Performance: what we're building

Each point above the top subdivision level stores only a 16-bit height along the surface. The graphics card rebuilds the surface from those heights as it draws. The model is split into clusters of triangles that are culled, uploaded and shaded as units, so a stroke re-sends only the clusters it changed, and only the clusters it changed get new shading.

When a Standard, Clay or Inflate dab spans many base edges (and no mask, hidden faces, alpha or layers are involved), it runs the plain mesh's brush code on the base. Detail heights are stored relative to the base, so they follow. Moved triangles are tracked in a bitset filled in parallel, only the runs of moved vertices go to the graphics card, and the dabs still waiting when you lift the pen go through the plain mesh's batched path. The stroke stays one undo step.

The re-shading a stroke leaves behind is split into pieces. While the view moves, each frame spends only a few milliseconds on it (the same time cap a stroke uses); when the view stops, the rest is done at once.

Each brush's falloff has a known slope, so during the stroke the new normal is computed directly from the old normal and that slope, without walking the neighbouring triangles. Exact normals are recomputed in the background and after pen-up. Computers with fewer threads keep the previous behaviour.

Very big dabs take their plane from an even sample of the vertices under them, and the waiting dabs are fused into one pass, so normals, the picking structure and the copy to the graphics card are updated once a frame instead of once a dab. Since 0.11.2 each batch is capped by time, so no single frame waits too long.

It stays off by default because in our measurements it's faster on some machines and scenes and slower on others. It turns on by default only when it wins on the low-end tier first.

The project keeps a small preview image, shown while the scene loads; edits wait until the model is in. The first frame lights each cluster by its flat facets until the detailed shading arrives. The cache holds the mesh's connectivity, the picking structure and the detail levels' attachment data, packed and checked against a hash, and is carried over on save while the mesh's connectivity doesn't change. It makes project files bigger; we're measuring by how much before it ships.

Each batch of changes builds the release packages beside the continuous-integration run, and the main branch moves forward only when the whole suite and CI pass. A release will then promote the packages already built and tested for the exact tagged commit instead of building them again.

sculptforge --benchmark standard-15m (or Help > Benchmark) sculpts the same scene with the same strokes and orbit on any machine and writes the results to a file. We measure lag from the pen sample to the frame that shows its dab, how long a stroke takes to finish after pen-up, orbit frame rate and memory. Runs start only when the machine is idle, and we compare builds in alternating runs.

Tiers, in the order a claim must pass them: the simulated low-end tier (also with only two efficiency cores, for a pessimistic CPU), the Apple M1 Pro laptop itself, then a Windows desktop with an NVIDIA RTX 3060. A gain that shows only on the desktop doesn't count.