implementation_defined.md

Implementation Specific Behavior

29.10.2.2 [simd.expos.abi] A single class template std::simd::_Abi<N, Regs, Var> is defined to cover all necessary ABI tag types.

• N is the number of elements in the basic_vec/basic_mask.
• Regs is the (minimum) number of registers needed to hold all N elements.
• Var is a bit-flag identifying mask and layout variants (vector masks vs bit masks; interleaved vs contiguous storage of real and imaginary parts of complex<T>)

29.10.2.2 [simd.expos.abi] There currently is no hard maximum on the number of elements in a data-parallel type. However, depending on target architecture and element type, certain numbers will be too inefficient to compile and use.

29.10.2.2 [simd.expos.abi] native-abi<T> is an alias for _Abi<N, 1, Var> such that N is maximal. Var prefers bit masks over vector masks and interleaved complex over contiguous complex.

29.10.3 [simd.syn] zero_element is INT_MIN and uninit_element is zero_element + 1.

29.10.7.1 [simd.overview] A basic_vec<T, _Abi<N, Regs, Var>> specialization is enabled if N >= 1, Regs >= 1, and

• Var identifies non-complex storage, vector masks, and T is an arithmetic type,
• Var identifies non-complex storage, bit masks, T is an arithmetic type, and the target supports AVX-512,
• Var identifies interleaved or contiguous complex storage, vector masks, and T is a specialization of complex<U>, or
• Var identifies interleaved or contiguous complex storage, bit masks, T is a specialization of complex<U>, and the target supports AVX-512,

29.10.7.1 [simd.overview]

• In general, basic_vec<T, _Abi<N, Regs, Var>> with arithmetic type T can be converted to and from GNU vector types (see GNU attribute vector_size(Bytes)) of element type T with std::bit_ceil(N) elements.
• On x86 targets basic_vec<T, _Abi<N, 1, Var>> can be converted to and from
  • __m128 if T is float or std::float32_t, and sizeof(basic_vec) is 16;
  • __m128d if T is double or std::float64_t, and sizeof(basic_vec) is 16;
  • __m128h if T is std::float16_t, and sizeof(basic_vec) is 16;
  • __m128i if T is integral, and sizeof(basic_vec) is 16;
  • __m256 if T is float or std::float32_t, and sizeof(basic_vec) is 32;
  • __m256d if T is double or std::float64_t, and sizeof(basic_vec) is 32;
  • __m256h if T is std::float16_t, and sizeof(basic_vec) is 32;
  • __m256i if T is integral, and sizeof(basic_vec) is 32;
  • __m512 if T is float or std::float32_t, and sizeof(basic_vec) is 64;
  • __m512d if T is double or std::float64_t, and sizeof(basic_vec) is 64;
  • __m512h if T is std::float16_t, and sizeof(basic_vec) is 64;
  • __m512i if T is integral, and sizeof(basic_vec) is 64.

29.10.9.1 [simd.mask.overview] A basic_mask<Bytes, _Abi<N, Regs, Var>> specialization is enabled if N >= 1, Regs >= 1, and

• TODO

29.10.9.1 [simd.mask.overview]

• If Var does not identify a vector mask, no additional conversions are available.

• If Var identifies non-complex element types, then basic_mask<Bytes, _Abi<N, Regs, Var>> can be converted to and from GNU vector types of element type T, where T is the first type of signed char, signed short, signed int, signed long long, or signed long with sizeof(T) == Bytes. The number of elements in the GNU vector type must be equal to std::bit_ceil(N).

• If Var identifies complex element types, basic_mask<Bytes, _Abi<N, Regs, Var>> can be converted to and from GNU vector types of element type T, where T is the first type of signed char, signed short, signed int, signed long long, or signed long with sizeof(T) == Bytes / 2 (matching the sizeof a single real/imag element in the complex value-type).

  • If Var identifies interleaved storage and N > 1, the number of elements in the GNU vector type must be equal to std::bit_ceil(N * 2).

  • Otherwise, the number of elements in the GNU vector type must be equal to std::bit_ceil(N).

License and Copyright

SPDX-License-Identifier: CC-BY-4.0 Copyright © 2026 GSI Helmholtzzentrum fuer Schwerionenforschung GmbH Matthias Kretz m.kretz@gsi.de