Automatic parallelization for graphics processing units

Abstract

Accelerated graphics cards, or Graphics Processing Units (GPUs), have become ubiquitous in recent years. On the right kinds of problems, GPUs greatly surpass CPUs in terms of raw performance. However, because they are difficult to program, GPUs are used only for a narrow class of special-purpose applications; the raw processing power made available by GPUs is unused most of the time.This paper presents an extension to a Java JIT compiler that executes suitable code on the GPU instead of the CPU. Both static and dynamic features are used to decide whether it is feasible and beneficial to off-load a piece of code on the GPU. The paper presents a cost model that balances the speedup available from the GPU against the cost of transferring input and output data between main memory and GPU memory. The cost model is parameterized so that it can be applied to different hardware combinations. The paper also presents ways to overcome several obstacles to parallelization inherent in the design of the Java bytecode language: unstructured control flow, the lack of multi-dimensional arrays, the precise exception semantics, and the proliferation of indirect references.

Leveraging GPUs using cooperative loop speculation

“…It will be even harder to extract the fine-grained parallelism necessary for efficient use of many-core systems like GPUs with thousands of threads. Therefore, several automatic static parallelization techniques for GPUs have been proposed to exploit more parallelism [Han and Abdelrahman 2010;Baskaran et al 2010;Wolfe 2010;Leung et al 2009;.…”

“…There are previous works that have focused on generating CUDA code from sequential input [Han and Abdelrahman 2010;Baskaran et al 2010;Wolfe 2010;Leung et al 2009;. HiCUDA [Han and Abdelrahman 2010] is a high-level directive-based compiler framework for CUDA programming where programmers need to insert directives into sequential C code to define the boundaries of kernel functions.…”

“…proposed accelerator, in which programmers use C# and a library to write their programs and let the compiler generate efficient GPU code. The work by Leung et al [2009] proposes an extension to a Java JIT compiler that executes program on the GPU. Delite [Chafi et al 2011] is another approach that aims at simplifying the creation of performance-oriented DSLs and compiling them for heterogeneous systems, including systems with GPUs.…”

RELEASE: A High-Level Paradigm for Reliable Large-Scale Server Software

“…Some of the more recent such research aims to effectively employ powerful dedicated and specialized co-processors like graphic cards. Notable VM designs in this direction are the CellVM [15], [30] for the Cell Broadband Engine, a VM with a distributed Java heap on a homogeneous TILE-64 system [28]; an extension of the JikesVM to detect and offload loops on CUDA devices [12]; and VMs for Intel's Larrabee GPGPU architecture [22]. Most of these designs are specific to VMs for Java-like languages which are based on a shared-memory concurrent programming model.…”

Mathematical Morphology in Computer Graphics, Scientific Visualization and Visual Exploration

“…However, in 2006 NVidia introduced a programming environment called CUDA [25], which allows the GPU to be programmed through more traditional means. At this moment a dedicated programming effort is still required to develop algorithms that perform efficiently on GPU hardware, but efforts are underway for automatic transformation of CPU programs into GPU counterparts [24].…”