GitHub

Architecture

This page explains how AppleAccelerate is structured internally, and why it integrates with the rest of the ecosystem through its own namespace rather than by hooking other packages' interfaces.

Two-layer design

AppleAccelerate is built in two layers:

  • Raw ABI layer — AppleAccelerate.LibAccelerate. This submodule (src/lib/LibAccelerate.jl) is auto-generated with Clang.jl directly from Apple's Accelerate C headers. It is a near 1:1 mirror of the C API — roughly a thousand @ccall wrappers plus the matching structs and enums — and is committed but never hand-edited (regenerate it with julia --project=gen gen/generate.jl). Because Accelerate's struct- and enum-heavy subframeworks (Quadrature, Sparse, BNNS) are painful and error-prone to bind by hand, generating this layer keeps every field offset and enum value in sync with the SDK automatically.

  • Idiomatic layer — the hand-written src/*.jl files. These provide the ergonomic Julia API (AppleAccelerate.exp, integrate, bnns_matmul, AASparseMatrix, …). They own the niceties — Julia functions, keyword options, tidy return values, broadcasting — and call into LibAccelerate for the actual ABI work, never naming a C field offset or enum integer themselves.

Most users only ever touch the idiomatic layer. When you need a symbol that has no idiomatic wrapper yet, the raw layer is available directly:

using AppleAccelerate
AppleAccelerate.LibAccelerate.some_unwrapped_symbol(args...)

No package extensions

AppleAccelerate ships no package extensions, and defines no methods on other packages' functions. Everything it provides lives under the AppleAccelerate. namespace, and the package exports nothing.

This is a deliberate choice. Earlier versions did hook AbstractFFTs.jl and NNlib.jl through extensions, which meant that merely loading AppleAccelerate — even for an unrelated feature like BLAS forwarding — silently changed what fft, plan_fft, or batched_mul! did elsewhere in the session. Because Accelerate's kernels only cover a subset of the input space (vDSP's FFT is power-of-2 and 1-D/2-D only), those methods won dispatch and then failed on inputs the general backend would have handled fine. See issue #139.

The consequence for you is simple and predictable:

using AppleAccelerate, FFTW

fft(x)                  # always FFTW — loading AppleAccelerate changes nothing
AppleAccelerate.fft(x)  # explicitly vDSP, for power-of-2 1-D/2-D input

You choose Accelerate per call site, by name. Nothing is intercepted, and benchmarking one against the other is a matter of changing the prefix.

Why LinearAlgebra and SparseArrays are hard dependencies

LinearAlgebra and SparseArrays are regular [deps], not weak dependencies:

  1. They are standard libraries that are effectively always present in the Julia sysimage, so making them weak would save nothing.
  2. BLAS/LAPACK forwarding needs LinearAlgebra at load time (__init__ installs Accelerate into libblastrampoline), so it cannot be deferred.
  3. Keeping SparseArrays a hard dependency lets AAFactorization subtype LinearAlgebra.Factorization and lets AASparseMatrix interoperate with SparseMatrixCSC directly.

BLAS/LAPACK forwarding is the one place AppleAccelerate does change global behavior — that is the documented purpose of loading it, and it goes through libblastrampoline, a mechanism designed for exactly this.