ffc team mailing list archive
-
ffc team
-
Mailing list archive
-
Message #03924
Re: FFC vs. SFC
On 21 June 2010 23:16, Kent Andre <kent-and@xxxxxxxxx> wrote:
>
> Strange. I used dorsal to compile the dev versions of the various
> packages. Dolfin, ffc, and sfc are only a few days old.
>
> But do you have any suggestions for how to optimize the FFC generated
> code ? I have used quadrature here, tensor representation is not good
> in this application.
I had a look at the generated code from FFC and SyFi
(which by the way is a real pain when PyDOLFIN is involved)
and realised that your comparison is not fair.
FFC uses the UFL algorithm estimate_total_polynomial_degree() to
compute the quadrature_degree.
SFC uses (in this case) the UFL algorithm estimate_quadrature_degree()
which is less conservative.
This results in FFC using 25 integration points to compute the
integral while SFC just uses 3.
If you use the option:
parameters ["form_compiler"]["quadrature_degree"] = "2"
the comparison is _more_ fair and the numbers are almost the same.
SFC still uses 3 points, while FFC now uses 4 points which leads to my
answer to your first question:
"How should I optimize FFC?" :)
https://blueprints.launchpad.net/ffc/+spec/simplex-quadrature
Kristian
> Kent
>
>
> On Mon, 2010-06-21 at 23:05 +0200, Kristian Oelgaard wrote:
>> I tried to run the code against the dev versions of FEniCS and got the
>> following error after successfully finishing the FFC part:
>>
>> Traceback (most recent call last):
>> File "demo.py", line 15, in <module>
>> V = FunctionSpace(mesh, "CG", 1)
>> File "/home/oelgaard/software/fenics/dolfin/local/lib/python2.6/site-packages/dolfin/function/functionspace.py",
>> line 230, in __init__
>> FunctionSpaceBase.__init__(self, mesh, element)
>> File "/home/oelgaard/software/fenics/dolfin/local/lib/python2.6/site-packages/dolfin/function/functionspace.py",
>> line 44, in __init__
>> ufc_element, ufc_dofmap = jit(self._ufl_element)
>> File "/home/oelgaard/software/fenics/dolfin/local/lib/python2.6/site-packages/dolfin/compilemodules/jit.py",
>> line 44, in mpi_jit
>> return local_jit(*args, **kwargs)
>> File "/home/oelgaard/software/fenics/dolfin/local/lib/python2.6/site-packages/dolfin/compilemodules/jit.py",
>> line 87, in jit
>> p = form_compiler.default_parameters()
>> AttributeError: 'module' object has no attribute 'default_parameters'
>>
>> Kristian
>>
>> On 21 June 2010 15:42, <kent-and@xxxxxxxxx> wrote:
>> >
>> > Hi, just portet SFC to dolfin-dev and ufl-dev and made some simple
>> > comparisons with FFC. I
>> > don't know exactly how representative these are, but I thought I'd try to
>> > compare
>> > them for fun.
>> >
>> > I have used the following code:
>> >
>> > from dolfin import *
>> > import time
>> >
>> > for compiler in ["ffc", "sfc"]:
>> >
>> > for N in [200, 400, 800]:
>> >
>> > parameters["form_compiler"]["name"]=compiler
>> > if compiler == "ffc":
>> > parameters["form_compiler"]["optimize"] = True
>> > parameters ["form_compiler"]["cpp_optimize"] = True
>> >
>> > mesh = UnitSquare(N, N)
>> > V = FunctionSpace(mesh, "CG", 1)
>> >
>> > P=6
>> > U = Function(V)
>> > energy = U**6*(U*U + inner(grad(U), grad(U)))*dx
>> >
>> > v = TestFunction(V)
>> > u = TrialFunction(V)
>> >
>> > F = derivative(energy, U, v)
>> > J = derivative(F, U, u)
>> >
>> > print ""
>> > print "First time"
>> > print ""
>> > t0 = time.time()
>> > A = assemble(J)
>> > t1 = time.time()
>> > print "First time for ", compiler, " was ", t1 - t0
>> >
>> > print ""
>> > print "Second time"
>> > print ""
>> > t0 = time.time()
>> > A = assemble(J)
>> > t1 = time.time()
>> >
>> > print "Second time for ", compiler, " was ", t1 - t0
>> >
>> >
>> > With these options SFC is about 3 times faster than FFC. How should I
>> > optimize FFC ?
>> >
>> > Kent
>> >
>> >
>> >
>> > _______________________________________________
>> > Mailing list: https://launchpad.net/~ffc
>> > Post to : ffc@xxxxxxxxxxxxxxxxxxx
>> > Unsubscribe : https://launchpad.net/~ffc
>> > More help : https://help.launchpad.net/ListHelp
>> >
>
>
>
Follow ups
References