SC17 Denver, CO

P10: HiCOO: A Hierarchical Sparse Tensor Format for Tensor Decompositions


Authors: Jiajia Li (Georgia Institute of Technology), Jimeng Sun (Georgia Institute of Technology), Richard Vuduc (Georgia Institute of Technology)

Abstract: This paper proposes a new Hierarchical COOrdinate (HiCOO) format for sparse tensors, which compresses its indices to units of sparse tensor blocks. HiCOO format does not favor one tensor mode over the others, thus can be used as a replacement of the traditional COOrdinate (COO) format. In this paper, we use HiCOO format for the Matriced Tensor Times Khatri-Rao Product (MTTKRP) operation, the most expensive computational core in the popular CANDECOMP/PARAFAC decomposition, then accelerate it on multicore CPU architecture using two parallel strategies for irregular shaped tensors. Parallel MTTKRP using HiCOO format achieves up to 3.5× (2.0× on average) speedup over COO format and up to 4.3× (2.2× on average) speedup over CSF format.
Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf


Poster Index