experimental solve_mp doesn't seem to work anymore

[mbakker7]

experimental solve_mp doesn't seem to work anymore. I wonder if it ever worked since the move to numba.

[mbakker7]

Maybe Huite can help.

[mbakker7]

Suggested solution (from Joost Delsman and in Dutch): Het lag aan het referencen van de module besselaesnumba vanuit elk element. Dan kan python m blijkbaar niet meer picklen er daarmee werkt de multiprocessing niet meer. Oplossing is de switch fy2p / numba niet meer intern in elk element te hebben, maar in de globale namespace. Heb de linesink.py aangepast, daarmee werkt het voor mij (gebruik geen andere elementen, maar daar geldt vast hetzelfde voor).

[jdelsman]

Solved it by moving the switch between numba f2py to the global namespace, instead of inline in the element classes:

from .besselaesnumba import besselaesnumba as bessel
bessel.initialize()
try:
from .src import besselaesnew
bessel = besselaesnew.besselaesnew
bessel.initialize()
#print('succes on f2py')
except:
pass

and then in the class remove the self.bessel, and instead refer to the global bessel

def __init__(self, .....):
    Element.__init__(self, model, nparam=1, nunknowns=0, layers=layers, \
                     name=name, label=label)
    .....
    # if self.model.f2py:
    #     self.bessel = besselaesnew.besselaesnew
    # else:
    #     self.bessel = besselaesnumba

def disvecinf(self, x, y, aq=None):
    '''Can be called with only one x,y value
    Returns array(nparam, self.aq.naq) with order
    order 0, layer[0]
    order 0, layer[1]
    ...
    order 1, layer[0]
    order 1, layer[1]
    etc
    '''
    if aq is None: aq = self.model.aq.find_aquifer_data(x, y)
    rv = np.zeros((2, self.nparam, aq.naq))
    if aq == self.aq:
        qxqyrv = rv.reshape((2, self.order + 1, self.nlayers, aq.naq))
        qxqy = np.zeros((2 * (self.order + 1), aq.naq))

--> qxqy[:, :] = bessel.disbeslsv(float(x), float(y), self.z1,
self.z2, aq.lab, self.order, aq.ilap, aq.naq)
qxqyrv[0, :] = self.aq.coef[self.layers] * qxqy[:self.order + 1,
np.newaxis, :]
qxqyrv[1, :] = self.aq.coef[self.layers] * qxqy[self.order + 1:,
np.newaxis, :]
return rv

[jdelsman]

linesink.zip