[Bf-committers] OpenMP issues with msvc2013 builds

[Campbell Barton]

Theres been some issues with OpenMP, but am wary of kicking out entire
technology just because we can't set an environment variable? (ok, of
course theres more to it - but seems a weak reasoning still).

Regarding OSX - Clang is preparing to use OpenMP from Intel, so
situation on OSX likely improves soon: http://clang-omp.github.io
clang's development moves at quite good pace so think its reasonable
to expect this is usable soon.


We had problems with Windows & setting env vars before, where we
weren't able to set CPython's stdout encoding.
https://developer.blender.org/T31555

It may be good to use some stub (BAT/COM/EXE) to launch blender on
Windows, then we don't run into a wall whenever some library needs an
env var set.

On Wed, Aug 6, 2014 at 10:32 PM, Martijn Berger
<martijn.berger at gmail.com> wrote:
> Excellent plan Bastien.
> From a semantic standpoint combining different threading strategies and or
> support libraries within one project is really far from ideal.
> For this to work we would need to write some kind of document or declare
> some already implemented part of code "the official way" to do this.
>
> Yes the code would be more verbose compared to openmp but also more
> explicit and likely better to control versus the current zoo of tricks in
> order to exploit parallelism.
>
>
>
>
> On Wed, Aug 6, 2014 at 2:15 PM, Bastien Montagne <montagne29 at wanadoo.fr>
> wrote:
>
>> We could also choose to drop OMP in our code (not all at once, but
>> gradually replacing it with real threading), since it’s also a pile of
>> crap on OSX currently afaik?
>>
>> Assuming our BLI_task 'lib' is ready for massive use (iirc Martijn
>> already had that in mind)…
>>
>> Or is there some serious issues with this approach (aside the slightly
>> more verbose [complicated?] code)?
>>
>> Le 06/08/2014 13:26, Sergey Sharybin a écrit :
>> > Switching to icc wouldn't happen any time soon anyway, so afraid we'll
>> need
>> > to have some shorter-term solution anyway... The question only is what
>> > exactly we'll choose.
>> >
>> > P.S. And yeah, there's nothing more permanent as a temporary solution..
>> >
>> >
>> > On Wed, Aug 6, 2014 at 4:48 PM, Martijn Berger <martijn.berger at gmail.com
>> >
>> > wrote:
>> >
>> >> Hi Campbell,
>> >>
>> >> I last compiled with intel c++ a few month's ago and without OSL /
>> Collada.
>> >> In general it works but as with any change it will again probably expose
>> >> lots of little issues once people actually use builds intensively.
>> >>
>> >> Main issue is the compiler is not free to use for windows not even for
>> open
>> >> source projects. ( the linux toolchain is free for open source ).
>> >>
>> >>
>> >>
>> >> On Wed, Aug 6, 2014 at 12:43 PM, Campbell Barton <ideasman42 at gmail.com>
>> >> wrote:
>> >>
>> >>> On Wed, Aug 6, 2014 at 6:23 PM, Sergey Sharybin <sergey.vfx at gmail.com>
>> >>> wrote:
>> >>>> Hi,
>> >>>>
>> >>>> We've spent quite a while trying to solve the upcoming stream of
>> >> reports
>> >>>> about high CPU usage with builds made with msvs2013 in certain
>> >>> situations.
>> >>>> Root of the issue goes to the change made to OMP inplementation back
>> to
>> >>>> msvc2010 days -- they've forced threads to spin for a while after they
>> >>> did
>> >>>> a work. This is a know issue in the library and nobody actually gonna
>> >> to
>> >>>> fix it [1].
>> >>>>
>> >>>> There's one woekaround to solve the issue which is to set
>> >> OMP_WAIT_POLICY
>> >>>> environment variable to PASSIVE. Unfortunately, since openmp is a
>> >> dynamic
>> >>>> library and can't be linked statically at all we can't modify
>> >> environment
>> >>>> variables from within blender using putenv(), this is to be done
>> >>>> externally, before blender.exe starts.
>> >>>>
>> >>>> Here's the list of possible solutions:
>> >>>>
>> >>>> - Declare msvc full of crap, switch to intel compilers
>> >>> Did anyone try Intel-c/c++ on Windows?
>> >>>
>> >>> A while back I got Blender building on Linux with Intel's compiler, it
>> >>> needed a few tweaks in CMake but wasn't really a big deal to get
>> >>> running.
>> >>> _______________________________________________
>> >>> Bf-committers mailing list
>> >>> Bf-committers at blender.org
>> >>> http://lists.blender.org/mailman/listinfo/bf-committers
>> >>>
>> >> _______________________________________________
>> >> Bf-committers mailing list
>> >> Bf-committers at blender.org
>> >> http://lists.blender.org/mailman/listinfo/bf-committers
>> >>
>> >
>> >
>>
>> _______________________________________________
>> Bf-committers mailing list
>> Bf-committers at blender.org
>> http://lists.blender.org/mailman/listinfo/bf-committers
>>
> _______________________________________________
> Bf-committers mailing list
> Bf-committers at blender.org
> http://lists.blender.org/mailman/listinfo/bf-committers



-- 
- Campbell