The University of Illinois at Urbana-Champaign
Year in School
REU Faculty Mentor
Research Area Interest
Architecture, Compilers, and Parallel Computing
Faster Sparse MTTKRP on Scalable Massively Parallel Architectures
Biography & Research Abstract
Tensor decomposition is one of the most fundamental strategies that could extract information from a sparse tensor. One of the most well-known tensor decomposition methods is the canonical polyadic decomposition (CPD). The matricized tensor times Khatri-Rao product (MTTKRP) operation takes most of the time and computing resources when computing CPD. Atomic operations on GPUs, which are heavily relied by many state-of-the-art MTTKRP kernels, are extremely expensive. We proposed a kernel that minimizes the number of atomic operations while maintaining parallelism, and our kernel can also be applied to a scalable architecture.
Hanwen Liu is a rising senior at University of Illinois Urbana-Champaign, where he is completing his major in computer engineering and his minor in mathematics. His interest lies in the area of computer microarchitecture and high-performance computing. Specifically, he keeps implementing and exploring applications of sparse tenser-tensor multiplication on passively parallel architectures. He has served as teaching assistant for several courses. His eventual career goal is to pass on what he has learned with others, as a professor.