Installation and compilation issues with anuga 3.2.0.dev version.

[yilou1]

Hello，
I'm having a problem compiling and installing anuga version 3.2.0.dev. I tried installing and compiling in both windows environment and linux environment, both show success without any error report or warning. But there is a problem when I actually use the CUDA method.

from anuga.shallow_water.sw_domain_cuda import nvtxRangePush, nvtxRangePop

ModuleNotFoundError Traceback (most recent call last)
Cell In[3], line 5
1 import anuga
2 import anuga.shallow_water
----> 5 from anuga.shallow_water.sw_domain_cuda import nvtxRangePush, nvtxRangePop

ModuleNotFoundError: No module named 'anuga.shallow_water.sw_domain_cuda'

dir(anuga.shallow_water)
['builtins',
'cached',
'doc',
'file',
'loader',
'name',
'package',
'path',
'spec',
'boundaries',
'checkpoint',
'forcing',
'shallow_water_domain',
'sw_domain_orig_ext',
'sww_interrogate',
'test']

Show compilation success

(anuga_env_3.11)H:\anuga_core>pip install --no-build-isolation -eObtaining file:///H:/anuga_coreChecking if build backend supports build_editable ... donePreparing editable metadata(pyproject.toml)... doneRequirement already satisfied: numpy>=1.24.0 in c:(usersllenovol.condalenvs(anuga_env_3.11\lib\site-packages (from anuga=3.2.0.dev0)(2.2.6)Building wheels for collectedpackages:anugaBuilding editable for anuga(pyproject.toml)... doneCreated wheel for anuga: filenameanuga-3.2.0.dev0-cp311-cp311-win_amd64.whl size=7822 sha256-2b56623eadaf2f0086e36336f095754dc6b05176ae12604501b8bf26d5cd2616Stored in directory: c: Users Lenovo\AppDatalLocallTemplpip-ephem-wheel-cache-4xlp4xzf\wheels(a1 cf\61\79593d87a34e12fb61ab63d77f9791bc76c763bec5acc0fцfa
Successfully built anugaInstalling collected packages:anugaSuccessfully installed anuga-3.2.0.dev

Is it possible to provide a reference to a successful compilation environment, or I did not see a package like cudatoolkit in the environment requirements to run the cuda base, is this similar to the requirements for the GPU, do I need to pay extra attention to anything for running with cuda.

[stoiver]

@yilou1 seems there is a bug when importing anuga.shallow_water.
We are working on the GPU implementation of anuga, but if cupy is not installed I didn't think anuga.shalkow_water would complain.

Probably can get around your problem by just importing directly from anuga

[stoiver]

@yilou1 how did you generate that error? The only place I can find that piece of code is in some of the run scripts in the shallow_water/tests directory (which probably shouldn't be there in the main branch). Is it one of those scripts that threw the exception? If you are just using the default branch (main) of anuga_core, then the file sw_domain_cuda.py is not even installed.

The actual branch we are working on is develop_gpu You would have to install the appropriate cuda libraries and install the cupy python module to compile anuga. But it is definitely a work in progress, and we dont have a version which is passing the unit tests at present.

[yilou1]

@stoiver Hello, using git to clone version 3.2.0 because I see scripts for cuda tests in the #anuga/examples/cuda directory, I compile and run the #anuga/examples/cuda script with this error. I followed your suggestion and anuga version 3.2.0 is working fine under main branch, similar to version 3.1.9 functionality. This one is a problem I'm having trying the cuda module to speed it up.

[stoiver]

@yilou1 we probably shouldn't have let that cuda example sneak into the example directory on the main branch. There is no cuda support in the main branch of anuga_core at the moment. We are working on it in the develop_gpu branch of anuga_core available from the anuga-community github repository.

If you are working on linux, there is some openmp support in the main branch. We can probably get openmp to work on apple. On windows we use the complier mingw32 which has only basic openmp support which is not good enough for our current code.

To use openmp, add the command

domain.set_multiprocessor_mode(2)

to your script. That should provide a speedup of your evolve loop.

[yilou1]

@stoiver Thanks for your help, I'll try the domain.set_multiprocessor_mode(2) speedup effect on linux. Since my simulation has a large simulation area, the mesh size is set to 1000 m^2 and the number of meshes is around 10 million, despite using 128 cores the speed is still not satisfactory. Very much looking forward to the cuda version of anuga. Thanks again!