dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Tue, Sep 27, 2005 at 09:06:53AM -0500, Robert C. Kirby wrote:
> >
> >Improving the assembly is interesting, but it would probably be more
> >worthwile to get the results for Stokes. (It would also put FFC and
> >DOLFIN to the test and perhaps we can nail down some bugs in the
> >process.)
> >
> I agree that these are the results that are more likely of interest  
> to the broader community, and we should push on them.
> 
> >Another point is that to improve the speed of the assembly (which I
> >think is already pretty good), it would be easier to improve on the
> >interaction with the mesh (which we need to redo with Sieve anyway).
> >
> 
> I claim that there is something also to be gained in batching  
> together computations.
>  Currently, for each element you
> - get the affine map
> - get the dof
> - build the matrix
> - insert it.
> 
> If you loop over N elements and get affine maps stored in a C array,  
> then loop over N elements and get the dof in a C array, then build  
> matrices for N arrays (this is level 3 blas instead of level 2), then  
> insert N matrices, you're likely to improve performance.
> 
> Of course, I recall that the geometry and insertion into PETSc are  
> the bottlenecks in the process?  However, when you go to Navier- 
> Stokes (or any trilinear operator), building the element matrix is  
> much more expensive while insertion time stays the say (relative to  
> Stokes, not Poisson, since there are more dof).  You'll see the  
> effects of level 3 BLAS much more in this regime.  But this is not  
> hard to try out and put into DOLFIN.

We should definitely do this. I'll probably do level 2 BLAS first,
since then I only have to modify the output from FFC to use BLAS
calls. This will be an immediate benefit for Johan Hoffman with 3D
Navier-Stokes since compiling the code generated by FFC (with gcc) is
currently the bottle-neck. Hopefully, I can implement this by the end
of this week.

/Anders