Software Tool Helps Tap Into The Power Of Graphics Processing
Today’s computers rely on powerful graphics processing units (GPUs) to create the spectacular graphics in video games. In fact, these GPUs are now more powerful than the traditional central processing units (CPUs) – or brains of the computer. As a result, computer developers are trying to tap into the power of these GPUs. Now a research team from North Carolina State University has developed software that could make it easier for traditional software programs to take advantage of the powerful GPUs, essentially increasing complex computing brainpower.
Taking advantage of a GPU’s processing ability is a big deal, because of the amount of computing power a GPU contains. The CPU from an average computer has about 10 gigaflops of computing power – or 10 billion operations per second. That sounds like a lot until you consider that the GPU from an average modern computer has 1 teraflop of computing power – which is 1 trillion operations per second.
But using a GPU for general computing functions isn’t easy. The actual architecture of the GPU itself is designed to process graphics, not other applications. Because GPUs focus on turning data into millions of pixels on a screen, the architecture is designed to have many operations taking place in isolation from each other. The operation telling one pixel what to do is separate from the operations telling other pixels what to do. This hardware design makes graphics processing more efficient, but presents a stumbling block for those who want to use GPUs for more complex computing processes.
A research team from NC state has developed software that could make it easier for traditional software programs to take advantage of GPUs. The research was funded by the National Science Foundation.
“We have developed a software tool that takes computer program A and translates it into computer program B – which ultimately does the same thing program A does, but does it more efficiently on a GPU,” says Dr. Huiyang Zhou, an associate professor of electrical and computer engineering at NC State and co-author of a paper describing the research. This sort of translation tool is called a compiler.
Program A, which is the user-provided input, is called a “naïve” version – it doesn’t consider GPU optimization, but focuses on providing a clear series of commands that tell the computer what to do. Zhou’s compiler software takes the naïve version and translates it into a program that can effectively utilize the GPU’s hardware so that the program operates a lot more quickly.
Zhou’s research team tested a series of standard programs to determine whether programs translated by their compiler software actually operated more efficiently than code that had been manually optimized for GPU use by leading GPU developers. Their results showed that programs translated by their compiler software ran approximately 30 percent more quickly than those optimized by the GPU developers.
“Tapping into your GPU can turn your personal computer into a supercomputer,” Zhou says.
The paper, “A GPGPU Compiler for Memory Optimization and Parallelism Management,” was co-authored by Zhou, NC State Ph.D. student Yi Yang, and University of Central Florida Ph.D. students Ping Xiang and Jingfei Kong. The paper will be presented June 7 at the Programming Language Design and Implementation conference in Toronto.
NC State’s Department of Electrical and Computer Engineering is part of the university’s College of Engineering.
Note to editors: The study abstract follows.
“A GPGPU Compiler for Memory Optimization and Parallelism Management”
Authors: Yi Yang, Huiyang Zhou, North Carolina State University; Ping Xiang, Jingei Kong, University of Central Florida
Presented: June 7, Programming Language Design and Implementation conference, Toronto
Abstract: This paper presents a novel optimizing compiler for general purpose computation on graphics processing units (GPGPU). It addresses two major challenges of developing high performance GPGPU programs: effective utilization of GPU memory hierarchy and judicious management of parallelism. The input to our compiler is a naïve GPU kernel function, which is functionally correct but without any consideration for performance optimization. The compiler analyzes the code, identifies its memory access patterns, and generates both the optimized kernel and the kernel invocation parameters. Our optimization process includes vectorization and memory coalescing for memory bandwidth enhancement, tiling and unrolling for data reuse and parallelism management, and thread block remapping or address-offset insertion for partition-camping elimination. The experiments on a set of scientific and media processing algorithms show that our optimized code achieves very high performance, either superior or very close to the highly fine-tuned library, NVIDIA CUBLAS 2.2, and up to 128 times speedups over the naive versions. Another distinguishing feature of our compiler is the understandability of the optimized code, which is useful for performance analysis and algorithm refinement.