Efficient convolution optimisation by composing micro-kernels

Nicolas Tollenaere (1) , Auguste Olivry (1) , Guillaume Iooss (1) , Hugo Brunie (2) , Albert Cohen (3) , P Sadayappan (4) , Fabrice Rastello (1)

1 CORSE - Compiler Optimization and Run-time Systems
2 LBNL - Lawrence Berkeley National Laboratory [Berkeley]
3 Google France
4 University of Utah

Tiling is a key loop transformation for optimizing tensor computations such as CNNs (Convolutional Neural Networks). Tile optimization involves an explosively large search space for multi-level tiling, including all possible permutations of the tiling loops and all possible valid tile sizes. In this paper, we develop a comprehensive methodology for finding optimized tile configurations with imperfectly nested micro-kernels ("beyond perfect") and outer tile loops optimized via analytical modeling. Experimental results on over 30 CNN benchmarks from three popular DNN pipelines demonstrate the effectiveness of the presented optimization approach by comparing with the Intel oneDNN library.

Domaines: Performance et fiabilité [cs.PF]; Calcul parallèle, distribué et partagé [cs.DC]

Nicolas Tollenaere, Auguste Olivry, Guillaume Iooss, Hugo Brunie, Albert Cohen, et al.. Efficient convolution optimisation by composing micro-kernels. 2021. ⟨hal-03149553v2⟩