DOI: https://doi.org/10.15368/theses.2021.157
Available at: https://digitalcommons.calpoly.edu/theses/2386
Date of Award
12-2021
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Christian Eckhardt
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
Real-time physics engines have seen recent performance improvements through techniques like hardware acceleration and artificial intelligence. However, state of the art physics simulation technology fails to account for the variation in simulation complexity over time. Sudden increases in contact frequency between simulated bodies can momentarily increase the processing time per frame. To solve this, we present a prediction-driven real-time dynamics method that uses a memory-efficient graph-based state buffer to minimize the cost of mispredictions. This buffer, which is generated by a separate thread running the physics pipeline, allows physics computation to temporarily run slower than real-time without affecting the frame rate of the host application. The main thread, whose role in dynamics computation gets limited to querying the simulation state and regenerating mispredicted state, sees a significant reduction in time spent per frame on dynamics computation when our multi-threaded prediction pipeline is enabled. Thus, our technique enables interactive multimedia applications to increase the computational budget for graphics at no cost perceptible to the end user. Furthermore, our method guarantees determinism and low input latency, making it suitable in competitive games and other real-time interactive applications. We also provide a C++ API to integrate custom game logic with the prediction engine to further minimize the frequency of mispredictions.