Abstract Details
Abstracts
Author: M. Cianciosa
Requested Type: Poster
Submitted: 2023-03-22 13:33:16
Co-authors: D. Batchelor, W. Elwasif
Contact Info:
Oak Ridge National Laboratory
PO Box 2008, MS6305
Oak Ridge, TN 37831-6305
United States
Abstract Text:
Optimizing a fusion pilot plant requires rapid exploration of whole-facility simulation. Ray tracing is a fast method to determine deposition profiles for heating. However, legacy codes do not support GPUs or modern exascale devices. Making effective use of GPU resources is often a challenge for Physicists. Compounding this problem is the new generation of exascale devices which use AMD and Intel GPUs instead of Nvidia. The GPU needs to be abstracted from the physics to support the current and future machines. Several frameworks and technologies have promised the ability to simplify cross-platform programming. However, these typically require hand tuning to extract efficient performance and support for non-CUDA architecture lags.
We present a new domain-specific compiler that abstracts the physics from the compute backend. Compilers are computer codes that translate between representations. For instance, a Fortran compiler translates human-readable source code to the binary code necessary to run on a specific processor. By building a graph representation of the physics equation, optimization reduces the graph complexity, transformations apply analytic derivatives, and optimized kernel functions generated. Using just-in-time (JIT) compilation, we can demonstrate efficient utilization of GPU resources for arbitrary dispersion relations. This framework shows correct ray behavior for different dispersion functions in a fully 3D geometry for Nvidia GPUs and the Apple M1 GPU.
Work supported by U.S. DOE under Contract DE-AC05-00OR22725 with UT-Battelle, LLC
Comments: