Repository navigation
sesmesh 1.3.0
A GPU back end for NVIDIA GPUs (CUDA), next to the AMD one from 1.2.0. Default meshes are unchanged from 1.2.0 (identical apart from the version in the OBJ header).
Added
- GPU back end, CUDA (#25):
Backend::Cuda(CudaOptions)/Backend::cuda(i)inSurfaceParams,--backend cudain the CLI (featurecuda; defaultCpu). It runs the same kernels as the HIP back end. They are compiled withnvccinto a fat binary for compute capability 7.5 (Turing) to 12.0, plus PTX. Only the NVIDIA driver is needed at run time; it is loaded when a device is opened, so a binary built with the feature still runs on the CPU where the driver is missing. Exact mode writes the CPU back end's files byte for byte, and--refine fastworks as with HIP. Batches live in device memory (discrete GPUs).Backend::Autotries HIP, then CUDA, then the CPU. - Tested on a laptop with a GeForce RTX 2070 Max-Q (sm_75, 8 GB) and a 6-core i7-9750H. The CPU/GPU gate found no difference in 774 SES and 212 SAS/vdW cases (3J3Q at 0.25 Å, 800 M triangles, included). 1AON at 0.5 Å takes 3.04 s exact and 1.70 s fast, against 5.48 s on the laptop's CPU, which bounds both modes (
docs/06-gpu-cuda.md). tools/gpu/cuda-build.Containerfilebuilds the CUDA back end on a machine without CUDA.tools/gpu/gate.shtakesBACKEND=cuda, and can make its CPU references on another machine (MODE=cpu|gpu).
Changed
sesmesh-gpu: the batch descriptions, encodings and self-test moved to a shared modulebatch, with traits both back ends implement. The kernel source iskernels/sesmesh.cuh, shared by HIP and CUDA.--gpu-deviceapplies to--backend hipandcuda.- The GPU gate uses no Python. It streams the OBJ files (tens of GB at 0.25 Å) through
sha256sumandgrep, compares statistics withjq, and prints the totals withsummary.sh. This release passed it on both back ends.
Binaries: macOS arm64 and Linux x86-64 (glibc 2.34 or newer), built without GPU features. On Linux arm64 and Windows, or for a GPU back end, use cargo install sesmesh-cli; add --features hip (needs ROCm's hipcc) or --features cuda (needs CUDA's nvcc) at build time.
Full details: CHANGELOG.md.