Hi,

I found Eigen when I was looking for a library to solve small (4x4 symmetric) Eigenvalue problems in Cuda kernels. Eigen looks like the perfect library for my purpose.

I tried the 2x2 example and it works fine. Unfortunately, it seems to me that only the computeDirect() method is supported in Cuda.
Trying to solve the 2x2 example with compute() does not work.
Is this the expected behaviour or am I missing something?

If this is not yet supported: Can you give me a rough estimate of the work that is needed to support 4x4 matrices.
I am at the point where I have to decide whether to write my own small 4x4 eigensolver to go on with my project or if I should put some effort in understanding Eigen to help porting it to Cuda.

yes, that's the current expected behavior. The support for nvcc is a work in progress stuff. Do you need all eigenvalues/vectors? or only the largest or smallest?

I need the eigenvector with the largest eigenvalue.

then a simple Power iteration method might be more GPU friendly. search for power iteration on this forum to get a source code example.

Ok. Thank you.