 Jukka Aho
|
84708cce2e
|
bench: Add basis function access pattern benchmark for Tet10
- Tests 5 strategies for accessing basis functions and derivatives
- Strategy 1: Tuple return + getindex (natural Julia)
- Strategy 2: @generated function (compile-time specialization)
- Strategy 3: Val dispatch (type-stable index)
- Strategy 4: BasisFunctions struct with getindex (most Julian)
- Strategy 5: Separate functions per basis (extreme specialization)
- Use case: Tet10 quadratic tetrahedron (10 nodes, workhorse for 3D FEM)
- Benchmarks both "access all" (element assembly) and "access single" (nodal assembly)
- Tests derivatives as Vec{3} tuples for gradient computation
- Focus: zero allocation, type stability, inlineability
- 412 lines evaluating access pattern performance for GPU-friendly design
|
2025-11-12 00:09:42 +02:00 |
|