Files
JuliaFEM.jl/docs/book/adr-004-integration-points-api.md
T

194 lines
5.1 KiB
Markdown
Raw Normal View History

---
title: "ADR-004: Zero-Allocation Integration Points API"
date: 2025-11-11
author: "Jukka Aho + AI Assistant"
status: "Accepted"
tags: ["adr", "integration", "performance", "api-design"]
---
## Context
Integration points (Gauss quadrature) are accessed millions of times during FEM assembly. The original implementation used runtime dispatch with mutable structs containing Dict fields, causing:
1. **Type instability** - Dict fields unknown at compile time
2. **Allocations** - New struct allocations every query
3. **~50× performance penalty** vs optimal approach
**Problem Statement:** How should integration points be accessed in assembly loops?
## Decision
**Adopt compile-time integration point API matching basis function design (eval_basis!).**
### New API
```julia
@inline function get_gauss_points!(::Type{T}, ::Type{S}) where {T<:AbstractTopology, S<:Gauss}
-> NTuple{N, Tuple{Float64, Vec{D}}}
```
**Key Properties:**
- **Compile-time generation**: Like `eval_basis!`, returns literal tuples
- **Vec{D} coordinates**: Tensors.jl Vec for efficient FEM math
- **Zero allocation**: Fully inlined, no runtime overhead
- **Type-stable**: All types known at compile time
### Usage Pattern
```julia
# Assembly loop - zero allocations:
for (weight, ξ) in get_gauss_points!(Triangle, Gauss{2})
N = eval_basis!(Lagrange{Triangle,1}, Float64, ξ)
dN = eval_dbasis!(Lagrange{Triangle,1}, ξ)
detJ = compute_jacobian(ξ)
K += weight * detJ * (dN' * D * dN)
end
```
**Why Vec{D}?**
- Natural for FEM: `dN/dξ ⋅ v`, tensor products, etc.
- GPU-friendly (immutable, stack-allocated)
- Matches golden standard (nodal assembly demos)
## Alternatives Considered
### Option A: Plain Tuples
```julia
get_gauss_points!(Triangle, Gauss{1})
# → ((0.5, (1/3, 1/3)),)
```
**Rejected:** Tuple coordinates less convenient for FEM math.
### Option B: Store in Element
```julia
struct Element{N,NIP,...}
ips::NTuple{NIP, IntegrationPoint{D}}
end
```
**Rejected:** Slight overhead, less flexible (fixed at construction).
### Option C: Global Constants
```julia
const TRI3_GAUSS1_IPS = ((0.5, Vec{2}((1/3, 1/3))),)
```
**Rejected:** Not composable (can't parameterize on topology/order).
### Option D: Runtime Dispatch (OLD)
```julia
get_integration_points(element::Seg2)
# → Vector{IP} # Mutable struct with Dict
```
**Rejected:** 50× slower, allocates, type-unstable.
## Performance Results
Benchmark: 1000 elements, 3 integration points each
| Approach | Time | Allocations | Speedup |
|----------|------|-------------|---------|
| OLD (runtime + Dict) | 53 μs | 515 KiB | 1× |
| NEW (compile-time + Vec) | **1.1 μs** | **0 bytes** | **48×** |
**Realistic FEM assembly:**
- OLD: 53 μs + 515 KiB allocations
- NEW: 1.1 μs + 0 allocations
## Implementation
### File Structure
```text
src/integration/
├── integration.jl # Abstract types (IntegrationPoint, AbstractIntegration)
├── gauss.jl # High-level Gauss{N} type
└── gauss_points.jl # NEW: Compile-time get_gauss_points!()
```
### Supported Topologies
**1D:**
- Segment: Gauss{1}, Gauss{2}, Gauss{3}
**2D:**
- Triangle: Gauss{1} (1 pt), Gauss{2} (3 pt), Gauss{3} (4 pt)
- Quadrilateral: Gauss{1} (1 pt), Gauss{2} (4 pt), Gauss{3} (9 pt)
**3D:**
- Tetrahedron: Gauss{1} (1 pt), Gauss{2} (4 pt), Gauss{3} (5 pt)
- Hexahedron: Gauss{1} (1 pt), Gauss{2} (8 pt), Gauss{3} (27 pt)
- Wedge: Gauss{1}, Gauss{2}
- Pyramid: Gauss{1}, Gauss{2}
## Consequences
### Positive
1. **50× faster** than old approach
2. **Zero allocations** in assembly loops
3. **Type-stable** - compiler knows everything
4. **Consistent with basis API** - same pattern as `get_basis_functions` (NOTE: `eval_basis!` is deprecated)
5. **GPU-ready** - Vec{D} immutable, can transfer to GPU
6. **Matches golden standard** - nodal assembly architecture
### Negative
1. **Breaking change** - old `get_integration_points(element)` deprecated
2. **Migration needed** - update assembly code to new API
3. **More verbose** - must specify topology and scheme explicitly
### Neutral
1. **Compile-time only** - dynamic integration orders need workaround
2. **Fixed quadrature rules** - pre-defined Gauss{1}, Gauss{2}, etc.
## Migration Strategy
### Phase 1: Add New API (✅ Complete)
- Implement `get_gauss_points!()` for all topologies
- Comprehensive tests
- Benchmark validation
### Phase 2: Update Assembly Code (In Progress)
- Fix `get_integration_points(element)` in elements.jl
- Update problem assembly functions
- Ensure tests pass
### Phase 3: Deprecate Old API
- Add deprecation warnings to old functions
- Document migration path
- Remove after one release cycle
## Related
- **ADR-002:** Basis function API (same pattern)
- **Golden Standard:** docs/book/multigpu_nodal_assembly.md
- **Nodal Assembly Demos:** demos/nodal_assembly_{cpu,gpu}.jl
## References
1. Benchmark: `benchmarks/integration_points_benchmark.jl`
2. Tests: `test/test_integration_points_api.jl`
3. Implementation: `src/integration/gauss_points.jl`
## Status History
- 2025-11-11: Accepted, implemented, tested
- Performance validated: 48× speedup, zero allocations