Profiling SIMD Instructions for the Acceleration of Audio Algorithms

Abstract

This thesis aims to contribute to the Lightweight Embedded Audio Framework (LEAF) by examining the performance of SIMD-related instructions on an STM32H750 microcontroller. While others have successfully used SIMD instructions to accelerate audio algorithms generally, there is little publicly-accessible research surrounding the ability to accelerate audio algorithms on embedded hardware. In this thesis, I run 24 SIMD-related instructions from the ARM Cortex-M7 instruction set on a Daisy Seed, an embedded platform for music that uses an STM32H750 microcontroller. I count the number of cycles required for each of these instructions and compare them to the number of cycles used by similar default instructions. The findings show that the SIMD instructions are consistently faster than the default instructions, although the difference in the number of cycles varies slightly depending on the application. The most noticeable difference occurs when comparing the SIMD instructions for signed 8-bit addition and subtraction (instructions SADD8 and SSUB8) against four default signed additions and subtractions, respectively. The most relevant acceleration for LEAF applications appears in the 30% speed up in the 16-bit SIMD operations compared to the 32-bit float operations. Future research could include performing a more rigorous timing analysis with different inputs and accounting for the time it takes to set up the SIMD operands.