MAGMA Forum

Hello,

I am currently working on a problem where I need to compute the LU factorization, on the GPU using MAGMA, of a large matrix (20000 x 20000 entries, or more), which is too big for the memory that we have on our GPU. Is there any support in MAGMA for doing the LU factorization of matrices of this size? For instance, by blocking the LU factorization into chunks that fit the GPU memory and overlapping computation and data transfer, or similar.

Afterwards I will also use the factorization to solve, so I am interested in the functions:
magma_dgetrf_gpu, and magma_dgetrs_gpu

Best Regards,
Jacob

Hi,

If the matrix is too large to be stored on your GPU, then magma_dgetrf automatically calls its out-of-GPU-memory implementation. So, the size of the matrix that can be factorized using magma_dgetrf is limited by the size of the CPU memory. If you are using MAGMA 1.4.0 beta, please see the bug reports in the MAGMA Forum post "magma-1.4.0-beta2: numerical problems [dz]getrf."

Since our GPU interface (e.g., magma_dgetrf_gpu) takes the input matrix on the GPU, it probably does not make sense if the matrix cannot be stored on the GPU? =)

We currently do not provide an out-of-GPU-memory implementation of dgetrs. For now, I will call LAPACK. If this becomes bottleneck, then please let us know.

Thank you,
Ichi