Bug with `SelfAdjointEigenSolver..eigenvalues().asDiagonal()` on CUDA

Summary

While Eigen::SelfAdjointEigenSolver works fine on CUDA and the eigenvalues() method returns the right results, SelfAdjointEigenSolver.eigenvalues().asDiagonal() returns an all-zero matrix.

Environment

The bug shows up on both Windows and Ubuntu (under WSL)

• Operating System : Windows/Linux
• Architecture : x64
• Eigen Version : from master
• Compiler Version : VS 2022/NVCC 12.6 and GCC 11.4.0/NVCC 12.5

Minimal Example

The following code shows a minimal code for the bug https://github.com/Ahdhn/SelfAdjointEigenSolver

__global__ void kernel()
{
    using MatT = Eigen::Matrix3f;

    MatT H;
    H << 1, 0, 0,  //
        0, 2, 0,   //
        0, 0, 3;

    Eigen::SelfAdjointEigenSolver<MatT> eig(H);

    printf("\n Eigenvalues: %f, %f, %f\n ",
           eig.eigenvalues()[0],
           eig.eigenvalues()[1],
           eig.eigenvalues()[2]);

    MatT D = eig.eigenvalues().asDiagonal();

    printf("\n Eigenvalues as diag: %f, %f, %f\n ", D(0, 0), D(1, 1), D(2, 2));
}

Steps to reproduce

git clone https://github.com/Ahdhn/SelfAdjointEigenSolver.git
cd SelfAdjointEigenSolver 
mkdir build
cd build 
cmake ..

Depending on the system, this will generate either a .sln project on Windows or a make file for a Linux system.

What is the current bug behavior?

eig.eigenvalues().asDiagonal() returns/prints an all zero matrix.

What is the expected correct behavior?

eig.eigenvalues().asDiagonal() should return/print a diagonal matrix with 1,2,3 on the diagonal.

• Have a plan to fix this issue.