General Distributed Framework for Contraction of Sparse Tensor with Tensor Network

Explore a 10-minute conference talk from ACM SIGPLAN that delves into a general distributed framework for contracting sparse tensors with tensor networks. Learn about algorithms and a runtime system designed to identify and execute the most efficient loop nest for any SpTTN (Sparse Tensor-Tensor Network) kernel. Discover how this approach optimizes performance in applications ranging from machine learning to computational quantum chemistry. Examine the framework's ability to enumerate loop nests for autotuning and find low-cost loop-nests based on metrics like buffer size or cache miss models. Gain insights into the runtime system's capability to identify optimal loop nests without user guidance and provide distributed-memory parallelization for SpTTN kernels. Evaluate the framework's performance using real-world and synthetic tensors, and compare its effectiveness against state-of-the-art libraries and specialized codes.

[CTSTA'23] A General Distributed Framework for Contraction of a Sparse Tensor with a Tensor Network