Skip to content

Latest commit

 

History

History
128 lines (89 loc) · 4.57 KB

RELEASE-NOTES.rst

File metadata and controls

128 lines (89 loc) · 4.57 KB
Raw

Release notes for Par4All 1.4.6, 2015/05/20

Par4All is an open-source environment to do source-to-source transformations on C and Fortran programs for parallelizing, optimizing, instrumenting, reverse-engineering, etc. on various targets, from embedded multiprocessor system-on-chip with hardware accelerators up to high performance computer and GPU.

Introduction

Par4All parallelizes C programs to OpenMP, CUDA and OpenCL, and Fortran programs to OpenMP.

This release presents new features of Par4All and PIPS that are described deeper in the changelog file. For all features and changes in the previous versions, please see the release notes archives from git.

Par4All is provided as Debian and Ubuntu Linux packages to ease installation but can be also compiled on Fedora Linux. More detailed information on installation can be found in the Par4All Installation guide.

More documentation and information is available on http://par4all.github.io

If you need more complex transformations, parallelizations and optimizations, you may contact SILKAN for professional support, even if it is no longer supported.

Additions

  • OpenCL:

    This version of Par4All generates kernel codes to be executed on GPU. To compile and to run these codes, CUDA environment should have been installed. For example you can transform a C program example.c to kernel codes in *.cl files and host codes in *.p4a.c files, and then compile the host codes to obtain an executable:

    p4a --opencl example.c -o example

  • C99 option for CUDA code generation:

    Par4All can deal with C99 code sources. Indeed nvcc doesn't support the following C99 syntax:

    f_type my_function(size_t n, size_t m,
                   float array[n][m]) {
    ...
}

    Using --c99 option, p4a will automatically generate the CUDA code in new C89 files (with no VLA but pointers with linearized accesses instead) that will be compiled by nvcc. A simple call to each kernel will be inserted into the original files *.p4a.c that can be compiled with your usual C99 compiler, and linked to the object files compiled from *.cu, to produce an executable.

    Example:

    p4a --cuda --c99 c99_example.c -o c99_example

  • SCMP:

    Generation of codes for the SCMP (Scalable Chip MultiProcessor) data-flow architecture from CEA, funded by SCALOPES ARTEMIS European project. For example the following command will generate an application for the SCMP architecture from a C program:

    p4a --scmp example.c

  • Easier management of compute capabilities of CUDA target, using --cuda-cc option

  • Atomic operations for parallelizing reductions on GPU

  • Can use PoCC to optimize some loop nests

Known issues

  • C:
    • you should avoid heavy use of pointers to enable a good parallelization. The point-to analysis is still on-going in PIPS, but in the general case it is untractable... So avoid if possible. Use nice C99 multidimensional arrays with dynamic size (as in Fortran for long time) and pointer to arrays instead. This is documented into the coding rules, found at http://download.par4all.org/doc/p4a_coding_rules
    • there are still issues in the part that analyzes the control graph (the PIPS controlizer). For example, right now we cannot deal well with a different hierarchy between the control graph and the variable scoping. For example if you have a block with a variable declaration within and a goto to/from outside, it may probably fail. This happens also if you have break or continue in loops with some local variables since those are internally represented with goto.
  • CUDA & OpenCL:
    • In this release the shared memory is not used yet, so to get the best performance, loop nests need to be quite compute intensive, typically with regular accesses to memory. Hopefully, the new cache architecture in Fermi GPU is interesting to balance this limitation.
    • Par4All does not generate additional tiling yet, so right now the iteration spaces of the 3 outer loops of a parallel loop nest are limited by hardware GPU limits.

This text is typeset according to the reStructuredText Markup Specification. For more information: http://docutils.sourceforge.net/docs/ref/rst/restructuredtext.html

This matters because this file is processed to build higher order documentation on Par4All.

%%% Local Variables: %%% mode: rst %%% ispell-local-dictionary: "american" %%% End: