CudaMalloc and CudaFree Optimization

[jtristano]

CudaMalloc and CudaFree Optimization

[pazner]

Hi @jtristano,

In general, many MFEM objects try to keep track of and reuse the workspace/temporary vectors that they need in order to minimize allocations. For example, MultigridBase::InitVectors is used to create the temporary vectors needed in the multigrid V-cycle so that they aren't allocated on every call to Mult.

Can you track down specifically where the allocations are happening in AddMult or AddMultTranspose? This seems like something that could be addressed.

Also, MFEM can use Umpire, which has support for memory pools, which can speed up allocations and memory management on the GPU.

[jtristano]

Thanks! After some review, I found that ConstrainedOperator doesnt use the z vector for AddMult or AddMultTranspose.. I'm overriding locally for now