[Users] Going from lapacke.h to mkl_lapacke.h

[Erik Schnetter]

Chris

I don't think the "mkl_" prefix is significant. I would look out for two things:

(1) I would compare the names, types, and order (!) of the arguments
to the functions you are calling, ensuring that they are what you are
expect. The Fortran versions of these routines are well standardized,
but the C versions might have a different convention. I don't expect
any difference there.

(2) I would check that the size of an int (32 or 64) is what you
expect. This can usually be chosen at compile time. I know that MKL
supports both, so you need to choose the right library at link time
that corresponds to the choice you made at compile time. I don't
recall how that choice is made at compile time -- it could be the
include path, the include file name, or a preprocessor flag that you
set. Of course, this needs to correspond to the actual int size that
you use in your code. Most likely, your int has 32 bit, which is
insufficient if you have really large vectors (with more than 2e9
elements). If you get this wrong, you either get completely
nonsensical results or a segfault.

-erik



On Sat, Mar 24, 2018 at 1:47 PM, Chris Stevens <c.stevens at ru.ac.za> wrote:
> Hi Erik,
>
> thanks for your quick reply. I'm not so worried about the error, it actually
> seems to be an MPI issue, even though MPI worked fine on the smaller
> cluster. Hopefully it isn't a pain to fix.
>
> What I really want to know is the difference between the standard lapacke
> and intel's mkl one. Is the only difference between the two, if all I used
> before was functions defined in lapacke.h, to replace this with
> mkl_lapacke.h? The code seems to run fine with only this change, but I
> remember seeing somewhere standard lapack function calls but with mkl__
> appended at the front. I want to know if I should also modify these function
> calls as well to make proper use of the mkl. The mkl_lapacke.h header file
> itself doesn't seem very insightful.
>
> I can't seem to find much useful documentation on this through google...
>
> Cheers,
>
> Chris
>
>
> On 03/24/2018 04:50 PM, Erik Schnetter wrote:
>
> Chris
>
> Can you say what is actually going wrong? We need to see an error message or
> similar to help out.
>
> -erik
>
> On Sat, Mar 24, 2018 at 03:38 Chris Stevens <c.stevens at ru.ac.za> wrote:
>>
>> Hi everyone,
>>
>> I have just ported my code from a local cluster to larger one which uses
>> intel mkl and wanted to clarify the changes I've made are the correct ones
>> wrt the lapacke library:
>>
>> -- In the local cluster, I had the option
>>
>> LAPACK_LIBS = lapack lapacke in my .ini file
>>
>> and used #include "lapacke.h" in the code.
>>
>> -- In the bigger cluster, I now have to use intel mkl, and I have the
>> following options in my .cfg
>>
>> LAPACK_DIR  =
>> /apps/compilers/intel/parallel_studio_xe_2016/mkl/lib/intel64
>>
>> LAPACK_LIBS = -mkl
>>
>> which will only compile my code if I replace #include "lapacke.h" with
>> #include "mkl_lapacke.h".
>>
>> Question: Is this the right way to generalise to mkl lapacke?
>>
>> Problem: The code runs fine on the local cluster, but breaks (not obvious
>> where) on the larger one with the above changes. The lapacke functions seem
>> to work OK at a glance, but seeing as this is the only thing that has
>> changed this may still be the culprit.
>>
>> Question: Do I perhaps need to change the lapacke function calls also? The
>> mkl_lapacke.h header file doesn't seem to give much insight... I think I
>> remember seeing somewhere calls replaced something like
>>
>> LAPACKE_dgels -> MKL_LAPACKE_dgels
>>
>> Thanks!
>>
>> Chris
>>
>> --
>> Dr Chris Stevens
>>
>> Claude Leon Postdoctoral Fellow
>>
>> Department of Mathematics
>>
>> Rhodes University
>>
>> Room 5
>>
>> Ph: +27 46 603 8932
>>
>> Web: www.chrisdoesmaths.com
>>
>> _______________________________________________
>> Users mailing list
>> Users at einsteintoolkit.org
>> http://lists.einsteintoolkit.org/mailman/listinfo/users
>
> --
> Erik Schnetter <schnetter at cct.lsu.edu>
> http://www.perimeterinstitute.ca/personal/eschnetter/
>
>
> --
> Dr Chris Stevens
>
> Claude Leon Postdoctoral Fellow
>
> Department of Mathematics
>
> Rhodes University
>
> Room 5
>
> Ph: +27 46 603 8932
>
> Web: www.chrisdoesmaths.com



-- 
Erik Schnetter <schnetter at cct.lsu.edu>
http://www.perimeterinstitute.ca/personal/eschnetter/