Re: [gentoo-user] openmp flag

[Florian Philipp]

Am 26.09.2012 21:46, schrieb Michael Mol:
> On Wed, Sep 26, 2012 at 2:25 PM, Florian Philipp <lists@×××××××××××.net> wrote:
>> Am 25.09.2012 17:01, schrieb Michael Mol:
>>> On Tue, Sep 25, 2012 at 10:42 AM, James <wireless@×××××××××××.com> wrote:
>>>> Hello,
>>>>
>>>> background:
>>>> It seems there is a major push now to put openmp:
>>>> [1,2] into  embedded systems [3].
>>>>
>>>> So I looked at these [4] packages to find something
>>>> interesting to look deeper into related to openMP.
>>>>
>>>> Blender immediately jumped out at me as a good example,
>>>> cause an old friend Ken Hughes is, imho, one of the
>>>> world's most amazing C programmers, and a stalwart at
>>>> the blender project.
>>>>
>>>>
>>>> OK, here's the question, I went to emerge blender
>>>> and found that the openmp flag is already set. {?}
>>>> Yet I looked everywhere and did not see the openmp flag
>>>> set (/etc/make.conf, /etc/portage/package.use)
>>>> so where is it getting set on my AMD workstation?
>>>>
>>>> [ebuild  N     ] media-gfx/blender-2.49b-r2  USE="ffmpeg
>>>>  nls ogg openmp -blender-game -openal -verse"
>>>>
>>>> I feel like I should know (profiles etc) but, I'm a little
>>>> bit brain_dead this am, so any help is appreciated.
>>>
>>> Packages can choose to have USE flags enabled or disabled for them by
>>> default. So blender likely has openmp enabled by default, without that
>>> affecting any other packages.
>>>
>>>>
>>>> OH, anyone is encouraged to "chime in" about openmp
>>>> and your thoughts as to it's viability and usefulness.
>>>> Do you believe it will become a core technology,
>>>> embedded into GCC? Used widely?
>>>
>>> If you can use it, use it. OpenMP is little more than a set of
>>> extensions to C (and C++) which allows the normally-scalar language to
>>> do some things in a parallel fashion without resorting to the costs of
>>> multithreading. This is good, because vector instructions have been
>>> available in x86 since MMX came out, and improvements to the vector
>>> instructions available to x86 still goes on.
>>>
>>
>> I guess this is just poorly phrased but to clarify: OpenMP *does* use
>> multithreading and nothing else. It does not, in any way, make more use
>> of vector instructions than GCC without -fopenmp. I guess what you mean
>> is avoiding the costs of *manual* multithreading using POSIX threads and
>> the like.
>
> Fair point.
>
>>
>> If you want to use vector instructions for your own code, you should
>> look into compiler intrinsics (i.e. vector instructions as built-in C
>> functions).
>> http://ds9a.nl/gcc-simd/
>
> Personally, I don't like compiler intrinsics; they're specific to
> given compilers. I've tended to write code which is supposed to
> compile on multiple compilers. (There's a world outside GCC...)
>

Yes. I haven't used it, either. I guess you could autoconf it and
replace it with vanilla C macros in most cases. Or as an easier
solution: #ifdef a vanilla C implementation together with the vector
code. Bonus points for added readability.

Kind of makes you wonder how well GCC can vectorize programs on its own
when you lay out your code in a way suitable for its own intrinsics
without actually using them.

>> And, just to nit-pick: OpenMP also works for Fortran.
>
> True; this slipped my mind. :)
>
>>
>>> Related are CUDA and OpenCL, which are two other systems for
>>> parallelizing code. CUDA assumes you have access to an nVidia GPU (and
>>> have a CUDA-enabled driver installed). OpenCL is a big more generic,
>>> and supports dispatching to CUDA, CPU vector instructions or even
>>> thread pools.
>>>
>>> Personally, my recommendation is to enable everything you can get
>>> working (be it, OpenMP, CUDA or OpenCL); vector processing is going to
>>> be generally more efficient than scalar processing. You don't need to
>>> worry about which is better unless you're a software developer. (And
>>> if you're a software developer, go study up on their differences;
>>> tradeoffs happen.)
>>>
>>
>> +1
>>
>> By the way: Did anyone get good results out of dev-util/intel-ocl-sdk
>> for OpenCL? Some time ago I tested it with a package that supported both
>> OpenMP and OpenCL (not sure which) and OpenCL didn't really make an
>> impact on my Core i5.
>
> Haven't tried it, no. I've got a Radeon 6870, and I can only have one
> OpenCL driver loaded at a time. (IBM has a middleman driver which
> supports dispatching to multiple backends, but I believe its a for-pay
> package.)
>

Isn't that what app-admin/eselect-opencl is for? I mean simple
switching, not dual application (which would be awesome, too).

Regards,
Florian Philipp