General purpose lattice QCD code set Bridge++ 2.0 for high performance computing

Abstract

Bridge++ is a general-purpose code set for a numerical simulation of lattice QCD aiming at a readable, extensible, and portable code while keeping practically high performance. The previous version of Bridge++ is implemented in double precision with a fixed data layout. To exploit the high arithmetic capability of new processor architecture, we extend the Bridge++ code so that optimized code is available as a new branch, i.e., an alternative to the original code. This paper explains our strategy of implementation and displays application examples to the following architectures and systems: Intel AVX-512 on Xeon Phi Knights Landing, Arm A64FX-SVE on Fujitsu A64FX (Fugaku), NEC SX-Aurora TSUBASA, and GPU cluster with NVIDIA V100.

Bridge++ 2.0: Benchmark results on supercomputer Fugaku

“…Figure 1: Data layout. The site degrees of freedom are packed to the vector register [4]. Two-dimensional tiling is used (the above is 4 × 4 for single precision case).…”

“…After the releas of version 1.0 in 2009, a lot of new architectures have appeared whose performance can be brought out only with specific optimization techniques. We have therefore decided to provide a major update of the code set that contains optimizations to new machines in the forthcoming version 2.0 [4]. The original fixed data layout with double precision floating point numbers is generalized to flexible data layouts in double or single precisions.…”

“…The original fixed data layout with double precision floating point numbers is generalized to flexible data layouts in double or single precisions. Exploratory implementations and tuning have been studied for the SIMD architecture with Intel AVX-512 [5][6][7], GPU architectures with OpenACC [8,9] or OpenCL [10? ], the PEZY-SC manycore accelerator [11], and a vector architecture of NEC SX-Aurora TSUBASA [4]. In the framework of Bridge++ version 2, these machine-specific implementations are provided as alternative codes to the original implementation which takes over the specific tasks such as calculation of fermion propagators.…”

The dynamics of zero modes in lattice gauge theory---difference between SU(2) and SU(3) in 4D

“…The computations were carried out on the PC clusters in KEK Computing Research Center and KEK Theory Center. We have used Bridge++ (http://bridge.kek.jp/Lattice-code/), which is a code set for numerical simulations of lattice gauge theories based on C++ [12,13]. We would like to thank Hideo Matsufuru for his help concerning the usage of this code set.…”

Multigrid Solver on Fugaku

“…Although combining with QWS is mainly for a performance reason, our second purpose is to provide an example of using QWS from an application code. The source code of the multigrid solver will be publicly available as a part of Bridge++ 2.0 [8].…”