ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

Andrey wrote:
> Hello again!
>
> Now I have another issue.
> I created a small project in order to use FFC and Dolfin:
>
> 1) ChorinTemamFenics.form (this is a part of evolution
> Navier - Stokes equations solver, in cylindrical coordinates
> with axial symmetry) 
> 2) main.cpp (this is only the sample to check how compiler
> will work with generated .h file and Dolfin headers)
> 3) Makefile and dolfin.conf (standard from Dolfin installation
> directory)
>
> The result is following:
> 1) Generated with “ffc -l dolfin ChorinTemamFenics.form” .h file
> has size 26Mb and there are some lines in the code with length 290Kb.
> 2) Compilation with “make” of main.cpp and .h takes more then 3 hours
> and then gcc finishes with “out of memory” message.

Yes, sometimes the default code generation strategy of FFC can generate 
quite a few lines of code. The generated code is completely unrolled so 
 this is natural.

There are a few things you can try:

1. Compile without optimization (g++ -O0) which gives g++ an easier job.

2. Avoid multiplications (if possible). If you have two coefficients f 
and g that appear in your form as f*g, then consider replacing them by h 
= f*g.

3. Use the (experimental) quadrature mode of FFC: ffc -r quadrature. 
This will use a different code generation strategy that should result in 
a much smaller code size (but possibly not as efficient).

There are a few other options but these are currently broken:

1. BLAS mode: ffc -f blas.

2. Project coefficients down to piecewise constants.

Both the BLAS mode and projections are currently broken.

/Anders

> Best regards,
> Andrey
>
>
> ------------------------------------------------------------------------
>
> _______________________________________________
> FFC-dev mailing list
> FFC-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/ffc-dev