MAGMA 0.2 does not seem to work with CUDA 3.0

[wsawyer]

We have successfully tested MAGMA 0.2 with CUDA 2.3, but have since upgraded to CUDA 3.0. MAGMA 0.2 does not seem to be compatible with 3.0. We can post error messages on request.

Will this problem be addressed in MAGMA 0.3. When will that be released?

Thanks much!

--Will and Vincenzo

[rnath]

Can you please tell us what kind of problem you are facing ?
did you recompiled all the BLAS routines? with sm_20 changes in the makefile ?

[wsawyer]

Thanks for the reply. Here some more details:

OS: Novell SUSE Linux Enterprise Server 11 Operating System

GPUs: 11xGTX285, 4xTesla

MKL: 11.1

Compilers: Intel icc/11.1 ifort/11.1


sm_20 seems to apply only to Fermi cards, therefore inappropriate for our GTX285 and Tesla cards. We tried compiling with both sm_13 and sm_20. There were also a number of changes which needed to be made to make.inc to get the tests to compile, w.r.t. get_nb.cpp for example, and adding all the mkl_* libraries individually to the LIB line (it appears that mkl is not bundled in mkl_em64t).

The behavior is:

1) In calls to magmablas_*, such as in testing_sgemv and testing_ssymv, it dies with

ERROR: CUDA Driver fails

2) In calls to magma_*, it dies with a segmentation fault.

3) In many instances there are MKL errors like

MKL ERROR: Parameter 7 was incorrect on entry to SGEQRF
MKL ERROR: Parameter 6 was incorrect on entry to SGEMV
MKL ERROR: Parameter 4 was incorrect on entry to SGEQLF

It is not clear what you mean by recompiling BLAS... magmablas?

Let me know if you need any more information. Comments welcome.

Vincenzo and Will