Automatic Parallelization
Automatic parallelization of general sequential, imperative code is
one of the ultimate goals of compiler construction. Numerical
algorithms in scientific computing can often be parallelized
efficiently with a data parallel approach. Basically there are three
related problems to address: First the data partition problem, second
the mapping problem where sets of partitions are mapped to a
processor, and third the data dependence analysis where to add
communication operations in a distributed memory environment and
synchronization operations in a shared memory environment. Both
partitioning and mapping can be hard problems depending on the data
dependence graph. The data dependence analysis itself can be
technically too complex for compilers. However, often there exist good
solutions for the problems known in the specific area of
application. Hence, parallelization of codes
written in a domain specific languages may be possible, while
parallelization in general is not feasible.
We present some examples codes, where an automatic parallelization can be done. Further, we have constructed a source-to-source C++ compilation system and some application libraries for this purpose.
for further details, download and additional material:
parallel for at SourceForge.net