Can I use cuBLAS directly in AMReX? (or are there alternatives?)

[tony3601000]

Howdy everyone,

I'm relatively new to AMReX. I want to implement a simple matrix-matrix multiplication function using cuBLAS within AMReX; however, I have concerns about memory management since this is usually done automatically with arenas. Otherwise, what alternatives are there for implementing GPU-accelerated matrix-matrix multiplication (using Nvidia GPUs in particular)? Thank you for the help!

[WeiqunZhang]

You should be able to use cuBLAS. As for memory management, you can replace

double* p;
cudaMalloc(&p, size_in_bytes);
dudaFree(p);

with

auto* p = (double*) The_Arena()->alloc(size_in_bytes);
The_Arena()->free(p);