Metal Backend type compatibility

[tudorgmo]

Hi,

I've been trying to use the package with the Metal (Apple Silicon) backend, but I run into issues due to the fact that Metal can only handle Float32 values, not Float64. Even when explicitly setting precision with set_precision(Float32), as well as defining all simulation parameters in the Float32 type, it seems like some of the internal parameters that get passed around are of type Float64 (such as time values t in field.jl). Would it be possible to look into fixing this error and making this compatible with Apple Silicon?

Thanks!

[ww1g11]

Hi @tudorgmo,

Yes, the Metal backend only supports Float32.

Do you get an error when setting set_precision(Float32)? We will run Metal tests in a future release.

You are right that even when the GPU is set to float32, some internal parameters such as the time in the effective_field are still Float64. The main reason is that the time is defined in the CPU and its time dependence function is also running on the CPU, so we haven't changed it to float32. Maybe it's a good idea to keep everything consistent, and we will look into this problem.

best wishes,