attempt to compile AOT function fails

[rbavery]

When I run the aquacrop-streamlit app, I get this error

RuntimeError: Attempted to compile AOT function without the compiler used by `numpy.distutils` present. If using conda try: #> conda install gcc_linux-64 gxx_linux-64

I'm setting up dependencies in a venv and not using conda. This may mean we need to add the gcc compiler dependency for a specific architecture family to the aquacrop dependencies: https://numba.pydata.org/numba-doc/dev/user/pycc.html

[rbavery]

It looks like the gcc dependencies can't be installed via pypi, only conda...

This indicates that if we distribute AOT compiled modules we will need to switch to a conda based setup and document that, instead of using pip in order to handle the compiler dependency. Or am I missing something?

[thomasdkelly]

Have you tried running the solution.py file locally first, then run pip install -e . again

Basically just overwriting the solution_aot file included in the package

[thomasdkelly]

I have also had issues when testing the new pypi upload, where it cant locate the solution_aot.so file even though it is inlcuded in the pip package

[rbavery]

I did not try that, but thanks for mentioning it. What I did instead was set up a conda environment that installed the gcc dependencies. Then I locally installed aquacrop with pip and I think I already had the compiled file so it worked. So I'm not sure but it seems like maybe gcc is a requirement to run the compiled functions?

[arongergely]

Shouldn't the code be compiled on target (user's system) during the pip install? Otherwise we would need to ship compiled code for each architecture and python version..

[thomasdkelly]

Have finally got the pypi package upload/download working. Was running into all kinds of errors while using the methods described here https://numba.pydata.org/numba-doc/dev/user/pycc.html.

One of which is similar to the point made by @arongergely that the compiled binarys only work for my local pc. pip was having a hard time building these into a wheels and so we now have to build a source distribution instead of wheels. The other reason to dot build a wheels is that we want to run a compilation script as the user pip installs the module so that the binary works for their architecture as @arongergely said.

Will create a pull request tomorrow that does this as well as creating a github action to push to pypi following a new version release.

Can go into more detail in that pull request if you have any more questions about the wheels erros etc.

[thomasdkelly]

And to your point @rbavery about gcc dependecy: I can pip install (and then compile) inside a fresh google colab using the code below without issue so I dont think we are assuming any crazy dependencies.

!pip install numba==0.55
!pip install -i https://test.pypi.org/simple/ aquacrop==0.4.18

from aquacrop.classes import    *
from aquacrop.core import       *
path = get_filepath('champion_climate.txt')
wdf = prepare_weather(path)

def run_model(smts,max_irr_season,year1,year2):
    maize = CropClass('Maize',PlantingDate='05/01') 
    loam = SoilClass('ClayLoam') 
    init_wc = InitWCClass(wc_type='Pct',value=[70])
    irrmngt = IrrMngtClass(IrrMethod=1,SMT=smts,) 

    # create and run model
    model = AquaCropModel(f'{year1}/05/01',f'{year2}/10/31',wdf,loam,maize,
                          IrrMngt=irrmngt,InitWC=init_wc)
    model.initialize()
    model.step(till_termination=True)
    return model.Outputs.Final

run_model([70]*4,300,2018,2018)

[thomasdkelly]

42