CODEWITHSUNDEEP
pythonpython-3.xdebuggingparallel-processingcython
flotzilla
flotzilla

Reputation: 1262

cython: run prange sequentially (profiling/debugging)

I have a meanwhile pretty large code base written mostly in Cython. Meanwhile, I've started parallelizing it by replacing "range"s by "prange"s. (So far more or less at random, as I still have to develop a gut feeling as to where I really profit from this and where not so much.)

One big question to which I haven't found an answer arose when profiling/debugging:

Is there a way to turn off parallelization (i.e. run "prange"s sequentially) at either a global or local level?

A global switch would be most convenient, but even if there's a local solution that should at least allow me to implement a global switch myself.

Just in case it's relevant: I'm using Python 3 (currently 3.4.5).

Upvotes: 0

Views: 80

Answers (1)

DavidW
DavidW

Reputation: 30913

There's at least three very easy ways:

  1. prange takes a num_threads argument. Set that equal to 1. This gives you local control.

  2. If you compile without OpenMP it'll still still run, but not in parallel. Remove extra_compile_args=['-fopenmp'] and extra_link_args=['-fopenmp'] (or equivalent) from setup.py. (OpenMP is implemented in terms of #pragmas so just gets ignored without compiler support). This gives you an easy global switch.

  3. OpenMP defines an environmental variable for the maximum number of threads OMP_NUM_THREADS. Set that to 1 in your operating system and then run your cython program. This gives a global switch (without recompiling)

Upvotes: 2

Related Questions