Thread-level Value Speculation for Image-processing Applications

Thread-level speculation (TLS) is a technique used to aggressively exploit the parallelism of a program at runtime. In this paper, we develop a programming model for software-based, thread-level value speculation that can be used for image processing. The proposed technique allows users to leverage thread-level value speculation to parallelize programs step-by-step. The whole implementation is built upon the POSIX (Portable Operating System Interface of UNIX) thread library. The experimental results show that the proposed programming model can generate performance benefits for the tested benchmark programs.