Flexible Approximate Computing for Mitigating Branch Divergence in GPUs

Graphics processing units (GPUs) have become prevalent across various domains due to their high throughput and energy efficiency. A major challenge in GPU computing is branch divergence, which occurs when the control flow paths of concurrently executing threads diverge. Branch divergence significantly degrades performance, as diverged threads are executed in a serialized, time-sliced manner. To mitigate this issue, we propose a novel approximate computing technique for GPUs. Our method approximates the results of diverged threads by selectively terminating their execution. Furthermore, it incorporates a dynamic mechanism that adjusts both the probability of termination and the number of instructions executed before termination, based on user-defined parameters. This technique allows users to balance execution speed and result accuracy, offering greater flexibility in controlling computational quality. Additionally, by changing the approximation aggressiveness at runtime, we can accommodate applications such as virtual reality/augmented reality, where the demand for accuracy changes in real-time.