Community Research and Development Information Service - CORDIS

H2020

LPGPU2 Report Summary

Project ID: 688759
Funded under: H2020-EU.2.1.1.

Periodic Reporting for period 1 - LPGPU2 (Low-Power Parallel Computing on GPUs 2)

Reporting period: 2016-01-01 to 2017-03-31

Summary of the context and overall objectives of the project

"Consumers today expect to be able to carry a supercomputer around with them, with detailed graphical displays, that are easy to use and last for more than a day on a single battery charge. Not only that, but now users expect their devices to see, listen and understand the world around them. This applies to devices from smartphones that can understand human spoken requests to all the way to cars that can drive themselves. To deliver on this capability requires very power efficient graphics processors (""GPUs"").

Power consumption is important to processing, smart power management algorithms such as dynamic voltage frequency scaling (DVFS) are used as a mitigation for high power consumption but this usually results in a less compelling user experience as the CPU and GPU are clocked down to conserve power resulting in less raw processing power. The strict power limitations means that these demands cannot be met through hardware improvements alone, the software must better exploit the available resources. Unfortunately, programmers are hindered when creating low-power GPU software by the quality of current performance analysis tools. As software becomes more complex it becomes increasingly unmanageable for programmers to optimize the software for low-power devices.

The LPGPU2 consortium has come together to work as a diverse team on delivering low-power GPUs from all the range of angles required. The consortium combines commercial tools, applications, platform and GPU designers with academic researchers to analyse GPU power and performance, define standard interfaces to reliably measure the power and performance, and create a tool chain to provide clear information and insights to software developers. The companies in the consortium are world leaders in power-efficient GPU design (Think Silicon), GPU / compute tools (Codeplay), graphics standards and applications (Samsung), video codecs and media players (Spin Digital), and the university in the consortium (Technical University of Berlin (TUB)) has leading experts on parallel applications and multi-core architectures.

This project proposes to aid the programmer in creating software for low-power GPUs by building on the results of the first LPGPU project to provide a complete performance analysis process for programmers. It will address all aspects of performance analysis, from hardware power and performance counters, to a tool chain that processes and visualizes information from these counters. The main objectives of this project are:

1. To help programmers to improve the energy efficiency of their applications, the LPGPU2 tool chain will provide hints and suggestions to GPU programmers showing ways to reduce power.
2. To enable programmers to be able to write their software once and run it on a variety of different low-power GPUs. The LPGPU2 project will work on standardizing power analysis and power-efficient programming models.
3. To increase the productivity in GPU software development, we propose an approach in which there are layers of technologies, that all work together via open standards, or open source software.
4. To bring technologies to market in a commercializable form, including productizing and commercializing the technologies developed in previous LPGPU (FP7 STREP) project. This includes bringing the SYCL standard into real-world AI applications and putting the LPGPU video decoders into commercial video systems."

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

The project has now been running for 15 months and all partners have been working hard, collaborating together and joining forces in addressing the project from different perspectives; from power models and tool development to applications and algorithm optimizations.

Regarding tools and power modelling, TUB in partnership with Think Silicon have developed counter-based power models that can be calibrated to different hardware platforms. TUB has also developed a flexible power measurement testbed, which was used in calibrating the power models and has a good potential of being commercially exploited. Codeplay ported CodeXL, an open-source profiler, to be the basis for the project's profiling and visualization tool. Samsung has proven the interception and performance counter collection through an API and a post-processing hosting environment. TUB and Think Silicon have been working on integrating their hardware and power models within the tool's visualizations and analysis layers.

On the applications side, Samsung has developed a range of applications showcasing font rendering, augmented reality as well as virtual reality. These will be further optimized using the LPGPU2 tool and help to improve Samsung's mobile graphics platform, which is used by millions of people worldwide. Think Silicon has developed a set of ISP applications using Vulkan and the NemaGFX API. An FPGA prototype has been implemented and the NemaGFX version of the ISP algorithms has been demonstrated at industrial exhibitions. Spin Digital has developed a complete media player using its H.265 codec and a new high-performance video rendering engine that uses the latest graphics APIs (Vulkan, DX12) and allows for the creation of next generation media playback applications (Ultra-HD support, HDR, etc). These were demonstrated at the world’s largest exhibitions of media technologies: NAB (Las Vegas), IBC (Amsterdam), and InterBEE (Tokyo).

The recent surge in the popularity of deep learning has captured interest from LPGPU2 on how to enable these algorithms on low power GPUs. Codeplay has been porting the TensorFlow machine learning framework to OpenCL and SYCL so that the most-used AI framework in the world can work on any AI accelerator that supports OpenCL and SYCL. Codeplay has also developed the VisionCpp open-source library enabling the use of the latest C++ techniques for achieving performance portability on AI and vision algorithms.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

The project will impact any part of European industry involved in visual applications software on portable devices. In particular, it will help European software developers working on videogames and apps for smartphones, by enabling them to analyse the power consumption of their applications and provide ways to reduce the battery usage. By driving down the power consumption of mobile devices when running highly-visual software, we further enable the widespread adoption and ease-of-use of smartphones impacting society at large.

So far some of the project's results could be considered beyond state of the art. TUB has devised novel power modelling techniques and published 2 scientific papers at renowned international conferences. Think Silicon has developed a methodology to insert monitoring hardware counters in embedded GPUs, so that the accuracy of the power estimations driven by the run-time measurements can be fully controlled trading off overheads vs quality. Samsung has enabled the use of cube map textures in mobile 360 video rendering using mobile GPUs instead of existing CPU, and the realistic placement of objects on mobile AR applications. Spin Digital implemented a new video rendering engine, which has better real-time performance than existing solutions and can obtain energy savings up to 25%. Codeplay has taken an established approach, C++ Embedded Domain Specific Languages, and turbo-charged it to work on a variety of parallel processors using open standards. This technique was proven in open-source software, such as VisionCpp, Eigen and TensorFlow.

Related information

Follow us on: RSS Facebook Twitter YouTube Managed by the EU Publications Office Top