Fix for alignment asserts from eigen when running with -mnative

[Jorgen Birkler]

Summary; FixedArray does not consider alignment to the 'T' type and in certain circumstances we got asserts and seg-faults from Eigen because the data was not aligned.

This patch seems to fix the issues.

[Sameer Agarwal]

Jorgen,

Thanks for the patch. Two things.

1. It would be useful to see what kind of errors you are getting and are being fixed by this. I am surprised because we do not put eigen arrays inside FixedArray anywhere in the code.

2. The path needs to be submitted to gerrit after signing a CLA, so that copyrights are handled correctly, following the instructions at 

http://ceres-solver.org/contributing.html#submitting-a-change

Sameer

On Fri, Jul 22, 2016 at 11:28 AM Jorgen Birkler <jor...@insidemaps.com> wrote:

Summary; FixedArray does not consider alignment to the 'T' type and in certain circumstances we got asserts and seg-faults from Eigen because the data was not aligned.

This patch seems to fix the issues.

--
You received this message because you are subscribed to the Google Groups "Ceres Solver" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ceres-solver...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/01fd9675-fef7-4ad8-951f-9e294ffa74b5%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Jorgen Birkler]

Thread #1 [interface_tests] 11582 [core: 1] (Suspended : Signal : SIGABRT:Aborted)

__GI_raise() at raise.c:54 0x7f0880940418

__GI_abort() at abort.c:89 0x7f088094201a

__assert_fail_base() at assert.c:92 0x7f0880938bd7

__GI___assert_fail() at assert.c:101 0x7f0880938c82

Eigen::internal::plain_array<double, 24, 0, 32>::plain_array() at DenseStorage.h:128 0x49c795

Eigen::DenseStorage<double, 24, 24, 1, 0>::DenseStorage() at DenseStorage.h:187 0x498b04

Eigen::PlainObjectBase<Eigen::Matrix<double, 24, 1, 0, 24, 1> >::PlainObjectBase() at PlainObjectBase.h:457 0x4954b0

Eigen::Matrix<double, 24, 1, 0, 24, 1>::Matrix() at Matrix.h:259 0x4921aa

ceres::Jet<double, 24>::Jet() at jet.h:182 0x48cd7a

ceres::internal::FixedArray<ceres::Jet<double, 24>, 8l>::InnerContainer::InnerContainer() at fixed_array.h:139 0x4862ba

ceres::internal::FixedArray<ceres::Jet<double, 24>, 8l>::FixedArray() at fixed_array.h:169 0x486344

ceres::internal::AutoDiff<im::InverseDepthAnchor2UVCostFunctor, double, 4, 4, 3, 4, 3, 6, 0, 0, 0, 0>::Differentiate() at autodiff.h:233 0x4807d9

ceres::AutoDiffCostFunction<im::InverseDepthAnchor2UVCostFunctor, 2, 4, 4, 3, 4, 3, 6, 0, 0, 0, 0>::Evaluate() at autodiff_cost_function.h:218 0x47efb7

ceres::internal::ResidualBlock::Evaluate(bool, double*, double*, double**, double*) const at 0x739a14

ceres::internal::ProgramEvaluator<ceres::internal::BlockEvaluatePreparer, ceres::internal::BlockJacobianWriter, ceres::internal::NullJacobianFinalizer>::Evaluate() at 0x7628d3

ceres::internal::TrustRegionMinimizer::Minimize() at 0x78085c

ceres::Solver::Solve() at 0x745e7d

ceres::Solve() at 0x747829

testTrackerBundleAdjusterRandom() at main_test.cpp:624 0x61714e

TestFixture_testTrackerBundleAdjusterMapPointParametrizationRandomNoInit_Test::TestBody() at main_test.cpp:770 0x61995a

<...more frames...>

Yes. I could not understand that either. T is a double. Inside the Jet it creates a Eigen::Matrix which is templated to use NoAlign bit. but it seems that when the constructor is being run the created Matrix is aligned.

Sorry I do not have any minimal test case that provokes the error at the moment.

BTW if we pass 

-march=native => ASSERT triggers

-march=native -mno-avx -mno-avx2  => No assert triggers

I'm on an i7 and 

gcc (Ubuntu 5.4.0-6ubuntu1~16.04.1) 5.4.0 20160609

Copyright (C) 2015 Free Software Foundation, Inc.

This is free software; see the source for copying conditions.  There is NO

warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

So -march=native will allow Eigen to generate avx vectorization (and eigen does...).

[Sameer Agarwal]

hmm fixed array of jets, I had forgotten about this.

I am guessing you are enabling c++11 when compiling your code, since that is the only path when the Eigen Array inside the Jet is aligned. or is that not the case?

Sameer

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/02e14a29-0c20-4b9c-9501-79904dab1c34%40googlegroups.com.

[Jorgen Birkler]

Correct. We are using C++11.

[Jorgen Birkler]

Update:

Just got a segfault in the autodiff. 

-mno-avx -mno-avx2 removes the segfault.

So the patch does not seem to cover all the cases.

/Jorgen

[Sameer Agarwal]

I think we have a lead on the fix for this, it will take a bit, will update later next week.

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/21518e29-dec4-4aea-8c27-d56d79c8ec30%40googlegroups.com.