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 (at) google.com (Keir Mierle) 30 // 31 // The ProgramEvaluator runs the cost functions contained in each residual block 32 // and stores the result into a jacobian. The particular type of jacobian is 33 // abstracted out using two template parameters: 34 // 35 // - An "EvaluatePreparer" that is responsible for creating the array with 36 // pointers to the jacobian blocks where the cost function evaluates to. 37 // - A "JacobianWriter" that is responsible for storing the resulting 38 // jacobian blocks in the passed sparse matrix. 39 // 40 // This abstraction affords an efficient evaluator implementation while still 41 // supporting writing to multiple sparse matrix formats. For example, when the 42 // ProgramEvaluator is parameterized for writing to block sparse matrices, the 43 // residual jacobians are written directly into their final position in the 44 // block sparse matrix by the user's CostFunction; there is no copying. 45 // 46 // The evaluation is threaded with OpenMP. 47 // 48 // The EvaluatePreparer and JacobianWriter interfaces are as follows: 49 // 50 // class EvaluatePreparer { 51 // // Prepare the jacobians array for use as the destination of a call to 52 // // a cost function's evaluate method. 53 // void Prepare(const ResidualBlock* residual_block, 54 // int residual_block_index, 55 // SparseMatrix* jacobian, 56 // double** jacobians); 57 // } 58 // 59 // class JacobianWriter { 60 // // Create a jacobian that this writer can write. Same as 61 // // Evaluator::CreateJacobian. 62 // SparseMatrix* CreateJacobian() const; 63 // 64 // // Create num_threads evaluate preparers. Caller owns result which must 65 // // be freed with delete[]. Resulting preparers are valid while *this is. 66 // EvaluatePreparer* CreateEvaluatePreparers(int num_threads); 67 // 68 // // Write the block jacobians from a residual block evaluation to the 69 // // larger sparse jacobian. 70 // void Write(int residual_id, 71 // int residual_offset, 72 // double** jacobians, 73 // SparseMatrix* jacobian); 74 // } 75 // 76 // Note: The ProgramEvaluator is not thread safe, since internally it maintains 77 // some per-thread scratch space. 78 79 #ifndef CERES_INTERNAL_PROGRAM_EVALUATOR_H_ 80 #define CERES_INTERNAL_PROGRAM_EVALUATOR_H_ 81 82 #ifdef CERES_USE_OPENMP 83 #include <omp.h> 84 #endif 85 86 #include "ceres/parameter_block.h" 87 #include "ceres/program.h" 88 #include "ceres/residual_block.h" 89 #include "ceres/internal/eigen.h" 90 #include "ceres/internal/scoped_ptr.h" 91 92 namespace ceres { 93 namespace internal { 94 95 template<typename EvaluatePreparer, typename JacobianWriter> 96 class ProgramEvaluator : public Evaluator { 97 public: 98 ProgramEvaluator(const Evaluator::Options &options, Program* program) 99 : options_(options), 100 program_(program), 101 jacobian_writer_(options, program), 102 evaluate_preparers_( 103 jacobian_writer_.CreateEvaluatePreparers(options.num_threads)) { 104 #ifndef CERES_USE_OPENMP 105 CHECK_EQ(1, options_.num_threads) 106 << "OpenMP support is not compiled into this binary; " 107 << "only options.num_threads=1 is supported."; 108 #endif 109 110 BuildResidualLayout(*program, &residual_layout_); 111 evaluate_scratch_.reset(CreateEvaluatorScratch(*program, 112 options.num_threads)); 113 } 114 115 // Implementation of Evaluator interface. 116 SparseMatrix* CreateJacobian() const { 117 return jacobian_writer_.CreateJacobian(); 118 } 119 120 bool Evaluate(const double* state, 121 double* cost, 122 double* residuals, 123 double* gradient, 124 SparseMatrix* jacobian) { 125 // The parameters are stateful, so set the state before evaluating. 126 if (!program_->StateVectorToParameterBlocks(state)) { 127 return false; 128 } 129 130 if (residuals != NULL) { 131 VectorRef(residuals, program_->NumResiduals()).setZero(); 132 } 133 134 if (jacobian != NULL) { 135 jacobian->SetZero(); 136 } 137 138 // Each thread gets it's own cost and evaluate scratch space. 139 for (int i = 0; i < options_.num_threads; ++i) { 140 evaluate_scratch_[i].cost = 0.0; 141 } 142 143 // This bool is used to disable the loop if an error is encountered 144 // without breaking out of it. The remaining loop iterations are still run, 145 // but with an empty body, and so will finish quickly. 146 bool abort = false; 147 int num_residual_blocks = program_->NumResidualBlocks(); 148 #pragma omp parallel for num_threads(options_.num_threads) 149 for (int i = 0; i < num_residual_blocks; ++i) { 150 // Disable the loop instead of breaking, as required by OpenMP. 151 #pragma omp flush(abort) 152 if (abort) { 153 continue; 154 } 155 156 #ifdef CERES_USE_OPENMP 157 int thread_id = omp_get_thread_num(); 158 #else 159 int thread_id = 0; 160 #endif 161 EvaluatePreparer* preparer = &evaluate_preparers_[thread_id]; 162 EvaluateScratch* scratch = &evaluate_scratch_[thread_id]; 163 164 // Prepare block residuals if requested. 165 const ResidualBlock* residual_block = program_->residual_blocks()[i]; 166 double* block_residuals = NULL; 167 if (residuals != NULL) { 168 block_residuals = residuals + residual_layout_[i]; 169 } else if (gradient != NULL) { 170 block_residuals = scratch->residual_block_residuals.get(); 171 } 172 173 // Prepare block jacobians if requested. 174 double** block_jacobians = NULL; 175 if (jacobian != NULL || gradient != NULL) { 176 preparer->Prepare(residual_block, 177 i, 178 jacobian, 179 scratch->jacobian_block_ptrs.get()); 180 block_jacobians = scratch->jacobian_block_ptrs.get(); 181 } 182 183 // Evaluate the cost, residuals, and jacobians. 184 double block_cost; 185 if (!residual_block->Evaluate( 186 &block_cost, 187 block_residuals, 188 block_jacobians, 189 scratch->residual_block_evaluate_scratch.get())) { 190 abort = true; 191 // This ensures that the OpenMP threads have a consistent view of 'abort'. Do 192 // the flush inside the failure case so that there is usually only one 193 // synchronization point per loop iteration instead of two. 194 #pragma omp flush(abort) 195 continue; 196 } 197 198 scratch->cost += block_cost; 199 200 // Store the jacobians, if they were requested. 201 if (jacobian != NULL) { 202 jacobian_writer_.Write(i, 203 residual_layout_[i], 204 block_jacobians, 205 jacobian); 206 } 207 208 // Compute and store the gradient, if it was requested. 209 if (gradient != NULL) { 210 int num_residuals = residual_block->NumResiduals(); 211 int num_parameter_blocks = residual_block->NumParameterBlocks(); 212 for (int j = 0; j < num_parameter_blocks; ++j) { 213 const ParameterBlock* parameter_block = 214 residual_block->parameter_blocks()[j]; 215 if (parameter_block->IsConstant()) { 216 continue; 217 } 218 MatrixRef block_jacobian(block_jacobians[j], 219 num_residuals, 220 parameter_block->LocalSize()); 221 VectorRef block_gradient(scratch->gradient.get() + 222 parameter_block->delta_offset(), 223 parameter_block->LocalSize()); 224 VectorRef block_residual(block_residuals, num_residuals); 225 block_gradient += block_residual.transpose() * block_jacobian; 226 } 227 } 228 } 229 230 if (!abort) { 231 // Sum the cost and gradient (if requested) from each thread. 232 (*cost) = 0.0; 233 int num_parameters = program_->NumEffectiveParameters(); 234 if (gradient != NULL) { 235 VectorRef(gradient, num_parameters).setZero(); 236 } 237 for (int i = 0; i < options_.num_threads; ++i) { 238 (*cost) += evaluate_scratch_[i].cost; 239 if (gradient != NULL) { 240 VectorRef(gradient, num_parameters) += 241 VectorRef(evaluate_scratch_[i].gradient.get(), num_parameters); 242 } 243 } 244 } 245 return !abort; 246 } 247 248 bool Plus(const double* state, 249 const double* delta, 250 double* state_plus_delta) const { 251 return program_->Plus(state, delta, state_plus_delta); 252 } 253 254 int NumParameters() const { 255 return program_->NumParameters(); 256 } 257 int NumEffectiveParameters() const { 258 return program_->NumEffectiveParameters(); 259 } 260 261 int NumResiduals() const { 262 return program_->NumResiduals(); 263 } 264 265 private: 266 // Per-thread scratch space needed to evaluate and store each residual block. 267 struct EvaluateScratch { 268 void Init(int max_parameters_per_residual_block, 269 int max_scratch_doubles_needed_for_evaluate, 270 int max_residuals_per_residual_block, 271 int num_parameters) { 272 residual_block_evaluate_scratch.reset( 273 new double[max_scratch_doubles_needed_for_evaluate]); 274 gradient.reset(new double[num_parameters]); 275 VectorRef(gradient.get(), num_parameters).setZero(); 276 residual_block_residuals.reset( 277 new double[max_residuals_per_residual_block]); 278 jacobian_block_ptrs.reset( 279 new double*[max_parameters_per_residual_block]); 280 } 281 282 double cost; 283 scoped_array<double> residual_block_evaluate_scratch; 284 // The gradient in the local parameterization. 285 scoped_array<double> gradient; 286 // Enough space to store the residual for the largest residual block. 287 scoped_array<double> residual_block_residuals; 288 scoped_array<double*> jacobian_block_ptrs; 289 }; 290 291 static void BuildResidualLayout(const Program& program, 292 vector<int>* residual_layout) { 293 const vector<ResidualBlock*>& residual_blocks = program.residual_blocks(); 294 residual_layout->resize(program.NumResidualBlocks()); 295 int residual_pos = 0; 296 for (int i = 0; i < residual_blocks.size(); ++i) { 297 const int num_residuals = residual_blocks[i]->NumResiduals(); 298 (*residual_layout)[i] = residual_pos; 299 residual_pos += num_residuals; 300 } 301 } 302 303 // Create scratch space for each thread evaluating the program. 304 static EvaluateScratch* CreateEvaluatorScratch(const Program& program, 305 int num_threads) { 306 int max_parameters_per_residual_block = 307 program.MaxParametersPerResidualBlock(); 308 int max_scratch_doubles_needed_for_evaluate = 309 program.MaxScratchDoublesNeededForEvaluate(); 310 int max_residuals_per_residual_block = 311 program.MaxResidualsPerResidualBlock(); 312 int num_parameters = program.NumEffectiveParameters(); 313 314 EvaluateScratch* evaluate_scratch = new EvaluateScratch[num_threads]; 315 for (int i = 0; i < num_threads; i++) { 316 evaluate_scratch[i].Init(max_parameters_per_residual_block, 317 max_scratch_doubles_needed_for_evaluate, 318 max_residuals_per_residual_block, 319 num_parameters); 320 } 321 return evaluate_scratch; 322 } 323 324 Evaluator::Options options_; 325 Program* program_; 326 JacobianWriter jacobian_writer_; 327 scoped_array<EvaluatePreparer> evaluate_preparers_; 328 scoped_array<EvaluateScratch> evaluate_scratch_; 329 vector<int> residual_layout_; 330 }; 331 332 } // namespace internal 333 } // namespace ceres 334 335 #endif // CERES_INTERNAL_PROGRAM_EVALUATOR_H_ 336