Expressive Vector Engine - SIMD in C++ Goes Brrrr
EVE is a re-implementation of the old EVE SIMD library by Falcou et al., which for a while was named Boost.SIMD. It's a C++20 and onward implementation of a type-based wrapper around SIMD extensions sets for most current architectures. It aims to show how C++20 can be used to design and implement an efficient, low-level, high-abstraction library suited for high performance.
It's a research project first and an open-source library second. We reserve the right to change the API and baseline compiler required until the first official 0.1 release. However, we'll try to minimise disruption. Semantic versioning will ensure API retro-compatibility if anything huge needs to change.
EVE is considered operational: it's usable and has a large feature set for a sensible number of instruction sets. However, some values or performance issues may still remain. Don't hesitate to report any funky code generation or bad optimisations so we can deliver the best performance around.
In terms of SIMD extension sets, we actively support (ie code is optimised and regularly tested) the following:
Full support with tests:
Partial/In-progress support with minimal checks:
We do not support ARM SVE with dynamic size or GPGPU; this is the job of another tool.
EVE requires a C++20-compliant compiler. The main features from C++17/20 we require are:
if constexpr<bits>Here is the current minimal compiler version supported:
| Compiler | Version |
|---|---|
| g++ | 13 or above |
| clang++ | 16 or above |
| Visual Studio | 19.44 or above |
The current recommended compiler version are:
| Compiler | Version |
|---|---|
| g++ | 14 or above |
| clang++ | 19 or above |
| Visual Studio | 19.44 or above |
Why C++20 ? We want to be able to use modern design components like Concepts and, later, Modules. C++14 is for 2014. We fully know it may hinder adoption in some situations, but we're not running a race. We value proper design and API choice over complex retro-compatibility.
Why this instead of std::simd or any other libraries? EVE is a playground for testing proper design and implementation decisions for such libraries. Vc and its standard version are fine work, but we aim to push the envelope regarding API and implementation without standard restrictions. We will try to keep up with standard naming as much as possible.
Is this Boost.SIMD v42.69? Yes and no. This reimplementation is based on slightly different premises and some radical new ideas. Our previous Boost exhibited all results.SIMD papers still hold, and we'll try to improve upon.
I want to help! That's Good. We need tests on exotic configurations, maybe a stable PPC test machine. Write code, make it fail, and fork and hack away. This will allow us to improve for use cases we may have missed. We would also value input (even harsh ones) on how we use our CI infrastructure or how our CMake looks. We tried our best, but those elements are complex all by themselves.
If you want to refer to EVE, you can use those papers (by order of preference in citation). A new, more up-to-date EVE-specific journal paper is in the works at the moment.
…No open issues yet, or sync has not completed.