Compiler Support for Sparse Tensor Computations in MLIR

Sparse tensors arise in problems science, engineering, machine learning, and data analytics. Programs that operate on such can exploit sparsity to reduce storage requirements computational time. Developing maintaining sparse software by hand, however, is a complex error-prone task. Therefore, we propose treating as property of tensors, not tedious implementation task, letting compiler generate code automatically from sparsity-agnostic definition the computation. This article discusses integrating this idea into MLIR.