RISC-V Audiomark (RISC-V Mentorship)

Mentor: Christian Herber

Company: NXP

EEMBC AudioMark is a benchmark which models a sophisticated, real-world audio pipeline that uses a neural net for keyword spotting. It combines digital signal processing and AI tasks in one workload.

This benchmark currently does not have dedicated RISC-V support. It is based on CMSIS-NN and CMSIS-DSP libraries, for which several vendor-specific ports to RISC-V by IP suppliers exist.

The goal of this mentorship is bootstrap the RISC-V support for AudioMark and the addition of architecture specific optimizations. The architectural extensions which benefit this benchmark are the RISC-V Vector extension, and the currently in-development specification for Packed SIMD (P).

Repository URL: https://github.com/eembc/audiomark

Learning Objectives:

1. RISC-V Toolchains and ecosystem: When working with cutting edge extension, it is necessary to build your own compiler toolchain. Additionally, you will need to work with simulators/emulators and debuggers.
2. Performance benchmarking: You will need to critically assess your own contributions, and
3. RISC-V architecture: You will learn about the RISC-V ISA, including advanced and bleeding-edge extensions.
4. Audio/DSP/AI applications: Working on this benchmark will give you hands on experience with a real Audio processing pipeline

Coding Challenge:

https://docs.google.com/document/d/1BLO9GU57161sGLYuBxm7MzcDJSVZIj5OYhqFli7t-Y0/edit?usp=sharing