Lattice QCD code Bridge++ on multi-thread and many core accelerators

Abstract

We are developing an object oriented code set "Bridge++" for simulation of lattice gauge theories. It aims to be an extensible, readable, and portable workbench while keeping sufficiently high performance in actual productive runs. This paper describes the status of two extensions. One is multi-threading with OpenMP in the recently released version 1.2. The other is a design to use arithmetic accelerator devices such as GPGPUs, which is still under developing version. Feasibility test is performed with OpenCL.

Bridge++ 2.0: Benchmark results on supercomputer Fugaku

“…Within this comparison, even solely solving one equation with multi-gird solver is faster than mixed BiCGStab including the large setup overhead to build the coarse system. The setup overhead becomes irrelevant in solving O (10) or more equations with the same background gauge field, as shown in the right panel of the The left panel is the elapsed time to solve 1 equation and the right panel is to solve 12 equations, including setup time denoted light colors, with 16 or null space vectors (denoted by 𝑛) to build the coarse system. For comparison, the timing with mixed precision BiCGStab solver is plotted.…”

“…Within this comparison, even solely solving one equation with multi-gird solver is faster than mixed BiCGStab including the large setup overhead to build the coarse system. The setup overhead becomes irrelevant in solving O (10) or more equations with the same background gauge field, as shown in the right panel of the…”

Accelerating lattice QCD simulations with 2 flavors of staggered fermions on multiple GPUs using OpenACC—A first attempt

“…An object oriented approach to lattice QCD simulations on accelerators has been under development for some time in Japan under the name Bridge++ [18][19] [20].…”