A Versatile Software Systolic Execution Model for GPU Memory-Bound Kernels

被引:11
作者
Chen, Peng [1 ,2 ]
Wahib, Mohamed [2 ]
Takizawa, Shinichiro [2 ]
Takano, Ryousei [3 ]
Matsuoka, Satoshi [1 ,4 ]
机构
[1] Tokyo Inst Technol, Tokyo, Japan
[2] Natl Inst Adv Ind Sci & Technol, AIST Tokyo Tech Real World Big Data Computat Open, Tsukuba, Ibaraki, Japan
[3] Natl Inst Adv Ind Sci & Technol, Tsukuba, Ibaraki, Japan
[4] RIKEN Ctr Computat Sci, Kobe, Hyogo, Japan
来源
PROCEEDINGS OF SC19: THE INTERNATIONAL CONFERENCE FOR HIGH PERFORMANCE COMPUTING, NETWORKING, STORAGE AND ANALYSIS | 2019年
关键词
Systolic Array; GPU; CUDA; Convolution; Stencil; OPTIMIZATION; ARRAYS; DESIGN;
D O I
10.1145/3295500.3356162
中图分类号
TP301 [理论、方法];
学科分类号
081202 ;
摘要
This paper proposes a versatile high-performance execution model, inspired by systolic arrays, for memory-bound regular kernels running on CUDA-enabled GPUs. We formulate a systolic model that shifts partial sums by CUDA warp primitives for the computation. We also employ register files as a cache resource in order to operate the entire model efficiently. We demonstrate the effectiveness and versatility of the proposed model for a wide variety of stencil kernels that appear commonly in HPC, and also convolution kernels (increasingly important in deep learning workloads). Our algorithm outperforms the top reported state-of-the-art stencil implementations, including implementations with sophisticated temporal and spatial blocking techniques, on the two latest Nvidia architectures: Tesla V100 and P100. For 2D convolution of general filter sizes and shapes, our algorithm is on average 2.5x faster than Nvidia's NPP on V100 and P100 GPUs.
引用
收藏
页数:81
相关论文
共 63 条
[1]   A Predictive Model for Solving Small Linear Algebra Problems in GPU Registers [J].
Anderson, Michael J. ;
Sheffield, David ;
Keutzer, Kurt .
2012 IEEE 26TH INTERNATIONAL PARALLEL AND DISTRIBUTED PROCESSING SYMPOSIUM (IPDPS), 2012, :2-13
[2]  
[Anonymous], 2010, SC10: Proceedings of the 2010 ACM/IEEE International Conference for High Performance Computing, Networking, Storage and Analysis, DOI [10.1109/SC.2010.2, DOI 10.1109/SC.2010.2]
[3]  
[Anonymous], 2018, ARXIV180406826 CS
[4]  
[Anonymous], 2013, GPR, DOI [DOI 10.1145/2400682.2400713, DOI 10.19476/J.YSXB.1004.0609.2013.09.014]
[5]  
[Anonymous], 2017, CUDA C PROGRAMMING G
[6]  
[Anonymous], 2014, ARXIV NEURAL EVOLUTI
[7]  
[Anonymous], 2019, CUDA Toolkit Documentation
[8]  
[Anonymous], 2011, P 4 WORKSH GEN PURP
[9]  
[Anonymous], 2015, ArrayFire - A high performance software library for parallel computing with an easy-to-use API
[10]  
[Anonymous], 2009, P 2 WORKSHOP GEN PUR