Using Numba's code generation implies the following:
users are encouraged to implement kernels assuming data is scalar type. vectorization is given to numba.vectorize
structured arrays are well supported at least for CPU version
all arrays with cuda_array_interface will be supported by the kernels.
wrappers need to be added to wrap methods so that the accessed attributes are exposed as the arguments to the function (e.g. Model.ode, Model.update...).
there is numba.jitclass decorator to handle class definitions but it's currently experimental only.
Using Numba's code generation implies the following:
numba.vectorize
Model.ode
,Model.update
...).numba.jitclass
decorator to handle class definitions but it's currently experimental only.