Complex dense matrix inversion using CUSP

[nicolar...@gmail.com]

Hello,

I'm a new CUSP user and I'm wondering if there's a way to evaluate the inverse of a dense complex valued matrix (by means of LU decomposition or other methods) on device. 

If there is one, can you provide some guidance on how to proceed?

Thank you.

Nicola

[Steven Dalton]

Hello Nicola,

  There are no methods in CUSP to efficiently invert dense matrices on the GPU. There are some interface routines to compute the LU of a batch of matrices directly from cublas[1] but I think that method works best on a collection of relatively small matrices and not one large matrix. Magma has some simple interfaces for computing factorizations on GPUs[2]. You can also take a look at cuSolver[3] since it has some new solvers for dense matrices. I can't recommend any particular library since I haven't tried them out.

Steve

[1] http://docs.nvidia.com/cuda/cublas/#cublas-lt-t-gt-getrfbatched

[2] http://icl.cs.utk.edu/projectsfiles/magma/doxygen/group__solvers.html

[3] http://docs.nvidia.com/cuda/cusolver/#axzz4EEPBx6s1

--
You received this message because you are subscribed to the Google Groups "cusp-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cusp-users+...@googlegroups.com.
To post to this group, send email to cusp-...@googlegroups.com.
Visit this group at https://groups.google.com/group/cusp-users.
For more options, visit https://groups.google.com/d/optout.

[nicolar...@gmail.com]

Hello Steve,

Thank you very much for your help. Are the cusp::lapack functions only available for host computations?

If I were to use one of those libraries you mentioned for the computation, can you use the cusp::array2d containers in some fashion?

I would still like to use cusp for some sparse operations, but in my application there are also dense operations to be computed. 

Thank you.

Nicola

[Shiyin Kang]

For large dense matrix inversion, you could use getrf + getrs from cuSolver.

This question and answer may help.

http://stackoverflow.com/questions/37731103/cublas-matrix-inverse-much-slower-than-matlab/37733544#37733544

[Steven Dalton]

Yes the cusp::lapack namespace is thin wrapper around traditional host/cpu routines in the Lapacke interface.

You can definitely connect cusp::array2d containers to any external libraries you find useful. It's a bit low level but you can see how cusp interacts with external routines by looking at functions defined in the cusp::lapack namespace[1]. Generally you need to make sure the layout of the array2d is consistent with the external library and then pass the data and leading dimensions.

Steve

[1] https://github.com/cusplibrary/cusplibrary/blob/develop/cusp/lapack/detail/generic.h#L37

--