solidcore vs SMLib: upstream test models

A static viewer at https://tangency-labs.github.io/solidcore-viewer/. It shows the 262 B-rep part files of the SMLib test corpus rendered through two kernels: SMLib, NVIDIA's C++ B-rep kernel (reference build), and solidcore, a Rust port of it. The page has a thumbnail gallery and a side-by-side, overlay and diff 3D view per model.

Method

Both kernels read each model and run the same sampling:

• Each edge curve is evaluated at 48 parameters over its interval.
• Each face gets an 88×88 grid over its UV domain. A grid point is kept where the face's point classification (UV classification, face tolerance) puts it inside the face, and the kept points are evaluated on the surface.
• Each face's trim loops are sampled in UV along their edge uses' UV trim curves (24 samples per edge use) and evaluated on the surface.

A model matches when both kernels have the same face and edge counts, the same inside/outside mask at every grid point, and every grid, edge and trim-loop sample within 1e-5 (outputs are written with 6 decimals).

Meshes. Grid cells are clipped to each face's trim loops (even-odd, with Clipper via pyclipper) and triangulated with earcut. Each new boundary vertex is evaluated on the surface by the same kernel. Loops that close only across a periodic seam are unwrapped. Edge uses with no UV trim curve, such as degenerate pole edges, are bridged in UV. A face whose loop polygon still disagrees with the kernel's inside mask, or that has no UV trim curves at all, uses the inside-mask cells instead (154 of 2,631 faces per kernel, 138 of them with no UV trim curves in the file). The published meshes use every 2nd grid line (45×45 per face), quantized to 16 bits and gzipped. All agreement numbers come from the full 88×88 run.

Timeouts and times. Each model runs per kernel in its own process with a 60 s limit, and solidcore runs that hit the limit were retried with 300 s. Times are wall-clock on one shared workstation with other jobs running, measured for this sampling workload only.

Layout

• index.html, app.js, core.js: the gallery and the model view. three.js r128 is loaded from a CDN, pinned with SRI.
• manifest.json: per-model status, timings and agreement metrics.
• data/<model>.bin.gz: quantized meshes and edges for both kernels. tools/build.py describes the format.
• thumb/<model>.webp: gallery thumbnails, rendered from tools/thumbs.html.
• tools/: the scripts that made the data, with paths set for the machine they ran on:
  • smlib_wire.cpp: C++, links the SMLib build.
  • wire.rs: the solidcore example mode.
  • run.py: one process per file and kernel, with timeouts.
  • mesh.py: trim-loop clipping.
  • build.py: data files and metrics.
  • thumb_server.py and thumbs.html: thumbnails.
  • make_pub.py: builds this copy.

Attribution and licence

Models: NVIDIA SMLib test files (TestFiles/pt_TestFiles), Apache-2.0. SMLib is NVIDIA's; solidcore is an independent Rust port. This repository holds only meshes derived from those files, the viewer and its scripts. It is licensed under Apache-2.0 (see LICENSE).

The viewer loads three.js (MIT) from a CDN. The data scripts use numpy (BSD), pyclipper (MIT; Clipper, Boost Software License) and mapbox_earcut (ISC).