NAME

KappaCUDA - Easy access to NVIDIA CUDA from Perl using the Kappa Library.

SYNOPSIS

  use strict;
  use warnings;
  use threads;
  # When Kappa keyword Perl is available, then variables can be
  # shared between this main perl interpreter and the perl interpreters
  # running as 'kernels' in the scheduled execution context.
  # Perl interpreters do not run on the GPU--they run on the CPU's
  # --but when they are running in the scheduled execution context they can
  # access the Kappa Variables and properties of of the CUDA GPU
  # for the given CUDA GPU context.
  use threads::shared;
  use KappaCUDA;

  my $kappa = KappaCUDA::Kappa::Instance("","",0);

  my $process = $kappa->GetProcess(0,0);

  # Setup CUDA context
  $process->ExecuteString ('<kappa>' . "\n" .
     '!Context ->  context;' . "\n" .
   '</kappa>' . "\n");

  # Setup dimensions
  $process->ExecuteString ('<kappa>' . "\n" .
    '!Value -> WA = (3 * {BLOCK_SIZE}); // Matrix A width' . "\n" .
    '!Value -> HA = (5 * {BLOCK_SIZE}); // Matrix A height' . "\n" .
    '!Value -> WB = (8 * {BLOCK_SIZE}); // Matrix B width' . "\n" .
    '!Value -> HB = #WA;                // Matrix B height' . "\n" .
    '!Value -> WC = #WB;                // Matrix C width' . "\n" .
    '!Value -> HC = #HA;                // Matrix C height' . "\n" .
  '</kappa>' . "\n");

  # Load matrixMul_kernel CUDA or PTX code
  $process->ExecuteString ('<kappa>' . "\n" .
    '!CUDA/Module MODULE_TYPE=%KAPPA{CU_MODULE} -> matrixMul = \'matrixMul_kernel.cu\';' . "\n" .
  '</kappa>' . "\n");

  # Create Variables
  $process->ExecuteString ('<kappa>' . "\n" .
    '!Variable -> A(#WA,#HA,%sizeof{float});' . "\n" .
    '!Variable -> B(#WB,#HB,%sizeof{float});' . "\n" .
    '!Variable VARIABLE_TYPE=%KAPPA{Device} DEVICEMEMSET=true ' . "\n" .
    '-> C(#WC,#HC,%sizeof{float});' . "\n" .
  '</kappa>' . "\n");

  # Free Variables
  $process->ExecuteString ('<kappa>' . "\n" .
    '!Free -> A;' . "\n" .
    '!Free -> B;' . "\n" .
    '!Free -> C;' . "\n" .
  '</kappa>' . "\n");

  # Get CUDA kernel attributes
  $process->ExecuteString ('<kappa>' . "\n" .
    '!CUDA/Kernel/Attributes MODULE=matrixMul -> matrixMul;' . "\n" .
  '</kappa>' . "\n");

  # Unload CUDA module
  $process->ExecuteString ('<kappa>' . "\n" .
    '!CUDA/ModuleUnload -> matrixMul;' . "\n" .
  '</kappa>' . "\n");

  # Reset CUDA context
  $process->ExecuteString ('<kappa>' . "\n" .
    '!ContextReset -> Context_reset;' . "\n" .
  '</kappa>' . "\n");

  # Stop and Finish
  $process->ExecuteString ('<kappa>' . "\n" .
    '!Stop;' . "\n" .
    '!Finish;' . "\n" .
  '</kappa>' . "\n");

  # wait for completion
  $kappa->WaitForAll();

DESCRIPTION

This module gives access (via the Kappa Library) to all NVIDA CUDA driver functionality (a superset of the NVIDIA CUDA runtime functionality). Since the Kappa Library is written in C++ and most (host) processing occurs within the Kappa library, the KappaCUDA module gives CUDA performance comparable to a C++ CUDA driver program. Since the Kappa Library automatically chooses NVIDIA suggested performance optimizations for the given NVIDIA GPU's and implements automatic concurrent CUDA kernel execution on FERMI GPU's, the KappaCUDA module can give CUDA performance better than C++ CUDA driver or runtime API programs (however, a hand tuned C++ Kappa Library or CUDA driver program will be faster).

The (Commercial) Kappa Library is required and is available from psilambda.com. The Kappa Library requires an appropriate NVIDIA CUDA GPU and driver--see the installation guides for the Kappa Library for other prerequisites.

Kappa classes, methods, and syntax

The following Kappa Library classes are available:

Kappa.h KappaConfig.h kappa/ArgumentsDirection.h kappa/ExceptionHandler.h kappa/Process.h kappa/Result.h kappa/Namespace.h kappa/Values.h kappa/Value.h kappa/Resource.h kappa/Instruction.h kappa/Attributes.h kappa/Arguments.h kappa/ProcessControlBlock.h kappa/Context.h

The Kappa User Guide and the Kappa Reference Manual which are available at http://psilambda.com/support/documentation document these classes, their methods, and the kappa syntax.

When using these classes, the class name in the documentation must be changed to the wrapped perl version by prepending the 'KappaCUDA::' prefix. For example: the C++ Kappa::Instance static method becomes the KappaCUDA::Kappa::Instance method.

Generation of KappaCUDA.pm and KappaCUDA_wrap.cxx.

This Perl module was generated using SWIG version 1.3.39. Assuming that the CUDA toolkit is installed in /usr/local/cuda and the Kappa Library header files are installed in /usr/include, the following command will regenerate the KappaCUDA.pm and KappaCUDA_wrap.cxx files:

swig -c++ -outcurrentdir -I/usr/local/cuda/include -I/usr/include -perl KappaCUDA.i

AUTHOR

Brian H. Dunford-Shore

Psi Lambda LLC <kappa@psilambda.com>.