GLASs: a GPU-accelerated Linear Algebra framework for exascaleproblems in physics and engineering

  • Gasparino Ferreira da Silva, Lucas (Barcelona Supercomputing Center)
  • Quintanas, Adria (Barcelona Supercomputing Center)
  • McDonnell, Cian (UPM)
  • Spiga, Filippo (NVIDIA)
  • Lehmkuhl, Oriol (Barcelona Supercomputing Center)

Please login to view abstract download link

We aim to present GLASs (GPU Library for Accelerated Solvers) as a lightweight, modular framework for implementing iterative solvers for large linear algebra problems on distributed GPU-accelerated systems. Its main design features are: •Easy to use templated class-based C++ interface, with Fortran bindings for easy integration with legacy codes•Lightweight client-library interaction through lambda expression arguments for solver operations •Custom BLAS-like independent modules for implementing compute-intensive operations •Options for performing mixed-precision computations, including NVIDIA’s bfloat16 half precision format •Easy switching between different communication backends, including NCCL •Communication calls performed on device data, enabling the use of GPUDirect RDMA where available