Room: BBL 201
Speaker: Levin Fritz
Title: Automatic Optimization of GPGPU Programs
One of the most important developments in processor technology in recent years
is the evolution of graphics processing units (GPUs) into general-purpose
coprocessors that can perform a wide range of tasks more efficiently than CPUs.
As GPU hardware has become more general, program environments have appeared
that support the development of small programs called kernels which implement
algorithms to be run on the GPU. In consequence, general-purpose GPU
programming, or GPGPU programming, has become accessible to a large number of
However, modern GPUs still retain some features that make it difficult to
optimize GPU programs. This talk presents an approach to solve this problem
through an optimizing compiler that takes as input straight-forward GPU kernel
code, performs low-level and hardware-specific optimizations and produces a
functionally-equivalent but efficient version of the kernel, so that developers
can focus on algorithms and writing clear and correct code.