CliMA / ClimaTimeSteppers.jl

A CPU- and GPU-friendly package for solving ordinary differential equations
Apache License 2.0
49 stars 4 forks source link

Improve fused increment design #251

Open charleskawczynski opened 7 months ago

charleskawczynski commented 7 months ago

We could probably improve the fused increment design in two ways:

Since the bit information is known at compile-time, we could use a recursion to create the broadcast expression (so that it remains statically inlined).

charleskawczynski commented 7 months ago

250 fixed the inference issue, and implemented the kernel fusion.