With the addition of C++11 support we can simplify the parallel for code by
removing the ifdef branching. Converts coordinate_descent_minimizer.cc to use
the thread_id ParallelFor API.
Tested by building with OpenMP, C++11 threads, TBB, and no threads. All tests
pass.
Also compared timing via the bundle adjuster.
./bin/bundle_adjuster --input=../problem-744-543562-pre.txt
With OpenMP num_threads=8
Head:
Time (in seconds):
Residual only evaluation 0.807753 (5)
Jacobian & residual evaluation 4.489404 (6)
Linear solver 41.826481 (5)
Minimizer 50.745857
Total 73.294424
CL:
Time (in seconds):
Residual only evaluation 0.970483 (5)
Jacobian & residual evaluation 4.647438 (6)
Linear solver 41.781892 (5)
Minimizer 50.848904
Total 73.089983
With OpenMP num_threads=1
HEAD:
Time (in seconds):
Residual only evaluation 2.990246 (5)
Jacobian & residual evaluation 14.132090 (6)
Linear solver 79.631951 (5)
Minimizer 100.281847
Total 122.946267
CL:
Time (in seconds):
Residual only evaluation 3.075178 (5)
Jacobian & residual evaluation 13.966451 (6)
Linear solver 77.005441 (5)
Minimizer 97.568712
Total 120.410454
Change-Id: I1857d7943073be7465b6c6476bf46ab11c5475a3
This adds a callback mechanism to for users to get notified just
before jacobian and residual evaluations. This will enable
aggressive caching and sharing of compute between cost functions.
Change-Id: I67993726920218edf71ab9ae70c34c204756c71a
Previously, the thread ID was acquired and released on every iteration
of the for loop. The C++11 concurrent queue implementation is much
slower than TBB's version and consequently this was a huge bottleneck.
This introduces another ParallelFor API which takes the thread ID as a
parameter in the evaluation function. This allows us to acquire and
release the thread ID for each block of work which drastically improves
the performance.
This change brings us on par with OpenMP and TBB. See below for a
timing comparison. Note: in this example this CLs C++11 version is
faster to compute the residuals because TBB still must acquire the
thread ID on every iteration, which has some overhead.
Tested by building and running tests for no threading, OpenMP, TBB, and
C++11 threads. Also ran bazel tests.
./bin/bundle_adjuster --input=problem-744-543562-pre.txt --num_threads=8
C++11 @Head
Time (in seconds):
Residual only evaluation 7.819692 (5)
Jacobian & residual evaluation 11.606063 (6)
Linear solver 47.860195 (5)
Minimizer 70.877072
Total 90.806338
---------------------------------------------------
C++11 (This CL)
Time (in seconds):
Residual only evaluation 1.217500 (5)
Jacobian & residual evaluation 5.796112 (6)
Linear solver 44.080873 (5)
Minimizer 54.635524
Total 77.640072
---------------------------------------------------
OpenMP
Time (in seconds):
Residual only evaluation 0.797023 (5)
Jacobian & residual evaluation 5.633916 (6)
Linear solver 43.280020 (5)
Minimizer 53.199058
Total 76.250861
---------------------------------------------------
TBB
Time (in seconds):
Residual only evaluation 1.911095 (5)
Jacobian & residual evaluation 5.557807 (6)
Linear solver 44.074680 (5)
Minimizer 55.002688
Total 78.052687
---------------------------------------------------
No Threads
Time (in seconds):
Residual only evaluation 2.939212 (5)
Jacobian & residual evaluation 18.519874 (6)
Linear solver 74.017837 (5)
Minimizer 98.980080
Total 122.216391
Change-Id: I3af959b0771bbdfe8cad8c13896191d6ac903181
A Ceres Context holds common global state that can be re-used within
Ceres. The Context current contains a thread pool if compiling with
C++11 threading support. Threads are expensive to create and destroy so
it is good to maintain across multiple Ceres solves.
Tested by compiling with and without TBB support and ran unit tests. Ran
bazel as well.
Change-Id: I82f598dfae642aa0e81a6039dc174608a5e8dbfb
This is in preparation for adding support for a c++11 based parallel
for implementation. The parallel for abstraction does not have the
ability to constrain the total number of threads in nested for loops.
This is solved by distributing the number of threads evenly between
the nested for loops. Adds a TODO to consolidate the next for loops
into a single loop that can be properly split between threads.
Tested by building with TBB and running tests.
Change-Id: I546973b9a4d19b9cdd53caff55d1c80bac8ea953
1. Replace two maps by 1.
2. Update number of calls and the time for the call at the cost
of a single map lookup.
3. Add Solver::Summary::num_linear_solves.
Fixes https://github.com/ceres-solver/ceres-solver/issues/340
Change-Id: I71eb9be7fb363a8cb066591c4c1761f256c81677
Fixes the current implementation where the desired number of threads may
not be honored if another tbb::task_scheduler_init is instantiated. We
are using tbb::task_arena to solve this which is only available in newer
versions of TBB.
Also increases the performance by not creating/destroying the TBB setup
via tbb::task_scheduler_init on every iteration evaluation. This
increases the performance in single threaded mode using TBB by 10x.
By not specifically calling tbb::task_scheduler_init, this will either
respect any active tbb::task_scheduler_init instantiations or use the
default TBB settings which is hardware dependent. Ceres will honor the
user's requested number of threads through the task_arenas.
Tested via compiling with TBB enabled and ran the unit tests.
Change-Id: I5538407563449cdb5a0eaf8b8ccab62263912110
There are platforms where OpenMP is not available. This
patch adds support for Intel Threading Building Blocks (TBB)
as an alternative threading backend.
Change-Id: I94497d7cba0c3cfaccfc992169236f17fe948ae9
Since Ceres is moving to using GitHub for issues, and the Google
Code URL in the current copyright header will soon become invalid,
update all the headers.
Change-Id: I1fce70375d1bcf098591f07b4d8f01a5c1e0789c
For historical reasons we had a "using namespace std;" in port.h. This
is generally a bad idea. So removing it and along the way doing a bunch
of cpplint cleanup.
Change-Id: Ia125601a55ae62695e247fb0250df4c6f86c46c6
- Previously we passed all compile options to Ceres via add_definitions
in CMake. This was fine for private definitions (used only by Ceres)
but required additional work for public definitions to ensure they
were correctly propagated to clients via CMake using
target_compile_definitions() (>= 2.8.11) or add_definitions().
- A drawback to these approaches is that they did not work for chained
dependencies on Ceres, as in if in the users project B <- A <- Ceres,
then although the required Ceres public compile definitions would
be used when compiling A, they would not be propagated to B.
- This patch replaces the addition of compile definitions via
add_definitions() with an autogenerated config.h header which
is installed with Ceres and defines all of the enabled Ceres compile
options.
- This removes the need for the user to propagate any compile
definitions in their projects, and additionally allows post-install
inspect of the options with which Ceres was compiled.
Change-Id: Idbdb6abdad0eb31e7540370e301afe87a07f2260
The standard sparse normal Cholesky solver assumes a fixed
sparsity pattern which is useful for a large number of problems
presented to Ceres. However, some problems are symbolically dense
but numerically sparse i.e. each residual is a function of a
large number of parameters but at any given state the residual
only depends on a sparse subset of them. For these class of
problems it is faster to re-analyse the sparsity pattern of the
jacobian at each iteration of the non-linear optimisation instead
of including all of the zero entries in the step computation.
The proposed solution adds the dynamic_sparsity option which can
be used with SPARSE_NORMAL_CHOLESKY. A
DynamicCompressedRowSparseMatrix type (which extends
CompressedRowSparseMatrix) has been introduced which allows
dynamic addition and removal of elements. A Finalize method is
provided which then consolidates the matrix so that it can be
used in place of a regular CompressedRowSparseMatrix. An
associated jacobian writer has also been provided.
Changes that were required to make this extension were adding the
SetMaxNumNonZeros method to CompressedRowSparseMatrix and adding
a JacobianFinalizer template parameter to the ProgramEvaluator.
Change-Id: Ia5a8a9523fdae8d5b027bc35e70b4611ec2a8d01
1. When a LAPACK implementation is present, then
DENSE_QR, DENSE_NORMAL_CHOLESKY and DENSE_SCHUR
can use it for doing dense linear algebra operations.
2. The user can switch dense linear algebra libraries
by setting Solver::Options::dense_linear_algebra_library_type.
3. Solver::Options::sparse_linear_algebra_library is now
Solver::Options::sparse_linear_algebra_library_type to be consistent
with all the other enums in Solver::Options.
4. Updated documentation as well as Solver::Summary::FullReport
to reflect these changes.
Change-Id: I5ab930bc15e90906b648bc399b551e6bd5d6498f
Evaluator now uses custom BLAS for gradient
computations.
Update the evaluator in trust_region_minimizer_test to compute
gradients.
Change-Id: I3f565bc203b47b2b795a0609d67f25775648653c
1. Add the ability to evaluate the problem without loss function.
2. Remove static Evaluator::Evaluate
3. Refactor the common code from problem_test.cc and
evaluator_test.cc into evaluator_test_utils.cc
Change-Id: I1aa841580afe91d288fbb65288b0ffdd1e43e827
1. Add an ExecutionSummary object to record execution
information about Ceres objects.
2. Add an EventLogger object to log events in a function call.
3. Add a ScopedExecutionTimer object to log times in ExecutionSummary.
4. Instrument ProgramEvaluator and all the linear solvers
to report their timing statistics.
5. Connect the timing statistics to Summary::FullReport.
6. Add high precision timer on unix systems using
gettimeofday() call.
7. Various minor clean ups all around.
Change-Id: I5e09804b730b09535484124be7dbc1c58eccd1d4
The buffers used to store the per thread value of the gradient
were not set to zero at the beginning of each call to evaluate.
Change-Id: I9c8afea54a4e2e0b805164025da3023166a309af
A non-linear generalization of Ruhe & Wedin's algorithm
for separable non-linear least squares problem. It is implemented
as coordinate descent on an independent subset of the parameter
blocks at the end of every successful Newton step. The resulting
algorithm has much improved convergence at the cost of some
execution time.
Change-Id: I8fdc5edbd0ba1e702c9658b98041b2c2ae705402
This extends the Evaluator interface to support evaluating the
gradient in addition to the residuals and jacobian, if requested.
bool Evaluate(const double* state,
double* cost,
double* residuals,
double* gradient, <----------- NEW
SparseMatrix* jacobian) = 0;
The ProgramEvaluator is extended to support the new gradient
evaluation. This required some gymnastics around the block
evaluate preparer, which now contains a scratch evaluate preparer
for the case that no jacobian is requested but the gradient is.
Gradient evaluation is a prerequisite for the planned suite of
first order methods, including nonlinear conjugate gradient,
CG_DESCENT, L-BFGS, trust region with line search, and more.
This also considerably refactors the evaluator_test to make it
shorter and check the results for all combinations of the optional
parameters [residuals, gradient, jacobian].
Change-Id: Ic7d0fec028dc5ffebc08ee079ad04eeaf6e02582
Mike Vitus