Bridging Portability and Performance in Sparse Tensor Computations Using SYCL

ABSTRACT Sparse tensors have become prevalent data structures in multiple applications, such as medical imaging and machine learning, making operations that decompose them, that is, creating smaller structures that retain most of the original information, essential. Two of the most commonly used tensor decomposition methods are the Canonical Polyadic and Tucker Decomposition, with the most time‐consuming operations being the MTTKRP and TTM‐chain, respectively. Modern computing platforms combine multiple devices with different architectures to achieve unprecedented levels of performance, creating an environment where portability is as important as performance. To tackle this challenge, this work proposes SYCL‐based MTTKRP and TTM‐chain approaches for sparse tensors, which are portable to any CPU or GPU, extending previous literature by handling mode‐4 and mode‐5 tensors and tackling the TTM‐chain operation as a whole, allowing for further optimizations. The experimental results show that the proposed approaches present linear to superlinear scalability as the problem size grows and outperform the portable state‐of‐the‐art by 4.9× on average.