LibHSA: One step towards mastering the era of heterogeneous hardware accelerators using FPGAs

Abstract

Various signal and image processing applications require vast acceleration in order to enable real-time processing and meet constraints in power consumption. On FPGAs these applications can be implemented as application-specific circuit. Although IP cores for various applications exist, even interfacing these usually requires experienced knowledge in hardware design. Using FPGAs or other accelerators in a heterogeneous system from a host CPU would simplify the usage of accelerator hardware for a common software developer. Recognizing this, several companies and partners from academia created the HSA Foundation (Heterogeneous System Architecture Foundation) to define a platform specification for heterogeneous system requirements as a macro-architecture for efficient and easy targeting heterogeneous processors from popular high-level languages like C/C++, Python, Java and other domain specific languages. In this paper we present an IP library (LibHSA), that greatly simplifies integration of hardware accelerator functions into existing HSA compliant systems. This allows accelerators to take advantage of the existing HSA programming model, libraries, compilers and toolchains. We will demonstrate the work of LibHSA utilizing a programmable image processor implementation on a Xilinx FPGA. The image processor supports low-level algorithms, e.g. Sobel, Median, Laplace, or Gauss. Our results show a substantial decrease integrating customized hardware accelerators using the LibHSA infrastructure. To our knowledge, our library is the first approach for integrating reconfigurable hardware into an HSA compliant system.