Repository navigation
v0.1.8
TreeAMR v0.1.8
Faster copy kernels, chiefly on a device (#18). On a device, scatter!, gather! and the ghost fill launch over one axis with precomputed 32-bit inverses instead of 64-bit divisions, and pure copies skip their weights. H200, 20 variables, G = 3, 512 × 32³: scatter! 0.335 → 0.140 ns/point, gather! 0.441 → 0.136, fill_ghosts! 0.399 → 0.296. On the CPU, one-thread scatter!/gather! are about 3× faster. Results are bit-identical to 0.1.7; no API change.
Merged pull requests: