Skip to content

sesmesh 1.2.0

Choose a tag to compare

@ugovaretto ugovaretto released this 21 Nov 01:20
· 8 commits to main since this release

A GPU back end for AMD GPUs (HIP / ROCm) and an opt-in fast mode for previews. Default meshes are byte-identical to 1.1.2.

Added

  • GPU back end, HIP (#24): new crate sesmesh-gpu, Backend::{Cpu, Auto, Hip} in SurfaceParams, --backend / --gpu-device in the CLI (feature hip; default Cpu). The device classifies, gathers and samples the grid in f32 and integer arithmetic only; refinement stays on the CPU in f64, so exact mode writes the CPU back end's meshes byte for byte. On a Ryzen AI Max+ 395 (16 Zen 5 cores, Radeon 8060S), 1AON at 0.25 Å takes 2.37 s instead of 3.40 s and 3J3Q at 1 Å takes 23.6 s instead of 28.9 s.
  • Fast mode (RefineMode::Fast, --refine fast): the device also refines the crossing edges, in f32 with brick-local double-float coordinates; the CPU finishes in f64 the edges the device did not converge on. It gives the same triangles as exact mode, and the vertices are as close to the surface, but positions differ in the last bits. 1AON at 0.25 Å takes 1.41 s, 3J3Q at 1 Å 12.3 s (1.98–2.47× the 16 cores).
  • GPU benchmarks and validation: hip-exact / hip-fast cases in benches/surface.rs and the CPU/GPU gate tools/gpu/gate.sh (980 cases, 0 differences for this release).

Changed

  • Releases are built, tested and uploaded from local machines; no workflow runs on GitHub (tools/release/package.sh, docs/release.md).
  • The paper and website give the time on the 16-core Zen 5 machine wherever they give an M4 time.

Binaries: macOS arm64 and Linux x86-64 (glibc 2.34 or newer), built without hip. On Linux arm64 and Windows, or for the GPU back end: cargo install sesmesh-cli (add --features hip; it needs ROCm's hipcc at build time).

Full details: CHANGELOG.md.