Section: New Results

Dynamic Granularity Adaptation of OpenCL Kernels on Heterogeneous Multi-device Systems

On-going work as part of the PhD of P. Huchant aims to transparently execute an OpenCL kernel, and further a complete task graph, on an heterogeneous multi-device system. We propose methods to split an OpenCL kernel at compile time and adapt its granularity dynamically to ensure load balance. If the kernel is executed multiple times, we propose to determine its granularity by using a linear program whose constraints are built from performance measurements collected during the first invocations of the kernel with predefined granularities. Splitting the execution of one kernel into different executions does not require additional information from the user, therefore increasing the level of portability of OpenCL codes. First experiments show the interest of our approach.