Skip to Main content Skip to Navigation
Preprints, Working Papers, ...

Efficient convolution optimisation by composing micro-kernels

Abstract : 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.
Complete list of metadata

https://hal.archives-ouvertes.fr/hal-03149553
Contributor : Guillaume Iooss <>
Submitted on : Friday, April 9, 2021 - 4:35:56 PM
Last modification on : Saturday, April 10, 2021 - 3:29:12 AM

File

HAL_Ttile_CNNopt_bibtex_update...
Files produced by the author(s)

Identifiers

  • HAL Id : hal-03149553, version 2

Collections

Citation

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

Share

Metrics

Record views

76

Files downloads

42