Inferring custom architectures from OpenCL

OpenCL has emerged as the de facto cross-platform standard in the GPU-based HPC computing domain. However, in FPGA-based HPC systems, OpenCL-to-FPGA compilers often yield suboptimal results due to the rigid architecture, limited shared-memory, and non-existent inter-work-item communication pathways implied by the OpenCL model. In this work, a methodology of inferring application-specific OpenCL “work-item” interfaces based on kernel code analysis is explored. A proof-of-concept prototype is implemented using an OpenCL source-to-source translator, which allows automated generation of the FPGA-based hardware accelerators directly from the OpenCL sources. The type and implementation of the inferred interface is tailored to match the data access patterns within the kernel. The inferred interface outperforms limitations of the OpenCL rigid architecture and communication model. The presented approach achieves a ~30x speedup over the generic memory-based approach for a 16 work-items application. A set of OpenCL coding patterns targeting FPGA-based HPC systems is also introduced. This technique is demonstrated on a popular bioinformatics algorithm, yet is applicable to any such algorithm with non-standard inter-cell communications.