search

juliamatlab / mexjulia

embedding Julia in the MATLAB process.
MIT License
52 stars 14 forks source link

extract common functionality from MATLAB.jl into a separate package and depend on that #20

Closed twadleigh closed 5 years ago

twadleigh commented 7 years ago

Probably called MxArrays or MxBase.

twadleigh commented 7 years ago

As the marshaling code copied over from MATLAB.jl evolves, the common ground between it and mexjulia is shrinking. It may be a bit before we can assess to what extent this is still a useful thing to do.

musm commented 7 years ago

I think it would also be nice to just rely on MATLAB to take care of the functionality needed? Looking over some of the files in mexjulia it looks like you based them on an older fork of MATLAB, which had many issues with segfaulting.

twadleigh commented 7 years ago

@musm, I will check out the recent patches on MATLAB.jl and apply them. However, mexjulia currently relies on patches in the mxarray code that don't make sense in the MATLAB.jl context.

twadleigh commented 7 years ago

For instance, one very invasive change that had to be made for mexjulia was the elimination of the own field. In the mexjulia context, MATLAB deletes the memory for unused objects itself when the mex function returns. To let Julia participate at all in the memory management of MxArrays will require careful object tracking and proper scoping constructs. For now, it leaves it all to MATLAB.

musm commented 7 years ago

MxArray( ..., false) would disable the own and let you manage it without having Julia call a finalizer.

https://github.com/JuliaInterop/MATLAB.jl/blob/master/src/mxarray.jl#L7

But perhaps that is not as sophisticated as you require.

twadleigh commented 7 years ago

The story gets more complicated, for instance, when you allow the user to make values persist across calls to mexjulia via mexMakeArrayPersistent. To make matters worse, mexMakeArrayPersistent makes values that can't be passed back to MATLAB (without a copy). Managing both scoped and persistent values in a user-friendly way will require a lot of machinery that doesn't apply to the MATLAB.jl context.