A Performance Portability Study Using Tensor Contraction Benchmarks

Driven by the end of Moore’s law, heterogeneous architectures, particularly GPUs, are experiencing a surge in demand and utilization. While these platforms hold the potential for achieving high performance, their programming remains challenging and requires extensive hardware knowledge. This complexity is further exacerbated by the different proprietary languages utilized by various vendors. In this paper, we conduct a performance-portability study on two portable languages, SYCL and Kokkos. Specifically, we focus on the case study of tensor contractions and employ COGENT, a DSL compiler for tensor contractions, to generate CUDA code for the 48 different tensor contractions in the TCCG benchmark suite. We extend COGENT to produce Kokkos code, and use Hipify and SycloMatic, which are tools that convert CUDA code to HIP and SYCL. Our analysis involves a comparison of the performance of each framework on both Nvidia and AMD GPUs. Our experiments show that identically tiled tensor contraction kernels in Kokkos and SYCL can exhibit significant performance differences compared to the corresponding CUDA/HIP program, respectively on Nvidia/AMD GPUs. The main reason for the performance differences arise from differences in register usage and the management of register spills to thread-private stack memory, affecting overall degree of thread-level concurrency and the volume of data movement to/from GPU DRAM.