A01: GEMM-Like Tensor-Tensor Contraction (GETT)
Authors
Event Type
ACM Student Research Competition
Poster
TimeWednesday, November 15th3:10pm -
3:20pm
Location701
DescriptionTensor contractions (TC) are a performance critical
component in numerous scientific computations. Despite
the close connection between matrix-matrix products
(GEMM) and TCs, the performance of the latter is in
general vastly inferior to that of an optimized GEMM. To
close such a gap, we propose a novel approach: GEMM-like
Tensor-Tensor multiplication (GETT). GETT mimics the
design of a high-performance GEMM implementation; as
such, it systematically reduces an arbitrary tensor
contractions to a highly-optimized "macro-kernel". This
macro-kernel operates on suitably "packed" sub-tensors
that reside in specified levels of the cache hierarchy.
GETT's decisive feature is its ability to pack
subtensors via tensor transpositions, yielding efficient
packing routines. In contrast to previous approaches to
TCs, GETT attains the same I/O cost as an equally-sized
GEMM, making GETT especially well-suited for
bandwidth-bound TCs. GETT's excellent performance is
highlighted across a wide range of random tensor
contractions.
