Speaker: Madhavan Seshadri, Software Developer Ste||ar Group
Developing massively parallel systems is restricted by the complex tasks which need to be managed by the programmer. GPU computing provides the opportunity to parallelize data parallel algorithms while CPU can run the sequential code. With increasing algorithmic development, some new algorithms require iterations of parallel computation on the GPUs (computation scale larger than GPU memory) while some require multiple different data parallel algorithms to run simultaneously, which are notorious to be managed by the programmer.
HPX is an open-source, general purpose C++ library for developing parallel and distributed applications with a broad community usage. This talk aims to discuss the development of HPX Compute language (HPXCL) API for the integration of GPU computation with asynchronous many task execution library HPX. Asynchronous functions are provided for kernel launch, kernel execution and data transfer with the capability to hide the communication latency through computation. To give an example, computation on multiple CPU nodes, GPU nodes can all occur in parallel and can be synchronized when the results are required by the user. This system unleashes the potential to take computation to the exa-scale level.
This development is currently spearheaded by the Stellar Group Community which is a consortium of global researchers. The presenter has been a contributor to this community since his Google Summer of Code participation in 2017.
Github Link: https://github.com/STEllAR-GROUP/hpxcl
Track: Kernel & Platform | Room: Training room 4-3 l Date: Friday, 23rd March, 2018
Produced by Engineers.SG