Neural Network Convolution Folding for Small-Channel Alignment

The patent applies dimension folding to transform the convolution operation by rearranging data dimensions. Specifically, it folds the width and/or height dimensions to the input channel dimension, changing the dimensional structure of the convolution kernel. This allows small output channel dimensions to be mapped to larger effective dimensions that align with hardware parallel operation unit quantities, thereby eliminating resource wastage while maintaining computing performance.

The patent folds the width and/or height dimensions of the convolution kernel to the input channel dimension. This dimensional transformation rearranges data so that the input channel dimension becomes larger and aligns with instruction set vectorization requirements. By doing so, it satisfies hardware alignment constraints and eliminates redundant computing while preserving the original convolution computation accuracy.

The patent performs dimension folding and data rearrangement as a preliminary action before the actual convolution operation. The convolution kernel is pre-transformed by folding width and/or height dimensions to the input channel dimension, creating a folded kernel that is ready for efficient hardware execution. This pre-processing approach simplifies the main convolution operation by ensuring all data is already aligned with hardware requirements, thereby improving parallel operation unit utilization without significantly increasing overall system complexity.