As far as I know, the math library currently in use by pyfr is gimmik and blas. I noticed that nvidia has a matrix multiplying tool that uses the sparse feature of the Ampere architecture called cuSPARSELt. Can it be used in PyFR?
No, rthis can’t be used in PyFR. There are two questions to consider here:
-
Which of the existing kernels in PyFR aren’t hitting the roofline model? (In my experience not many, you generally need to be pushing the order quite high to see performance degradation from the model.)
-
Which elements are you thinking are targeting? Only really the hex, quad and to some extend prism elements are sparse, but sparse libraries typically require some special representation like CSR. Converting to this representation has a non-zero overhead which will erode any possible wins (again refering to point 1, if there even any wins to be had).
In my view, if you are interested in how sparse matrices can be handled better in pyfr (which generally means when you are highly strong scaled), the best place to focus effort is in gimmik.