A deterministic Anisotropic Mesh Adaptation Method for Multicore CPUs and GPUs: The Same Mesh, Bit for Bit, On Every Thread Count and Device.

Data package of the paper "A deterministic anisotropic mesh adaptation method for multicore CPUs and GPUs: the same mesh, bit for bit, on every thread count and device" (Y. Mesri). The paper presents a metric-based adaptation method for triangular and tetrahedral meshes that produces the same mesh, to the last bit, on any number of threads and on CPU or GPU. This package lets readers check that claim directly on the adapted meshes, and contains the data behind the paper's tables and figures. Contents: Adapted meshes (NumPy .npz): four cases, each from a GPU run (NVIDIA A100) and a CPU run (32 threads): a part with curved boundaries at two sizes (70 thousand and 0.88 million tetrahedra), and the polar-2 and delta-wing cases of the Unstructured Grid Adaptation Working Group parallel benchmark (4.8 and 4.9 million tetrahedra). manifest.json: SHA-256, element counts and provenance of each mesh. verify.py: checks, with NumPy only, that each file has its recorded SHA-256, that the CPU and GPU meshes are identical, that no tetrahedron is inverted, that no face is shared by more than two tetrahedra, and reports the volume. Results: the data behind the tables and figures, including the comparisons with other remeshers, thread scaling, sensitivity to the numbering, and the applications to supersonic flows (cone, wing-body). Input: the input mesh of the part. To verify: unzip the archive and run python verify.py in its directory. The SOFIA source code is not included and is not publicly available at this time, The inputs of the parallel benchmark come from github.com/UGAWG/parallel-adapt-results; see README.md for the exact commits.