@techreport{oai:ipsj.ixsq.nii.ac.jp:00237576,
 author = {Yanchen, Li and Fumihiko, Ino and Yanchen, Li and Fumihiko, Ino},
 issue = {15},
 month = {Aug},
 note = {Deep neural network (DNN) pruning is a popular method for accelerating computations in DNNs by removing unimportant parameters. Among pruning methods, tile-wise pruning (TWP) with sparse matrix multiplication achieves significant acceleration with minimal pruning loss. However, sparse matrix multiplication based on TWP suffers from load imbalance when important weight elements in the matrices of the DNN are unevenly distributed. To address this issue, we propose Adaptive Tile Pruning (ATP) and Split-Tiled Sparse Matrix Multiplication (STSpMM). ATP constructs sparse matrices with flexibly balanced workloads while preserving DNN model accuracy. Meanwhile, STSpMM efficiently handles ATP-generated sparse matrices on GPUs by splitting and redistributing large workloads. We evaluated our approach on pruned ResNet-34 model using ImageNet, and BERT-Small on QNLI tasks. Results demonstrate that ATP-pruned models processed via STSpMM achieve greater acceleration than previous methods while maintaining accuracy., Deep neural network (DNN) pruning is a popular method for accelerating computations in DNNs by removing unimportant parameters. Among pruning methods, tile-wise pruning (TWP) with sparse matrix multiplication achieves significant acceleration with minimal pruning loss. However, sparse matrix multiplication based on TWP suffers from load imbalance when important weight elements in the matrices of the DNN are unevenly distributed. To address this issue, we propose Adaptive Tile Pruning (ATP) and Split-Tiled Sparse Matrix Multiplication (STSpMM). ATP constructs sparse matrices with flexibly balanced workloads while preserving DNN model accuracy. Meanwhile, STSpMM efficiently handles ATP-generated sparse matrices on GPUs by splitting and redistributing large workloads. We evaluated our approach on pruned ResNet-34 model using ImageNet, and BERT-Small on QNLI tasks. Results demonstrate that ATP-pruned models processed via STSpMM achieve greater acceleration than previous methods while maintaining accuracy.},
 title = {Enhancing Sparse DNN Inference on GPUs: Adaptive Tile Pruning and Split-Tiled Sparse Matrix Multiplication},
 year = {2024}
}