CuGPU not defined error #244
Comments
|
Did you try |
|
Yes, this is with |
|
Ok may need DynamicGrids.CuGPU |
|
I just tried I simplified it to the simplest code that gives me the error, which is: and results in: The same thing happens with either |
|
If you are on julia-1.10.2, then I could reproduce, and it seems to require the same workaround as in #243: (@v1.10) pkg> activate --temp
Activating new project at `/tmp/jl_7aYPFi`
(jl_G6yIEw) pkg> add DynamicGrids#dev, CUDA
julia> DynamicGrids.CuGPU()
CuGPU{32}() |
|
But if you are trying the README, you'll then face julia> using CUDAKernels, CUDA
│ Package CUDAKernels not found, but a package named CUDAKernels is available from a registry.
│ Install package?
│ (jl_G6yIEw) pkg> add CUDAKernels
└ (y/n/o) [y]:
Resolving package versions...
ERROR: Unsatisfiable requirements detected for package CUDAKernels [72cfdca4]:
CUDAKernels [72cfdca4] log:
├─possible versions are: 0.1.0-0.4.7 or uninstalled
├─restricted to versions * by an explicit requirement, leaving only versions: 0.1.0-0.4.7
└─restricted by compatibility requirements with KernelAbstractions [63c18a36] to versions: uninstalled — no versions left
└─KernelAbstractions [63c18a36] log:
├─possible versions are: 0.1.0-0.9.18 or uninstalled
└─restricted to versions 0.9 by DynamicGrids [a5dba43e], leaving only versions: 0.9.0-0.9.18
└─DynamicGrids [a5dba43e] log:
├─possible versions are: 0.21.4 or uninstalled
└─DynamicGrids [a5dba43e] is fixed to version 0.21.4 |
|
We should really release a new version. The problem is its a breaking version and I'm not sure all the breaks make sense, and there are likely a few more on the way. But I have been working on branches for over a year now, So |
|
Indeed, the breaking changes made me give up for now, as I just wanted to show your awesome package to a colleague.
Nice, and just to be clear, #244 (comment) was with |
|
Yeah... my aim is to get work back on main and registered soon. But my other more general packages have been taking most of my time, plus PhD constraints. I actually use this constantly in all of my work, but not in a stable-releases kind of way... |
|
Thank you! I had been using an older version of Julia (1.7.3) because it was available on the cluster I was using, but I installed 1.10.2 myself, and used |
|
Reopening because I thought my further errors were unrelated, but now I'm thinking they are related. If it's not related, I'm glad to separate this out. Any insight on this? Am I missing some change I need to make to use the GPU? All I have changed is the value of I'm using this code, which works with Full error: |
|
Ok. So first make sure your model runs fine with e.g. Then you need to be 100% sure all of the code inside your Those errors all suggest that things like this are happening. You really dont want to allocate inside To fix this I suggest running your CPU simulations for a few seconds using ProfileView.jl Any red (instability) or yellow (allocations) patches in the inner function calls will mean that your model wont run on a GPU. Cthulhu.jl can help, you can click on the ProfileView.jl profile and call There is a method As for the cluster, get your manager to add 1.10, its much better :) (honestly here I never end up using a cluster because with no allocations my desktop is happy to run my sims a million times!) |
|
Oops you had linked it. This is certainly a problem, allocating an upordown=sample([1,-1])use StaticArrays.jl, and: sample(@SVector[1, -1])As is: collect(1:8)Just use Basically, you should never allocate or write to any array inside a rule other than writing to a grid with e.g. |
|
This is also bad, what is the reason to do this? The neighborhood is alread a StaticArrays nbrvalues = [neighborhood[1],neighborhood[2],neighborhood[3],neighborhood[4],neighborhood[5],neighborhood[6],neighborhood[7],neighborhood[8]] |
|
This nv=p*nbrvalues./(p*nbrvalues+fill(cell*(1-p),8)) nv=p.*neighborhood./(p.*neighborhood.+cell.*(1-p) Also broadcast everything else so it fuses. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I am trying to change my simulation to use GPUs rather than just CPUs. I changed
sim!(output, spec_rule,nbr_rule; proc = ThreadedCPU(), boundary=Wrap())to use "CuGPU" instead of "ThreadedCPU", but then I get the error:nested task error: UndefVarError: CuGPU not definedWhat other changed are required for
CuGPU()to be recognized?The text was updated successfully, but these errors were encountered: