While trying out Eigen 3.3 (ubuntu 16.04) I ran into:

/usr/include/eigen3/Eigen/src/SparseCore/SparseTranspose.h:59:63: error: no type named ‘ReverseInnerIterator’ in ‘struct Eigen::internal::evaluator<Eigen::CwiseUnaryOp<Eigen::internal::scalar_cast_op<unsigned char, float>, const Eigen::Block<Eigen::Map<Eigen::SparseMatrix<unsigned char, 1, int>, 0, Eigen::Stride<0, 0> >, 1, -1, true> > >’

With the code below.
Admittedly, the code isn't exactly a beaty.
So I'm wondering, did I just shoot myself in the foot, or is this actually supposed to work ?

Code: Select all

template<typename Vector>
void SolveNonLinear(Vector& b)
{
   b.squaredNorm();
   // more code snipped
}

void demo(int row)
{
Eigen::MappedSparseMatrix<Tdata, Eigen::RowMajor, Spectrum::DynamicSparseMatrix::Index> X(slice.rows, slice.cols, slice.nnz, slice.pOuterIdx, slice.pInnerIdx, slice.pValue);
auto bSparse = X.row(row).template cast<float>();
auto pTrans = bSparse.transpose();
SolveNonLinear(pTrans);
}


If I remove the transpose(), it compiles and links.
With the transpose() the code compiles and works fine with eigen 3.2.4

good catch, the issue is triggered only with the combinaison of both cast and transpose.... I'll fix it asap.

in the meantime, you can workaround by swapping the cast and transpose:

X.row(i).transpose().cast<float>().squaredNorm()

ggael wrote:in the meantime, you can workaround by swapping the cast and transpose:

X.row(i).transpose().cast<float>().squaredNorm()

Thanks!

For the record, this precise issue has been fixed prior to the 3.3 release.