Exploiting System-Level Concurrency Abstractions for Hardware Descriptions

This technical report explores the idea of using an existing concurrent programming language and its associated tools for the compilation and debugging for modeling parallel computations which can be implemented on FPGAs to yield systems that significantly outperform their sequential software counterparts on conventional processors. An important application of such an approach is to make FPGA-based co-processors more accessible to software developers and other scientist because it removes the need to describe and implement parallel algorithms in terms of conventional hardware descriptions languages like Verilog and VHDL.

Previous work has focused on automatically translating sequential programs into hardware which is a problem which is equivalent to automatic software parallelization. There is no known satisfactory solution for this problem. Other researchers have developed new languages or made modifications to existing languages to add special features for expressing concurrency to help model parallelism in hardware. A distinguishing aspect of our work is that we restrict ourselves to the use of an existing language and its concurrency mechanisms and libraries. By doing so we make it possible for developers to use existing compilers, debuggers and analysis tools to help develop and debug their designs. Furthermore, developers do not need to learn a new language and can rely on mature tools which are well documented.

Another advantage of our approach is that it gives the developer greater control over the quality of results because the synthesized parallel architecture and communication infrastructure is directly related to the original parallel description. This allows the developer to make space/time trade-offs with greater control compared to techniques which rely on more indirect methods for influencing the structure of the output e.g. the use of pragmas.