Numerics.NET.Native.Accelerate 10.8.0

Numerics.NET Native Accelerate

This package adds native Numerics.NET acceleration on macOS 12 or later for:

• Intel x64 (osx-x64);
• Apple silicon ARM64 (osx-arm64).

It contains one RID-specific Numerics.NET.Native.Accelerate.dylib for each architecture. Both single- and double-precision entry points are in each library. The libraries dynamically link Apple's system Accelerate framework; Apple framework binaries are not included in this package.

Each dylib implements the shared Numerics.NET native-provider ABI 1.0 bootstrap. Before constructing either managed provider, Numerics.NET validates the native provider identity, ABI version, required precision, and required capabilities. Older native binaries without xoGetNativeProviderInfo are rejected. The single- and double-precision providers share one validated dylib registration. Optional diagnostic build information, including the native bridge product version, compiler, and build architecture, is available through the bootstrap ABI.

The dylib advertises dense BLAS, LAPACK, selected real array functions, and 1D FFT as native capability groups. It does not advertise complex array functions, sparse operations, provider configuration, or direct 2D FFTs. Those capabilities describe native bridge exports, not behavior supplied by managed fallback or separable composition.

Dense real and complex BLAS and LAPACK operations are accelerated. Real vector add, subtract, multiply, divide, absolute value, and selected contiguous transcendental functions use vDSP and vForce when the stride and overlap layout is safe. Unsupported vector operations and layouts inherit the managed implementation.

One-dimensional real and complex FFTs are accelerated for directly supported vDSP lengths in both precisions. The interleaved backend is preferred whenever it supports the exact length. Qualifying contiguous complex buffers execute directly in caller storage; strided and overlapping layouts use reusable interleaved scratch. The split-complex backend handles split-only lengths. Exact lengths outside the union of both native sets use a provider-assisted Bluestein transform without changing the requested mathematical length. Its convolution work FFT uses a directly supported Accelerate complex length. Convolution and correlation may safely pad to the smallest union-supported length selected by the provider. Two-dimensional transforms use separable 1D transforms, so individual dimensions may select interleaved, split, or Bluestein execution. Rader orchestration and direct native 2D transforms are not included. Sparse operations remain managed.

Direct-length support queries continue to describe primitive vDSP execution, not the exact lengths that can be completed through Bluestein orchestration. The generic Bluestein core owns immutable chirp and kernel-spectrum data. Each active transform owns one child work FFT and one reusable complex workspace, so committed exact-length wrappers are reentrant without global plan sharing or internal parallelism.

Accelerate controls its own threading. Intel MKL configuration options, including MKL thread-count and conditional-reproducibility controls, do not apply.