different behavior for eigen::matrix inversion when compiled with gcc and nvcc

Question

I do matrix inversion(code provided below) using Eigen(Lets call it inversion.cpp). Inversion.cpp compiles with g++ and gives correct inverse.

Now I change the code name to Inversion.cu and try to compile with nvcc. The compile fails with a long list of errors.

#include <Eigen/Geometry>
using namespace Eigen;
using namespace std;

typedef Matrix<float, 6, 6> Matrix6f;

int main()
{

Matrix6f M;

M<< 0.932322,  0.125848,  -0.85106, -0.313612,  -1.50979, -0.691102,
0.125848,  0.663803, -0.555993,  0.117918, -0.645694, -0.625737,
-0.85106, -0.555993,   1.26442,   0.39079,   1.89422,  0.959306,
-0.313612,  0.117918,   0.39079,   2.70718,   1.84228,   2.12945,
-1.50979, -0.645694,   1.89422,   1.84228,   3.89408,    2.5017,
-0.691102, -0.625737,  0.959306,   2.12945,    2.5017,   3.39874;

cout << "Inverse= " << M.inverse() << endl;

return 0;
}

The errors I get when I try to compile code with nvcc is:

Scanning dependencies of target matrix_inverse
[100%] Linking CXX executable matrix_inverse
CMakeFiles/matrix_inverse.dir/matrix_inverse_generated_matrix_inverse.cu.o: In function `Eigen::internal::compute_inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>, Eigen::Matrix<float, 6, 6, 0, 6, 6>, 6>::run(Eigen::Matrix<float, 6, 6, 0, 6, 6> const&, Eigen::Matrix<float, 6, 6, 0, 6, 6>&)':
/usr/include/eigen3/Eigen/src/LU/InverseImpl.h:28: undefined reference to `Eigen::MatrixBase<Eigen::Matrix<float, 6, 6, 0, 6, 6> >::partialPivLu() const'
collect2: error: ld returned 1 exit status
CMakeFiles/matrix_inverse.dir/build.make:79: recipe for target 'matrix_inverse' failed
make[2]: *** [matrix_inverse] Error 1
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/matrix_inverse.dir/all' failed
make[1]: *** [CMakeFiles/matrix_inverse.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

My cmakelists:

cmake_minimum_required(VERSION 2.8 FATAL_ERROR)

project(inversion)
INCLUDE_DIRECTORIES (/usr/include/eigen3)

FIND_PACKAGE(CUDA REQUIRED)
INCLUDE_DIRECTORIES(${CUDA_INCLUDE_DIRS})

set(
    CUDA_NVCC_FLAGS
    ${CUDA_NVCC_FLAGS};
    -O3 -gencode arch=compute_52,code=sm_52;
    )

cuda_add_executable (matrix_inverse matrix_inverse.cu)

The compile log is huge and I add only some relevant parts below:

/usr/include/eigen3/Eigen/src/LU/InverseImpl.h(335): warning: calling a __host__ function from a __host__ __device__ function is not allowed
          detected during instantiation of "const Eigen::Inverse<Derived> Eigen::MatrixBase<Derived>::inverse() const [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>]" 
/home/krr/temp_codes/eigen/matrix_inverse/src/matrix_inverse.cu(21): here

/usr/include/eigen3/Eigen/src/Core/Inverse.h(58): warning: calling a __host__ function from a __host__ __device__ function is not allowed
          detected during:
            instantiation of "Eigen::Index Eigen::Inverse<XprType>::cols() const [with XprType=Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(733): here
            instantiation of "void Eigen::internal::call_assignment_no_alias(Dst &, const Src &, const Func &) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>, Func=Eigen::internal::assign_op<float>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(712): here
            instantiation of "void Eigen::internal::call_assignment(Dst &, const Src &, const Func &, Eigen::internal::enable_if<<expression>, void *>::type) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>, Func=Eigen::internal::assign_op<float>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(693): here
            instantiation of "void Eigen::internal::call_assignment(Dst &, const Src &) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>]" 
/usr/include/eigen3/Eigen/src/Core/PlainObjectBase.h(682): here
            instantiation of "Derived &Eigen::PlainObjectBase<Derived>::_set(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, OtherDerived=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>]" 
/usr/include/eigen3/Eigen/src/Core/Matrix.h(225): here
            [ 4 instantiation contexts not shown ]
            instantiation of "Derived &Eigen::PlainObjectBase<Derived>::_set_noalias(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, OtherDerived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/PlainObjectBase.h(783): here
            instantiation of "void Eigen::PlainObjectBase<Derived>::_init1<T,OtherDerived>(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, T=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, OtherDerived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/Matrix.h(296): here
            instantiation of "Eigen::Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::Matrix(const T &) [with _Scalar=float, _Rows=6, _Cols=6, _Options=0, _MaxRows=6, _MaxCols=6, T=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/DenseBase.h(410): here
            instantiation of "Eigen::DenseBase<Derived>::EvalReturnType Eigen::DenseBase<Derived>::eval() const [with Derived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/IO.h(248): here
            instantiation of "std::ostream &Eigen::operator<<(std::ostream &, const Eigen::DenseBase<Derived> &) [with Derived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/home/krr/temp_codes/eigen/matrix_inverse/src/matrix_inverse.cu(21): here

/usr/include/eigen3/Eigen/src/Core/Inverse.h(57): warning: calling a __host__ function from a __host__ __device__ function is not allowed
          detected during:
            instantiation of "Eigen::Index Eigen::Inverse<XprType>::rows() const [with XprType=Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(733): here
            instantiation of "void Eigen::internal::call_assignment_no_alias(Dst &, const Src &, const Func &) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>, Func=Eigen::internal::assign_op<float>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(712): here
            instantiation of "void Eigen::internal::call_assignment(Dst &, const Src &, const Func &, Eigen::internal::enable_if<<expression>, void *>::type) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>, Func=Eigen::internal::assign_op<float>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(693): here
            instantiation of "void Eigen::internal::call_assignment(Dst &, const Src &) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>]" 
/usr/include/eigen3/Eigen/src/Core/PlainObjectBase.h(682): here
            instantiation of "Derived &Eigen::PlainObjectBase<Derived>::_set(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, OtherDerived=Eigen::Inverse<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>>]" 
/usr/include/eigen3/Eigen/src/Core/Matrix.h(225): here
            [ 4 instantiation contexts not shown ]
            instantiation of "Derived &Eigen::PlainObjectBase<Derived>::_set_noalias(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, OtherDerived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/PlainObjectBase.h(783): here
            instantiation of "void Eigen::PlainObjectBase<Derived>::_init1<T,OtherDerived>(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, T=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, OtherDerived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/Matrix.h(296): here
            instantiation of "Eigen::Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::Matrix(const T &) [with _Scalar=float, _Rows=6, _Cols=6, _Options=0, _MaxRows=6, _MaxCols=6, T=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/DenseBase.h(410): here
            instantiation of "Eigen::DenseBase<Derived>::EvalReturnType Eigen::DenseBase<Derived>::eval() const [with Derived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/IO.h(248): here
            instantiation of "std::ostream &Eigen::operator<<(std::ostream &, const Eigen::DenseBase<Derived> &) [with Derived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/home/krr/temp_codes/eigen/matrix_inverse/src/matrix_inverse.cu(21): here

/usr/include/eigen3/Eigen/src/Core/Solve.h(72): warning: calling a __host__ function from a __host__ __device__ function is not allowed
          detected during:
            instantiation of "Eigen::Index Eigen::Solve<Decomposition, RhsType>::rows() const [with Decomposition=Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, RhsType=Eigen::CwiseNullaryOp<Eigen::internal::scalar_identity_op<float>, Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(733): here
            instantiation of "void Eigen::internal::call_assignment_no_alias(Dst &, const Src &, const Func &) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Solve<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_identity_op<float>, Eigen::Matrix<float, 6, 6, 0, 6, 6>>>, Func=Eigen::internal::assign_op<float>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(712): here
            instantiation of "void Eigen::internal::call_assignment(Dst &, const Src &, const Func &, Eigen::internal::enable_if<<expression>, void *>::type) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Solve<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_identity_op<float>, Eigen::Matrix<float, 6, 6, 0, 6, 6>>>, Func=Eigen::internal::assign_op<float>]" 
/usr/include/eigen3/Eigen/src/Core/AssignEvaluator.h(693): here
            instantiation of "void Eigen::internal::call_assignment(Dst &, const Src &) [with Dst=Eigen::Matrix<float, 6, 6, 0, 6, 6>, Src=Eigen::Solve<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_identity_op<float>, Eigen::Matrix<float, 6, 6, 0, 6, 6>>>]" 
/usr/include/eigen3/Eigen/src/Core/PlainObjectBase.h(682): here
            instantiation of "Derived &Eigen::PlainObjectBase<Derived>::_set(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, OtherDerived=Eigen::Solve<Eigen::PartialPivLU<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_identity_op<float>, Eigen::Matrix<float, 6, 6, 0, 6, 6>>>]" 
/usr/include/eigen3/Eigen/src/Core/Matrix.h(225): here
            [ 10 instantiation contexts not shown ]
            instantiation of "Derived &Eigen::PlainObjectBase<Derived>::_set_noalias(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, OtherDerived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/PlainObjectBase.h(783): here
            instantiation of "void Eigen::PlainObjectBase<Derived>::_init1<T,OtherDerived>(const Eigen::DenseBase<OtherDerived> &) [with Derived=Eigen::Matrix<float, 6, 6, 0, 6, 6>, T=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>, OtherDerived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/Matrix.h(296): here
            instantiation of "Eigen::Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::Matrix(const T &) [with _Scalar=float, _Rows=6, _Cols=6, _Options=0, _MaxRows=6, _MaxCols=6, T=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/DenseBase.h(410): here
            instantiation of "Eigen::DenseBase<Derived>::EvalReturnType Eigen::DenseBase<Derived>::eval() const [with Derived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/usr/include/eigen3/Eigen/src/Core/IO.h(248): here
            instantiation of "std::ostream &Eigen::operator<<(std::ostream &, const Eigen::DenseBase<Derived> &) [with Derived=Eigen::Inverse<Eigen::Matrix<float, 6, 6, 0, 6, 6>>]" 
/home/krr/temp_codes/eigen/matrix_inverse/src/matrix_inverse.cu(21): here

What versions/platform of nvcc/g++/Eigen? The combination has been known in the past not to play nicely with each other... As an added bonus, can you throw in your exact command line flags? — Avi Ginsburg, Nov 22 '18 at 06:20
You say in the question that you get a "long list of errors, but I see only one error, and that is a linking error, not a compilation error. It that really what is happening? A linking error for that class member function? Because it appears to be a fully documented limitation of Eigen -- http://eigen.tuxfamily.org/bz/show_bug.cgi?id=1089 — talonmies, Nov 22 '18 at 06:50
@talonmies Except that OP has `#include ` which starts by including `Eigen/Core` and `LU`, at least in the version I checked (3.3.4). So that documented limitation shouldn't be relevant. — Avi Ginsburg, Nov 22 '18 at 07:14
All of that addition output you have added only contains compiler warnings, not errors — talonmies, Nov 22 '18 at 15:21

ggael · Accepted Answer · 2018-11-23T15:24:06.237

1

This is simply because inversion of matrices larger than 4 are not supported yet on CUDA.

To answer talonmies's comment, since the OP's file is compiled with nvcc, Eigen automatically enables CUDA support. In particular the method MatrixBase::inverse() is declared as host device because if the underlying expression is 4x4 or smaller then it will branch, at compile-time, to CUDA compatible code. For larger ones, it branches to non-cuda compatible code... So the consequence is that CUDA and non-CUDA non-trivial code better has to be well separated in .cu and .cpp files.

edited Nov 23 '18 at 15:24

answered Nov 22 '18 at 20:31

ggael

28,425
2
65
71

I completed the analysis. – ggael Nov 23 '18 at 15:24

different behavior for eigen::matrix inversion when compiled with gcc and nvcc

1 Answers1

Linked