Home > Archive > Fortran > December 2005 > MKL on Non-Intel processors
You are viewing an archived Text-only version of the thread.
To view this thread in it's original format and/or if you want to reply to
this thread please [click here]
| Author |
MKL on Non-Intel processors
|
|
| Stig Kildegård Andersen 2005-12-06, 7:05 pm |
| With IVF one has to patch the executables with a tool such as the one from
http://www.swallowtail.org/naughty-intel.html in order to be able to use
SSE2, SSE3, and so on (I think). Are there any similar tricks needed to get
the routines from the Intel MKL library to run as fast as possible on
non-Intel chips, or does the Intel MKL always run optimised out of the box.
I am comparing some custom code against MKL routines on an Athlon64X2, and I
do not want to put the MKL routines at a di vantage by not using the/any
needed tricks.
Best Regards and Thank You in Advance,
Stig Kildegård Andersen
| |
| Greg Lindahl 2005-12-07, 4:01 am |
| In article <43959d0b$0$191$edfadb0f@dread11.news.tele.dk>,
Stig Kildegård Andersen <SearchForMyFullName@google.com> wrote:
> Are there any similar tricks needed to get
> the routines from the Intel MKL library to run as fast as possible on
> non-Intel chips, or does the Intel MKL always run optimised out of the box.
Rumor has it that there are multiple checks deeper in this library.
Are you sure the license allows it to be run on AMD chips? And does
AMD's ACML library have what you need in it?
-- greg
| |
| Stig Kildegård Andersen 2005-12-07, 7:04 pm |
| >> Are there any similar tricks needed to get
>
> Rumor has it that there are multiple checks deeper in this library.
> Are you sure the license allows it to be run on AMD chips? And does
> AMD's ACML library have what you need in it?
The MKL Release Notes show Athlon and Opteron processors as supported
hardware, so I supposed that it was also allowed to use it on an Athlon.
Specifically I need LAPACK and PARDISO. LAPACK I can get from ACML but I do
not think that PARDISO is included. The routines do appear to run OK, but
the PARDISO solver apparently scales very poorly to two cores for my
particular matrix layout. It can certainly keep both cores busy when run
with two threads. Also if I divide the Mflop count provided by PARDISO with
the running time it goes somewhere between 1.5 and 1.7 GFlops which does
sound very fast.. But it just does not speed up much when going from one to
two cores. Could just be my specific problem, a memory bandwith bottleneck
with the dual core chip, or something else...
Thank You and Best Regards,
Stig Kildegård
| |
| Steve Lionel 2005-12-08, 7:15 pm |
| On Wed, 7 Dec 2005 20:50:48 +0100, "Stig Kildegård Andersen"
<SearchForMyFullName@google.com> wrote:
>The MKL Release Notes show Athlon and Opteron processors as supported
>hardware, so I supposed that it was also allowed to use it on an Athlon.
If you have questions about MKL, please feel free to ask them in the MKL User
Forum at http://softwareforums.intel.com/ids/board?board.id=MKL Knowledgeable
Intel engineers as well as users participate there.
Steve Lionel
Software Products Division
Intel Corporation
Nashua, NH
User communities for Intel Software Development Products
http://softwareforums.intel.com/
Intel Fortran Support
http://developer.intel.com/software/products/support/
| |
|
|
> And does
> AMD's ACML library have what you need in it?
but which version should one link to intel binaries. They seem to
provide only gnu pathscale and portland acmls. Is it sufficient to link
the gnu version and libg2c or are there binary incompatibilities (e.g.
complex function return values and the like) ?
Joost
|
|
|
|
|