Reducing parameter space usage by `LatitudeLongitudeGrid`

[glwagner]

It's come to light that the LatitudeLongitudeGrid consumes almost 1 kb of parameter space as an argument on the GPU. This is a problem because at least for some versions of CUDA + GPUs (unsure how much is due to the GPU itself vs the CUDA version) we only have 4 kb of parameter space.

I'm wondering if we can save space by removing the latitude and longitude spacings Δλ and Δφ:

Oceananigans.jl/src/Grids/latitude_longitude_grid.jl

Lines 369 to 398 in 4076823

	function Adapt.adapt_structure(to, grid::LatitudeLongitudeGrid)
	TX, TY, TZ = topology(grid)
	return LatitudeLongitudeGrid{TX, TY, TZ}(nothing,
	grid.Nx, grid.Ny, grid.Nz,
	grid.Hx, grid.Hy, grid.Hz,
	grid.Lx, grid.Ly, grid.Lz,
	Adapt.adapt(to, grid.Δλᶠᵃᵃ),
	Adapt.adapt(to, grid.Δλᶜᵃᵃ),
	Adapt.adapt(to, grid.λᶠᵃᵃ),
	Adapt.adapt(to, grid.λᶜᵃᵃ),
	Adapt.adapt(to, grid.Δφᵃᶠᵃ),
	Adapt.adapt(to, grid.Δφᵃᶜᵃ),
	Adapt.adapt(to, grid.φᵃᶠᵃ),
	Adapt.adapt(to, grid.φᵃᶜᵃ),
	Adapt.adapt(to, grid.Δzᵃᵃᶠ),
	Adapt.adapt(to, grid.Δzᵃᵃᶜ),
	Adapt.adapt(to, grid.zᵃᵃᶠ),
	Adapt.adapt(to, grid.zᵃᵃᶜ),
	Adapt.adapt(to, grid.Δxᶠᶜᵃ),
	Adapt.adapt(to, grid.Δxᶜᶠᵃ),
	Adapt.adapt(to, grid.Δxᶠᶠᵃ),
	Adapt.adapt(to, grid.Δxᶜᶜᵃ),
	Adapt.adapt(to, grid.Δyᶠᶜᵃ),
	Adapt.adapt(to, grid.Δyᶜᶠᵃ),
	Adapt.adapt(to, grid.Azᶠᶜᵃ),
	Adapt.adapt(to, grid.Azᶜᶠᵃ),
	Adapt.adapt(to, grid.Azᶠᶠᵃ),
	Adapt.adapt(to, grid.Azᶜᶜᵃ),
	grid.radius)
	end

these are used when metrics are not precomputed. So, removing them would require either discontinuing support for non-precomputed-metrics (I'm ok with this), or at least making it so that we can't build a grid on a GPU without precomputing metrics (fine too).

We don't need either Δλ or Δφ inside kernels, right?

Unfortunately, I do not think that the same kind of savings is possible for OrthogonalSphericalShellGrid.

[glwagner]

cc @simone-silvestri

[simone-silvestri]

hmmm, I don't think that precompute_metrics=false is extensively used. Also, the memory benefit obtained by not precomputing the metrics does not outweigh the speedup obtained by precomputing the metrics.

For this reason, and the added advantage of removing maintenance burden, I tend to prefer discontinuing the support for precompute_metrics=false.

Probably the only reason to keep support is for debugging reasons?

[glwagner]

It'd be easy as an intermediate step to discontinue GPU support only I suppose. Depends how much we want to simplify the code...

Also this is a reason not to eliminate LatitudeLongitudeGrid because it seems like it will always have fewer parameters than OrthogonalSphericalShellGrid.