A High-Performance Sparse Tensor Algebra Compiler in Multi-Level IR

Abstract: Tensor algebra is widely used in many applications, such as scientific computing, machine learning, and data analytics. The tensors represented real-world data are usually large and sparse. There are tens of storage formats designed for sparse matrices and/or tensors and the performance of sparse tensor operations depends on a particular architecture and/or selected sparse format, which makes it challenging to implement and optimize every tensor operation of interest and transfer the code from one architecture to another. We propose a tensor algebra domain-specific language (DSL) and compiler infrastructure to automatically generate kernels for mixed sparse-dense tensor algebra operations, named COMET. The proposed DSL provides high-level programming abstractions that resemble the familiar Einstein notation to represent tensor algebra operations. The compiler performs code optimizations and transformations for efficient code generation while covering a wide range of tensor storage formats. COMET compiler also leverages data reordering to improve spatial or temporal locality for better performance. Our results show that the performance of automatically generated kernels outperforms the state-of-the-art sparse tensor algebra compiler, with up to 20.92x, 6.39x, and 13.9x performance improvement, for parallel SpMV, SpMM, and TTM over TACO, respectively.