Page 1 of 1

clMagma and acml 6.1.0 questions

Posted: Fri Feb 13, 2015 10:19 am
by Jan Kucera
Hallo,

The acml 6.1.0 libraries now include the clblas library
("libclBLAS.so.2.3.0") and claim to be using the GPU:
"The key feature of ACML is that the library transparently dispatches appropriate tasks to a GPU if it is available, with no additional code development on your part, enabling your application to easily take advantage of the available resources...."
http://developer.amd.com/community/blog ... ates-code/


I, a COMPLETE OUTSIDER, have several broad questions which the developers of
clMagma might wish to answer.
I believe the answers might be of interest to other prospective users of
clMagma and/or the acml libraries.
  • 1) Is there any connection/overlap between the acml 6.1.0 libraries and the
    clMagma library?
  • 2) Is clMagma going to be developed further as an independent project?
  • 3) Is it possible/does it make any sense to link clMagma against the
    "acml 6.1.0" libraries (and the included libclBLAS.so.2.3.0). Will it work?
    Should one only use the combination of a previous version of acml
    (like acml 5.3.1) and latest version of clblas available independently
    (clBLAS 2.2.0)?
  • 3b) When limited to routines just replacing the standard LAPACK calls in
    the usual CPU-centric code (i.e, not those versions of routines which
    operate on data stored in the GPU memory), will it have any advantage over the
    "acml 6.1.0" alone?
  • 4) I believe linking against other optimized lapack libraries (mkl, openblas)
    and clBLAS 2.2.0 still makes sense... Am I right?
Thanks for any response!

Jan Kucera

Re: clMagma and acml 6.1.0 questions

Posted: Fri Feb 13, 2015 5:41 pm
by mgates3
Some brief notes:
1) clMagma is not currently part of ACML. We can use ACML.
2) Yes, clMagma is still being developed.
3) Unknown. clMagma with ACML 6.1.0 is untested at the moment.
3b) Since we explicitly manage CPU and GPU memory, we expect clMagma to have better performance than offloading individual BLAS calls.
4) Yes, you can use any CPU BLAS & LAPACK libraries with clMagma. We need clBLAS for the GPU functions, of course.

As we test it further, we will update you on the status.
-mark