The PGI Accelerator Programming Model for GPUs and Beyond

Colloq: Speaker: 
Michael Wolfe
Colloq: Speaker Institution: 
The Portland Group, Inc.
Colloq: Date and Time: 
Fri, 2010-03-12 10:00
Colloq: Location: 
Building 5100, Room 128 (JICS Auditorium)
Colloq: Host: 
Jeffrey Vetter
Colloq: Host Email: 
vetter@ornl.gov
Colloq: Abstract: 
For the past decade, the top HPC systems have been massively parallel clusters of commodity processors; current and future HPC systems increase performance and strong scaling by adding one or more highly parallel accelerators to each multicore node. We present PGI's roadmap to address the programming challenges for such systems. PGI Accelerator Fortran and C99 compilers offer a proven method to program host+accelerator (such as x64+GPU) platforms, using techniques employed by classical vectorizing and OpenMP compilers. Automatic compiler analysis, assisted and augmented by directives, enables users to generate efficient accelerator code without resorting to new languages or rewriting in new C++ class/template libraries. Compiler feedback allows the programmer to drive the optimization process and tune performance incrementally. These same techniques can be used to program massively multicore CPUs, supporting high degrees of parallelism, preserving portability across a wide range of system architectures, and increasing programmer productivity on heterogeneous manycore systems.We look at the current status and future directions.
Colloq: Speaker Bio: 
Michael Wolfe has actively developed compilers and tools for parallel computing in industry and academia for over 30 years. He has been with PGI as a compiler engineer since 1996. He has published one textbook and several dozen technical papers.