Ginkgo: A Modern Linear Operator Algebra Framework for High Performance Computing

Abstract

In this paper, we present G , a modern C++ math library for scienti c high performance computing. While classical linear algebra libraries act on matrix and vector objects, G 's design principle abstracts all functionality as "linear operators, " motivating the notation of a "linear operator algebra library. " G 's current focus is oriented towards providing sparse linear algebra functionality for high performance GPU architectures, but given the library design, this focus can be easily extended to accommodate other algorithms and hardware architectures. We introduce this sophisticated so ware architecture that separates core algorithms from architecture-speci c backends and provide details on extensibility and sustainability measures. We also demonstrate G 's usability by providing examples on how to use its functionality inside the MFEM and deal.ii nite element ecosystems. Finally, we o er a practical demonstration of G 's high performance on state-of-the-art GPU architectures.CCS Concepts: •Mathematics of computing → Mathematical so ware; •Computing methodologies → Massively parallel algorithms; •So ware and its engineering → So ware creation and management;

Accelerating an iterative eigensolver for nuclear structure configuration interaction calculations on GPUs using OpenACC

“…Second, the use of atomics does introduce some overhead to memory accesses, the number of which could potentially be reduced but not eliminated by employing some techniques such as those used in Ref. [32].…”

Porting a sparse linear algebra math library to Intel GPUs

“…GINKGO is a GPU-focused cross-platform linear operator library focusing on sparse linear algebra [3,2]. The library design is guided by combining ecosystem extensibility with heavy, architecture-specific kernel optimization using the platform-native languages CUDA (NVIDIA GPUs), HIP (AMD GPUs), or OpenMP (Intel/AMD/ARM multicore) [4].…”

Ginkgo -- A Math Library designed for Platform Portability

“…This concept is convenient when considering devices such as CUDA or HIP accelerators, which feature their own separate memory space. Although implementing a GINKGO executor that leverages features such as unified virtual memory (UVM) is possible via the interface, in order to attain higher performance we decided to manage all copies by direct calls to the underlying APIs [12].…”