1 .. _chapter-version-history: 2 3 =============== 4 Version History 5 =============== 6 7 1.7.0 8 ===== 9 10 Backward Incompatible API Changes 11 --------------------------------- 12 13 #. ``Solver::Options::sparse_linear_algebra_library`` has been renamed 14 to ``Solver::Options::sparse_linear_algebra_library_type``. 15 16 New Features 17 ------------ 18 19 #. Sparse and dense covariance estimation. 20 #. A new Wolfe line search. (Alex Stewart) 21 #. ``BFGS`` line search direction. (Alex Stewart) 22 #. C API 23 #. Speeded up the use of loss functions > 17x. 24 #. Faster ``DENSE_QR``, ``DENSE_NORMAL_CHOLESKY`` and ``DENSE_SCHUR`` 25 solvers. 26 #. Support for multiple dense linear algebra backends. In particular 27 optimized ``BLAS`` and ``LAPACK`` implementations (e.g., Intel MKL, 28 ACML, OpenBLAS etc) can now be used to do the dense linear 29 algebra for ``DENSE_QR``, ``DENSE_NORMAL_CHOLESKY`` and 30 ``DENSE_SCHUR`` 31 #. Use of Inner iterations can now be adaptively stopped. Iteration 32 and runtime statistics for inner iterations are not reported in 33 ``Solver::Summary`` and ``Solver::Summary::FullReport``. 34 #. Improved inner iteration step acceptance criterion. 35 #. Add BlockRandomAccessCRSMatrix. 36 #. Speeded up automatic differentiation by 7\%. 37 #. Bundle adjustment example from libmv/Blender (Sergey Sharybin) 38 #. Add the ability to turn shared library compilation on and off 39 #. No more dependence on Protocol Buffers. 40 #. Incomplete LQ factorization. 41 #. Ability to write trust region problems to disk. 42 #. Add sinh, cosh, tanh and tan functions to automatic differentiation 43 (Johannes Schnberger) 44 #. Simplifications to the cmake build file. 45 #. ``miniglog`` can now be used as a replacement for ``google-glog`` 46 on non Android platforms. (This is NOT recommended). 47 48 Bug Fixes 49 --------- 50 51 #. Fix ``ITERATIVE_SCHUR`` solver to work correctly when the schur 52 complement is of size zero. (Soohyun Bae) 53 #. Fix the ``spec`` file for generating ``RPM`` packages (Brian Pitts 54 and Taylor Braun-Jones). 55 #. Fix how ceres calls CAMD (Manas Jagadev) 56 #. Fix breakage on old versions of SuiteSparse. (Fisher Yu) 57 #. Fix warning C4373 in Visual Studio (Petter Strandmark) 58 #. Fix compilation error caused by missing suitesparse headers and 59 reorganize them to be more robust. (Sergey Sharybin) 60 #. Check GCC Version before adding -fast compiler option on 61 OSX. (Steven Lovegrove) 62 #. Add documentation for minimizer progress output. 63 #. Lint and other cleanups (William Rucklidge and James Roseborough) 64 #. Collections port fix for MSC 2008 (Sergey Sharybin) 65 #. Various corrections and cleanups in the documentation. 66 #. Change the path where CeresConfig.cmake is installed (Pablo 67 Speciale) 68 #. Minor errors in documentation (Pablo Speciale) 69 #. Updated depend.cmake to follow CMake IF convention. (Joydeep 70 Biswas) 71 #. Stablize the schur ordering algorithm. 72 #. Update license header in split.h. 73 #. Enabling -O4 (link-time optimization) only if compiler/linker 74 support it. (Alex Stewart) 75 #. Consistent glog path across files. 76 #. ceres-solver.spec: Use cleaner, more conventional Release string 77 (Taylor Braun-Jones) 78 #. Fix compile bug on RHEL6 due to missing header (Taylor Braun-Jones) 79 #. CMake file is less verbose. 80 #. Use the latest upstream version of google-test and gmock. 81 #. Rationalize some of the variable names in ``Solver::Options``. 82 #. Improve Summary::FullReport when line search is used. 83 #. Expose line search parameters in ``Solver::Options``. 84 #. Fix update of L-BFGS history buffers after they become full. (Alex 85 Stewart) 86 #. Fix configuration error on systems without SuiteSparse installed 87 (Sergey Sharybin) 88 #. Enforce the read call returns correct value in ``curve_fitting_c.c`` 89 (Arnaud Gelas) 90 #. Fix DynamicAutoDiffCostFunction (Richard Stebbing) 91 #. Fix Problem::RemoveParameterBlock documentation (Johannes 92 Schnberger) 93 #. Fix a logging bug in parameter_block.h 94 #. Refactor the preconditioner class structure. 95 #. Fix an uninitialized variable warning when building with ``GCC``. 96 #. Fix a reallocation bug in 97 ``CreateJacobianBlockSparsityTranspose``. (Yuliy Schwartzburg) 98 #. Add a define for O_BINARY. 99 100 101 1.6.0 102 ===== 103 104 New Features 105 ------------ 106 107 #. Major Performance improvements. 108 109 a. Schur type solvers (``SPARSE_SCHUR``, ``DENSE_SCHUR``, 110 ``ITERATIVE_SCHUR``) are significantly faster due to custom BLAS 111 routines and fewer heap allocations. 112 113 b. ``SPARSE_SCHUR`` when used with ``CX_SPARSE`` now uses a block 114 AMD for much improved factorization performance. 115 116 c. The jacobian matrix is pre-ordered so that 117 ``SPARSE_NORMAL_CHOLESKY`` and ``SPARSE_SCHUR`` do not have to 118 make copies inside ``CHOLMOD``. 119 120 d. Faster autodiff by replacing division by multplication by inverse. 121 122 e. When compiled without threads, the schur eliminator does not pay 123 the penalty for locking and unlocking mutexes. 124 125 #. Users can now use ``linear_solver_ordering`` to affect the 126 fill-reducing ordering used by ``SUITE_SPARSE`` for 127 ``SPARSE_NORMAL_CHOLESKY``. 128 #. ``Problem`` can now report the set of parameter blocks it knows about. 129 #. ``TrustRegionMinimizer`` uses the evaluator to compute the gradient 130 instead of a matrix vector multiply. 131 #. On ``Mac OS``, whole program optimization is enabled. 132 #. Users can now use automatic differentiation to define new 133 ``LocalParameterization`` objects. (Sergey Sharybin) 134 #. Enable larger tuple sizes for Visual Studio 2012. (Petter Strandmark) 135 136 137 Bug Fixes 138 --------- 139 140 #. Update the documentation for ``CostFunction``. 141 #. Fixed a typo in the documentation. (Pablo Speciale) 142 #. Fix a typo in suitesparse.cc. 143 #. Bugfix in ``NumericDiffCostFunction``. (Nicolas Brodu) 144 #. Death to BlockSparseMatrixBase. 145 #. Change Minimizer::Options::min_trust_region_radius to double. 146 #. Update to compile with stricter gcc checks. (Joydeep Biswas) 147 #. Do not modify cached CMAKE_CXX_FLAGS_RELEASE. (Sergey Sharybin) 148 #. ``<iterator>`` needed for back_insert_iterator. (Petter Strandmark) 149 #. Lint cleanup. (William Rucklidge) 150 #. Documentation corrections. (Pablo Speciale) 151 152 153 1.5.0 154 ===== 155 156 Backward Incompatible API Changes 157 --------------------------------- 158 159 #. Added ``Problem::Evaluate``. Now you can evaluate a problem or any 160 part of it without calling the solver. 161 162 In light of this the following settings have been deprecated and 163 removed from the API. 164 165 - ``Solver::Options::return_initial_residuals`` 166 - ``Solver::Options::return_initial_gradient`` 167 - ``Solver::Options::return_initial_jacobian`` 168 - ``Solver::Options::return_final_residuals`` 169 - ``Solver::Options::return_final_gradient`` 170 - ``Solver::Options::return_final_jacobian`` 171 172 Instead we recommend using something like this. 173 174 .. code-block:: c++ 175 176 Problem problem; 177 // Build problem 178 179 vector<double> initial_residuals; 180 problem.Evaluate(Problem::EvaluateOptions(), 181 NULL, /* No cost */ 182 &initial_residuals, 183 NULL, /* No gradient */ 184 NULL /* No jacobian */ ); 185 186 Solver::Options options; 187 Solver::Summary summary; 188 Solver::Solve(options, &problem, &summary); 189 190 vector<double> final_residuals; 191 problem.Evaluate(Problem::EvaluateOptions(), 192 NULL, /* No cost */ 193 &final_residuals, 194 NULL, /* No gradient */ 195 NULL /* No jacobian */ ); 196 197 198 New Features 199 ------------ 200 #. Problem now supports removal of ParameterBlocks and 201 ResidualBlocks. There is a space/time tradeoff in doing this which 202 is controlled by 203 ``Problem::Options::enable_fast_parameter_block_removal``. 204 205 #. Ceres now supports Line search based optimization algorithms in 206 addition to trust region algorithms. Currently there is support for 207 gradient descent, non-linear conjugate gradient and LBFGS search 208 directions. 209 #. Added ``Problem::Evaluate``. Now you can evaluate a problem or any 210 part of it without calling the solver. In light of this the 211 following settings have been deprecated and removed from the API. 212 213 - ``Solver::Options::return_initial_residuals`` 214 - ``Solver::Options::return_initial_gradient`` 215 - ``Solver::Options::return_initial_jacobian`` 216 - ``Solver::Options::return_final_residuals`` 217 - ``Solver::Options::return_final_gradient`` 218 - ``Solver::Options::return_final_jacobian`` 219 220 #. New, much improved HTML documentation using Sphinx. 221 #. Changed ``NumericDiffCostFunction`` to take functors like 222 ``AutoDiffCostFunction``. 223 #. Added support for mixing automatic, analytic and numeric 224 differentiation. This is done by adding ``CostFunctionToFunctor`` 225 and ``NumericDiffFunctor`` objects to the API. 226 #. Sped up the robust loss function correction logic when residual is 227 one dimensional. 228 #. Sped up ``DenseQRSolver`` by changing the way dense jacobians are 229 stored. This is a 200-500% improvement in linear solver performance 230 depending on the size of the problem. 231 #. ``DENSE_SCHUR`` now supports multi-threading. 232 #. Greatly expanded ``Summary::FullReport``: 233 234 - Report the ordering used by the ``LinearSolver``. 235 - Report the ordering used by the inner iterations. 236 - Execution timing breakdown into evaluations and linear solves. 237 - Effective size of the problem solved by the solver, which now 238 accounts for the size of the tangent space when using a 239 ``LocalParameterization``. 240 #. Ceres when run at the ``VLOG`` level 3 or higher will report 241 detailed timing information about its internals. 242 #. Remove extraneous initial and final residual evaluations. This 243 speeds up the solver a bit. 244 #. Automatic differenatiation with a dynamic number of parameter 245 blocks. (Based on an idea by Thad Hughes). 246 #. Sped up problem construction and destruction. 247 #. Added matrix adapters to ``rotation.h`` so that the rotation matrix 248 routines can work with row and column major matrices. (Markus Moll) 249 #. ``SCHUR_JACOBI`` can now be used without ``SuiteSparse``. 250 #. A ``.spec`` file for producing RPMs. (Taylor Braun-Jones) 251 #. ``CMake`` can now build the sphinx documentation (Pablo Speciale) 252 #. Add support for creating a CMake config file during build to make 253 embedding Ceres in other CMake-using projects easier. (Pablo 254 Speciale). 255 #. Better error reporting in ``Problem`` for missing parameter blocks. 256 #. A more flexible ``Android.mk`` and a more modular build. If binary 257 size and/or compile time is a concern, larger parts of the solver 258 can be disabled at compile time. 259 260 Bug Fixes 261 --------- 262 #. Compilation fixes for MSVC2010 (Sergey Sharybin) 263 #. Fixed "deprecated conversion from string constant to char*" 264 warnings. (Pablo Speciale) 265 #. Correctly propagate ifdefs when building without Schur eliminator 266 template specializations. 267 #. Correct handling of ``LIB_SUFFIX`` on Linux. (Yuliy Schwartzburg). 268 #. Code and signature cleanup in ``rotation.h``. 269 #. Make examples independent of internal code. 270 #. Disable unused member in ``gtest`` which results in build error on 271 OS X with latest Xcode. (Taylor Braun-Jones) 272 #. Pass the correct flags to the linker when using 273 ``pthreads``. (Taylor Braun-Jones) 274 #. Only use ``cmake28`` macro when building on RHEL6. (Taylor 275 Braun-Jones) 276 #. Remove ``-Wno-return-type-c-linkage`` when compiling with 277 GCC. (Taylor Braun-Jones) 278 #. Fix ``No previous prototype`` warnings. (Sergey Sharybin) 279 #. MinGW build fixes. (Sergey Sharybin) 280 #. Lots of minor code and lint fixes. (William Rucklidge) 281 #. Fixed a bug in ``solver_impl.cc`` residual evaluation. (Markus 282 Moll) 283 #. Fixed varidic evaluation bug in ``AutoDiff``. 284 #. Fixed ``SolverImpl`` tests. 285 #. Fixed a bug in ``DenseSparseMatrix::ToDenseMatrix()``. 286 #. Fixed an initialization bug in ``ProgramEvaluator``. 287 #. Fixes to Android.mk paths (Carlos Hernandez) 288 #. Modify ``nist.cc`` to compute accuracy based on ground truth 289 solution rather than the ground truth function value. 290 #. Fixed a memory leak in ``cxsparse.cc``. (Alexander Mordvintsev). 291 #. Fixed the install directory for libraries by correctly handling 292 ``LIB_SUFFIX``. (Taylor Braun-Jones) 293 294 1.4.0 295 ===== 296 297 Backward Incompatible API Changes 298 --------------------------------- 299 300 The new ordering API breaks existing code. Here the common case fixes. 301 302 **Before** 303 304 .. code-block:: c++ 305 306 options.linear_solver_type = ceres::DENSE_SCHUR 307 options.ordering_type = ceres::SCHUR 308 309 **After** 310 311 312 .. code-block:: c++ 313 314 options.linear_solver_type = ceres::DENSE_SCHUR 315 316 317 **Before** 318 319 .. code-block:: c++ 320 321 options.linear_solver_type = ceres::DENSE_SCHUR; 322 options.ordering_type = ceres::USER; 323 for (int i = 0; i < num_points; ++i) { 324 options.ordering.push_back(my_points[i]) 325 } 326 for (int i = 0; i < num_cameras; ++i) { 327 options.ordering.push_back(my_cameras[i]) 328 } 329 options.num_eliminate_blocks = num_points; 330 331 332 **After** 333 334 .. code-block:: c++ 335 336 options.linear_solver_type = ceres::DENSE_SCHUR; 337 options.ordering = new ceres::ParameterBlockOrdering; 338 for (int i = 0; i < num_points; ++i) { 339 options.linear_solver_ordering->AddElementToGroup(my_points[i], 0); 340 } 341 for (int i = 0; i < num_cameras; ++i) { 342 options.linear_solver_ordering->AddElementToGroup(my_cameras[i], 1); 343 } 344 345 346 New Features 347 ------------ 348 349 #. A new richer, more expressive and consistent API for ordering 350 parameter blocks. 351 #. A non-linear generalization of Ruhe & Wedin's Algorithm II. This 352 allows the user to use variable projection on separable and 353 non-separable non-linear least squares problems. With 354 multithreading, this results in significant improvements to the 355 convergence behavior of the solver at a small increase in run time. 356 #. An image denoising example using fields of experts. (Petter 357 Strandmark) 358 #. Defines for Ceres version and ABI version. 359 #. Higher precision timer code where available. (Petter Strandmark) 360 #. Example Makefile for users of Ceres. 361 #. IterationSummary now informs the user when the step is a 362 non-monotonic step. 363 #. Fewer memory allocations when using ``DenseQRSolver``. 364 #. GradientChecker for testing CostFunctions (William Rucklidge) 365 #. Add support for cost functions with 10 parameter blocks in 366 ``Problem``. (Fisher) 367 #. Add support for 10 parameter blocks in ``AutoDiffCostFunction``. 368 369 370 Bug Fixes 371 --------- 372 373 #. static cast to force Eigen::Index to long conversion 374 #. Change LOG(ERROR) to LOG(WARNING) in ``schur_complement_solver.cc``. 375 #. Remove verbose logging from ``DenseQRSolve``. 376 #. Fix the Android NDK build. 377 #. Better handling of empty and constant Problems. 378 #. Remove an internal header that was leaking into the public API. 379 #. Memory leak in ``trust_region_minimizer.cc`` 380 #. Schur ordering was operating on the wrong object (Ricardo Martin) 381 #. MSVC fixes (Petter Strandmark) 382 #. Various fixes to ``nist.cc`` (Markus Moll) 383 #. Fixed a jacobian scaling bug. 384 #. Numerically robust computation of ``model_cost_change``. 385 #. Signed comparison compiler warning fixes (Ricardo Martin) 386 #. Various compiler warning fixes all over. 387 #. Inclusion guard fixes (Petter Strandmark) 388 #. Segfault in test code (Sergey Popov) 389 #. Replaced ``EXPECT/ASSERT_DEATH`` with the more portable 390 ``EXPECT_DEATH_IF_SUPPORTED`` macros. 391 #. Fixed the camera projection model in Ceres' implementation of 392 Snavely's camera model. (Ricardo Martin) 393 394 395 1.3.0 396 ===== 397 398 New Features 399 ------------ 400 401 #. Android Port (Scott Ettinger also contributed to the port) 402 #. Windows port. (Changchang Wu and Pierre Moulon also contributed to the port) 403 #. New subspace Dogleg Solver. (Markus Moll) 404 #. Trust region algorithm now supports the option of non-monotonic steps. 405 #. New loss functions ``ArcTanLossFunction``, ``TolerantLossFunction`` 406 and ``ComposedLossFunction``. (James Roseborough). 407 #. New ``DENSE_NORMAL_CHOLESKY`` linear solver, which uses Eigen's 408 LDLT factorization on the normal equations. 409 #. Cached symbolic factorization when using ``CXSparse``. 410 (Petter Strandark) 411 #. New example ``nist.cc`` and data from the NIST non-linear 412 regression test suite. (Thanks to Douglas Bates for suggesting this.) 413 #. The traditional Dogleg solver now uses an elliptical trust 414 region (Markus Moll) 415 #. Support for returning initial and final gradients & Jacobians. 416 #. Gradient computation support in the evaluators, with an eye 417 towards developing first order/gradient based solvers. 418 #. A better way to compute ``Solver::Summary::fixed_cost``. (Markus Moll) 419 #. ``CMake`` support for building documentation, separate examples, 420 installing and uninstalling the library and Gerrit hooks (Arnaud 421 Gelas) 422 #. ``SuiteSparse4`` support (Markus Moll) 423 #. Support for building Ceres without ``TR1`` (This leads to 424 slightly slower ``DENSE_SCHUR`` and ``SPARSE_SCHUR`` solvers). 425 #. ``BALProblem`` can now write a problem back to disk. 426 #. ``bundle_adjuster`` now allows the user to normalize and perturb the 427 problem before solving. 428 #. Solver progress logging to file. 429 #. Added ``Program::ToString`` and ``ParameterBlock::ToString`` to 430 help with debugging. 431 #. Ability to build Ceres as a shared library (MacOS and Linux only), 432 associated versioning and build release script changes. 433 #. Portable floating point classification API. 434 435 436 Bug Fixes 437 --------- 438 439 #. Fix how invalid step evaluations are handled. 440 #. Change the slop handling around zero for model cost changes to use 441 relative tolerances rather than absolute tolerances. 442 #. Fix an inadvertant integer to bool conversion. (Petter Strandmark) 443 #. Do not link to ``libgomp`` when building on 444 windows. (Petter Strandmark) 445 #. Include ``gflags.h`` in ``test_utils.cc``. (Petter 446 Strandmark) 447 #. Use standard random number generation routines. (Petter Strandmark) 448 #. ``TrustRegionMinimizer`` does not implicitly negate the 449 steps that it takes. (Markus Moll) 450 #. Diagonal scaling allows for equal upper and lower bounds. (Markus Moll) 451 #. TrustRegionStrategy does not misuse LinearSolver:Summary anymore. 452 #. Fix Eigen3 Row/Column Major storage issue. (Lena Gieseke) 453 #. QuaternionToAngleAxis now guarantees an angle in $[-\pi, \pi]$. (Guoxuan Zhang) 454 #. Added a workaround for a compiler bug in the Android NDK to the 455 Schur eliminator. 456 #. The sparse linear algebra library is only logged in 457 Summary::FullReport if it is used. 458 #. Rename the macro ``CERES_DONT_HAVE_PROTOCOL_BUFFERS`` 459 to ``CERES_NO_PROTOCOL_BUFFERS`` for consistency. 460 #. Fix how static structure detection for the Schur eliminator logs 461 its results. 462 #. Correct example code in the documentation. (Petter Strandmark) 463 #. Fix ``fpclassify.h`` to work with the Android NDK and STLport. 464 #. Fix a memory leak in the ``levenber_marquardt_strategy_test.cc`` 465 #. Fix an early return bug in the Dogleg solver. (Markus Moll) 466 #. Zero initialize Jets. 467 #. Moved ``internal/ceres/mock_log.h`` to ``internal/ceres/gmock/mock-log.h`` 468 #. Unified file path handling in tests. 469 #. ``data_fitting.cc`` includes ``gflags`` 470 #. Renamed Ceres' Mutex class and associated macros to avoid 471 namespace conflicts. 472 #. Close the BAL problem file after reading it (Markus Moll) 473 #. Fix IsInfinite on Jets. 474 #. Drop alignment requirements for Jets. 475 #. Fixed Jet to integer comparison. (Keith Leung) 476 #. Fix use of uninitialized arrays. (Sebastian Koch & Markus Moll) 477 #. Conditionally compile gflag dependencies.(Casey Goodlett) 478 #. Add ``data_fitting.cc`` to the examples ``CMake`` file. 479 480 481 1.2.3 482 ===== 483 484 Bug Fixes 485 --------- 486 487 #. ``suitesparse_test`` is enabled even when ``-DSUITESPARSE=OFF``. 488 #. ``FixedArray`` internal struct did not respect ``Eigen`` 489 alignment requirements (Koichi Akabe & Stephan Kassemeyer). 490 #. Fixed ``quadratic.cc`` documentation and code mismatch 491 (Nick Lewycky). 492 493 1.2.2 494 ===== 495 496 Bug Fixes 497 --------- 498 499 #. Fix constant parameter blocks, and other minor fixes (Markus Moll) 500 #. Fix alignment issues when combining ``Jet`` and 501 ``FixedArray`` in automatic differeniation. 502 #. Remove obsolete ``build_defs`` file. 503 504 1.2.1 505 ===== 506 507 New Features 508 ------------ 509 510 #. Powell's Dogleg solver 511 #. Documentation now has a brief overview of Trust Region methods and 512 how the Levenberg-Marquardt and Dogleg methods work. 513 514 Bug Fixes 515 --------- 516 517 #. Destructor for ``TrustRegionStrategy`` was not virtual (Markus Moll) 518 #. Invalid ``DCHECK`` in ``suitesparse.cc`` (Markus Moll) 519 #. Iteration callbacks were not properly invoked (Luis Alberto Zarrabeiti) 520 #. Logging level changes in ConjugateGradientsSolver 521 #. VisibilityBasedPreconditioner setup does not account for skipped camera pairs. This was debugging code. 522 #. Enable SSE support on MacOS 523 #. ``system_test`` was taking too long and too much memory (Koichi Akabe) 524 525 1.2.0 526 ===== 527 528 New Features 529 ------------ 530 531 #. ``CXSparse`` support. 532 #. Block oriented fill reducing orderings. This reduces the 533 factorization time for sparse ``CHOLMOD`` significantly. 534 #. New Trust region loop with support for multiple trust region step 535 strategies. Currently only Levenberg-Marquardt is supported, but 536 this refactoring opens the door for Dog-leg, Stiehaug and others. 537 #. ``CMake`` file restructuring. Builds in ``Release`` mode by default, and now has platform specific tuning flags. 538 #. Re-organized documentation. No new content, but better 539 organization. 540 541 542 Bug Fixes 543 --------- 544 545 #. Fixed integer overflow bug in ``block_random_access_sparse_matrix.cc``. 546 #. Renamed some macros to prevent name conflicts. 547 #. Fixed incorrent input to ``StateUpdatingCallback``. 548 #. Fixes to AutoDiff tests. 549 #. Various internal cleanups. 550 551 552 1.1.1 553 ===== 554 555 Bug Fixes 556 --------- 557 558 #. Fix a bug in the handling of constant blocks. (Louis Simard) 559 #. Add an optional lower bound to the Levenberg-Marquardt regularizer 560 to prevent oscillating between well and ill posed linear problems. 561 #. Some internal refactoring and test fixes. 562 563 1.1.0 564 ===== 565 566 New Features 567 ------------ 568 569 #. New iterative linear solver for general sparse problems - ``CGNR`` 570 and a block Jacobi preconditioner for it. 571 #. Changed the semantics of how ``SuiteSparse`` dependencies are 572 checked and used. Now ``SuiteSparse`` is built by default, only if 573 all of its dependencies are present. 574 #. Automatic differentiation now supports dynamic number of residuals. 575 #. Support for writing the linear least squares problems to disk in 576 text format so that they can loaded into ``MATLAB``. 577 #. Linear solver results are now checked for nan and infinities. 578 #. Added ``.gitignore`` file. 579 #. A better more robust build system. 580 581 582 Bug Fixes 583 --------- 584 585 #. Fixed a strict weak ordering bug in the schur ordering. 586 #. Grammar and typos in the documents and code comments. 587 #. Fixed tests which depended on exact equality between floating point values. 588 589 1.0.0 590 ===== 591 592 Initial Release. Nathan Wiegand contributed to the Mac OSX port. 593
