 Jukka Aho
|
600a2eeb0a
|
docs: Add GPU assembly proof-of-concept summary
Complete working proof-of-concept for GPU-accelerated finite element
assembly demonstrating entire solve staying on GPU.
Implementation features:
- Element-parallel GPU kernel for 2D linear elasticity
- Quad4 elements with 2×2 Gauss quadrature
- Matrix-free Jacobian-vector product (finite difference on GPU)
- Complete Newton-Krylov loop on GPU (no CPU escapes)
- Boundary condition enforcement
Validation results:
- GPU assembly matches CPU (relative error < 1e-15)
- Entire solve pipeline stays on GPU
- Only transfers: mesh (once), u0 (input), u_final (output)
Test case: 10×10 Quad4 mesh (100 elements, 242 DOFs), steel properties
(E=200 GPa, ν=0.3), fixed left edge, displacement on right edge.
Architecture: u0 → GPU → [Newton loop: residual + Jv + GMRES + update] → u_final
Reference: demos/gpu_assembly_poc.jl (212 lines documentation)
|
2025-11-12 00:20:49 +02:00 |
|