5. Implementation

The framework is fully automatic, in order to rebuild new results when necessary. However, such procedures take quite a while. POLYBENCH consists of several shell scripts and C programs which take advantage of $C^3$ available parsing mechanism. We use GNU make to generate POLYBENCH executable from the source.

For polyhedral inputs, we use the $C^3$ sparse format, i.e. ASCII, for every implementation, thus an internal format conversion is needed for every new library. Examples of this format can be found in 6_subsec:examples_of_problems.

For each new library, we need to:

• write a wrapper including the internal format conversion and compile it;
• write a simple pb_2time batch and launch it with the polybench_create_2time_criteria_data.sh script;
• write a simple pb_hg batch and launch it with the polybench_create_histograms.sh script.

Outputs are graphs, created by gnuplot with the histogram and datastring enabled patches, and tables in ASCII, that illustrate differences in execution times and in numbers of exceptions. Examples of wrappers and batches are available.