Grant: Collaborative Research and Development Grants
University of Toronto
$39,857.00
Performance auto-tuning of GPU programs using machine learning
EF 2013-14
Résumé
There is a phenomenal growth in software applications for mobile devices such as smart phones and tablets. This growth increases the pressure on mobile processor manufacturers such as Intel, AMD and Qualcomm to make more powerful processors that consume less power. Today, these manufacturers are making processors with multiple cores (CPUs) combined with a general-purpose graphics-processing unit (GPU) on the same chip. The use of a GPU is attractive because the GPU can deliver a ten-fold increase in speed at a fraction of the power, compared to a CPU. However, this GPU benefit can be achieved only if programs are carefully written and tuned for the GPU--a process that is complex and error prone. The goal of this research is to use Machine Learning (ML) to automatically tune programs for better performance on GPUs. We will collaborate with Qualcomm, Inc. in Markham, Ontario to achieve this goal. Qualcomm is a world leader in next generation mobile technologies. They produce processors that integrate several cores with a GPU and are used in many mobile phones and tablets. The proposed work has three thrusts. The first is to develop ML models to determine what code optimizations should be applied based on program characteristics. The large space of GPU code optimizations poses challenges to existing ML models. Thus, we will assess and extend these existing models. The second thrust is to develop compiler support to facilitate the extraction of the program features and to automatically apply the optimizations determined by the models. We will extend our hiCUDA compiler infrastructure to do so. The third thrust is to accelerate key mobile applications to train and evaluate the ML models. Our end goal is to build a framework and integrates ML models with a compiler for GPU performance auto-tuning. Achieving this goal will make Qualcomm's GPUs more appealing to mobile application developers, giving Qualcomm an edge in existing marketplaces and the potential to create new ones. This will lead to job creation, particularly in Ontario, where Qualcomm's mobile application development is done.
Programme
Collaborative Research and Development Grants
Bénéficiaire
Emplacement
Ontario, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
451889-2013
Name nom
Abdelrahman, Tarek
Department département
Electrical & Computer Engineering
Organization id
31
Institution établissement
University of Toronto
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2013
Competition year année de concours
2013
Award amount
39857
Program id
CRDPJ
Program name en
Collaborative Research and Development Grants
Program name fr
Subventions de recherche et développement coopérative
Group en
Research partnerships programs
Group fr
Programmes de partenariat
Committee code
96
Committee name en
RPP Internal Decision Cttee
Committee name fr
Cté de décision interne - PPR
Area of application code
801
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Computer software
Area of application fr
Logiciel
Research subject code
2704
Research subject group en
Information technology
Research subject group fr
Technologie de l'information
Research subject en
Distributed and parallel processing
Research subject fr
Traitement réparti et simultané
Partie
1
Nb partie
3
Application title
Performance auto-tuning of GPU programs using machine learning
Keyword
NSERC
Application summary
There is a phenomenal growth in software applications for mobile devices such as smart phones and tablets. This growth increases the pressure on mobile processor manufacturers such as Intel, AMD and Qualcomm to make more powerful processors that consume less power. Today, these manufacturers are making processors with multiple cores (CPUs) combined with a general-purpose graphics-processing unit (GPU) on the same chip. The use of a GPU is attractive because the GPU can deliver a ten-fold increase in speed at a fraction of the power, compared to a CPU. However, this GPU benefit can be achieved only if programs are carefully written and tuned for the GPU--a process that is complex and error prone. The goal of this research is to use Machine Learning (ML) to automatically tune programs for better performance on GPUs. We will collaborate with Qualcomm, Inc. in Markham, Ontario to achieve this goal. Qualcomm is a world leader in next generation mobile technologies. They produce processors that integrate several cores with a GPU and are used in many mobile phones and tablets. The proposed work has three thrusts. The first is to develop ML models to determine what code optimizations should be applied based on program characteristics. The large space of GPU code optimizations poses challenges to existing ML models. Thus, we will assess and extend these existing models. The second thrust is to develop compiler support to facilitate the extraction of the program features and to automatically apply the optimizations determined by the models. We will extend our hiCUDA compiler infrastructure to do so. The third thrust is to accelerate key mobile applications to train and evaluate the ML models. Our end goal is to build a framework and integrates ML models with a compiler for GPU performance auto-tuning. Achieving this goal will make Qualcomm's GPUs more appealing to mobile application developers, giving Qualcomm an edge in existing marketplaces and the potential to create new ones. This will lead to job creation, particularly in Ontario, where Qualcomm's mobile application development is done.