OpenCL applications may present tight constraints on work-group size due to algorithm design or chosen implementation strategy. This may hamper functional or performance portability across different platforms, due to lack of resources. The current solution is to re-design the implementation, optimizing it for the new platform. However, this can become a showstopper for new platforms, for which a large manual optimization effort is needed to port benchmark suites and applications. In this work, we aim at tackling such issues by applying work-item coalescing techniques to optimize the mapping of the work-items to the processing elements. However, this is generally not sufficient to achieve good performance as different design patterns may be applied to exploit the specific features of the target architecture. We show how additional target specific transformations can improve the performance with respect to the work-items coalescing baseline. We employ a Matrix Multiply case study to show how the work-item coalescing transformations can impact functional portability, together with providing an opportunity of automatically inserting the use of asynchronous copies on embedded many-core platforms endowed with such a feature.
