Magma_1.3 multi-gpu routines

[jeremiahpalmer]

Hello! I see in the magma_1.3 presentation pdf mentions of multi-gpu dsytrd routines. Where can I find the routines that were used to produce those results? Are they referenced in the testing directory?

Thanks,
Jeremiah

[jeremiahpalmer]

...Or maybe they will appear in a later magma version? I am eager to use this functionality.

[mgates3]

Sorry, they weren't quite ready and didn't make the release. They should come out in another release soon.
-mark

[jeremiahpalmer]

Thanks for the info.

[neuron]

Hi,
is it present a multi-gpu implementation of SGEMM in Magma 1.3 ?
If not do you think to add it in the future?

Regards

[mgates3]

No, though a multi-GPU gemm can be easily fashioned out of regular cublas gemm calls on each GPU. It really depends on how the matrices A, B, and C are distributed, e.g., by block-rows or block-columns.

Whereas for the symmetric case, we have a multi-GPU symv and are working on symm, because those cannot be easily composed from cublas symm or gemm calls.
-mark