avec
May 13, 2026 · View on GitHub
avec is a little library for using SIMD instructions on both x86 and ARM.
It features containers for aligned memory, with views, allocators and interleaving/deinterleaving functionality.
The API is designed to work seamlessly with Agner Fog's vectorclass, which is included as a submodule.
Since vectorclass only supports x86, avec reimplements a subset of its functionality for ARM using NEON. See the section ARM support for details.
Containers and views
In vectorclass each SIMD type has its own class: Vec4f for __m128, Vec8f for __m256, Vec4d for __m256d and so on.
In avec, the template classes VecBuffer<Vec> and VecView<Vec> are used to manage blocks of aligned memory and convert it to and from the SIMD classes of vectorclass.
Interleaving
The template class InterleavedBuffer<Scalar> (where Scalar can be either float or double) is used to interleave a buffer of any number of audio channels into a set of VecBuffer<Vec8f>, VecBuffer<Vec4f> and VecBuffer<Vec2f> (when Scalar is float), or of VecBuffer<Vec8d>, VecBuffer<Vec4d> and VecBuffer<Vec2d> (when Scalar is double).
Only the VecBuffers whose underlying vectorclass type is supported by the hardware will be used, in order to easily abstract over the many SIMD instruction sets.
ARM support
On ARM, Vec4f and Vec2d are implemented for float32x4_t and float64x2_t, with most of their member functions, all of their operators overloaded, and some math function overloads (exp, log, sin, cos, sincos, tan).
Vec2d math implementations on ARM
Three implementations of the double-precision NEON math overloads (exp, log, sin, cos, sincos, tan) are available, selected at compile time:
- Default (no flag): NEON port of the double-precision math functions from Agner Fog's vectorclass library. Vectorized, full f64 precision. See
NeonMathDoubleVCL.hpp. AVEC_USE_NEON_PD: Pommier-style NEON math fromNeonMathDouble.hpp. Vectorized, but only ~f32-grade precision (~7 decimal digits) since the algorithms were originally written for single precision and ported to double without strengthening the polynomial approximations.AVEC_USE_SCALAR_PD: per-lane scalarlibmcalls (std::sinetc.). Full f64 precision; roughly 2× the per-call cost of the vectorized variants.
The two flags are mutually exclusive. The VCL-derived default is the recommended choice; the Pommier path is kept around for cases where the lower precision is acceptable in exchange for slightly leaner codegen, and the scalar path as a portability/debug fallback.
The single-precision Vec4f math overloads always go through Pommier's NeonMathFloat.hpp — no flag involved.
Credits
avec includes code from Boost.Align by Joseph Fernandes, without depending on the whole Boost library. See the file BoostAlign.hpp.
The default Vec2d math implementation is a NEON port of the double-precision math functions (exp, log, sin, cos, sincos, tan) from Agner Fog's vectorclass library (Apache License 2.0). Algorithms (VDT / Cephes / Moshier-derived) are unchanged from VCL's vectormath_exp.h / vectormath_trig.h; only the SIMD plumbing is rewritten in NEON. See NeonMathDoubleVCL.hpp.
The single-precision NEON math overloads (Vec4f exp, log, sin, cos, sincos, tan), and the alternative double-precision Pommier-style overloads available under AVEC_USE_NEON_PD, were written by Julien Pommier and are available at http://gruntthepeon.free.fr/ssemath/neon_mathfun.html.
Namespace
vectorclass declarations live inside namespace vcl (forced via VCL_NAMESPACE=vcl
before the vectorclass.h includes in avec/Simd.hpp). The widely-used identifiers
— Vec2d, Vec4f, Vec4d, Vec8d, Vec8f, Vec16f, their *b mask cousins,
permute2/4/8, select — are re-exported at global scope so existing consumer
code (audio-dsp, individual plugin DSP TUs) keeps compiling without vcl::
qualifiers. Free functions found via ADL (operator overloads, exp(Vec2d), …)
still resolve correctly because the operand types live in vcl::.
The reason this matters: vectorclass's vector_convert.h declares free functions
named extend() for the AVX-2 / AVX-512 vector widths. On macOS, that name
collides with the anonymous-enum value extend = 0x40 declared at global scope
by <MacTypes.h>, which iPlug2's IControls.h pulls in transitively via Cocoa
headers. Without the namespace wrapper, no iPlug2 + avec plugin builds for
x86_64 on macOS.
Tests
$ cd test
$ ./run-tests.sh
builds and runs the doctest-based test suite for every architecture this host
supports. On Apple Silicon that's both arm64 (NEON path) and x86_64
(vectorclass path via Rosetta). On Intel and Linux it's just the host arch.
What's covered:
- SIMD math (
exp,log,sin,cos,tan,sqrt) onVec2dper lane, compared againststd::within a per-function ULP tolerance — catches drift between the x86 vectorclass implementation and the NEON shim. - Arithmetic operators on
Vec2dandVec4f. - Aligned and unaligned
load/storeround-trips. permute2<i, j>— every {0,1}×{0,1} combination, verifies the indexing convention matches across architectures.InterleavedBuffer<Float>::interleave+deinterleaveround-trip, parameterised over channel counts that exercise both the Vec-aligned path and the tail path.
Each REQUIRE / CHECK is a doctest assertion, so a regression produces a
non-zero exit code and a precise file:line failure report — suitable for CI.
Documentation
The documentation, available at https://unevens.github.io/avec/, can be generated with Doxygen running
$ doxygen doxyfile.txt