Slow evaluation of vector-shaped Expressions

In master (today's dev docker image), interpolation of vector shape Expressions is significantly slower than with 2017.1.0.r3.

MWE1:

from dolfin import *

mesh = UnitCubeMesh(16, 16, 16)
V = VectorFunctionSpace(mesh, 'CG', 1)

exp = Expression(('1', '2', '3'), degree=1)

t0 = Timer('Zi interpolate')
interpolate(exp, V)
print(t0.stop())

on master: 0.007603146 on stable: 0.004061956

Evaluating a subclassed Expression with shape (3, 3) on a large mesh, this difference was clearly noticeable. Placing a Timer inside the eval() reveals that this function is called dim times as often on master as on stable, while the solution is the same.

MWE2:

from dolfin import *

class TensorExpr(Expression):
    def eval(self, values, x):
        t0 = Timer('Z call eval')
        for i in range(9):
            values[i] = i
        del t0

    def value_shape(self):
        return (3, 3)                                                                                                                                                                                                                                                                                                                               

G = TensorExpr(degree=1)                                                                                                                                                                                                                                                                                                                            mesh = UnitCubeMesh(16, 16, 16)                                                                                                                                           
V = TensorFunctionSpace(mesh, 'CG', 1)                                                                                   
t0 = Timer('Z interpolate')                                                                                                                                               
u = interpolate(G, V)                                                                                                                                                     
del t0                                                                                                       

list_timings(TimingClear_keep, [TimingType_wall])

Output stable:

Z call eval                       |  98304  3.6987e-06      0.3636
Z interpolate                     |      1      1.2807      1.2807

Output master:

Z call eval                       |  884736  3.8758e-06      3.4291
Z interpolate                     |       1      11.965      11.965

9x more calls of eval and time in interpolation.

Best wishes

Comments (2)