search

ORNL / ReSolve

Library of GPU-resident linear solvers
Other
57 stars 2 forks source link

Abstract GPU kernels from the algorithm code #87

Closed pelesh closed 7 months ago

pelesh commented 11 months ago

Prototypes for GPU kernel wrappers should not depend on vendor-specific compiler. See comment below:

I don't think we should manage vendor specific includes at this level. The random sketching function should be hardware agnostic.

_Originally posted by @pelesh in https://github.com/ORNL/ReSolve/pull/83#discussion_r1408081274_

pelesh commented 10 months ago

@kswirydo, consider implementing FWHT kernels in a similar way we implemented VectorKernels (i.e. the vector kernel; we have only one here :)). The downside of this approach is we can build only one backend per build (i.e. cannot build HIP and CUDA at the same time).

Alternatively, we could use similar approach we used for handlers. That would be a little more work, but would be a more flexible solution.

pelesh commented 7 months ago

Addressed in #147