[Numpy-discussion] [SciPy-user] conforming to Python GIL...

[Anne Archibald]

On 03/04/2008, Travis E. Oliphant <oliphant@enthought.com> wrote:
> fred wrote:
>  > Hi,
>  >
>  > I use a lot of ConVeX OPTimsation and fortran (via f2py) routines in my
>  > Traits app.
>  >
>  > As I want to compute the data and want to display them, I use threads.
>  >
>  > The issue I get is that data displayed (using Chaco2) are not updated
>  > (app is frozen) while computing the input data.
>  >
>  >  From D. Morrill's answer (the Traits guru ;-)), it appears that cvxopt
>  > (and solve() from scipy, in fact) and fortran modules does not release
>  > the "Python GIL (Global Intepreter Lock)".
>  >
>
> This requires a bit of effort to solve.   We need to in multiple places...
>
>  1) release the GIL
>  2) put semaphores into code that is not thread-safe.
>
>  f2py should be modified to handle this.    Other could should be
>  modified to handle it as well.
>
>  I suspect NumPy should grow an API to handle the semaphore thing easily
>  (I think Python already has one), so these may be able to be wrapped
>  into the MACROS already available for releasing the GIL.

What is the story on f2py's releasing of the GIL? I had understood
that wrapper code was able to declare the wrapped code to be
thread-safe, so that the GIL was released while it executed. Is the
problem here that cvxopt is not threadsafe? I guess the best solution
then is to make a cvxopt- (and related routines)-specific lock?

Anne