Tensor computations present signi cant performance challenges that impact a wide spectrum of applications ranging from machine learning, healthcare analytics, social network analysis, data mining to quantum chemistry and signal processing. E orts to improve the performance of tensor computations include exploring data layout, execution scheduling, and parallelism in common tensor kernels. is work presents a benchmark suite for arbitrary-order sparse tensor kernels using state-of-the-art tensor formats: coordinate (COO) and hierarchical coordinate (HiCOO) on CPUs and GPUs. It presents a set of reference tensor kernel implementations that are compatible with real-world tensors and power law tensors extended from synthetic graph generation techniques. We also propose Roo ine performance models for these kernels to provide insights of computer platforms from sparse tensor view.

No abstract

Optimizing scientific applications to take full advantage of modern memory subsystems is a continual challenge for application and compiler developers. Factors beyond working set size affect performance. A benchmark framework that explores the performance in an application-specific manner is essential to characterize memory performance and at the same time inform memory-efficient coding practices. We present AdaptMemBench, a configurable benchmark framework that measures achieved memory performance by emulating application-specific access patterns with a set of kernel-independent driver templates. This framework can explore the performance characteristics of a wide range of access patterns and can be used as a testbed for potential optimizations due to the flexibility of polyhedral code generation. We demonstrate the effectiveness of AdaptMemBench with case studies on commonly used computational kernels such as triad and multidimensional stencil patterns.

No abstract