Because they are so small and inexpensive, processors are now included in all but the smallest hardware designs. This grants flexibility to hardware designers because the non-performance-critical components can be implemented in software and the performance-critical components can be implemented in hardware. Using software for parts of the design can decrease the effort required to implement configuration and orchestration logic (for example). It can also offer hardware developers greater adaptability in meeting new project requirements or supporting additional applications.

As a system evolves through design exploration, the boundary between the software and hardware pieces can change substantially. The old paradigm of ``separate hardware and software designs before the project starts’’ is no longer sustainable, and hardware teams are increasingly responsible for delivering significant software components.

Despite this trend, hardware engineers find themselves with surprisingly poor support for the development of the software that is so integral to their project’s success. They are often required to manually develop the necessary software and hardware to connect the two environments. In the software world, this is equivalent to manually re-creating header files from the prose description of an interface implemented by a library. Such ad hoc solutions are tedious, fragile, and difficult to maintain. Without a consistent framework and toolchain for jointly managing the components of the hardware/software boundary, designers are prone to make simple errors which can be expensive to debug.

The goal of our work is to support the flexible and consistent partitioning of designs across hardware and software components. We have identified the following four goals as central to this endeavor:

  • Connect software and hardware by compiling interface declarations.
  • Enable concurrent access to hardware accelerators from software.
  • Enable high-bandwidth sharing of system memory with hardware accelerators.
  • Provide portability across platforms (CPU, OS, bus types, FPGAs).

In this paper, we present a software-driven hardware development framework called Connectal. Connectal consists of a fully-scripted tool-chain and a collection of libraries which can be used to develop production quality applications comprised of software components running on CPUs communicating with hardware components implemented in FPGA or ASIC.

When designing Connectal, our primary goal was to create a collection of components which are easy to use for simple implementations and which can be configured or tuned for high performance in more complicated applications. To this end, we adopted a decidedly minimalist approach, attempting to provide the smallest viable programming interface which can guarantee consistent access to shared resources in a wide range of software and hardware execution environments. Because our framework targets the implementation of performance-critical systems rather than their simulation, we have worked hard to remove any performance penalty associated with its use.

We wrote the hardware components of the Connectal libraries in Bluespec System Verilog (BSV) because it enables a higher level of abstraction than the alternatives and supports parameterized types. The software components are implemented in C/C++. We chose Bluespec interfaces as the interface definition language (IDL) for Connectal’s interface compiler.

This paper describes the Connectal framework, and how it can be used to flexibly move between a variety of software environments and communication models when mapping applications to platforms with connected FPGAs and CPUs.

Document Organization

In Section Accelerating String Search, we present an example running in a number of different execution environments. In Section The Connectal Framework, we give an overview of the Connectal framework and its design goals. In Section Sec-Impl we discuss the details of Connectal and how it can be used to implement the example. Section Sec-ToolChain describes the implementation of Connectal, supported platforms, and the tool chain used to coordinate the various parts of the framework. The paper concludes with a discussion of performance metrics and related work.