- On GCC 4.9+ although GCC supports LTO, it requires use of the
non-default gcc-ar & gcc-ranlib. Whilst we can ensure Ceres is
compiled with these, doing so with GCC 4.9 causes multiple definition
linker errors of static ints inside Eigen when compiling the tests
and examples when they are not also built with LTO.
- On OS X (Xcode 6 & 7) after the latest update to gtest, if LTO
is used when compiling the tests (& examples), two tests fail
due to typeinfo::operator== (things are fine if only Ceres itself is
compiled with LTO).
- This patch disables LTO for all compilers. It should be revisited when
the performance is more stable across our supported compilers.
Change-Id: I17b52957faefbdeff0aa40846dc9b342db1b02e3
- If Ceres is built as a shared library, and LTO is enabled for Ceres
and the tests, then type_info::operator==() incorrectly returns false
in gtests' CheckedDowncastToActualType() in the following tests:
-- levenberg_marquardt_strategy_test.
-- gradient_checking_cost_function_test.
on at least Xcode 6 & 7 as reported here:
https://github.com/google/googletest/issues/595.
- This does not appear to be a gtest issue, but is perhaps an LLVM bug
or an RTTI shared library issue. Either way, disabling the use of
LTO when compiling the test application resolves the issue.
- Allow LTO to be enabled for GCC, if it is supported.
- Add CMake function to allow easy appending to target properties s/t
Ceres library-specific compile flags can be iteratively constructed.
Change-Id: I923e6aae4f7cefa098cf32b2f8fc19389e7918c9
1. Move common test infrastructure into test_util.
2. system_test now only contains powells function.
3. Add bundle_adjustment_test.
Instead of a single function which computes everything,
there is now a test for each solver configuration which
uses the reference solution computed by the fixture.
Change-Id: I16a9a9a83a845a7aaf28762bcecf1a8ff5aee805
- Increasing the inline threshold results in very variable performance
improvements, and could potentially confuse users if they are trying
to set the inline threshold themselves.
- As such, we no longer export our inline threshold configuration for
Clang, but instead document how to change it in the FAQs.
Change-Id: I88e2e0001e4586ba2718535845ed1e4b1a5b72bc
The test for CompressedRowSparseMatrix::AppendRows tries to add
a matrix of size zero, which results in an invalid pointer deferencing
even though that pointer is never written to.
Change-Id: I97dba37082bd5dad242ae1af0447a9178cd92027
The outer product computation logic in SparseNormalCholeskySolver
does not work well with dynamic sparsity. The overhead of computing
the sparsity pattern of the normal equations is only amortized if
the sparsity is constant. If the sparsity can change from call to call
SparseNormalCholeskySolver will actually be more expensive.
For Eigen and for CXSparse we now explicitly compute the normal
equations using their respective matrix-matrix product routines and solve.
Change-Id: Ifbd8ed78987cdf71640e66ed69500442526a23d4
- When compiled with Clang, Ceres and all of the examples are compiled
with an increased inlining-threshold, as the default value can result
in poor Eigen performance.
- Previously, client code using Ceres would typically not use an
increased inlining-threshold (unless the user has specifically added
it themselves). However, increasing the inlining threshold can result
in significant performance improvements in auto-diffed CostFunctions.
- This patch adds the inlining-threshold flags to the interface flags
for the Ceres CMake target s/t any client code using Ceres (via
CMake), and compiled with Clang, will now be compiled with the same
increased inlining threshold as used by Ceres itself.
Change-Id: I31e8f1abfda140d22e85bb48aa57f028a68a415e
This method numerically computes function derivatives in different
scales, extrapolating between intermediate results to conserve function
evaluations. Adaptive differentiation is essential to produce accurate
results for functions with noisy derivatives.
Full changelist:
-Created a new type of NumericDiffMethod (RIDDERS).
-Implemented EvaluateRiddersJacobianColumn in NumericDiff.
-Created unit tests with f(x) = x^2 + [random noise] and
f(x) = exp(x).
Change-Id: I2d6e924d7ff686650272f29a8c981351e6f72091
- Previously, when Ceres was built as a static library we did not
compile position independent code. This means that the resulting
static library could not be linked against shared libraries, but
could be used by executables.
- To enable the use of a static Ceres library by other shared libraries
as reported in [1], the static library must be generated from
position independent code (except on Windows, where PIC does not
apply).
[1] https://github.com/Itseez/opencv_contrib/pull/290#issuecomment-130389471
Change-Id: I99388f1784ece688f91b162d009578c5c97ddaf6
The logic for determing static/dynamic f-block size in
DetectStructure was broken in a corner case, where the very first
row block which was used to initialize the f_block_size contained
more than one f blocks of varying sizes. The way the if block
was structured, no iteration was performed on the remaining
f-blocks and the loop failed to detect that the f-block size
was actually changing.
If in the remaining row blocks, there were no row blocks
with varying f-block sizes, the function will erroneously
return a static f-block size.
Thanks to Johannes Schonberger for providing a reproduction for this
rather tricky corner case.
Change-Id: Ib442a041d8b7efd29f9653be6a11a69d0eccd1ec
The schur eliminator treats rows with e blocks and row with
no e blocks separately. The template specialization logic only
applies to the rows with e blocks.
So, in cases where the rows with e-blocks have a fixed size f-block
but the rows without e-blocks have f-blocks of varying sizes,
DetectStructure will return a static f-block size, but we need to be
careful that we do not blindly use that static f-block size everywhere.
This patch fixes a bug where such care was not being taken, where
it was assumed that the static f-block size could be assumed for all
f-block sizes.
A new test is added, which triggers an exception in debug mode. In
release mode this error does not present itself, due to a peculiarity
of the way Eigen works.
Thanks to Werner Trobin for reporting this bug.
Change-Id: I8ae7aabf8eed8c3f9cf74b6c74d632ba44f82581
When the user provides an ordering which starts at a non-zero group id,
or has gaps in the groups, then CAMD, the algorithm used to reorder
the program can crash or return garbage results.
The solution is to map the ordering into grouping constraints, and then
to re-number the groups to be contiguous using a call to
MapValuesToContiguousRange. This was already done for CAMD based
ordering for Schur type solvers, but was not done for SPARSE_NORMAL_CHOLESKY.
Thanks to Bernhard Zeisl for not only reporting the bug but also
providing a reproduction.
Change-Id: I5cfae222d701dfdb8e1bda7f0b4670a30417aa89
The previous implementation incorrectly cached the outer product matrix
pattern even when `dynamic_sparsity = true`.
Change-Id: I1e58315a9b44f2f457d07c56b203ab2668bfb8a2
- Updated to new CMake style where function names are all lowercase,
this will be backwards compatible as CMake function names are
case insensitive.
- Updated using Emacs' M-x unscreamify-cmake-buffer.
Change-Id: If7219816f560270e59212813aeb021353a64a0e2
1. Push the boundary handling logic into the underlying array
object. This has two very significant impacts:
a. The interpolation code becomes extremely simple to write
and to test.
b. The user has more flexibility in implementing how out of bounds
values are handled. We provide one default implementation.
Change-Id: Ic2f6cf9257ce7110c62e492688e5a6c8be1e7df2
The reason this rather serious looking typo has not
caused any problems uptil now is because NUM_ROW_B is
computed but never actually used.
Thanks to Werner Trobin for pointing this out.
Change-Id: Id2b4d9326ec21baec8a85423e3270aefbafb611e
Often a parameter block is the Cartesian product of a number of
manifolds. For example, a rigid transformation SE(3) = SO(3) x R^3
In such cases, where you have the local parameterization
of the individual manifolds available,
ProductParameterization can be used to construct a local
parameterization of the cartesian product.
Change-Id: I4b5bcbd2407a38739c7725b129789db5c3d65a20
And undo the last two botched CLs.
Thanks to Chris Sweeney and Taylor Braun Jones for saving my bacon.
I will do appropriate penance to repend for my sins.
Change-Id: I14de958e651f85e4c1741fba2cb46ffe7e873346
- On at least some compilers, -std=c++11 is required in order to compile
against std::shared_ptr & std::unordered_map, which resulted in our
checks failing to find them and using the TR1 versions instead, which
causes conflicts for users using C++11.
- Now, if the compiler supports it and the user enables the CXX11
option, we explicitly enable C++11 before searching for shared_ptr &
unordered_map, which means we should always find the C++11 versions
if they are available.
- As use of CXX11 results in a version of Ceres that must be used with
-std=c++11 for GCC & Clang, we roll this into the Ceres target when
the version of CMake supports this, otherwise we warn the user they
will have to do this themselves.
- CXX11 is OFF by default, to ensure that the behaviour of Ceres is
unchanged from before.
Change-Id: I157ea7a4fadc6bc02da176b8e771f1f327ccaf78
This adds a new wrapper class called DynamicCostFunctionToFunctor
that closes a gap in the current API: the existing
CostFunctionToFunctor can only be used with a SizedCostFunction, where
the number and sizes of all parameter vectors are known at compile-time.
The DynamicCostFunctionToFunctor allows you to wrap a generic
CostFunction into a templated functor which can then be used in a
DynamicAutoDiffCostFunction.
Also updates the existing CostFunctionToFunctor class to internally use
DynamicCostFunctionToFunctor.
Change-Id: I088adc3271c58d2519126c27037c3576965a36d6
Before this change, the default step size
for a function F(x) at x was
step_size = |x| * relative_step_size
if step_size was exactly zero, then to prevent
division by zero we would fall back to relative_step_size.
This however is not good enough, as values of x say 1e-64
would lead to step sizes ~ 1e-70 and dividing by such numbers
leads to inaccurate results. For even smaller numbers, like
1e-300, which I have observed can occur as the optimization
algorithm makes progress, this leads to NaNs.
The key change in this CL is to change the fallback mechanism
to be
step_size = max(|x| * relative_step_size, min_step_size)
where
min_step_size = sqrt(DBL_EPSILON)
This is the recommended minimum value for the step size
for double precision arithmetic on the interwebs.
This results in a small loss of precision in the transcendental
functions test, but that is unavoidable as we are not taking
sufficiently small steps anymore.
On the whole though this will improve the numerical performance
of the algorithm.
To validate this approach, one of the parameter values for the
EasyFunctorTest has been set to 1e-64, which causes the test
to start failing without the corrected fallback logic.
This change should also address some if not all of
https://github.com/ceres-solver/ceres-solver/issues/121
Change-Id: I4a9013ef358626c1ba7b8abad60b3904163d63f6
The test makes sure the Rosenbrock function is correctly minimized from the
canonical starting point using the default settings.
Change-Id: Iea820f976707bde37162981c5db87fac5167ba9e
I think this is all of the cases. These cases arise because pow(a,b) is limited
to real valued results, if the argument and result were complex valued then
these cases would disappear.
NOTE: Since there is so much special casing here, it is worth checking to see
if cpow() is implemented in terms of pow(), and what might be the consequences
of using cpow() on the type std::complex<Jet<double, N> >. It is *possible*
that a separate implementation of cpow might be required also.
Also some comment fixes.
Change-Id: Ia1e38df4cdcb548f778304c2854cacba6e1556ff
The code seemed to imply that its possible to call the Write
method with a null pointer which is never the case. There would
be no point to calling Write.
Thanks to Michael Vitus for pointing this out.
Change-Id: Ic9a276856d0a7e65d53a1cc8742d4831c1a52615
Eigen upstream was broken a little while ago, and it seemed to be
the case that we needed a fix for using the LLT factorization on
ARM.
This has been fixed and AFAIK there are no stable eigen releases
with this bug in it.
For full gore, see
http://eigen.tuxfamily.org/bz/show_bug.cgi?id=992
In light of the fix, the extra layer of indirection introduced earlier
is not needed and we are reverting to normal programming.
Change-Id: I16929d2145253b38339b573b27b6b8fabd523704
The call to llt in backsubstitute seems to be using one
of the fixed size specializations which is best done with
an inline call to llt/ldlt rather than introducing yet another
variant of the SolverUpperTriangularUsingCholesky and calling it.
Also change the way SolverUpperTriangularUsingCholesky handles
error. It always computes the solution even if it is garbage
and then returns the error code.
This ensures that the previous code that depends on unconditional
computation still works.
Change-Id: Idb1e6efdae9a3775a072e3b87cde02e0bbddb319
When solving a linear system using Eigen's dense Cholesky factorization
if the right hand side of the linear system is the same vector
that will store the solution, call solveInPlace instead of solve.
Change-Id: I3e6d2f21ff420c25217cd87ee5d269fdfabbf19a
CERES_EIGEN_VERSION was being defined by the CMakeList.txt file
but it is needed by the android build too. So this change
directly constructs the CERES_EIGEN_VERSION string out of the
raw Eigen version numbers.
Change-Id: I65309805a59076c3082141d9042ab7e0e1b972bc
It seems that Eigen's LLT factorization is broken on ARM.
This patch enables the use of LDLT factorization instead of LLT
factorization. The switch is controlled at compile time using a
preprocessor define - CERES_USE_EIGEN_LDLT.
By default we continue to use LLT factorization though.
To make the switching easier without introducing the Cholesky factorization
based inversion and linear system solve routines have been abstracted into
two new functions.
Android.mk has been updated to enable the LDLT factorization, but
the cmake file has not been updated as I will leave it to Alex's
capable hands to do proper detection of ARM as a target platform.
Change-Id: Iffe3abd2ce894de2a388b454df3da909b482d5e5
- Previous tolerance of 2.0 * std::numeric_limits<double>::epsilon()
was too tight for Cygwin, worked on all other known platforms.
Change-Id: Ia79ad8961272dbb608d8e8ddd3f6d52e5f0735f4
Simon Rutishauser