AUTH-AUTOGPU

Automatic code generation for Graphics Processing Units

 Partecipanti

Mappa

 Word cloud

Esplora la "nuvola delle parole (Word Cloud) per avere un'idea di massima del progetto.

 Obiettivo del progetto (Objective)

'An interesting trend in multi-core processors is the Graphics Processing Units (GPUs). GPUs consist of hundreds of very simple processing units specialized in efficient parallel computations. Until the recent past, GPUs were seen and used exclusively for the computations required in the processing of ray tracing for graphical life-like scenes mostly used in computer games. But recently there is considerable interest in utilizing the GPUs for high performance numerical computation. GPUs come at a substantial cost discount with respect to attainable performance, due to the ``price subsidization'' by the huge and profitable market for computer games on gaming consoles and personal computers, especially when compared to the tiny market of high performance scientific computing. The dependence of high performance on a specific GPU architecture and the short life due to short product cycle, impose a very heavy burden on programmers responsible for code optimization and maintenance since it has to be redone every few months. With this work, we propose a methodology and a system we call ``AutoGPU'' to extend the life cycle of optimized numerical codes on GPUs by 1) defining a high-level explicit parallelism language for programming numerical algorithms in the domain of signal and image processing applications, 2) automating the code generation with the use of a special compiler, code templates and algorithm libraries and 3) automating the tuning of performance critical software components using advanced optimization and searching techniques. Despite the lack of tools, there is an explosion of effort to utilize the GPU architecture for general purpose computing tasks with very promising results. With the availability of program generators and optimizers like AutoGPU, we expect the promise of supercomputing capability on a desktop computer to become closer to reality.'

Introduzione (Teaser)

Advancing the technology of graphics processing units (GPUs) promises to support a paradigm shift in manufacturing computer processors. One enterprising EU-funded project has made significant headway in the field.

Descrizione progetto (Article)

Powering some of the most sophisticated computer games to date, GPUs are raising the eyebrows of computer scientists and researchers for use in high-performance numerical computation. Thanks to widespread demand for computer gaming, GPU technology has gone down in price to an extent where it is now being seen as a cost-effective avenue to advance high-speed computing.

The EU-funded project 'Automatic code generation for graphics processing units' (AUTH-AUTOGPU) worked on extending the life cycle of optimised numerical codes on GPUs to make them more efficient for everyday computing. One area of focus was on developing a high-level explicit parallelism language for programming numerical algorithms used for signal and image processing applications.

AUTH-AUTOGPU worked on automating code generation by using a special compiler, code templates and algorithm libraries. Using advanced optimisation and searching techniques, project members also targeted automating the tuning of performance-critical software components.

To achieve its aims, the project team successfully developed a platform dubbed AUToGPU, facilitating the design and prototype of parallel algorithms for digital signal and image processing. This has helped accelerate the development cycle, particularly with respect to GPUs, automating performance optimisation, adapting to GPU hardware updates and rendering the software more agile.

AUToGPU exploits special-purpose compiler techniques and novel mathematical abstractions to advance high-performance computation by manipulating domain-specific mathematical structures and matching them to a GPU architecture. It enables better processing of a unified abstraction of the algorithms in high-level mathematical expression.

Rules are then used to transform the algorithm abstraction into equivalent variants using mathematical identities. Ongoing exploration of algorithmic variants helps further applications in Compute Unified Device Architecture, the parallel computing platform and programming model invented by graphics giant NVIDIA.

Against this backdrop, the phenomenal success of NVIDIA and its GPU architecture in the high-performance market is opening new possibilities for big microprocessor provider companies. Both AMD and Intel have entered the GPU accelerator market as interest in this technology increases. In this light, AUToGPU can potentially serve a much larger domain of processors very soon, particularly since GPU architecture is set to become the norm in microprocessor design.