Summary
In recent years, GPUs have been increasingly utilized not only for graphics processing but also for general-purpose computing tasks such as numerical simulations and machine learning. A typical GPU execution model requires instructions to be transferred from the CPU. However, as GPU computational power has improved, the cost of these transfers has become a significant overhead. To address this issue, a GPU-driven computing paradigm has emerged, aiming to minimize CPU intervention.
One existing GPU-driven approach, indirect commands, allows instructions to be preloaded into a GPU buffer, reducing the overhead of instruction transfers. However, this method has a fundamental limitation: it does not allow dynamic changes to the types of commands written to the buffer. As a result, it struggles to efficiently handle cases where instruction execution must adapt based on GPU computation results.
To overcome this limitation, the GPU work graph has been introduced. This feature represents a set of GPU-executable commands and their dependencies as a graph. By referring to this graph, the GPU can dynamically resolve dependencies and execute new instructions in sequence without CPU intervention.
In this study, we explore the application of GPU work graphs to fluid simulation and resolution control. The use of GPUs to accelerate fluid simulations has been widely studied. Additionally, adaptive resolution control—where the simulation’s discretization interval is adjusted for efficiency—has been investigated as a means to further optimize performance. However, implementing resolution control efficiently on a GPU is challenging, as it requires branching computations based on the simulation state. By leveraging the GPU work graph, we aim to develop a more efficient implementation for this process.
(a) and (b) show the visualization results of the simulation. At the beginning, a small time step is set, and as the flow stabilizes, a larger time step is applied.
(c) illustrates the graph structure constructed using the implemented GPU work graph.
Members
| Name | Affiliation | Web site |
|---|---|---|
 Keigo Miyashita | Keio University | |
| Shumpei Sugita | Keio University |
Publications
Presentation
Unrefereed
- Keigo Miyashita，Shumpei Sugita，Issei Fujishiro：”Adaptive time stepping for SPH using GPU work graph”，in Proceedings of SIG Technical Reports, ExaWizards Inc., Vol. 2025-CG-197, No. 1, pp.1-6, March 5, 2025, (in Japanese).
- Keigo Miyashita, Shumpei Sugita, Issei Fujishiro: “Adaptive time stepping for SPH using GPU work graph,” in Proceedings of the 87th National Convention of International Processing Society of Japan, Vol. 2, pp. 483―484 (5S-06), The University of Ritsumeikan, March 14, 2025 (in Japanese).
Grants
Grant-in-Aid for Scientific Research (A): 21H04916 (2021-)