Support for user-defined forcing functions. #85
Conversation
|
Should we just try to resolve #59 now? |
src/forcing.jl
Outdated
| S::TS | ||
| end | ||
|
|
||
| @inline zero_func(u, v, w, T, S, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) = 0 |
There was a problem hiding this comment.
Is it slower to write @inline zero_func(args...) = 0? This is less explicit, but does not require maintenance.
| @inbounds Gu[i, j, k] = -u∇u(u, v, w, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) + fCor*avg_xy(v, Nx, Ny, i, j, k) - δx_c2f(pHY′, Nx, i, j, k) / (Δx * ρ₀) + 𝜈∇²u(u, 𝜈h, 𝜈v, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) | ||
| @inbounds Gv[i, j, k] = -u∇v(u, v, w, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) - fCor*avg_xy(u, Nx, Ny, i, j, k) - δy_c2f(pHY′, Ny, i, j, k) / (Δy * ρ₀) + 𝜈∇²v(v, 𝜈h, 𝜈v, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) | ||
| @inbounds Gw[i, j, k] = -u∇w(u, v, w, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) + 𝜈∇²w(w, 𝜈h, 𝜈v, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) | ||
| @inbounds Gu[i, j, k] = -u∇u(u, v, w, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) + fCor*avg_xy(v, Nx, Ny, i, j, k) - δx_c2f(pHY′, Nx, i, j, k) / (Δx * ρ₀) + 𝜈∇²u(u, 𝜈h, 𝜈v, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) + F.u(u, v, w, T, S, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) |
There was a problem hiding this comment.
What happens if you try to pass grid instead of Nx, Ny, Nz, Δx, Δy, Δz?
There was a problem hiding this comment.
On the GPU it complains that the arguments are not isbitstype :(
There was a problem hiding this comment.
Got it. So both Fields and Grids will have to be worked on.
|
I think there are already a couple of challenges for this pull request:
As #59 will result in a ton of refactoring, I think it might be more appropriate to address it in a different branch after #73 is resolved, so that everything ( If doing so results in large performance drops (which I think might have happened here) then it might take a while to figure out #59. |
src/forcing.jl
Outdated
| @inline zero_func(u, v, w, T, S, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k) = 0 | ||
|
|
||
| function Forcing(Tu, Tv, Tw, TT, TS) | ||
| if Tu == nothing |
There was a problem hiding this comment.
Use === when comparing against nothing
|
Performance is fine now so merging in. Thanks @glwagner! |
Higher-order advection schemes
Just hacked something together that allows for user-defined forcing functions for the CPU. Have not tested on the GPU yet.
Basically there's a struct
Forcingthat stores the user-defined forcing functions. It will replace the oldForcingFieldsstruct. Seeexamples/deep_convection_3d.jlfor how I switched to using a forcing function for T to enforce a cooling surface heat flux.A big issue is that the current implementation slows down the time stepping by a factor of 2-3x. So we'll have to figure out why before merging.
The function must have a signature like
F(u, v, w, T, S, Nx, Ny, Nz, Δx, Δy, Δz, i, j, k)right now so this won't produce a nice solution as we will have to figure out #59 before the function signature can look as nice assurface_cooling_disk(grid, velocities, tracers, i, j, k). This is work for another branch.Will keep working on this before merging. Just wanted to start something.
Resolves #73