In this paper, we describe a new software package for sparse tensor decompositions that targets emerging parallel computing architectures. We are focused on the canonical polyadic tensor decomposition of sparse tensors, whose key kernel is the matricized tensor times Khatri-Rao product (MTTKRP). Our goal is to develop software that is portable to a variety of multicore and manycore computing architectures, such as multicore CPUs, the Intel Xeon Phi, and NVIDIA GPUs. To do this, we use the Kokkos framework, which is designed to enable high performance across multiple architectures with a single code even though these architectures handle fine-grained parallelism in varying ways. Not only are the specifics of the implementation interesting to persons tuning tensor computations, but also potentially to other high-performance software developers looking to understand ways to make their codes portable across architectures without supporting multiple code bases. Kokkos is an evolving but extensible system, and we explain one problem we had in achieving performance and how we surmounted it by encapsulating fine-grained parallelism through the use of a special compile-time polymorphic array type. We also introduce a new sparse tensor format based on a variation of the traditional coordinate storage format. This format reorders the traversal of tensor nonzeros in the MTTKRP to reduce atomic-write contention; this new format leads to significantly improved performance on all architectures with minimal increase in memory footprint. Performance of our software is measured on several computing platforms, including the Xeon Phi and NVIDIA GPUs. We show that we are competitive with state-of-the-art approaches available in the literature while having the advantage of being able to run on a wider of variety of architectures with a single code.
from cs updates on arXiv.org https://ift.tt/2OaY4G5
//
0 comments:
Post a Comment