An Exploration of OpenCL on Multiple Hardware Platforms for a Numerical Relativity Application

Abstract

Currently there is considerable interest in making use of many-core processor architectures, such as Nvidia and AMD graphics processing units (GPUs) for scientific computing. In this work we explore the use of the Open Computing Language (OpenCL) for a typical Numerical Relativity application: a time-domain Teukolsky equation solver (a linear, hyperbolic, partial differential equation solver using finite-differencing). OpenCL is the only vendor-agnostic and multi-platform parallel computing framework that has been adopted by all major processor vendors. Therefore, it allows us to write portable source-code and run it on a wide variety of compute hardware and perform meaningful comparisons. The outcome of our experimentation suggests that it is relatively straightforward to obtain order-of-magnitude gains in overall application performance by making use of many-core GPUs over multi-core CPUs and this fact is largely independent of the specific hardware architecture and vendor. We also observe that a single high-end GPU can match the performance of a small-sized, message-passing based CPU cluster.

Generic effective source for scalar self-force calculations

“…It seems likely that the embarrassingly parallel nature of the effective source calculation on a grid of points is an ideal candidate for implementation in a GPU programming framework such as CUDA or OpenCL. Given other applications have seen speed-ups by 1 to 2 orders of magnitude [73], it is not unreasonable to expect similar performance gains for effective source calculations.…”

High accuracy gravitational waveforms from black hole binary inspirals using OpenCL

“…In this section we detail our approach taken towards parallelism, not only to take advantage of the many cores of a single GPU, but also those on multiple GPUs. We describe here the different ideas we have implemented and their final performance outcomes [13,14]. The lessons learned have ultimately helped us converge towards a rather optimal implementation.…”

“…In addition, it is necessary to establish the appropriate data communication between the GPU cores and the remaining code that is executing on the CPU -we use clEnqueueReadBuffer, clEnqueueWriteBuffer instructions to transfer data backand-forth from main memory and we only use global memory on the GPU to simplify communication between the GPU cores. We make this simplification with the goal of keeping the code's portability intact, even if it impacts performance to some extent [14].…”