PatDNN: Achieving Real-Time DNN Execution on Mobile Devices with Pattern-based Weight Pruning

December 2020

tl;dr: Pattern based fine-grained pruning method and use compiler to speed up inference again.

Overall impression

Current pruning methods exhibits two extremes in the design space.

PatDNN introduces a new dimension in design space: fine-grained pruning patterns (thus PatDNN) inside coarse-grained structures. The direct output is still unstructured and does not lend to speedup in HW. Then compiler is used to reorder kernels and compress weight storage and convert the unstructured network to a structured one that is more suitable for HW acceleration.

Key ideas

Technical details