Overview

CMSIS-DSP is an optimized compute library for embedded systems (DSP is in the name for legacy reasons). It provides optimized compute kernels for Arm Cortex-M and Cortex-A. Different variants are available according to the core and most of the functions are using a vectorized version when vector instructions are available:

• Armv7-M DSP Extensions
• Helium vector instructions
• Neon vector instructions

Vector instructions can provide a huge boost to the performances of an algorithm.

This CMSIS-Pack contains the CMSIS-DSP library.

CMSIS-DSP Kernels

Kernels provided by CMSIS-DSP (list not exhaustive):

• Basic mathematics (real, complex, quaternion, linear algebra, fast math functions)
• DSP (filtering like FIR, Biquad, IIR ...)
• Transforms (CFFT, RFFT, MFCC ...)
• Statistics
• Classical ML (Support vector machine, distance functions for clustering ...)

Kernels are provided with several data types: f64, f32, f16, q31, q15, q7.

C++ API (DSP++)

DSP++ makes it easier to combine existing kernels into efficient Cortex-M algorithms. This optional, header-only C++ API uses automatic loop fusion to avoid temporary arrays and repeated passes over data. It is included in the pack and has no impact on applications using only the C API.

Experimental autodiff extension

Autodiff builds on DSP++ to enable on-device machine learning with a focus on fine-tuning using a subset of existing CMSIS-DSP kernels. It is experimental and not a new ML framework.

CMSIS-DSP and Python

A Python package is available for CMSIS-DSP. It provides access to CMSIS-DSP functions (including fixed point) from Python with NumPy compatibility.

Links

• Documentation
• Examples
• Repository
• Issues