tabulate_tensor() slow

Question

When profiling my code, I noticed that a lot of time is spent the tabulate_tensor(...) calls in ./dolfin/fem/Assembler.cpp::assemble_cells(...). Is that a function produced by FFC? What could I do to get further down to the cause of this bottle neck?

Answer 1

This is the code that is generated by FFC.

You can try set some ffc related parameters. Try them in random order and see what makes the most difference.

parameters["form_compiler"]["cpp_optimize"] = True
parameters["form_compiler"]["optimize"] = True
parameters["form_compiler"]["cpp_optimize_flags"] = "-O3 --fast-math" if you_dare else "-O3"
parameters["form_compiler"]["quadrature_degree"] = 2

If you have uflacs installed you can also try:

parameters["form_compiler"]["representation"] = "uflacs"

Comments

This helped me finding out that it is the quadrature degree that is killing performance here. I'll go and find out why FEniCS thinks it needs such a high quadrature degree and fix the cause for it.