[ANN] DynamicQuantities.jl: type stable physical quantities

[MilesCranmer]

@non-Jedi

Interestingly, operations on heterogeneous arrays only have a 27x advantage for DynamicQuantities for g. I was expecting something more similar to the 1,500x gap for the non-inferred homogeneous case.

In general I think this would be best done with StructArrays.jl, so that the dimension-specific operations can be vectorized too. I’d be very curious to see how that changes your measurement.

Perhaps this is also because the dimensions are stored as SimpleRatio{Int} rather than an Int8 (which is more amenable to vectorization)? @Oscar_Smith was suggesting this in the initial thread based on some measurements:

@Oscar_Smith

Using @btime randn(5) .* Dimensions(mass=2/5, length=2) as a benchmark, I see a pretty big improvement going to Int8 instead of Int. It won’t make much difference for scalars, but for vectors, it cuts allocations by 75% and time by 19%.

It’s likely that few people need more than an Int8 for physical dimensions. The tricky part is dealing with integer overflows if we stick with SimpleRatio: even four repeated + (2//3) - (2//3) would be enough to overflow Int8. (Since SimpleRatio doesn’t repeatedly divide by the gcd)