Block-based Fast Differentiable IIR in PyTorch
Published:
I recently came across a presentation by Andres Ezequiel Viso from GPU Audio at ADC 2022, in which he talked about how they accelerate IIR filters on the GPU. The approach they use is to formulate the IIR filter as a state-space model (SSM) and augment the transition matrix so that each step processes multiple samples at once. The primary speedup stems from the fact that GPUs are very good at performing large matrix multiplications, and the SSM formulation enables us to leverage this capability.