aa wrapper and KLU migration to PNM for PFs

.claude/claude.md

@@ -75,7 +75,6 @@ Efficient computation of power network matrices for large-scale systems. Support
 ### `ext/`
 
   - **MKLPardisoExt.jl**: MKL-Pardiso sparse solver (Windows/Linux)
-  - **AppleAccelerateExt.jl**: native macOS BLAS acceleration
 
 ### `test/`
 
@@ -112,7 +111,7 @@ Documentation source
 ### Optional Extensions
 
   - **MKL + Pardiso**: high-performance sparse factorization
-  - **AppleAccelerate**: native macOS dense linear algebra
+  - **AppleAccelerate**: native macOS sparse direct solver (libSparse `SparseFactorizationLU`, built-in via `src/AccelerateWrapper/`, macOS 15.5+)
 
 ## Core Abstractions
 
@@ -160,7 +159,7 @@ Documentation source
   - **Indexing**: `matrix[bus_num, branch_tuple]` auto-maps to internal indices
   - **Subnetworks**: `subnetwork_axes` Dict maps reference buses to island components
   - **Caching**: VirtualPTDF/LODF use LRU cache (default 100 MiB) for row storage
-  - **Solvers**: KLU (default), Dense, MKLPardiso, AppleAccelerate via extensions
+  - **Solvers**: KLU, Dense, MKLPardiso (ext), and built-in AppleAccelerateLU (libSparse `SparseFactorizationLU` + Inf-norm equilibration scaling, macOS 15.5+) via `src/AccelerateWrapper/`; KLU is the default on non-Apple / macOS < 15.5; AppleAccelerateLU is the default on macOS ≥ 15.5.
 
 ## Test Patterns
 

Project.toml

@@ -15,15 +15,12 @@ SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 SuiteSparse_jll = "bea87d4a-7f5b-5778-9afe-8cc45184846c"
 
 [weakdeps]
-AppleAccelerate = "13e28ba4-7ad8-5781-acae-3021b1ed3924"
 Pardiso = "46dd5b70-b6fb-5a00-ae2d-e8fea33afaf2"
 
 [extensions]
-AppleAccelerateExt = "AppleAccelerate"
 MKLPardisoExt = "Pardiso"
 
 [compat]
-AppleAccelerate = "^0.4"
 DataStructures = "^0.19"
 DocStringExtensions = "~0.8, ~0.9"
 HDF5 = "0.17"

docs/src/reference/internals.md

@@ -14,3 +14,14 @@ symbols are exported from `PowerNetworkMatrices`.
 ```@autodocs
 Modules = [PowerNetworkMatrices.KLUWrapper]
 ```
+
+## `AccelerateWrapper`
+
+`PowerNetworkMatrices.AccelerateWrapper` is a thin, allocation-aware wrapper over Apple's
+`libSparse.dylib` (provided by the system Accelerate framework) used internally for sparse
+linear solves on macOS. Non-Apple builds load stub fallbacks that throw on use. None of
+these symbols are exported from `PowerNetworkMatrices`.
+
+```@autodocs
+Modules = [PowerNetworkMatrices.AccelerateWrapper]
+```

ext/MKLPardisoExt.jl

@@ -171,28 +171,15 @@ function PNM._calculate_LODF_matrix_MKLPardiso(
     a::SparseArrays.SparseMatrixCSC{Int8, Int},
     ptdf::Matrix{Float64},
 )
+    # The demand matrix `diag(1 - PTDF·A[i,i])` is diagonal, so the
+    # "solve" is a row-wise element-wise division. The Pardiso path that
+    # used to factor it and back-solve via `_pardiso_sequential_LODF!` /
+    # `_pardiso_single_LODF!` is no longer needed for this stage.
     linecount = size(ptdf, 2)
     ptdf_denominator_t = a * ptdf
-    m_I = Int[]
-    m_V = Float64[]
-    for iline in 1:linecount
-        if (1.0 - ptdf_denominator_t[iline, iline]) < PNM.LODF_ENTRY_TOLERANCE
-            push!(m_I, iline)
-            push!(m_V, 1.0)
-        else
-            push!(m_I, iline)
-            push!(m_V, 1 - ptdf_denominator_t[iline, iline])
-        end
-    end
-    lodf_t = zeros(linecount, linecount)
-    A = SparseArrays.sparse(m_I, m_I, m_V)
-    if linecount > PNM.DEFAULT_LODF_CHUNK_SIZE
-        PNM._pardiso_sequential_LODF!(lodf_t, A, ptdf_denominator_t)
-    else
-        PNM._pardiso_single_LODF!(lodf_t, A, ptdf_denominator_t)
-    end
-    lodf_t[LinearAlgebra.diagind(lodf_t)] .= -1.0
-    return lodf_t
+    m_V = PNM._build_lodf_demand(ptdf_denominator_t, linecount)
+    PNM._apply_lodf_demand!(ptdf_denominator_t, m_V)
+    return ptdf_denominator_t
 end
 
 end # module