native_f16x8_sse2.hpp

 
 
#pragma once
 
#include "native_simd_utility.hpp"
#include "../utility/utility.hpp"
#include "../macros.hpp"
 
 
 
namespace hi {
inline namespace v1 {
 
#ifdef HI_HAS_SSE2
 
 
template<>
struct native_simd<float16,8> {
    using value_type = float16;
    constexpr static size_t size = 8;
    using register_type = __m128i;
 
    register_type v;
 
    native_simd(native_simd const&) noexcept = default;
    native_simd(native_simd &&) noexcept = default;
    native_simd &operator=(native_simd const &) noexcept = default;
    native_simd &operator=(native_simd &&) noexcept = default;
 
    native_simd() noexcept : v(_mm_setzero_si128()) {}
 
    explicit native_simd(native_simd<float,8> const &a) noexcept;
    native_simd(native_simd<float,4> const &a, native_simd<float,4> const &b) noexcept;
 
    [[nodiscard]] explicit native_simd(register_type other) noexcept : v(other) {}
 
    [[nodiscard]] explicit operator register_type () const noexcept {
        return v;
    }
 
    [[nodiscard]] native_simd(value_type a, value_type b = value_type{}, value_type c = value_type{}, value_type d = value_type{},
                             value_type e = value_type{}, value_type f = value_type{}, value_type g = value_type{},
                             value_type h = value_type{}) noexcept :
        v(_mm_set_epi16(h.v, g.v, f.v, e.v, d.v, c.v, b.v, a.v)) {}
 
    [[nodiscard]] explicit native_simd(value_type const *other) noexcept : v(_mm_loadu_si128(reinterpret_cast<register_type const *>(other))) {}
 
    void store(value_type *out) const noexcept
    {
        hi_axiom_not_null(out);
        _mm_storeu_si128(reinterpret_cast<register_type *>(out), v);
    }
 
    [[nodiscard]] explicit native_simd(void const *other) noexcept : v(_mm_loadu_si128(static_cast<register_type const *>(other))) {}
 
    void store(void *out) const noexcept
    {
        hi_axiom_not_null(out);
        _mm_storeu_si128(static_cast<register_type *>(out), v);
    }
 
    [[nodiscard]] explicit native_simd(std::span<value_type const> other) noexcept
    {
        hi_axiom(other.size() >= size);
        v = _mm_loadu_si128(reinterpret_cast<register_type const *>(other.data()));
    }
 
    void store(std::span<value_type> out) const noexcept
    {
        hi_axiom(out.size() >= size);
        _mm_storeu_si128(reinterpret_cast<register_type *>(out.data()), v);
    }
 
    template<size_t N>
    [[nodiscard]] explicit native_simd(std::array<value_type, N> other) noexcept requires (N >= size) : v(_mm_loadu_si128(reinterpret_cast<register_type const *>(other.data()))) {}
 
    template<size_t N>
    [[nodiscard]] explicit operator std::array<value_type, N> () const noexcept requires (N >= size)
    {
        auto r = std::array<value_type, size>{};
        _mm_storeu_si128(reinterpret_cast<register_type *>(r.data()), v);
        return r;
    }
 
 
    [[nodiscard]] static native_simd broadcast(int16_t a) noexcept
    {
        return native_simd{_mm_set1_epi16(a)};
    }
 
    [[nodiscard]] static native_simd broadcast(native_simd a) noexcept
    {
#ifdef HI_HAS_AVX2
        return native_simd{_mm_broadcastw_epi16(a.v)};
#else
        // Create a mask for 1 word each dword, AND it with a.v.
        auto tmp = _mm_undefined_si128();
        tmp = _mm_cmpeq_epi32(tmp, tmp);
        tmp = _mm_slli_epi32(tmp, 16);
        tmp = _mm_and_si128(tmp, a.v);
 
        // Broadcast the first word to all the bytes in the first dword.
        tmp = _mm_or_si128(tmp, _mm_slli_epi32(tmp, 16));
 
        // Broadcast the first dword to all 4 dwords.
        tmp = _mm_shuffle_epi32(tmp, 0b00'00'00'00);
        return native_simd{tmp};
#endif
    }
 
    [[nodiscard]] static native_simd from_mask(size_t mask) noexcept
    {
        hi_axiom(mask <= 0b1111'1111);
 
        return native_simd{
            mask & 0b0000'0001 ? value_type{} : value_type::from_uint16_t(0xffff),
            mask & 0b0000'0010 ? value_type{} : value_type::from_uint16_t(0xffff),
            mask & 0b0000'0100 ? value_type{} : value_type::from_uint16_t(0xffff),
            mask & 0b0000'1000 ? value_type{} : value_type::from_uint16_t(0xffff),
            mask & 0b0001'0000 ? value_type{} : value_type::from_uint16_t(0xffff),
            mask & 0b0010'0000 ? value_type{} : value_type::from_uint16_t(0xffff),
            mask & 0b0100'0000 ? value_type{} : value_type::from_uint16_t(0xffff),
            mask & 0b1000'0000 ? value_type{} : value_type::from_uint16_t(0xffff)};
    }
 
    [[nodiscard]] size_t mask() const noexcept
    {
        auto tmp = _mm_movemask_epi8(v);
        tmp &= 0b0101'0101;
        tmp |= tmp >> 1;
        tmp &= 0b0011'0011;
        tmp |= tmp >> 2;
        tmp &= 0b0000'1111;
        return narrow_cast<size_t>(tmp);
    }
 
 
    [[nodiscard]] friend native_simd operator==(native_simd a, native_simd b) noexcept
    {
        return native_simd{_mm_cmpeq_epi16(a.v, b.v)};
    }
 
    [[nodiscard]] friend native_simd operator!=(native_simd a, native_simd b) noexcept
    {
        return ~(a == b);
    }
 
    [[nodiscard]] friend native_simd operator&(native_simd a, native_simd b) noexcept
    {
        return native_simd{_mm_and_si128(a.v, b.v)};
    }
 
    [[nodiscard]] friend native_simd operator|(native_simd a, native_simd b) noexcept
    {
        return native_simd{_mm_or_si128(a.v, b.v)};
    }
 
    [[nodiscard]] friend native_simd operator^(native_simd a, native_simd b) noexcept
    {
        return native_simd{_mm_xor_si128(a.v, b.v)};
    }
 
    [[nodiscard]] friend native_simd operator~(native_simd a) noexcept
    {
        auto ones = _mm_undefined_si128();
        ones = _mm_cmpeq_epi32(ones, ones);
        return native_simd{_mm_andnot_si128(a.v, ones)};
    }
 
    template<size_t Mask>
    [[nodiscard]] friend native_simd set_zero(native_simd a) noexcept
    {
        static_assert(Mask <= 0b1111'1111);
        hilet mask = from_mask(Mask);
        return not_and(mask, a);
    }
 
    template<size_t Index>
    [[nodiscard]] friend native_simd insert(native_simd a, value_type b) noexcept
    {
        static_assert(Index < 4);
        return native_simd{_mm_insert_epi16(a, b.v, narrow_cast<int>(Index))};
    }
 
    template<size_t Index>
    [[nodiscard]] friend value_type extract(native_simd a) noexcept
    {
        return std::bit_cast<value_type>(_mm_extract_epi16(a, Index));
    }
 
    template<size_t Mask>
    [[nodiscard]] friend native_simd blend(native_simd a, native_simd b) noexcept
    {
#ifdef HI_HAS_SSE4_1
        return native_simd{_mm_blend_epi16(a, b, Mask)};
#else
        hilet mask = from_mask(Mask);
        return not_and(mask, a) | (mask & b);
#endif
    }
 
    //template<fixed_string SourceElements>
    //[[nodiscard]] static native_simd permute(native_simd a) noexcept
    //{
    //    constexpr auto order = detail::native_swizzle_to_packed_indices<SourceElements, size>();
    //
    //    if constexpr (order == 0b111'110'101'100'011'010'001'000) {
    //        return a.v;
    //    } else {
    //        auto tmp = _mm_shufflelo(a.v, 
    //        return native_simd{_mm_shuffle_epi16(a.v, order)};
    //    }
    //}
 
    template<fixed_string SourceElements>
    [[nodiscard]] friend native_simd swizzle(native_simd a) noexcept
    {
        constexpr auto one_mask = detail::native_swizzle_to_mask<SourceElements, size, '1'>();
        constexpr auto zero_mask = detail::native_swizzle_to_mask<SourceElements, size, '0'>();
        constexpr auto number_mask = one_mask | zero_mask;
 
        if constexpr (number_mask == 0b11111111) {
            // Swizzle was /[01][01][01][01]/.
            return swizzle_numbers<SourceElements>();
 
        } else if constexpr (number_mask == 0b00000000) {
            // Swizzle was /[^01][^01][^01][^01]/.
            return permute<SourceElements>(a);
 
#ifdef HI_HAS_SSE4_1
        } else if constexpr (number_mask == zero_mask) {
            // Swizzle was /[^1][^1][^1][^1]/.
            hilet ordered = permute<SourceElements>(a);
            return set_zero<zero_mask>(ordered);
#endif
 
        } else {
            hilet ordered = permute<SourceElements>(a);
            hilet numbers = swizzle_numbers<SourceElements>();
            return blend<number_mask>(ordered, numbers);
        }
    }
 
    [[nodiscard]] friend native_simd not_and(native_simd a, native_simd b) noexcept
    {
        return native_simd{_mm_andnot_si128(a.v, b.v)};
    }
 
    template<fixed_string SourceElements>
    [[nodiscard]] static native_simd swizzle_numbers() noexcept
    {
        constexpr auto one_mask = detail::native_swizzle_to_mask<SourceElements, size, '1'>();
        constexpr auto zero_mask = detail::native_swizzle_to_mask<SourceElements, size, '0'>();
        constexpr auto number_mask = one_mask | zero_mask;
        constexpr auto alpha_mask = ~number_mask & 0b11111111;
       
        if constexpr ((zero_mask | alpha_mask) == 0b11111111) {
            return native_simd{_mm_setzero_si128()};
 
        } else if constexpr ((one_mask | alpha_mask)== 0b11111111) {
            return native_simd::broadcast(value_type::from_uint16_t(0x3c00)); // 1.0
 
        } else {
            return native_simd{
                to_bool(one_mask & 0b00000001) ? value_type::from_uint16_t(0x3c00) : value_type{},
                to_bool(one_mask & 0b00000010) ? value_type::from_uint16_t(0x3c00) : value_type{},
                to_bool(one_mask & 0b00000100) ? value_type::from_uint16_t(0x3c00) : value_type{},
                to_bool(one_mask & 0b00001000) ? value_type::from_uint16_t(0x3c00) : value_type{},
                to_bool(one_mask & 0b00010000) ? value_type::from_uint16_t(0x3c00) : value_type{},
                to_bool(one_mask & 0b00100000) ? value_type::from_uint16_t(0x3c00) : value_type{},
                to_bool(one_mask & 0b01000000) ? value_type::from_uint16_t(0x3c00) : value_type{},
                to_bool(one_mask & 0b10000000) ? value_type::from_uint16_t(0x3c00) : value_type{}
            };
        }
 
    }
 
};
 
#endif
 
 
}}