From MATLAB algorithm to energy : exploring and evaluating HLS and DSP workflows

Abstract

Smartphones and IoT devices are energy-constrained systems which require a focus on energy-efficient solutions in computation. To maintain energy-efficient communication, hardware-level acceleration features are essential for meeting performance and power requirements. A baseband processor includes various acceleration features in modern wireless communication systems like 5G and future 6G, which are crucial for energy-efficient designs. MATLAB is a common tool for modeling such systems, providing effective tools for quick prototyping and analysis.

This thesis explores workflows that allow algorithm designers to assess the energy consumption of designs created in MATLAB. Ideally, this should be a fast and straightforward process with short feedback loop. However, the explored workflows involve several steps and ultimately do not provide energy consumption numbers in an ideal format. Nevertheless, the thesis serves as a manual for introduced workflows for generating resource estimates of designs developed in MATLAB. It also increases understanding of input requirements for the workflows and raises awareness of related challenges. The scope of the workflows was narrowed to high-level synthesis tools and DSP architecture. The thesis evaluates seven workflows using qualitative and quantitative metrics while estimating resources for two different benchmarks. Workflows are compared to identify those that are easy to use, flexible, and scalable. Emphasis is on qualitative characteristics, but quantitative evaluation characteristics demonstrate the information the workflows can provide and the conclusions that can be drawn.

The thesis highlights three workflows, one from each of the main target tools. The first is based on an integrated toolchain that includes Cadence Stratus and HDL Coder, the second combines Siemens Catapult with an open-source library called HLSLibs, and the third combines the Cadence Xtensa Xplorer development environment for Tensilica DSP with MATLAB Coder.