I/O-Aware Reformulation Reduces Memory Traffic in Wavelet Convolutions

A new study on arXiv (2608.10805) presents a reformulation of wavelet convolutions (WTConv) that takes I/O into account, making it a popular choice over traditional convolutions in neural networks. WTConv boosts a network's receptive field with each decomposition level while keeping the parameter count linear. However, the typical implementation struggles with memory due to extensive data movement in high-bandwidth memory (HBM). The researchers develop an I/O model to tackle this and suggest three algebraic changes: (1) executing the Haar analysis butterfly on-chip, (2) combining multi-level synthesis into a single pass based on output-coordinate bits, and (3) adding learned per-channel scales to convolution weights. These modifications yield a fused implementation that significantly reduces HBM traffic. The study evaluates the WTConvNeXt setup across various levels and conditions, demonstrating a considerable drop in memory usage. This work is crucial for digital art and AI technology, improving neural network efficiency and enabling more advanced models on limited hardware.