May 29, 2012 -- December 2011 saw the kick-off of an ambitious research project called “CARP: Correct and Efficient Accelerator Programming”, which aims to boost the programmability of accelerator hardware, such as graphics processing units (GPUs), by innovating in programming language design and implementation, as well as formal verification techniques. Funded by the European Commission’s Seventh Framework Programme (FP7), the consortium, which consists of four industrial (ARM and three SMEs) and four academic partners, seeks to provide a unified flow for developing correct and efficient accelerator software, thus increasing reliability and energy efficiency of computing systems.
“I view accelerator programming as a challenge that must be tackled both from the top-down, via programming frameworks allowing software developers to synthesise efficient platform-specific code from platform-neutral algorithm representation, and from the bottom-up, via effective tools for debugging and verifying low-level code,” said Dr Alastair Donaldson, project coordinator and lecturer at Imperial College London.
“Effective programming tools are essential to help broaden the adoption of heterogeneous systems, such as systems-on-chip accelerated by ARM® Mali™ graphics processing units (GPUs). We aim to provide software developers with a variety of programming technologies that range from industry standards, such as OpenCL™, to domain-specific frameworks. The emphasis is on efficiency, performance portability and productivity,” said Dr Anton Lokhmotov, staff engineer, ARM.
“Parallel programming is becoming increasingly synonymous with accelerator programming. The CARP project is a unique opportunity for programming tools researchers to contribute practical solutions to the productivity, performance, and energy consumption challenges of accelerated computing systems, in close collaboration with hardware vendors and domain experts”, said Dr Albert Cohen, senior research scientist at INRIA.
“Analysing accelerator software both qualitatively and quantitatively, as well as accurately modelling accelerator hardware using stochastic techniques, is key to optimising energy efficiency of computing systems, ranging from embedded devices to supercomputer installations”, said Prof Joost-Pieter Katoen, professor at RWTH Aachen University. “The CARP project opens up a whole new application domain for my research, as verifying low-level software for massively parallel accelerators is the ideal test-bed for scaling up my verification techniques for concurrent software,” said Dr Marieke Huisman, associate professor at University of Twente.
“Whilst I believe that the CARP project will make a lot of progress in minimising the need for low-level accelerator programming, developers of performance-critical code will find invaluable our automatic source code analysis tools which will ensure correctness portability of accelerator software,” said Dr Dino Distefano, CEO of Monoidics.
“Tracking facial features using highly computationally intensive algorithms at real-time often requires specialising software to each platform of interest. We hope to considerably reduce our software development and maintenance costs by minimising the need to write platform-specific code and making it easier to port our software to future platforms,” said Dr Elnar Hajiyev, technology director at Realeyes.
“The ability to perform many compute-intensive tasks on fast and energy-efficient accelerators opens up the door for amazing functionality and user experience improvements. Our Basemark CL benchmarking product is an excellent means for validating tools and techniques developed in the CARP project,” said Ville-Veikko Helppi, marketing director at Rightware.
Links for more information: