Evaluation of Open-Source Linear Algebra Libraries in Embedded Applications

Many emerging applications based on technologies like machine learning or cryptography mandate increased numerical computations being carried out close to the sensors in respective embedded processing units. Rather than reinventing the wheel, adoption of existing mathematical software libraries from standard desktop computing seems viable. To that end, this paper evaluates existing implementations of the Basic Linear Algebra Subprograms (BLAS) interface for their use on embedded computing devices. In particular, various implementations are benchmarked with regard to their performance and memory consumption on a comparatively small RISC CPU. To facilitate classfication on a broader range we use a diverse selection of test applications from small synthetic benchmarks like simple vector and matrix operations to real-world applications using Artificial Neural Networks (ANNs).