LPGPU2 by EU Research

Higher performance for low-power system-on-chips Graphics processing units (GPUs) are an important part of many modern technologies, including smartphones and automotive systems. We spoke to Dr Georgios Keramidas about the work of the LPGPU2 project in developing a tool to help developers optimise the software for GPUs, opening up a path towards improved power efficiency. A type of

electronic circuit, low-power graphic processing units (GPUs) are essential to the performance of many modern technologies, including smartphones, wearable technology and certain automotive systems. With GPUs used ever-more widely across a range of different applications, demand is growing for improved performance and power efficiency, an issue central to the work of the recently-concluded LPGPU2 project. “The target of the project was to build a development environment to optimise code, the software for embedded GPUs,” says Dr Georgios Keramidas, the project’s technical coordinator. This work holds clear relevance to the commercial sector, so the project consortium included both academic and industrial partners. “The industrial partners provided some commercial software use cases they wanted to build, which were

LPGPU2 power measurement testbed.

developed as part of the project,” continues Dr Keramidas. “The project delivered an opensource toolsuite, that can be downloaded and used to optimise the software in GPUs. The target was embedded GPUs.” This research was driven to a large extent by the needs of the commercial sector, with companies looking to improve the

performance of low-power devices like smartphones or certain processing devices used in cars for example, and to ensure they can meet evolving market needs without consuming more power. While modern smartphones can of course perform a wide variety of tasks, the market never stands still and more complex applications continue to emerge; one area of interest in the sector is augmented reality applications, which Dr Keramidas says raises new challenges. “If you are going to run augmented reality applications on your smartphone, then you have two main problems. The first is that it’s going to be very slow. The second problem – the more important one – is that if you try to run virtual reality or augmented reality applications on your smartphone, then the battery will drain in a couple of hours,” he outlines. The LPGPU2 Team.

www.euresearcher.com

Example Visualization Features of the LPGPU2 tool.

LPGPU Tool The LPGPU2 tool itself is designed to help extend the battery life of GPU software, providing programmers a means to analyse performance and power consumption. A monitoring framework collects information from the smartphone (or other mobile device), giving programmers a basis on which they can then look to analyse performance. “This includes information from the hardware, from the GPU, from the operating system, from the API, and from the application itself. We have a standardised way of collecting this kind of information,” says Dr Keramidas. This information is then rapidly transferred to a host machine for processing. “We developed various different algorithms in order to process this information. There is a lot of data here, we are talking about big data sets,” stresses Dr Keramidas. “We developed sophisticated algorithms to process this information, and to identify certain parts of the application and visualise their performance. The next step was to provide feedback to the programmers, to enable them to then optimise those parts of the application.” There is a lot of data to gather, so Dr Keramidas and his colleagues in the project developed sampling techniques in order to collect it efficiently. Post-processing tools were developed to analyse the data, along with highly accurate power measurement tools and power models, mainly developed by TUB, from which developers can gain important insights into the power consumption of different components of the GPU. “These

developers are developing applications for embedded devices, for battery-operated devices. For example, this could be people working on smartphone applications,” outlines Dr Keramidas. The project’s research holds relevance for industrial sectors beyond smartphones however, so Dr Keramidas says the consortium also included partners from several other areas, including Codeplay and Spin Digital, a company based in Berlin. “Spin Digital looked at applications for ultrahigh resolution video decoding, rendering, and media players. Then Codeplay provided applications related to machine learning,” he says.

reality applications and augmented reality applications – running for example in S9 or S8 Samsung smartphones – as well as certain computationally intensive applications,” says Dr Keramidas. For their part, Think Silicon are providing applications related to image processing and computational photography. “We are developing the LPGPU2 tool, and we will try to optimise these different applications using this tool. All those applications are commercial applications,” continues Dr Keramidas.

Rigorous framework This

use-case

driven

approach

led

The industrial

partners have provided some commercial software use cases they want to build, which could be developed as part of the project. The project will deliver an open-source tool, that can be downloaded and used to optimise the software in embedded Graphics Processing Units. The other two technology companies involved in the project were Samsung and Think Silicon, a company specialising in the provision of high performance, low power graphics IP semiconductor modules, at which Dr Keramidas is the Chief Scientific Officer. As a major player in the technology industry, Samsung are keen to maintain their position at the forefront of development. “Samsung are using the tool to optimise virtual

to a rigorous framework that allows programmers to optimise applications in these commercial use cases. With minor modifications, the tool could also be used to optimise code from parallel architectures and parallel processors, for example in a desktop machine, yet Dr Keramidas says this is not a priority in research. “In this project we targetted the low-power domain,” he stresses. The tool can be used

EU Research

to optimise software for low-power GPUs, even those built by different vendors, which Dr Keramidas says is an important point for the project’s industrial partners. “If a code is developed for the GPU that is running on your smartphone, and a developer takes this code and then supports it in another GPU, then this code is going to be slow,” he explains. The LPGPU2 tool however is vendoragnostic, which means that it can understand what is running in the other GPU in this type of situation and deal with the required documentations, widening its potential applications. Another important part of the project’s overall agenda related to their work in standardisation, part of the wider goal of enabling cross-platform technology. “We play a role in standardisation efforts. So, we are involved in various standardisation activities, to standardise the way that we collect the data from the device,” outlines Dr Keramidas. In particular, the project played an active role in the Khronos group, an industrial consortium dedicated to creating open standards for various different applications, including graphics and general purpose applications. “We are promoting these standardisation activities in the Khronos group. The Khronos group develops open standards, like OpenGL,

OpenCL, and SYCL for example,” says Dr Keramidas. The wider goal in this research is to develop an open-source tool, available to developers to download, who can then use it to optimise the software running on a GPU. This could help to extend the lifetime of a battery and to improve efficiency, which of course is an important issue in the smartphone sector, where consumers increasingly want to use advanced graphics applications. “For example, a smartphone battery would now last for longer, according to the current power consumption model, as our tool can help to reduce the overall power consumption of the graphics applications,” outlines Dr Keramidas. Important advances have been achieved in these terms, and with the project nearing the end of its term, the focus now is on improving the tool. Open Access Repository: HYPERLINK “https://github.com/codeplaysoftware/ L P G PU2- Co d e X L”h t t p s ://g i t h u b.co m/ codeplaysoftware/LPGPU2-CodeXL Acknowledgement: “The LPGPU2 project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 688759.”

LPGPU2 Low-Power Parallel Computing on GPUs 2 Project Objectives

The objective of the LPGPU2 project is to offer a toolset which delivers significantly longer battery life in mobile devices, while delivering high performance and graphics quality. The LPGPU2 tool will help programmers develop powerefficient code for GPUs by identifying bottlenecks relating to performance (for example in terms of frames-per-second) and power (for example in terms of energy per instruction). The tool supports various state-of-the-art APIs (e.g., OpenGL, Vulkan, and OpenCL) and has been validated in various commercial applications (including VR, AR, multi-API video player, and Neural Network applications) offered by the project patterns.

Project Funding

LPGPU2 has received funding from the European Union’s Horizon 2020 programme.

Project Partners

• Technical University of Berlin (TU Berlin), DE, Project Coordinator • Samsung Research UK, UK • Codeplay Software Ltd., UK • Think Silicon S.A., GR, Project Technical Coordinator • Spin Digital GMBH, DE

Contact Details

Technical Coordinator, Georgios Keramidas Patras Science Park, Rion Achaias, 26504, Greece T: +30 2610 911543 E: g.keramidas@think-silicon.com W: www.think-silicon.com W: www.lpgpu.org/wp Ben Juurlink, Jan Lucas, Nadjib Mammeri, Georgios Keramidas, Katerina Pontzolkova, Ignacio Aransay, Chrysa Kokkala, Martyn Bliss, Andrew Richards. “Enabling GPU software developers to optimize their applications—The LPGPU2 approach,” in Proceedings of IEEE International Conference on Design and Architectures for Signal and Image Processing (DASIP), 2017.

Dr Georgios Keramidas

Dr Georgios Keramidas is the CSO of Think Silicon S.A and Technical Coordinator of the LPGPU2 project. He has a successful track record in delivering commercial projects as well as national and collaborative programmes, while he has also published one book and more than 60 scientific papers on low power processors.