Setting vectors is slow

[jrmaddison]

The construct Function.vector()[:] = is very slow. The following example:

from firedrake import *

import time

mesh = UnitIntervalMesh(1000000)
space = FunctionSpace(mesh, "Lagrange", 1)
F = Function(space, name="F")
G = Function(space, name="G")

t0 = time.time()
with F.dat.vec_wo as F_v, G.dat.vec_ro as G_v:
    for i in range(10):
        G_v.copy(result=F_v)
print(f"time = {time.time() - t0:.3f} s")

t0 = time.time()
for i in range(10):
    F.vector().set_local(G.vector().get_local())
print(f"time = {time.time() - t0:.3f} s")

t0 = time.time()
for i in range(10):
    F.vector()[:] = G.vector()
print(f"time = {time.time() - t0:.3f} s")

leads, in a test on my machine, to

time = 0.010 s
time = 0.023 s
time = 164.478 s

[dham]

F.vector() is basically only a FEniCS compatibility fudge. The native implementation is F.dat, why not just use that?

[wence-]

This is because the first two do an all-at-once copy, the last one does, for every entry, a getitem on G.vector().

[jrmaddison]

I was caught by this when porting FEniCS code, and the cost wasn't obvious until it dominated the calculation. Perhaps a warning could be added? Or this could removed, as it is so slow when used like this.

[dham]

I'm open to the option of putting in a warning on Vector.__setitem__. Can anyone see a downside? I suggest guarding it with a variable so it only triggers once per Vector.

[wence-]

You could also

if isinstance(value, Vector):
    self.dat.data[idx] = value[idx]

In __setitem__?