Skip to content

std::experimental::simd for GCC [ISO/IEC TS 19570:2018]

License

Notifications You must be signed in to change notification settings

VcDevel/std-simd

Repository files navigation

std::experimental::simd

portable, zero-overhead C++ types for explicitly data-parallel programming

Development here is going to move on to std::simd for C++26. For the TS implementation reach for GCC/libstdc++. std::experimental::simd is shipping with GCC since version 11.

This package implements ISO/IEC TS 19570:2018 Section 9 "Data-Parallel Types". The implementation derived from https://github.com/VcDevel/Vc.

By default, the install.sh script places the std::experimental::simd headers into the directory where the standard library of your C++ compiler (identified via $CXX) resides.

It is only tested and supported with GCC trunk, even though it may work with older GCC versions.

Target support

  • x86_64 is the main development platform and thoroughly tested. This includes support from SSE-only up to AVX512 on Xeon Phi or Xeon CPUs.
  • aarch64, arm, and ppc64le was tested and verified to work. No significant performance evaluation was done.
  • In any case, a fallback to correct execution via builtin arithmetic types is available for all targets.

Installation Instructions

$ ./install.sh

Use --help to learn about the available options.

Example

Scalar Product

Let's start from the code for calculating a 3D scalar product using builtin floats:

using Vec3D = std::array<float, 3>;
float scalar_product(Vec3D a, Vec3D b) {
  return a[0] * b[0] + a[1] * b[1] + a[2] * b[2];
}

Using simd, we can easily vectorize the code using the native_simd<float> type (Compiler Explorer):

using std::experimental::native_simd;
using Vec3D = std::array<native_simd<float>, 3>;
native_simd<float> scalar_product(Vec3D a, Vec3D b) {
  return a[0] * b[0] + a[1] * b[1] + a[2] * b[2];
}

The above will scale to 1, 4, 8, 16, etc. scalar products calculated in parallel, depending on the target hardware's capabilities.

For comparison, the same vectorization using Intel SSE intrinsics is more verbose, uses prefix notation (i.e. function calls), and neither scales to AVX or AVX512, nor is it portable to different SIMD ISAs:

using Vec3D = std::array<__m128, 3>;
__m128 scalar_product(Vec3D a, Vec3D b) {
  return _mm_add_ps(_mm_add_ps(_mm_mul_ps(a[0], b[0]), _mm_mul_ps(a[1], b[1])),
                    _mm_mul_ps(a[2], b[2]));
}

Build Requirements

none. It's header-only.

However, to build the unit tests you will need:

  • cmake >= 3.0
  • GCC >= 9.1

To execute all AVX512 unit tests, you will need the Intel SDE.

Building the tests

$ make test

This will create a build directory, run cmake, compile the tests, and execute the tests.

Documentation

https://en.cppreference.com/w/cpp/experimental/simd

Publications

License

The simd headers, tests, and benchmarks are released under the terms of the 3-clause BSD license.

Note that the code in libstdc++ is distributed under GPL3 with runtime library exception.