High-level synthesis with SIMD units

This paper presents novel techniques to integrate the use of Single Instruction Multiple Data (SIMD) functional units in a high-level synthesis (HLS) design methodology. SIMD functional units can be configured to operate in one or more SIMD modes, in which they process multiple sets of smaller bitwidth operands in parallel. Conceptually, the use of SIMD functional units enables HLS to (i) exploit parallelism to a higher degree without using additional resources, (ii) improve resource utilization by enabling hardware re-use at a fine-grained level, and (iii) improve energy efficiency for a given area and/or performance constraint. We illustrate the issues involved in performing high-level synthesis with SIMD functional units, and discuss how algorithms involved in a typical high-level synthesis flow can be enhanced to result in maximal performance and energy improvements. These techniques are not restricted to specific high-level synthesis tools/algorithms, and can be plugged into any generic high-level synthesis system.