PGI releases PGI Accelerator compiler with OpenACC support

The Portland Group, a leading provider of graphics processing units (GPUs) and graphics accelerators, has announced the availability of its PGI Accelerator™ Fortran and C compilers that support the OpenACC® 1.0 standard, a new imperative programming model for graphics processing units (GPUs) and graphics accelerators. The OpenACC® programming standard lets programmers provide hints or instructions in the programs they write.

"PGI continues to make GPU programming easier and more efficient for experts in science and engineering, and the OpenACC standard, as a subset of the PGI Accelerator programming model, is a perfect fit for our existing compiler products," said Douglas Miles, director of the Portland Group. "PGI's Fortran 2003 and C versions of the PGI Accelerator compilers support high-level imperative x64+ GPU programming and GPU programming using the CUDA C/C++/Fortran extensions."

The OpenACC standard was co-founded by PGI, Cray and NVIDIA, and supported by CAPS. It was first released at the SC11 conference in November 2011. It is fully compatible with and interoperable with the NVIDIA® CUDA® parallel programming architecture. The OpenACC 1.0 technical specification was jointly developed by the three founding companies, and most of the content comes from the PGI Accelerator programming model. The OpenACC application program interface (API) describes a series of compiler instructions that transfer loop scheduling statements and code segments written in standard C, C++ and Fortran languages ​​from the main CPU to the connected accelerator, allowing developers to port code across operating systems, main CPUs and accelerators. After adopting the instruction compilation method, developers only need a multi-platform and multi-vendor compatible code base, which brings many advantages and conveniences to cross-platform and multi-generation application development.

By adopting the OpenACC standard, parallel programmers can move code from the main CPU to the connected accelerator by indicating (also known as directive identification) the code segments suitable for acceleration. In addition to bringing parallel computing opportunities to the compiler, directives can also be used to explain how to efficiently copy loop statements to a specific accelerator and how to optimize data movement. The instruction programming model provides developers with more development resources, allowing the compiler to perform the delicate process of offloading the main CPU and transferring computing tasks to the accelerator, allowing developers to focus on optimizing algorithms and application functions. The main benefit of the instruction programming model is that it can significantly improve application software performance without modifying the underlying source code.

PGI Accelerator Compiler with OpenACC Support

First released in 2008, the PGI Accelerator compiler further enhances the OpenACC standard in the following areas, providing users with greater functionality and design flexibility:

Automatically generate optimal loop scheduling instructions

Automatically occupy shared memory space

Automatic sum reductions

Interoperable with CUDA Fortran and CUDA C/C++

PGI Unified Binary™ technology executables work properly with or without an accelerator

In addition, the PGI Accelerator compiler includes PGI's full set of x86 host performance optimization technologies, including automatic SIMD vectorization, automatic parallelization, interprocedural analysis, intrinsic functions, etc.

Currently, the PGI Accelerator Fortran and C99 compilers are supported on x64+NVIDIA systems running on Linux, OS X, and Windows. All computer systems with Intel and AMD x64 processors and built-in NVIDIA GPUs with CUDA capabilities support the PGI Accelerator Fortran and C99 compilers. PGI 2012 compilers version 12.6 and later will support OpenACC.