[pythran] Re: Bug report - "# omp..." is silently ignored (whilst "#omp..." is good)
- From: Pierrick Brunet <pierrick.brunet@xxxxxxxx>
- To: pythran@xxxxxxxxxxxxx
- Date: Fri, 14 Mar 2014 20:39:17 +0100 (CET)
Hum, sorry about it. Here is a fix :
https://github.com/serge-sans-paille/pythran/pull/219 .
I think the "how many OpenMP does I have" is too much. If we display all
possible usefull information, you screen will be overflooded.
To debug this kind of issue, you should look in the cpp file using the -E
flags. In the cpp you will see your transformed code so if you search "#pragma
omp" you will see how many are present. I think there are no other possible
approche. May be we can add a flags to display the final transformed Python
code which will show this kind of issue?
Cheers,
Pierrick
----- Mail original -----
> De: "Ian Ozsvald" <ian@xxxxxxxxxxxxxx>
> À: "pythran" <pythran@xxxxxxxxxxxxx>
> Envoyé: Vendredi 14 Mars 2014 20:10:48
> Objet: [pythran] Bug report - "# omp..." is silently ignored (whilst
> "#omp..." is good)
>
> I've just discovered the bug that I had at PyDataLondon (phew, I'm not
> mad). I couldn't get an omp reduction to work - it looks like I
> ignored the specification and prepended a whitespace which then causes
> the openmp line to be silently ignored.
>
> In the midst of my critical function if I have:
> #omp parallel for reduction(+:nbr_trials_in_unit_circle)
> then my runtime is:
> 0.118036031723 seconds # parallel!
> but if I have:
> # omp parallel for reduction(+:nbr_trials_in_unit_circle)
> then my runtime is:
> 0.221085786819 seconds # single threaded
>
> I'll make a feature suggestion - could pythran (perhaps with -v)
> output the number of openmp declarations it has found? This would have
> helped me with debugging the issue.
>
> As an intermediate step for now - maybe edit the docs to note the
> exact syntax that's required, noting that extra whitespace will cause
> the omp declaration to be silently ignored?
>
> For future reference - what's the easiest way to debug this problem? I
> see (now) that if I use the -E flag I can investigate the generated
> C++ and indeed I see the omp declaration appearing/not appearing, so I
> can definitely verify it this way. Is there any other approach?
>
> Cheers, i.
>
> --
> Ian Ozsvald (A.I. researcher)
> ian@xxxxxxxxxxxxxx
>
> http://IanOzsvald.com
> http://MorConsulting.com/
> http://Annotate.IO
> http://SocialTiesApp.com/
> http://TheScreencastingHandbook.com
> http://FivePoundApp.com/
> http://twitter.com/IanOzsvald
> http://ShowMeDo.com
>
>
Other related posts:
