Optimizing Technologies has released a beta version of Auto-Parallelizer for Linux on x86, IA64, and PowerPC platforms. Auto-Parallelizer is a tool that looks for parallelizable parts of source code of applications written in a sequential manner and language. In some cases it modifies the source code to maximize the number of such parts or to extend them. For example, users can write a computational program in a sequential manner, then feed the source code to Auto-Parallelizer and the tool will produce code, some parts of which are executable in parallel. This usually speeds up the program execution by several times.
Auto-Parallelizer uses libraries that support OpenMP standard for executing code in parallel. The current implementation uses libgomp, however, it is portable to any existing library that supports all needed interface functions. The implementation currently parallelizes loops only. In the future it will also parallelize sections of sequential code, e.g. several calls.
At present Auto-Parallelizer works only with GCC with the modification the company made in its sorce code. To make it operate you must have:
- A modified gcc-based compiler that reads and writes program semantics description files
- Auto-Parallelizer addendum that reads files, analyzes, modifies and writes them back
The modified gcc-based part is available as built for x86 Linux flavors binaries and as a source code. The addendum is isolated from gcc -- it is based on Universal Translating Library. The addendum is available as binary package built for x86 Linux.
The company says it has compared performance results of Auto-Parallelizer beta version with the results icc 11.0.074 (one of the most efficient compiler for x86 and IA64 platforms) and gcc 4.3.1 (one of the most efficient compiler for PowerPC platform. The results of the comparisons are available here: