Added new optimised path for setup phase

NVIDIA / AMGX

Distributed multigrid linear solver library on GPU

468 stars 136 forks source link

Added new optimised path for setup phase #280

Closed mattmartineau closed 7 months ago

mattmartineau commented 8 months ago

Add use_opt_kernels behind which many of the classical and aggressive classical setup routines have optimised versions (tested up to H100)
Add use_cusparse_kernels to override all custom SpGEMM operations with cuSPARSE
- Perf currently excellent, but memory utilisation can be high
Restructured to remove many references to old arch versions
- Ideally, all future specialisation will be handled with preprocessor

mattmartineau commented 7 months ago

There are many ongoing optimisations within the sparse operators. For now, those sparse operators are tested with our internal test suite of matrices, but I am hiding the optimisations behind a new generic flag use_opt_kernels, until they are externally tested and promoted for the release of v2.5.0.

I am removing support for CUDA 10.