T2S-Tensor: Productively Generating High-Performance Spatial Hardware for Dense Tensor Computations

We present a language and compilation framework for productively generating high-performance systolic arrays for dense tensor kernels on spatial architectures, including FPGAs and CGRAs. It decouples a functional specification from a spatial mapping, allowing programmers to quickly explore various spatial optimizations for the same function. The actual implementation of these optimizations is left to a compiler. Thus, productivity and performance are achieved at the same time. We used this framework to implement several important dense tensor kernels. We implemented dense matrix multiply for an Arria-10 FPGA and a research CGRA, achieving 88% and 92% of the performance of manually written, and highly optimized expert (ninja") implementations in just 3% of their engineering time. Three other tensor kernels, including MTTKRP, TTM and TTMc, were also implemented with high performance and low design effort, and for the first time on spatial architectures."