Speaker
Description
OpenCL is a well-established and widely-supported standard for executing parallel workloads
on accelerator devices such as conventional multicore CPUs as well as GPUs and FPGAs.
In this presentation, detail is given on a modern Fortran library which wraps calls to the OpenCL API with a higher abstraction level aimed at scientists and engineers looking to execute highly-parallel OpenCL kernels from Fortran. Modern Fortran features, including derived types, generics, operator-overloading and the iso c binding, are exploited to bring the Fortran style to OpenCL by: abstracting away pointers; providing a level of type-safety for device memory; detecting and handling program errors in a user-friendly manner; and providing a concise but feature-rich interface.
Device kernels are written in the OpenCL C dialect and the Fortran library provides routines to: initialize the accelerator, allocate device memory, enqueue kernels for execution, perform memory transfers and manage device synchronisation.
Code extracts and results are presented for two fluid dynamics codes implementing a lattice Boltzmann method and a multigrid finite volume Euler solver. Background is given on the challenges and design choices for programming GPU hardware from a Fortran perspective, followed by discussion on the future of accelerator offloading from the Fortran language.
