9 Replies Latest reply on Apr 4, 2013 9:19 AM by kknox

    clMAGMA v0.1 Beta now available on UTK's ICL website

    kknox

      For user's interested in finding an OpenCL accelerated LAPACK inspired library, a beta of clMAGMA is now available to download on ICL's MAGMA web page. 

      This release provides OpenCL implementations for MAGMA's one-sided dense matrix factorizations (LU, QR, and Cholesky), and thus extending MAGMA's support to include AMD GPUs.

      More information, graphs and the download link are available on the UTK ICL website

       

      In support of the clMAGMA beta release, AMD is making available a beta release of our linux BLAS library v1.7.257, which you can read about and download on the AMD APPML website

       

      Thank you for your interest.

        • Re: clMAGMA v0.1 Beta now available on UTK's ICL website
          oscarbarenys1

          Hi,

          some questions:

          first is I have ported clmagma to Windows.. seems to work with current amd blas windows 1.6 version on a 5850.. is 1.7 expected to bring more performance to blas routines to 5xxx hw or only is focused only on 7xxx series..

          also I see from changelog you seem to finally have completed full blas 2 and blas 3 routines.. as my port confirms this new functions are not needed for clmagma compilation, right?

          Do you plan now on focusing on blas 1 implementations?.. and finally have similar feature set to cublas (which is an impressive feat)..

          is I see clmagma in Macos would compile easily if we had APPML BLAS for MacOS.. are you working on compiling this and FFT library for MACOS? well.. perhaps problem seems Macos doesn't have full cl_khr_fp64 ext and only some APPLE_basic_fp64ops extension but that would be right enough for linear algebra right? perhaps not for FFT which would require double precision implementations of trigonometric functions..

          and finally once compiled windows port seems that the simple dgemm function doesn't work properly on Nvidia HW.. From what I understand your library should work (altough not achieve top perf) on all OpenCL GPU imps right?

          Thanks.

            • Re: clMAGMA v0.1 Beta now available on UTK's ICL website
              kknox

              Hi Oscar~

              is 1.7 expected to bring more performance to blas routines to 5xxx hw or only is focused only on 7xxx series..

              v1.7 adds performance optimization's that help with smaller rectangular panels, which helps clMAGMA performance because it factorizes bigger problems into a series of calculations on panels.  These optimization's are not specific to 7xxx family of cards.

              also I see from changelog you seem to finally have completed full blas 2 and blas 3 routines.. as my port confirms this new functions are not needed for clmagma compilation, right?

              That's right, the new routines are not currently needed in clMAGMA.  The routines in v1.6 are sufficient for Cholesky, LU and QR.

              Do you plan now on focusing on blas 1 implementations?

              I'm not prepared to announce anything with respect to this yet.

              are you working on compiling this and FFT library for MACOS?

              We have made no effort to compile the APPML libraries for MacOS; I don't know what the demand for this is like.  However, I will say that since our libraries can already cross-compile for both windows/linux, I don't forsee it being difficult to add MacOS support.  Maybe we could create a poll to see what the demand from our users would be like?

              and finally once compiled windows port seems that the simple dgemm function doesn't work properly on Nvidia HW.  From what I understand your library should work (altough not achieve top perf) on all OpenCL GPU imps right?

              Correct, our library should work on any OpenCL device.  Could you send or attach a simple repro case for the failure?  It would be great to see the output of clinfo <or equivalent> to see what hardware and platform you are running on, and the software versions of the SDK and drivers you have installed.

            • Re: clMAGMA v0.1 Beta now available on UTK's ICL website
              oscarbarenys1

              Thanks for detailed response..

              Regarding poll for Macos release yes please go ahead.. I think than better than a fully supported release can do a quick port (hopefully:-)) and say *UNSUPPORTED by AMD* provided for experimental usage.. since at least your competition offers blas and fft optimized libraries on MacOs platforms tough CUDA only..

               

              regarding dgemm failure I'm executing "testing\testing_dgemm.cpp" sample from clmagma 0.1. It fails at

              magma_dgemm which calls your clAmdBlasDgemm routine..