Analysis of OpenCL Work-Group Reduce for Intel GPUs

As hardware becomes more flexible in terms of programming, software APIs must expose hardware features ina portable way. Additions in the OpenCL 2.0 API expose thread communication through the newly defined work-group functions. In this paper we focus on two implementations of the work-group functions in the OpenCL compiler backend for Intel's GPUs. Wefirst describe the particularities of Intel's GEN GPU architecture and the Beignet OpenCL open source project. Both work-group implementations are then detailed, one based on thread to thread message passing while the other on thread to shared local memory read/write. The focus is around choosing an optimal variant based on how each implementation maps to the hardware and its impact on performance.