Kerf.Kernel.ModelingData 0.4.0

Kerf.Kernel

A managed .NET CAD kernel derived from Open CASCADE Technology (OCCT) 8.0.1, ported to C# and redesigned for .NET. See docs/api for the public API, docs/design.md for the design and docs/port-status.md for what has been ported.

• Pure managed code: no native libraries, no P/Invoke.
• .NET names: the public API is in namespaces by area (Kerf.Geometry, Kerf.Topology, Kerf.Booleans…), without OCCT's package prefixes, and members follow .NET's naming. The algorithms are internal, in Kerf.Engine.<Area>. docs/api/kerf.map says what each OCCT name became, so OCCT's documentation can still be found; each derived file names its OCCT source in a // Port of: header.
• Immutable: every value you get back is immutable.
• Results: an operation that can fail for an expected reason returns a Result, with diagnostics.

using Kerf.Geometry;

Transform3D rotation = Transform3D.Rotation(Axis3D.ZAxis, Math.PI / 2);
Point3D point = new Point3D(1, 0, 0).Transformed(rotation);   // (0, 1, 0); the original is unchanged

Assemblies

Runtime settings for a host

These belong to the application, not the library. For work on many threads, such as a spreadsheet recalculating workbooks at once:

<PropertyGroup>
  <ServerGarbageCollection>true</ServerGarbageCollection>
  <ConcurrentGarbageCollection>true</ConcurrentGarbageCollection>
  <TieredPGO>false</TieredPGO>
</PropertyGroup>
<ItemGroup>
  <RuntimeHostConfigurationOption Include="System.Runtime.TieredCompilation.CallCountingDelayMs" Value="0" />
</ItemGroup>

Measured on 24 logical processors, a round of 24 copies of Bento's examples takes 620 ms against 2,700 ms with the runtime's defaults, and a single recalculation 180 ms against 360, at a peak of about 400 MB against 200. Leave the server collector's dynamic adaptation (DATAS) on: turning it off gains little and takes about 700 MB. Without PGO, code that has settled for minutes is about 6% slower, so a long batch job may prefer to leave it on. See docs/performance.md, "Unit 11D".

No tiering setting beyond these helps opening (unit 11C tried quick JIT, quick JIT for loops, on-stack replacement and the background worker's timeout). An application that ships its own runtime can go further by publishing self-contained with PublishReadyToRunComposite: opening 419 ms against 509 on Bento's examples, at 169 MB for the whole application, measured on assemblies built from source rather than from the packages. If it publishes ReadyToRun at all, setting PublishReadyToRunCrossgen2ExtraArgs to --opt-cross-module:*;--Ot, as Kerf's packages do, should help its own generics over Kerf's types (not measured).

Building

Requirements: the .NET 10 SDK and git.

./tools/get-occt.ps1          # OCCT V8_0_1 source into .deps/occt-src (for the corpus test and port status)
./tools/get-occt-testdata.ps1 # OCCT's test data (99 MB) into .deps/occt-testdata, for the Draw corpus
dotnet test

tests/Kerf.Kernel.Oracle.Tests compares the port with native OCCT's results, recorded when native OCCT was still built (unit 10A) and frozen since it went (unit 10B): see docs/oracle-testing.md.

Contributing

docs/design-rules.md gives the rules for new and redesigned code (phase 10); docs/translation-rules.md says how C++ became C# until phase 10, and docs/porting-playbook.md the steps for porting a package.

Licence

GNU Lesser General Public License 2.1 with the Open CASCADE exception, as OCCT: see LICENSE, OCCT_LGPL_EXCEPTION.txt and NOTICE. Kerf.Kernel is a modified version of OCCT 8.0.1 and isn't endorsed by Open Cascade SAS. Delabella's port (BRepMesh/delabella.cs) keeps its MIT licence: see THIRD-PARTY-NOTICES.md.