5 Replies Latest reply on Jan 2, 2008 2:09 PM by chipf

    ACML 4.0.0 now available

    chipf
      ACML 4.0.0 is now posted on http://developer.amd.com/acml.jsp.

      The primary changes in 4.0.0 include support for the Barcelona processors and an update of the LAPACK to the new 3.1.1 version.

      Barcelona support right now consists of new DGEMM, SGEMM, and CFFT kernels.
      A CPUID processor family check determines which version of the kernel will be used.
      More kernels will be added in future releases of ACML.

      Other changes include:
      1. A fix in the DGEBRD routine which affected certain problems in the MOLPRO application.
      2. The addition of some LAPACK utility routines to the windows DLL libraries. The routines added are: *lassq, *lassrt, ilaver, *isnan. Other new LAPACK routines added are dsgesv, zcgesv, dstemr, sstemr, cstemr, and zctemr, which are all part of the 3.1.1 update.
      3. Use of less aggressive optimization flags when compiling the DGEMM FORTRAN cleanup kernel. This resolves a problem where partitioning a DGEMM call into smaller N problem sizes would give slightly different answers compared to calling DGEMM with the full problem. This fix is applied to the gfortran and pgi linux64 libraries.

      The Barcelona kernels only operate in 64-bit mode. Because of this, we have not provided 32-bit versions of the new library, except for the ifort-win32 library.
        • ACML 4.0.0 now available
          illenseer
          The ACML 4.0.0 for Windows in booth versions 32bit and 64bit compiled with Intel FORTRAN compilers is available and I tested it.

          Some time ago I contacted ACML support with the following issue:

          Dear Sir or Madam,

          we are currently trying to get ACML 3.6.0 (compiled with IFort for Windows
          64 bit) to link statically with our product. Unfortunately I cannot get
          this running to our satisfaction, because I always get an unresolved symbol

          libacml.lib(acmlputenv.obj) :
          error LNK2019: unresolved external symbol __imp_putenv referenced in function acmlputenv_
          OLDNAMES.lib(putenv.obi) :
          error LNK2001: unresolved external symbol __imp_putenv

          I am linking with multithreaded static CRT (i.e. /MT) resulting in the
          usage of "libcmt.lib" which indeed has this export "putenv" which is
          obviously needed by acml. But as the object file

          "BUILD_ifort_64/amd/OBJECTS/acmlputenv.obj"

          inside "libacml.lib" is obviously compiled with

          Linker Directives
          -----------------
          /DEFAULTLIB:"MSVCRT"
          /DEFAULTLIB:"OLDNAMES"

          as the tool "DUMPBIN.EXE" shows, this always introduces a dependency on
          the new MSVCRT 8.0 (msvcr80.dll) which I have to omit. Thus is it possible
          to get a library "libacml.lib" where this is linked with

          /DEFAULTLIB:"LIBCMT"

          so that I do not get the dependency to link with MSVCRT? Or is there any
          other possibility to get a staic linking?


          Please feel free to come back to me, if you need further information to
          provide an answer.


          Best Regards
          Frank Illenseer


          ACML support answered:

          [...]
          n the future versions of ACML we will consider releasing the static libraries in such a way that the user do not have to have an external dll dependency.
          [...]


          So I was very happy to see an "MT" subdirectory for the 64bit version of ACML 4.0.0 and this one linked as expected and the dependency on "msvcr80.dll" keeps gone.

          But then at the same time I was very frustrated as this static version of ACML 4.0.0 is not there for 32bit!! I.e. there the very same problem as of my initial complaint still persists!

          Please provide the same static version (MT) for 32bit as you did for 64bit.
          So that the two versions of ACML 4.0.0 are shipped in a consistent way. (I.e. both packages contain the same versions.)


          Thanks and regards
          Frank
          • ACML 4.0.0 now available
            zz
            I just test the eigenvalue calculation part use my opetron machine to solve
            for all a eigenvalue of a 2178X2178 float complex matrix with cgeev subroutines.

            I found no improvement. According to AMD website, the new ACML should
            incorporate the lapack 3.6.1, which should be significant fast (up to 10 times
            faster), anybody know the reason behind it?
            • ACML 4.0.0 now available
              mrshurik
              Hello,

              Is there any GPU support?

              http://ati.amd.com/products/streamprocessor/specs.html

              Math functions can be implemented using a new release of AMD Core Math Library (ACML), which accelerates these functions on the GPU, and tools like ATI Shader Analyzer and AMD CodeAnalyst help identify and correct performance issues.
                • ACML 4.0.0 now available
                  chipf
                  The released versions of ACML do not have GPU support yet. We are working on porting routines to run on the FireStream GPUs. We have have encouraging results on DGEMM and SGEMM for large problems. We expect to release a GPU enabled ACML supporting certain key routines at the end of Q1.
                • ACML 4.0.0 now available
                  joseph1110
                  Hi chipf,


                  I currently use an AMD Athlon 64 X2 Dual-Core Processor 5600+ based PC under Red Hat Linux Fedora 7 and
                  do Math., et caetera, as a hobby.
                  Will a Phenom/Barcelona FPU give me ~ 34 - NOT 19 - digits of floating-point precision for "long double"
                  C/C++ variables, please?
                  Could you run a simple test using GCC for me which I could supply, please?!