Vectorizing IIR Filters
Making Recursive Filters Parallel Utilizing SIMD
IIR filters dominate real-time audio, yet their feedback paths make straightforward SIMD optimization difficult. This talk distills practical ways to vectorize IIR filters on mainstream CPUs like ARM and x86 without sacrificing numerical stability. We will explore three levels of vectorization of IIR filters: (1) batch multiples channels for lane-wise filtering, (2) restructure multi-band crossovers into a tree-like structure, (3) apply algebraic splits to turn cascaded biquad into concurrent ones. (4) [bonus] Explore a hybrid architecture that seamlessly blends parallel and cascaded filtering to maximize SIMD throughput.
Ayan Shafqat
A dynamic and innovative person with many years of professional experience specializing in software development, algorithm optimization, and embedded systems. Expert in digital signal processing and systems architecture, blending a strong technical background with a profound passion for music and high-performance computing. Adept in a variety of programming languages and operating system kernels, with a particular fondness for exploring and contributing to libre and open-source technologies. Known for a curious and creative approach to problem-solving, excelling in environments that challenge skills and push the boundaries of engineering and human knowledge.