GXBIT: Combining polyhedral model with dynamic binary translation

During the last decade, polyhedral model has been widely used as a mathematical model for autoparallelization, and in recent years, with the development of multi-core architecture, polyhedral model has been employed to transform sequential code to parallel code that can run simultaneously on different cores. At the same time, the rapid development of GPU makes CPU/GPU architecture become increasingly popular because of GPU's powerful parallel processing capabilities. However, we have no other methods of using GPU except for CUDA and Stream SDK, which are all based on explicit programming. Apart from this, there are two constraints for explicit programming: binary incompatibility among different GPUs as well as the cost of rewriting source code. Considering these constraints, we use polyhedral model and a dynamic binary translator to build a virtual execution environment: GXBit. GXBit is composed of analysis and execution stage, and the former one is the main focus of this paper. Analysis stage uses binary instrumentation and binary analysis to probe potential parallel parts (usually nested loop) of a binary executable and then polyhedral model is employed to detect whether there is data dependence or not among all iterations of a nested loop. Execution stage is discussed briefly in this paper. Although there are performance loss in binary translation, GXBit has an 8x speedup on average to X86 compute-intensive version through the result of running two applications taken from the CUDA SDK Sample and one test case from the UIUC Parboil Benchmark Suite. © 2005 - 2012 JATIT & LLS. All rights reserved.