Evaluating shared virtual memory in an OpenCL framework for embedded systems on FPGAs

There is now significant interest in OpenCL for FPGAs because it is the first time the FPGA vendors have provided a programming model and a computing platform with integrated high-level synthesis. OpenCL is intended for heterogenous platforms, not just FPGAs, and the standard continues to evolve. Recently, OpenCL has introduced Shared Virtual Memory (SVM) with the goal of simplifying the programming model by allowing hosts and devices to access the same memory space more easily. In this paper, we propose different approaches to implement SVM in an OpenCL framework built specifically to study OpenCL in the context of embedded applications running on FPGAs. We evaluate these different approaches and compare the trade-offs between an OpenCL framework with SVM support and without SVM support. Our results show that the approach that implements the virtual address to physical address translation with a dedicated Memory Management Unit (MMU) performs better than the other approaches. Our results also show that, for input sizes less than 1MB for a vector addition benchmark, the OpenCL framework with SVM support performs better than the OpenCL framework without SVM support until the SVM handling in the kernel starts to dominate.