POPA: Expressing High and Portable Performance across Spatial and Vector Architectures for Tensor Computations
Open Access
- 2 April 2024
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 199-210
Abstract
This paper aims at high and portable performance for tensor computations across spatial (e.g., FPGAs) and vector architectures (e.g., GPUs). The state-of-the-art usually address performance portability across vector architectures (CPUs and GPUs). However, they either miss FPGAs or do not achieve high performance. Without a common architectural abstraction, they program and optimize spatial and vector devices separately, causing low portability. We propose a unified programming framework, POPA, which achieves portability via architectural abstraction and performance via specialization. A parallel dataflow machine is proposed as a unified, abstract hardware target that hides differences of concrete architectures. The machine consists of software-defined systolic arrays and a tensor-specific cache hierarchy, which captures pipeline parallelism and customizable memories on FPGAs, as well as multithreading parallelism on GPUs. The machine is specified in a unified programming model as two dataflow graphs for scheduling compute and data movement, respectively. A compiler then specializes the abstract machine to exploit the properties of FPGAs and GPUs, bridging the gap between the abstract machine and a concrete architecture. We evaluate POPA on several Intel FPGAs and GPUs with high-profile tensor kernels, and this is the first system that achieves >=80% performance of expert-written code or machine peak across architectures, to the best of our knowledge.Keywords
Funding Information
- the National Science Foundation of China (T2293700,T2293701,T2325001)
This publication has 36 references indexed in Scilit:
- Gemmini: Enabling Systematic Deep-Learning Architecture Evaluation via Full-Stack Integration2021
- Stateful dataflow multigraphs2019
- A versatile software systolic execution model for GPU memory-bound kernels2019
- Swizzle Inventor2019
- Tiramisu: A Polyhedral Compiler for Expressing Fast and Portable Code2019
- PolySA2018
- SODA2018
- Spatial: a language and compiler for application accelerators2018
- Halide2013
- A practical dynamic single assignment transformationACM Transactions on Design Automation of Electronic Systems, 2007