simd_algorithm.hpp

Go to the documentation of this file.

// -----------------------------------------------------------------------------------------------------
// Copyright (c) 2006-2020, Knut Reinert & Freie Universität Berlin
// Copyright (c) 2016-2020, Knut Reinert & MPI für molekulare Genetik
// This file may be used, modified and/or redistributed under the terms of the 3-clause BSD-License
// shipped with this file and also available at: https://github.com/seqan/seqan3/blob/master/LICENSE.md
// -----------------------------------------------------------------------------------------------------
 
#pragma once
 
#include <array>
#include <utility>
 
#include <seqan3/core/simd/concept.hpp>
#include <seqan3/core/simd/detail/builtin_simd.hpp>
#include <seqan3/core/simd/detail/simd_algorithm_sse4.hpp>
#include <seqan3/core/simd/detail/simd_algorithm_avx2.hpp>
#include <seqan3/core/simd/detail/simd_algorithm_avx512.hpp>
#include <seqan3/core/simd/simd_traits.hpp>
#include <seqan3/std/concepts>
 
namespace seqan3::detail
{
 
template <simd::simd_concept simd_t, size_t... I>
constexpr simd_t fill_impl(typename simd_traits<simd_t>::scalar_type const scalar, std::index_sequence<I...>) noexcept
{
    return simd_t{((void)I, scalar)...};
}
 
template <simd::simd_concept simd_t, typename scalar_t, scalar_t... I>
constexpr simd_t iota_impl(scalar_t const offset, std::integer_sequence<scalar_t, I...>)
{
    return simd_t{static_cast<scalar_t>(offset + I)...};
}
 
template <size_t divisor, simd_concept simd_t>
constexpr simd_t extract_impl(simd_t const & src, uint8_t const mask)
{
    simd_t dst{};
    constexpr size_t chunk = simd_traits<simd_t>::length / divisor;
    size_t offset = chunk * mask;
    for (size_t i = 0; i < chunk; ++i)
        dst[i] = src[i + offset];
 
    return dst;
}
 
template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
constexpr target_simd_t upcast_signed(source_simd_t const & src)
{
    static_assert(simd_traits<target_simd_t>::max_length == simd_traits<source_simd_t>::max_length,
                  "Target vector has different byte size.");
 
    if constexpr (simd_traits<source_simd_t>::max_length == 16) // SSE4
        return upcast_signed_sse4<target_simd_t>(src);
    else if constexpr (simd_traits<source_simd_t>::max_length == 32) // AVX2
        return upcast_signed_avx2<target_simd_t>(src);
    else if constexpr (simd_traits<source_simd_t>::max_length == 64) // AVX512
        return upcast_signed_avx512<target_simd_t>(src);
    else
        static_assert(simd_traits<source_simd_t>::max_length <= 32, "simd type is not supported.");
}
 
template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
constexpr target_simd_t upcast_unsigned(source_simd_t const & src)
{
    static_assert(simd_traits<target_simd_t>::max_length == simd_traits<source_simd_t>::max_length,
                  "Target vector has different byte size.");
 
    if constexpr (simd_traits<source_simd_t>::max_length == 16) // SSE4
        return upcast_unsigned_sse4<target_simd_t>(src);
    else if constexpr (simd_traits<source_simd_t>::max_length == 32) // AVX2
        return upcast_unsigned_avx2<target_simd_t>(src);
    else if constexpr (simd_traits<source_simd_t>::max_length == 64) // AVX512
        return upcast_unsigned_avx512<target_simd_t>(src);
    else
        static_assert(simd_traits<source_simd_t>::max_length <= 32, "simd type is not supported.");
}
 
template <uint8_t index, simd::simd_concept simd_t>
constexpr simd_t extract_half(simd_t const & src)
{
    static_assert(index < 2, "The index must be in the range of [0, 1]");
 
    return detail::extract_impl<2>(src, index);
}
 
template <uint8_t index, simd::simd_concept simd_t>
    requires detail::is_builtin_simd_v<simd_t> &&
             detail::is_native_builtin_simd_v<simd_t>
constexpr simd_t extract_half(simd_t const & src)
{
    static_assert(index < 2, "The index must be in the range of [0, 1]");
 
    if constexpr (simd_traits<simd_t>::length < 2) // In case there are less elements available return unchanged value.
        return src;
    else if constexpr (simd_traits<simd_t>::max_length == 16) // SSE4
        return detail::extract_half_sse4<index>(src);
    else if constexpr (simd_traits<simd_t>::max_length == 32) // AVX2
        return detail::extract_half_avx2<index>(src);
    else // Anything else
        return detail::extract_impl<2>(src, index);
}
 
template <uint8_t index, simd::simd_concept simd_t>
constexpr simd_t extract_quarter(simd_t const & src)
{
    static_assert(index < 4, "The index must be in the range of [0, 1, 2, 3]");
 
    return detail::extract_impl<4>(src, index);
}
 
template <uint8_t index, simd::simd_concept simd_t>
    requires detail::is_builtin_simd_v<simd_t> &&
             detail::is_native_builtin_simd_v<simd_t>
constexpr simd_t extract_quarter(simd_t const & src)
{
    static_assert(index < 4, "The index must be in the range of [0, 1, 2, 3]");
 
    if constexpr (simd_traits<simd_t>::length < 4) // In case there are less elements available return unchanged value.
        return src;
    else if constexpr (simd_traits<simd_t>::max_length == 16) // SSE4
        return detail::extract_quarter_sse4<index>(src);
    else if constexpr (simd_traits<simd_t>::max_length == 32) // AVX2
        return detail::extract_quarter_avx2<index>(src);
    else // Anything else
        return detail::extract_impl<4>(src, index);
}
 
template <uint8_t index, simd::simd_concept simd_t>
constexpr simd_t extract_eighth(simd_t const & src)
{
    return detail::extract_impl<8>(src, index);
}
 
template <uint8_t index, simd::simd_concept simd_t>
    requires detail::is_builtin_simd_v<simd_t> &&
             detail::is_native_builtin_simd_v<simd_t>
constexpr simd_t extract_eighth(simd_t const & src)
{
    static_assert(index < 8, "The index must be in the range of [0, 1, 2, 3, 4, 5, 6, 7]");
 
    if constexpr (simd_traits<simd_t>::length < 8) // In case there are less elements available return unchanged value.
        return src;
    else if constexpr (simd_traits<simd_t>::max_length == 16) // SSE4
        return detail::extract_eighth_sse4<index>(src);
    else if constexpr (simd_traits<simd_t>::max_length == 32) // AVX2
        return detail::extract_eighth_avx2<index>(src);
    else  // Anything else
        return detail::extract_impl<8>(src, index);
}
 
} // namespace seqan3::detail
 
namespace seqan3
{
 
inline namespace simd
{
 
template <simd::simd_concept simd_t>
constexpr simd_t fill(typename simd_traits<simd_t>::scalar_type const scalar) noexcept
{
    constexpr size_t length = simd_traits<simd_t>::length;
    return detail::fill_impl<simd_t>(scalar, std::make_index_sequence<length>{});
}
 
template <simd::simd_concept simd_t>
constexpr simd_t iota(typename simd_traits<simd_t>::scalar_type const offset)
{
    constexpr size_t length = simd_traits<simd_t>::length;
    using scalar_type = typename simd_traits<simd_t>::scalar_type;
    return detail::iota_impl<simd_t>(offset, std::make_integer_sequence<scalar_type, length>{});
}
 
template <simd::simd_concept simd_t>
constexpr simd_t load(void const * mem_addr)
{
    assert(mem_addr != nullptr);
    simd_t tmp{};
 
    for (size_t i = 0; i < simd_traits<simd_t>::length; ++i)
        tmp[i] = *(static_cast<typename simd_traits<simd_t>::scalar_type const *>(mem_addr) + i);
 
    return tmp;
}
 
template <simd::simd_concept simd_t>
    requires detail::is_builtin_simd_v<simd_t> &&
             detail::is_native_builtin_simd_v<simd_t>
constexpr simd_t load(void const * mem_addr)
{
    assert(mem_addr != nullptr);
 
    if constexpr (simd_traits<simd_t>::max_length == 16)
        return detail::load_sse4<simd_t>(mem_addr);
    else if constexpr (simd_traits<simd_t>::max_length == 32)
        return detail::load_avx2<simd_t>(mem_addr);
    else if constexpr (simd_traits<simd_t>::max_length == 64)
        return detail::load_avx512<simd_t>(mem_addr);
    else
        static_assert(simd_traits<simd_t>::max_length >= 16 && simd_traits<simd_t>::max_length <= 64,
                      "Unsupported simd type.");
}
 
template <simd::simd_concept simd_t>
constexpr void transpose(std::array<simd_t, simd_traits<simd_t>::length> & matrix)
{
    std::array<simd_t, simd_traits<simd_t>::length> tmp{};
 
    for (size_t i = 0; i < matrix.size(); ++i)
        for (size_t j = 0; j < matrix.size(); ++j)
            tmp[j][i] = matrix[i][j];
 
    std::swap(tmp, matrix);
}
 
// Implementation for seqan builtin simd.
template <simd::simd_concept simd_t>
    requires detail::is_builtin_simd_v<simd_t> &&
             detail::is_native_builtin_simd_v<simd_t> &&
             (simd_traits<simd_t>::max_length == simd_traits<simd_t>::length)
constexpr void transpose(std::array<simd_t, simd_traits<simd_t>::length> & matrix)
{
    if constexpr (simd_traits<simd_t>::length == 16) // SSE4 implementation
        detail::transpose_matrix_sse4(matrix);
    else if constexpr (simd_traits<simd_t>::length == 32) // AVX2 implementation
        detail::transpose_matrix_avx2(matrix);
    else
        transpose(matrix);
}
 
template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
constexpr target_simd_t upcast(source_simd_t const & src)
{
    static_assert(simd_traits<target_simd_t>::length <= simd_traits<source_simd_t>::length,
                  "The length of the target simd type must be greater or equal than the length of the source simd type.");
 
    target_simd_t tmp{};
    for (unsigned i = 0; i < simd_traits<target_simd_t>::length; ++i)
        tmp[i] = static_cast<typename simd_traits<target_simd_t>::scalar_type>(src[i]);
 
    return tmp;
}
 
template <simd::simd_concept target_simd_t, simd::simd_concept source_simd_t>
    requires detail::is_builtin_simd_v<target_simd_t> &&
             detail::is_builtin_simd_v<source_simd_t> &&
             detail::is_native_builtin_simd_v<source_simd_t>
constexpr target_simd_t upcast(source_simd_t const & src)
{
    static_assert(simd_traits<target_simd_t>::length <= simd_traits<source_simd_t>::length,
                  "The length of the target simd type must be greater or equal than the length of the source simd type.");
 
    if constexpr (simd_traits<source_simd_t>::length == simd_traits<target_simd_t>::length)
    {
        static_assert(simd_traits<target_simd_t>::max_length == simd_traits<source_simd_t>::max_length,
                    "Target vector has a different byte size.");
        return reinterpret_cast<target_simd_t>(src);  // Same packing so we do not cast.
    }
    else if constexpr (std::signed_integral<typename simd_traits<source_simd_t>::scalar_type>)
    {
        return detail::upcast_signed<target_simd_t>(src);
    }
    else
    {
        static_assert(std::unsigned_integral<typename simd_traits<source_simd_t>::scalar_type>,
                      "Expected unsigned scalar type.");
        return detail::upcast_unsigned<target_simd_t>(src);
    }
}
 
} // inline namespace simd
 
} // namespace seqan3