
547       //
548       // which breaks schur elimination. Introducing a temporary by removing the
549       // .noalias() annotation causes the issue to disappear. Tracking this
550       // issue down was tricky, since the test suite doesn't run when built with
551       // the non-standalone toolchain.
552       //
553       // TODO(keir): Make a reproduction case for this and send it upstream.
554       block -= b1_transpose_inverse_ete * b2;
555 #else
556       block.noalias() -= b1_transpose_inverse_ete * b2;
557 #endif  // CERES_WORK_AROUND_ANDROID_NDK_COMPILER_BUG
558     }
559   }
560 }
561 
562 // For rows with no e_blocks, the schur complement update reduces to S
563 // += F'F. This function iterates over the rows of A with no e_block,
564 // and calls NoEBlockRowOuterProduct on each row.
565 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
566 void
567 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
NoEBlockRowsUpdate(const BlockSparseMatrixBase * A,const double * b,int row_block_counter,BlockRandomAccessMatrix * lhs,double * rhs)568 NoEBlockRowsUpdate(const BlockSparseMatrixBase* A,
569                    const double* b,
570                    int row_block_counter,
571                    BlockRandomAccessMatrix* lhs,
572                    double* rhs) {
573   const CompressedRowBlockStructure* bs = A->block_structure();
574   for (; row_block_counter < bs->rows.size(); ++row_block_counter) {
575     const CompressedRow& row = bs->rows[row_block_counter];
576     const double *row_values = A->RowBlockValues(row_block_counter);
577     for (int c = 0; c < row.cells.size(); ++c) {
578       const int block_id = row.cells[c].block_id;
579       const int block_size = bs->cols[block_id].size;
580       const int block = block_id - num_eliminate_blocks_;
581       VectorRef(rhs + lhs_row_layout_[block], block_size).noalias()
582           += (ConstMatrixRef(row_values + row.cells[c].position,
583                              row.block.size, block_size).transpose() *
584               ConstVectorRef(b + row.block.position, row.block.size));
585     }
586     NoEBlockRowOuterProduct(A, row_block_counter, lhs);
587   }
588 }
589 
590 
591 // A row r of A, which has no e_blocks gets added to the Schur
592 // Complement as S += r r'. This function is responsible for computing
593 // the contribution of a single row r to the Schur complement. It is
594 // very similar in structure to EBlockRowOuterProduct except for
595 // one difference. It does not use any of the template
596 // parameters. This is because the algorithm used for detecting the
597 // static structure of the matrix A only pays attention to rows with
598 // e_blocks. This is becase rows without e_blocks are rare and
599 // typically arise from regularization terms in the original
600 // optimization problem, and have a very different structure than the
601 // rows with e_blocks. Including them in the static structure
602 // detection will lead to most template parameters being set to
603 // dynamic. Since the number of rows without e_blocks is small, the
604 // lack of templating is not an issue.
605 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
606 void
607 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
NoEBlockRowOuterProduct(const BlockSparseMatrixBase * A,int row_block_index,BlockRandomAccessMatrix * lhs)608 NoEBlockRowOuterProduct(const BlockSparseMatrixBase* A,
609                      int row_block_index,
610                      BlockRandomAccessMatrix* lhs) {
611   const CompressedRowBlockStructure* bs = A->block_structure();
612   const CompressedRow& row = bs->rows[row_block_index];
613   const double *row_values = A->RowBlockValues(row_block_index);
614   for (int i = 0; i < row.cells.size(); ++i) {
615     const int block1 = row.cells[i].block_id - num_eliminate_blocks_;
616     DCHECK_GE(block1, 0);
617 
618     const int block1_size = bs->cols[row.cells[i].block_id].size;
619     const ConstMatrixRef b1(row_values + row.cells[i].position,
620                             row.block.size, block1_size);
621     int r, c, row_stride, col_stride;
622     CellInfo* cell_info = lhs->GetCell(block1, block1,
623                                        &r, &c,
624                                        &row_stride, &col_stride);
625     if (cell_info != NULL) {
626       CeresMutexLock l(&cell_info->m);
627       MatrixRef m(cell_info->values, row_stride, col_stride);
628       m.block(r, c, block1_size, block1_size)
629           .selfadjointView<Eigen::Upper>()
630           .rankUpdate(b1.transpose(), 1.0);
631     }
632 
633     for (int j = i + 1; j < row.cells.size(); ++j) {
634       const int block2 = row.cells[j].block_id - num_eliminate_blocks_;
635       DCHECK_GE(block2, 0);
636       DCHECK_LT(block1, block2);
637       int r, c, row_stride, col_stride;
638       CellInfo* cell_info = lhs->GetCell(block1, block2,
639                                          &r, &c,
640                                          &row_stride, &col_stride);
641       if (cell_info == NULL) {
642         continue;
643       }
644 
645       const int block2_size = bs->cols[row.cells[j].block_id].size;
646       CeresMutexLock l(&cell_info->m);
647       MatrixRef m(cell_info->values, row_stride, col_stride);
648       m.block(r, c, block1_size, block2_size).noalias() +=
649           b1.transpose() * ConstMatrixRef(row_values + row.cells[j].position,
650                                           row.block.size,
651                                           block2_size);
652     }
653   }
654 }
655 
656 // For a row with an e_block, compute the contribition S += F'F. This
657 // function has the same structure as NoEBlockRowOuterProduct, except
658 // that this function uses the template parameters.
659 template <int kRowBlockSize, int kEBlockSize, int kFBlockSize>
660 void
661 SchurEliminator<kRowBlockSize, kEBlockSize, kFBlockSize>::
EBlockRowOuterProduct(const BlockSparseMatrixBase * A,int row_block_index,BlockRandomAccessMatrix * lhs)662 EBlockRowOuterProduct(const BlockSparseMatrixBase* A,
663                       int row_block_index,
664                       BlockRandomAccessMatrix* lhs) {
665   const CompressedRowBlockStructure* bs = A->block_structure();
666   const CompressedRow& row = bs->rows[row_block_index];
667   const double *row_values = A->RowBlockValues(row_block_index);
668   for (int i = 1; i < row.cells.size(); ++i) {
669     const int block1 = row.cells[i].block_id - num_eliminate_blocks_;
670     DCHECK_GE(block1, 0);
671 
672     const int block1_size = bs->cols[row.cells[i].block_id].size;
673     const typename EigenTypes<kRowBlockSize, kFBlockSize>::ConstMatrixRef
674         b1(row_values + row.cells[i].position,
675            row.block.size, block1_size);
676     {
677       int r, c, row_stride, col_stride;
678       CellInfo* cell_info = lhs->GetCell(block1, block1,
679                                          &r, &c,
680                                          &row_stride, &col_stride);
681       if (cell_info == NULL) {
682         continue;
683       }
684 
685       CeresMutexLock l(&cell_info->m);
686       MatrixRef m(cell_info->values, row_stride, col_stride);
687 
688       Eigen::Block<MatrixRef, kFBlockSize, kFBlockSize>
689           block(m, r, c,  block1_size, block1_size);
690       block.template selfadjointView<Eigen::Upper>()
691           .rankUpdate(b1.transpose(), 1.0);
692     }
693 
694     for (int j = i + 1; j < row.cells.size(); ++j) {
695       const int block2 = row.cells[j].block_id - num_eliminate_blocks_;
696       DCHECK_GE(block2, 0);
697       DCHECK_LT(block1, block2);
698       const int block2_size = bs->cols[row.cells[j].block_id].size;
699       int r, c, row_stride, col_stride;
700       CellInfo* cell_info = lhs->GetCell(block1, block2,
701                                          &r, &c,
702                                          &row_stride, &col_stride);
703       if (cell_info == NULL) {
704         continue;
705       }
706 
707       const typename EigenTypes<kRowBlockSize, kFBlockSize>::ConstMatrixRef
708           b2(row_values + row.cells[j].position,
709              row.block.size,
710              block2_size);
711 
712       CeresMutexLock l(&cell_info->m);
713       MatrixRef m(cell_info->values, row_stride, col_stride);
714       Eigen::Block<MatrixRef, kFBlockSize, kFBlockSize>
715           block(m, r, c,  block1_size, block2_size);
716       block.noalias() += b1.transpose() * b2;
717     }
718   }
719 }
720 
721 }  // namespace internal
722 }  // namespace ceres
723 
724 #endif  // CERES_INTERNAL_SCHUR_ELIMINATOR_IMPL_H_
725