mirror of
https://github.com/JuliaFEM/JuliaFEM.jl.git
synced 2026-08-06 04:21:33 +00:00
Jukka Aho
267 lines
9.9 KiB
Julia
267 lines
9.9 KiB
Julia
# This file is a part of JuliaFEM.
|
||
# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
|
||
|
||
"""
|
||
Float32 (single-precision) `apply_K!` through the precision-parametric
|
||
`DOFBasedCOOCacheKA` path added in F+++.
|
||
|
||
Locks in the contract that closes the "Apple Metal can't store Float64"
|
||
gap from the previous round:
|
||
|
||
1. **`to_float32(cpu_cache)` produces a `DOFBasedCOOCacheKA` whose
|
||
storage is fully Float32**: `detJ_w_batch`, `N_batch`, `∇N_batch`,
|
||
`X_batch`, and `qp_buffers` (a `SymmetricTensor{4,3,Float32,36}`
|
||
for `ContinuumKernel`, `SymmetricTensor{2,3,Float32,6}` for
|
||
`HeatKernel`) all carry `Float32` element type.
|
||
|
||
2. **The same `apply_K_kernel!` runs in Float32**: extracts `F`
|
||
dynamically from `eltype(y)` and accumulates in `zero(F)`. The
|
||
KernelAbstractions launcher routes the F32 cache + F32 vectors to
|
||
the same generic kernel without any duplicated code path.
|
||
|
||
3. **F32 results agree with the F64 reference to single-precision
|
||
accuracy** (≤ `1e-5` relative on well-conditioned problems) for
|
||
both `ContinuumKernel` and `HeatKernel`.
|
||
|
||
4. **Round-trip back-compat**: the default
|
||
`DOFBasedCOOCacheKA(cpu_cache)` (Float64) still produces
|
||
bit-identical results to the direct CPU `apply_K!`.
|
||
|
||
5. **No interference with the CPU path**: existing CPU `apply_K!`
|
||
and `assemble!` remain Float64 end-to-end.
|
||
|
||
Once a Float32 cache is on the device via `Adapt.adapt(MetalBackend(),
|
||
cache_f32)`, the same `apply_K!` call runs natively on Apple GPUs. The
|
||
device-specific test for that lives in `test/backend/metal/`; this file
|
||
locks in the *precision-parametric* contract on the CPU backend so the
|
||
Metal port has a green oracle to compare against.
|
||
"""
|
||
|
||
using Test
|
||
using JuliaFEM
|
||
using JuliaFEM: ContinuumFormulation, FullThreeD, Vertex, Temperature
|
||
using JuliaFEM: @DOFSet, DOF
|
||
using JuliaFEM: LinearElastic, Displacement, ContinuumKernel
|
||
using JuliaFEM: HeatConductivity, HeatKernel
|
||
using JuliaFEM: DOFBasedCOOAssembler, DOFBasedCOOCache, DOFBasedCOOCacheKA
|
||
using JuliaFEM: extract_system, apply_K!, sync_from_cpu!, to_float32, create_elements!
|
||
using LinearAlgebra
|
||
using SparseArrays
|
||
using Tensors
|
||
using Random
|
||
using KernelAbstractions
|
||
|
||
# ----------------------------------------------------------------------------
|
||
# Mesh helpers
|
||
# ----------------------------------------------------------------------------
|
||
|
||
function _hex8_box(nx::Int, ny::Int, nz::Int)
|
||
nodes = Vec{3,Float64}[]
|
||
nidx(i, j, k) = (i - 1) + (j - 1) * (nx + 1) + (k - 1) * (nx + 1) * (ny + 1) + 1
|
||
for k in 1:(nz + 1), j in 1:(ny + 1), i in 1:(nx + 1)
|
||
push!(nodes, Vec{3}((Float64(i - 1) / nx,
|
||
Float64(j - 1) / ny,
|
||
Float64(k - 1) / nz)))
|
||
end
|
||
conns = NTuple{8,UInt32}[]
|
||
for k in 1:nz, j in 1:ny, i in 1:nx
|
||
n1 = nidx(i, j, k)
|
||
n2 = nidx(i + 1, j, k)
|
||
n3 = nidx(i + 1, j + 1, k)
|
||
n4 = nidx(i, j + 1, k)
|
||
n5 = nidx(i, j, k + 1)
|
||
n6 = nidx(i + 1, j, k + 1)
|
||
n7 = nidx(i + 1, j + 1, k + 1)
|
||
n8 = nidx(i, j + 1, k + 1)
|
||
push!(conns, (UInt32(n1), UInt32(n2), UInt32(n3), UInt32(n4),
|
||
UInt32(n5), UInt32(n6), UInt32(n7), UInt32(n8)))
|
||
end
|
||
return Mesh{8,Hexahedron{8}}(nodes, conns)
|
||
end
|
||
|
||
function _setup_elasticity(mesh)
|
||
material = LinearElastic(E = 210e9, ν = 0.3)
|
||
kernel = ContinuumKernel(ContinuumFormulation{FullThreeD}(),
|
||
material, Displacement{3}())
|
||
S = @DOFSet{u::DOF{Displacement{3}, Vertex}}
|
||
elements, dof_mgr = create_elements!(mesh, Element{Hexahedron{8}, Lagrange{1}, S})
|
||
asm = DOFBasedCOOAssembler()
|
||
cache = DOFBasedCOOCache(elements, dof_mgr, mesh, kernel)
|
||
return cache, asm, kernel, mesh
|
||
end
|
||
|
||
function _setup_heat(mesh; k_value::Float64 = 50.2)
|
||
material = HeatConductivity(k = k_value)
|
||
kernel = HeatKernel(ContinuumFormulation{FullThreeD}(), material)
|
||
S = @DOFSet{T::DOF{Temperature, Vertex}}
|
||
elements, dof_mgr = create_elements!(mesh, Element{Hexahedron{8}, Lagrange{1}, S})
|
||
asm = DOFBasedCOOAssembler()
|
||
cache = DOFBasedCOOCache(elements, dof_mgr, mesh, kernel)
|
||
return cache, asm, kernel, mesh
|
||
end
|
||
|
||
# ----------------------------------------------------------------------------
|
||
# 1. Storage typing of the F32 cache
|
||
# ----------------------------------------------------------------------------
|
||
|
||
@testset "to_float32: storage element types are Float32" begin
|
||
println("\n" * "=" ^ 70)
|
||
println("F+++ Float32 KA cache storage typing")
|
||
println("=" ^ 70)
|
||
|
||
@testset "Elasticity: ContinuumKernel" begin
|
||
mesh = _hex8_box(2, 1, 1)
|
||
cache, asm, kernel, m = _setup_elasticity(mesh)
|
||
assemble!(cache, asm, kernel, m)
|
||
|
||
cache_f32 = to_float32(cache)
|
||
@test eltype(cache_f32.X_batch) === Vec{3,Float32}
|
||
@test eltype(cache_f32.N_batch) === Float32
|
||
@test eltype(cache_f32.∇N_batch) === Vec{3,Float32}
|
||
@test eltype(cache_f32.detJ_w_batch) === Float32
|
||
@test eltype(cache_f32.qp_buffers) === SymmetricTensor{4,3,Float32,36}
|
||
println(" Elasticity F32 cache types ✓ (qp_buffers = " *
|
||
"$(eltype(cache_f32.qp_buffers)))")
|
||
end
|
||
|
||
@testset "Heat: HeatKernel" begin
|
||
mesh = _hex8_box(2, 1, 1)
|
||
cache, asm, kernel, m = _setup_heat(mesh)
|
||
assemble!(cache, asm, kernel, m)
|
||
|
||
cache_f32 = to_float32(cache)
|
||
@test eltype(cache_f32.X_batch) === Vec{3,Float32}
|
||
@test eltype(cache_f32.detJ_w_batch) === Float32
|
||
@test eltype(cache_f32.qp_buffers) === SymmetricTensor{2,3,Float32,6}
|
||
println(" Heat F32 cache types ✓ (qp_buffers = " *
|
||
"$(eltype(cache_f32.qp_buffers)))")
|
||
end
|
||
end
|
||
|
||
# ----------------------------------------------------------------------------
|
||
# 2. F32 apply_K! agrees with F64 reference to single-precision accuracy
|
||
# ----------------------------------------------------------------------------
|
||
|
||
@testset "F32 apply_K!: agrees with F64 to single precision" begin
|
||
println("\n" * "=" ^ 70)
|
||
println("F+++ Float32 apply_K! correctness (CPU backend)")
|
||
println("=" ^ 70)
|
||
|
||
Random.seed!(20260508)
|
||
|
||
@testset "Elasticity $(nx)×$(ny)×$(nz)" for (nx, ny, nz) in
|
||
[(2, 1, 1), (3, 2, 2), (4, 3, 3)]
|
||
|
||
mesh = _hex8_box(nx, ny, nz)
|
||
cache, asm, kernel, m = _setup_elasticity(mesh)
|
||
n = cache.ndofs
|
||
|
||
assemble!(cache, asm, kernel, m)
|
||
K, _ = extract_system(cache)
|
||
|
||
# F64 KA reference (aliased storage with the CPU cache)
|
||
cache_f64 = DOFBasedCOOCacheKA(cache)
|
||
sync_from_cpu!(cache_f64, cache)
|
||
|
||
# F32 KA mirror
|
||
cache_f32 = to_float32(cache)
|
||
sync_from_cpu!(cache_f32, cache)
|
||
|
||
max_rel = 0.0
|
||
for trial in 1:5
|
||
x64 = randn(n)
|
||
y_ref = K * x64
|
||
|
||
y64 = zeros(n)
|
||
apply_K!(y64, cache_f64, kernel, x64)
|
||
@test norm(y64 - y_ref) / norm(y_ref) < 1e-12
|
||
|
||
x32 = Float32.(x64)
|
||
y32 = zeros(Float32, n)
|
||
apply_K!(y32, cache_f32, kernel, x32)
|
||
rel = norm(Float64.(y32) - y_ref) / norm(y_ref)
|
||
@test rel < 1e-5
|
||
max_rel = max(max_rel, rel)
|
||
end
|
||
|
||
nelems = length(m.connectivity)
|
||
println(" Elast $(nx)×$(ny)×$(nz) $(lpad(nelems,3)) elem " *
|
||
"ndof=$(lpad(n,4)) max(F32 vs F64)=$(round(max_rel; sigdigits = 3))")
|
||
end
|
||
|
||
@testset "Heat $(nx)×$(ny)×$(nz)" for (nx, ny, nz) in
|
||
[(2, 1, 1), (3, 2, 2), (4, 3, 3)]
|
||
|
||
mesh = _hex8_box(nx, ny, nz)
|
||
cache, asm, kernel, m = _setup_heat(mesh)
|
||
n = cache.ndofs
|
||
|
||
assemble!(cache, asm, kernel, m)
|
||
K, _ = extract_system(cache)
|
||
|
||
cache_f32 = to_float32(cache)
|
||
sync_from_cpu!(cache_f32, cache)
|
||
|
||
max_rel = 0.0
|
||
for trial in 1:5
|
||
x64 = randn(n)
|
||
y_ref = K * x64
|
||
|
||
x32 = Float32.(x64)
|
||
y32 = zeros(Float32, n)
|
||
apply_K!(y32, cache_f32, kernel, x32)
|
||
rel = norm(Float64.(y32) - y_ref) / max(norm(y_ref), 1.0)
|
||
@test rel < 1e-5
|
||
max_rel = max(max_rel, rel)
|
||
end
|
||
|
||
nelems = length(m.connectivity)
|
||
println(" Heat $(nx)×$(ny)×$(nz) $(lpad(nelems,3)) elem " *
|
||
"ndof=$(lpad(n,4)) max(F32 vs K*x F64)=$(round(max_rel; sigdigits = 3))")
|
||
end
|
||
end
|
||
|
||
# ----------------------------------------------------------------------------
|
||
# 3. F32 cache enforces precision agreement (helpful error)
|
||
# ----------------------------------------------------------------------------
|
||
|
||
@testset "F32 cache rejects mismatched-precision input vectors" begin
|
||
mesh = _hex8_box(2, 1, 1)
|
||
cache, asm, kernel, m = _setup_elasticity(mesh)
|
||
assemble!(cache, asm, kernel, m)
|
||
|
||
cache_f32 = to_float32(cache)
|
||
sync_from_cpu!(cache_f32, cache)
|
||
n = cache.ndofs
|
||
|
||
# Passing Float64 vectors at an F32 cache must be flagged early
|
||
# (otherwise we'd silently accumulate junk because the kernel reads
|
||
# F = eltype(y) and assumes the cache matches).
|
||
y64 = zeros(n); x64 = randn(n)
|
||
@test_throws AssertionError apply_K!(y64, cache_f32, kernel, x64)
|
||
end
|
||
|
||
# ----------------------------------------------------------------------------
|
||
# 4. Default Float64 KA path is unchanged (round-trip back-compat)
|
||
# ----------------------------------------------------------------------------
|
||
|
||
@testset "Default DOFBasedCOOCacheKA stays bit-identical to CPU apply_K!" begin
|
||
Random.seed!(20260508)
|
||
|
||
mesh = _hex8_box(3, 2, 2)
|
||
cache, asm, kernel, m = _setup_elasticity(mesh)
|
||
n = cache.ndofs
|
||
assemble!(cache, asm, kernel, m)
|
||
K, _ = extract_system(cache)
|
||
|
||
cache_f64 = DOFBasedCOOCacheKA(cache)
|
||
sync_from_cpu!(cache_f64, cache)
|
||
|
||
for _ in 1:3
|
||
x = randn(n)
|
||
y_ref = K * x
|
||
y = zeros(n); apply_K!(y, cache_f64, kernel, x)
|
||
@test norm(y - y_ref) / norm(y_ref) < 1e-12
|
||
end
|
||
end
|