mirror of
https://github.com/JuliaFEM/JuliaFEM.jl.git
synced 2026-09-22 18:52:16 +00:00
Jukka Aho
a49de72aaf
src/assemblers/caches/material_cache.jl | 421 ++++++++++++++++++++++++++++++++ 1 file changed, 421 insertions(+)
src/assemblers/caches/
Pre-allocated per-element scratch and global COO/CSC caches reused by every assembly strategy. Keeping them in one place makes it easy to see what is allocated up-front and what is consumed by the hot loops.
Files
geometry_cache.jl—GeometryCache. Per-element shape values, derivatives, and Jacobian data sized for one quadrature rule. Parametric on storage so the same struct works heap-owned (element-based assemblers) or as a view into batched SoA storage (DOF-based assembler).element_cache.jl—ElementCache. Assembly-time per-element scratch (local stiffness, force, dof index buffers).material_cache.jl—AssemblyMaterialWorkspaceplus thecreate_material_cachehelper. Holds the per-IP material state buffers (stress, tangent, history) used by the kernels.coo_cache.jl—COOCache, the global I/J/V triplet workspace used by the element-based assembler.
The hot-path invariant for all of these is zero allocation per element
on a steady-state assemble. The regression test for that is
test/assemblers/test_dof_based_zero_alloc.jl.