PetscFEIntegrateJacobian_Basic has four layers of loops:

   e - element
   f - component of field being differentiated
   g - component of field we're differentiating against
   q - quadrature point

It looks like g3_func and friends are being called once for each
combination of (e,f,g,q), even though the calls to g3_func are
independent of (f,g).  If g3_func is expensive, this duplication seems
a bit slow.

It seems like the loops should be flipped to be (e,q,f,g) major order,
with the user functions hoisted up to the (e,q) level.  Is there a
reason not to do this?

Geoffrey

Reply via email to