Tiled Matrix Matrix Multiply

[eportillo2]

Hello,

I have been trying to find a parallel (GPU+CUDA) tiled SGEMM, to solve big matrices which otherwise would not fit in GPU memory.

A friend of mine said he once saw a Magma version of tiled SGEMM, but I can't find it.

Does such implementation exist?

Thanks for your time...

[mgates3]

Yes, there were tiled algorithms in MAGMA. These have been split off into the MORSE project.
http://icl.cs.utk.edu/morse/software/

Note that the regular MAGMA gesv (getrf) and posv (potrf) will handle matrices larger than the GPU memory, as well as multiple GPUs.

-mark