hgrecco
commented
4 years ago With current Numba (0.51.2) the following code:
import numba as nb
@nb.njit(cache=True)
def func(t, y):
return y
@nb.njit(cache=True)
def stepper(f, a, b):
return 3 * f(a, b)
print(stepper(func, 4., 2.))fails with the following error:
Traceback (most recent call last):
[...]
TypeError: cannot pickle 'weakref' objectdue to cache=True.
If this works in the near future, instantiation time can drop dramatically as compilation will be done once. This might require some code reorganization.
For example, right now the stepper is organized in two layers:
def step_builder(*outer_args):
"""Build a stepper.
This outer function should only contains attributes
associated with the solver class not with the solver instance.
"""
@numba.njit
def _step(*inner_args):
"""Perform a single step.
This inner function should only contains attributes
associated with the solver instance not with the solver class.
"""
# code to step
return _stepThe step_builder is called at class instantiation, _step at each step. This separation makes the code very clear and organized. But it might not work and the code will need to be flattened to something like:
@numba.njit
def _step(*inner_args, *outer_args):
"""Perform a single step.
This outer function should only contains attributes
associated with the solver class not with the solver instance.
"""
maurosilber
Illviljan
Instantiating the solver takes quite long due to compilation. It would be interesting to improve de code (without loosing clarity) so that numba is able to use the compiled code cache.