MKL and numpy/scipy

classic Classic list List threaded Threaded
10 messages Options
Reply | Threaded
Open this post in threaded view
|

MKL and numpy/scipy

Brian Granger-2
Hi,

I am playing with the latest EPD that has numpy and scipy linked
against Intel's MKL.  So nice!  I saw this pages that lists some of
the routines that will use MKL:

http://www.enthought.com/epd/mkl/

Is there a simple way of getting a complete list of the MKL optimized
functions in numpy/scipy?

Thanks,

Brian
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

Ilan Schnell
Hello Brian,

numpy and scipy in EPD are build using the
standard build process, linking against MKL
instead of ATLAS, accelerate (OSX), or LAPACK,
as the backend.
This means that the things which use these
linear algebra routines will use MKL.  Note
however, that there is much more in MKL, which
numpy/scipy could in principle use if one would
tweak the source.

- Ilan


On Tue, Jun 22, 2010 at 10:16 PM, Brian Granger <[hidden email]> wrote:

> Hi,
>
> I am playing with the latest EPD that has numpy and scipy linked
> against Intel's MKL.  So nice!  I saw this pages that lists some of
> the routines that will use MKL:
>
> http://www.enthought.com/epd/mkl/
>
> Is there a simple way of getting a complete list of the MKL optimized
> functions in numpy/scipy?
>
> Thanks,
>
> Brian
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

Brian Granger-2
Ilan,

On Tue, Jun 22, 2010 at 8:35 PM, Ilan Schnell <[hidden email]> wrote:
> Hello Brian,
>
> numpy and scipy in EPD are build using the
> standard build process, linking against MKL
> instead of ATLAS, accelerate (OSX), or LAPACK,
> as the backend.

That was my understanding, and maybe this is a question for the numpy list.

> This means that the things which use these
> linear algebra routines will use MKL.

I am wondering if there is a simpe way of see all the functions in
numpy/scipy that fall into this category (other than sifting through
the source code of numpy+scipy).   I may repost this on the
numpy/scipy lists...

> Note
> however, that there is much more in MKL, which
> numpy/scipy could in principle use if one would
> tweak the source.

But EPD does not include the MKL header files, right?

Cheers,

Brian

> - Ilan
>
>
> On Tue, Jun 22, 2010 at 10:16 PM, Brian Granger <[hidden email]> wrote:
>> Hi,
>>
>> I am playing with the latest EPD that has numpy and scipy linked
>> against Intel's MKL.  So nice!  I saw this pages that lists some of
>> the routines that will use MKL:
>>
>> http://www.enthought.com/epd/mkl/
>>
>> Is there a simple way of getting a complete list of the MKL optimized
>> functions in numpy/scipy?
>>
>> Thanks,
>>
>> Brian
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

Ilan Schnell
I'll think about how it would be easy to check if some
function uses MKL.

> But EPD does not include the MKL header files, right?

No.  We are allowed to redistribute the MKL binaries,
because we own a developer licenses.  However, we
are not allows to give away the headers, manuals, etc...
Sorry, but it's not our fault.

- Ilan


On Tue, Jun 22, 2010 at 11:37 PM, Brian Granger <[hidden email]> wrote:

> Ilan,
>
> On Tue, Jun 22, 2010 at 8:35 PM, Ilan Schnell <[hidden email]> wrote:
>> Hello Brian,
>>
>> numpy and scipy in EPD are build using the
>> standard build process, linking against MKL
>> instead of ATLAS, accelerate (OSX), or LAPACK,
>> as the backend.
>
> That was my understanding, and maybe this is a question for the numpy list.
>
>> This means that the things which use these
>> linear algebra routines will use MKL.
>
> I am wondering if there is a simpe way of see all the functions in
> numpy/scipy that fall into this category (other than sifting through
> the source code of numpy+scipy).   I may repost this on the
> numpy/scipy lists...
>
>> Note
>> however, that there is much more in MKL, which
>> numpy/scipy could in principle use if one would
>> tweak the source.
>
> But EPD does not include the MKL header files, right?
>
> Cheers,
>
> Brian
>
>> - Ilan
>>
>>
>> On Tue, Jun 22, 2010 at 10:16 PM, Brian Granger <[hidden email]> wrote:
>>> Hi,
>>>
>>> I am playing with the latest EPD that has numpy and scipy linked
>>> against Intel's MKL.  So nice!  I saw this pages that lists some of
>>> the routines that will use MKL:
>>>
>>> http://www.enthought.com/epd/mkl/
>>>
>>> Is there a simple way of getting a complete list of the MKL optimized
>>> functions in numpy/scipy?
>>>
>>> Thanks,
>>>
>>> Brian
>>> _______________________________________________
>>> Enthought-Dev mailing list
>>> [hidden email]
>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

Brian Granger-2
On Tue, Jun 22, 2010 at 9:53 PM, Ilan Schnell <[hidden email]> wrote:
> I'll think about how it would be easy to check if some
> function uses MKL.
>
>> But EPD does not include the MKL header files, right?
>
> No.  We are allowed to redistribute the MKL binaries,
> because we own a developer licenses.  However, we
> are not allows to give away the headers, manuals, etc...
> Sorry, but it's not our fault.

That is what I thought.  The other package that can use MKL is
numexpr.  But, it needs the header files in the build process.  Would
it make sense to ship numexpr with EPD?  I know Enthought is being
more picky these days about what packages get into EPD, but numexpr is
pretty nice and with MKL support, it provides a nice (and simple) way
of parallelizing certain types of things on a multicore system.  Maybe
I can bring this up at SciPy.

Cheers,

Brian

> - Ilan
>
>
> On Tue, Jun 22, 2010 at 11:37 PM, Brian Granger <[hidden email]> wrote:
>> Ilan,
>>
>> On Tue, Jun 22, 2010 at 8:35 PM, Ilan Schnell <[hidden email]> wrote:
>>> Hello Brian,
>>>
>>> numpy and scipy in EPD are build using the
>>> standard build process, linking against MKL
>>> instead of ATLAS, accelerate (OSX), or LAPACK,
>>> as the backend.
>>
>> That was my understanding, and maybe this is a question for the numpy list.
>>
>>> This means that the things which use these
>>> linear algebra routines will use MKL.
>>
>> I am wondering if there is a simpe way of see all the functions in
>> numpy/scipy that fall into this category (other than sifting through
>> the source code of numpy+scipy).   I may repost this on the
>> numpy/scipy lists...
>>
>>> Note
>>> however, that there is much more in MKL, which
>>> numpy/scipy could in principle use if one would
>>> tweak the source.
>>
>> But EPD does not include the MKL header files, right?
>>
>> Cheers,
>>
>> Brian
>>
>>> - Ilan
>>>
>>>
>>> On Tue, Jun 22, 2010 at 10:16 PM, Brian Granger <[hidden email]> wrote:
>>>> Hi,
>>>>
>>>> I am playing with the latest EPD that has numpy and scipy linked
>>>> against Intel's MKL.  So nice!  I saw this pages that lists some of
>>>> the routines that will use MKL:
>>>>
>>>> http://www.enthought.com/epd/mkl/
>>>>
>>>> Is there a simple way of getting a complete list of the MKL optimized
>>>> functions in numpy/scipy?
>>>>
>>>> Thanks,
>>>>
>>>> Brian
>>>> _______________________________________________
>>>> Enthought-Dev mailing list
>>>> [hidden email]
>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>
>>> _______________________________________________
>>> Enthought-Dev mailing list
>>> [hidden email]
>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

Ilan Schnell
While it is true that we are picky about what goes
into the EPD installers, we are not so picky about
what we put into the subscriber repository.  I can
take a look at numexpr and maybe put it into the
scipy2010 repository (alongside pyzmq).

- Ilan


On Wed, Jun 23, 2010 at 12:01 AM, Brian Granger <[hidden email]> wrote:

> On Tue, Jun 22, 2010 at 9:53 PM, Ilan Schnell <[hidden email]> wrote:
>> I'll think about how it would be easy to check if some
>> function uses MKL.
>>
>>> But EPD does not include the MKL header files, right?
>>
>> No.  We are allowed to redistribute the MKL binaries,
>> because we own a developer licenses.  However, we
>> are not allows to give away the headers, manuals, etc...
>> Sorry, but it's not our fault.
>
> That is what I thought.  The other package that can use MKL is
> numexpr.  But, it needs the header files in the build process.  Would
> it make sense to ship numexpr with EPD?  I know Enthought is being
> more picky these days about what packages get into EPD, but numexpr is
> pretty nice and with MKL support, it provides a nice (and simple) way
> of parallelizing certain types of things on a multicore system.  Maybe
> I can bring this up at SciPy.
>
> Cheers,
>
> Brian
>
>> - Ilan
>>
>>
>> On Tue, Jun 22, 2010 at 11:37 PM, Brian Granger <[hidden email]> wrote:
>>> Ilan,
>>>
>>> On Tue, Jun 22, 2010 at 8:35 PM, Ilan Schnell <[hidden email]> wrote:
>>>> Hello Brian,
>>>>
>>>> numpy and scipy in EPD are build using the
>>>> standard build process, linking against MKL
>>>> instead of ATLAS, accelerate (OSX), or LAPACK,
>>>> as the backend.
>>>
>>> That was my understanding, and maybe this is a question for the numpy list.
>>>
>>>> This means that the things which use these
>>>> linear algebra routines will use MKL.
>>>
>>> I am wondering if there is a simpe way of see all the functions in
>>> numpy/scipy that fall into this category (other than sifting through
>>> the source code of numpy+scipy).   I may repost this on the
>>> numpy/scipy lists...
>>>
>>>> Note
>>>> however, that there is much more in MKL, which
>>>> numpy/scipy could in principle use if one would
>>>> tweak the source.
>>>
>>> But EPD does not include the MKL header files, right?
>>>
>>> Cheers,
>>>
>>> Brian
>>>
>>>> - Ilan
>>>>
>>>>
>>>> On Tue, Jun 22, 2010 at 10:16 PM, Brian Granger <[hidden email]> wrote:
>>>>> Hi,
>>>>>
>>>>> I am playing with the latest EPD that has numpy and scipy linked
>>>>> against Intel's MKL.  So nice!  I saw this pages that lists some of
>>>>> the routines that will use MKL:
>>>>>
>>>>> http://www.enthought.com/epd/mkl/
>>>>>
>>>>> Is there a simple way of getting a complete list of the MKL optimized
>>>>> functions in numpy/scipy?
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Brian
>>>>> _______________________________________________
>>>>> Enthought-Dev mailing list
>>>>> [hidden email]
>>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>
>>>> _______________________________________________
>>>> Enthought-Dev mailing list
>>>> [hidden email]
>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>
>>> _______________________________________________
>>> Enthought-Dev mailing list
>>> [hidden email]
>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

bdforbes
Is it possible to use MKL FFTs in Enthought?
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

Ilan Schnell
Hello,
we have created an binding to the MKL FFTs, which can be
used as a replacement for numpy's own FFT.  This turned
out to be quite fast, in particular on multi core machines.
However, this package was only created for an internal
customer project, and we haven't made it available yet.
I'm not sure what the plans are.

- Ilan


On Sun, Apr 10, 2011 at 9:57 PM, bdforbes <[hidden email]> wrote:

> Is it possible to use MKL FFTs in Enthought?
>
> --
> View this message in context: http://enthought-dev.117412.n3.nabble.com/MKL-and-numpy-scipy-tp915891p2805056.html
> Sent from the Enthought Dev mailing list archive at Nabble.com.
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

bdforbes
Thanks, hopefully it will eventually make it in.

Ben
Reply | Threaded
Open this post in threaded view
|

Re: MKL and numpy/scipy

Chris Colbert
I've got a side project to create python bindings to MKL and IPP which will make it into future versions of EPD. How soon this happens depends on how much free time I can devote to the project. But it is definitely on the radar.


On Sun, Apr 10, 2011 at 11:18 PM, bdforbes <[hidden email]> wrote:
Thanks, hopefully it will eventually make it in.

Ben


--
View this message in context: http://enthought-dev.117412.n3.nabble.com/MKL-and-numpy-scipy-tp915891p2805097.html
Sent from the Enthought Dev mailing list archive at Nabble.com.
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev


_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev