Branch Divergence-Aware Flexible Approximating Technique on GPUs.

Graphics Processing Units (GPUs) are widely adopted in various fields, attributed to their high throughput and energy efficiency. One of the primary challenges in GPU computing is branch divergence, which is a scenario that arises when the control flow among simultaneously processing threads diverges. This phenomenon leads to substantial performance degradation, as diverged threads are processed in a time-slice manner. To address this issue, we introduce a novel approximate computing approach in GPUs. Our proposed method approximates the outcomes of diverged threads by selectively terminating their execution. Additionally, this approach features a dynamic mechanism that adjusts both the likelihood of termination and the number of instructions executed prior to termination, according to user-defined criteria. This method enables users to finely tune the balance between execution speed and result fidelity, thereby providing enhanced flexibility in managing computational quality.