Shard: The Next Generation GPGPU Compiler

Overview:

In recent years, General-Purpose Graphics Processing Units (GPGPUs) possess an innate parallelism architecture, making them highly effective for non-graphics parallel computations. And a multitude of studies and real-world applications have underscored the immense value of GPGPUs in both academic research and the industrial sector such as machine learning, deep learning, computational fluid dynamics, and bioinformatics, to name a few. These units have empowered scientists and engineers to tackle problems that were previously deemed intractable.

Yet despite the immense computational benefits GPGPUs bring, they pose novel challenges to programming models, compilations, optimizations, and code generations, for which a whole new discipline of theory and implementation techniques are required.

The Shard project investigate these prominent gaps, with a distinct emphasis on the GPGPU compiler research, architectural design, optimization algorithms, and novel security mechanisms. Specifically, this project aims to:

investigate compiler design and implementation techniques for GPGPU compilers;
improve robustness and trustworthiness of GPGPU Compilers; and
optimization algorithms (especially post-link binary optimizations) to improve performance.

Publications:

Members:

Qiliang Fan
Baojian Hua
Jiacheng Zhao

Computer Systems & Security Group, USTC