We present novel techniques for executing signal processing operations such as frequency domain transforms, spectrogram generation, and mel-frequency cepstral coefficients (MFCC) extraction using massively parallel operations supported by IMG Series4 neural network accelerators (NNA). By enabling the execution of a wide variety of such operations on our NNAs, we allow audio and signal preprocessing tasks, that would otherwise have to be executed elsewhere in the system, to be executed on the same device as the neural network, which has potential to reduce overall bandwidth, power and latency requirements. 


Fourier analysis and other signal processing operations have not typically been supported to date on NNAs. It is anticipated that the dedicated convolution support and highly power- and area-efficient SIMD processing available on our Series4 NNA will make it a good fit for these kinds of operations.

It is often the case that some of these operations are computed as a step before a neural network (i.e. data preprocessing) or as a layer inside the neural network itself. Examples of such applications include radar, audio, and medical image preprocessing, all of which typically involve Fourier analysis. In such applications, a lack of support for these preprocessing stages would mean that the computation for the unsupported operations would need to be done on another device. The switch between two different devices is undesirable, firstly because it implies the necessity to have both the devices available, and secondly, because transporting data between the two devices introduces latency, bandwidth and power overheads. These problems are solved by supporting these operations on the NNA