C++26 Data-parallel types (simd)

Build, install, use?

Parts of this implementation are already integrated into GCC 16. If you want to try the development version you can either install this library via:

make prefix=~/.local install

or simply point your compiler's include path to this repository.

Alternatively you can also install into the compiler's standard library directory:

make install-system

There is no support for uninstalling, since install-system overwrites the std::simd headers of libstdc++.

Build your own project with C++26 (latest GCC, PRs to support Clang are welcome). Just include <simd> and you're good to go:

#include <simd>

To build the tests there are multiple targets available. make help will list all of them.

Feature macros

By default all vectorizable types (including complex) are supported. More features can be enabled by defining the following macros:

Macro	Description
VIR_EXTENSIONS	Enable several optimizations and warnings on guaranteed precondition violations.
VIR_PATCH_PERMUTE_DYNAMIC	Implements [simd.permute.dynamic].
VIR_PATCH_MATH	Implements [simd.math].
VIR_PATCH_IMPROVE_CX	Implements abs and norm for vec<complex<T>>. Three different approaches =1, =2, and =3 make different optimization/code-gen trade-offs.
VIR_PATCH_MISSED_OPT	Enable hand-written instruction selection for optimization patterns the compiler misses. Includes ktest-based mask reductions and pshufb-based type conversions on x86.
VIR_PATCH_TEST_STORES	Fix masked stores.
VIR_CONSTEVAL_BROADCAST	Use consteval broadcast constructor for value-preserving conversions: Either the value doesn't change or the program is ill-formed. Peace of mind.

Implementation status

Feature	Status
P1928R15 std::simd — merge data-parallel types from the Parallelism TS 2	✅ done (except math)
P2663R7 Interleaved complex values support in std::simd	✅ done (except math)
P2664R11 Proposal to extend std::simd with permutation API	🟡 partial
P2876R3 Proposal to extend std::simd with more constructors and accessors	✅ done
P2929R2 simd_invoke	🔴 open (still in design phase)
P2933R4 Extend ⟨bit⟩ header function with overloads for std::simd	✅ done
P2933R4 std::simd overloads for <bit> header	✅ done (not optimized)
P2964R2 Allowing user-defined types in std::simd	🔴 open (still in design phase)
P3287R3 Exploration of namespaces for std::simd	✅ done
P3430R3 simd issues: explicit, unsequenced, identity-element position, and members of disabled simd	✅ done
P3440R2 Add n_elements named constructor to std::simd	🔴 open (still in design phase)
P3441R2 Rename simd_split to simd_chunk	✅ done
P3480R6 std::simd is a range	✅ done
P3691R1 Reconsider naming of the namespace for 'std::simd'	✅ done
P3844R4 Reword [simd.math] for consteval conversions	✅ done
P3932R0 Fix LWG4470: Fix integer-from in [simd]	✅ done
P3973R0 bit_cast_as: Element type reinterpretation for std::simd	🔴 open (still in design phase)
P3983R0 simd object representation	🔴 open (still in design phase)
P3985R0 Concepts for std::simd	🔴 open (still in design phase)
P4012R1 value-preserving consteval broadcast to simd::vec	✅ done
P4042R0 Fix LWG4543: incorrect cast between simd::vec and simd::mask via conversion to and from impl-defined vector types	✅ done

Issue	Status
LWG4230 simd<complex>::real/imag is overconstrained	✅ done
LWG4231 datapar::chunk<N> should use simd-size-type instead of size_t	✅ done
LWG4232 datapar::resize does not resize	✅ done
LWG4238 simd_mask<complex<double>>::operator+/-/~ return a disabled simd specialization	✅ done
LWG4280 simd::partial_load uses undefined identifier T	🔴 unclear
LWG4375 std::simd::bit_ceil should not be noexcept	🔴 unclear
LWG4376 ABI tag in return type of [simd.mask.unary] is overconstrained	✅ done
LWG4382 The simd::basic_mask(bool) overload needs to be more constrained	🔴 unclear
LWG4385 Including <simd> doesn't provide std::begin/end	✅ done (via inclusion of <span>)
LWG4386 std::simd::select(bool c, const T& a, const U& b) is underconstrained	🔴 unclear
LWG4390 simd::basic_vec(U&&) default template parameter	⚪ treated as NAD
LWG4391 Ambiguities of simd::basic_vec constructor	🔴 unclear
LWG4392 simd::unchecked_load misses difference type casting	🔴 unclear
LWG4393 simd::unchecked_scatter_to is underconstrained	🔴 unclear
LWG4394 simd::unchecked_load(I first, S last) construct span maybe ill-formed	🔴 unclear
LWG4402 List-initialization of iterators in [simd.mask.overview]	🔴 unclear
LWG4403 simd::basic_vec CTAD misses difference type casting	✅ done
LWG4407 constexpr-wrapper-like needs remove_cvref_t in simd::basic_vec constructor	✅ done
LWG4408 Hardening simd::vec::operator[]	🔴 unclear
LWG4409 Constant expression ranges::size(r) Constraints and Mandates in [simd]	🔴 open
LWG4412 Fix declaration of zero_element and uninit_element	✅ done
LWG4413 Unused/left-over simd::alignment specialization for basic_mask	✅ done
LWG4414 §[simd.expos.abi] deduce-abi-t is underspecified and incorrectly referenced from rebind and resize	✅ done (P3932R0)
LWG4420 §[simd] conversions (constructor, load, stores, gather, and scatter) are incorrectly constrained for <stdfloat> types	✅ done
LWG4436 simd broadcast is overconstrained — std::cw<0.f> is not convertible to simd::vec<float16_t>	🔴 unclear
LWG4470 The use of integer-from<Bytes> all over [simd] is incorrect for Bytes=sizeof(complex<double>)	✅ done (P3932R0)
LWG4518 simd::cat return type requires inefficient ABI tag change/conversion	✅ done (P3932R0)
LWG4535 Disallow user specialization of <simd> templates	✅ done (not actionable)
LWG4543 Incorrect cast between simd::vec and simd::mask via conversion to and from impl-defined vector types	✅ done (P4042R0)