diff options
Diffstat (limited to 'third_party/xsimd/include/xsimd/arch/xsimd_avx2.hpp')
| -rw-r--r-- | third_party/xsimd/include/xsimd/arch/xsimd_avx2.hpp | 38 |
1 files changed, 38 insertions, 0 deletions
diff --git a/third_party/xsimd/include/xsimd/arch/xsimd_avx2.hpp b/third_party/xsimd/include/xsimd/arch/xsimd_avx2.hpp index a5b07ec9da..a4881778e0 100644 --- a/third_party/xsimd/include/xsimd/arch/xsimd_avx2.hpp +++ b/third_party/xsimd/include/xsimd/arch/xsimd_avx2.hpp @@ -76,6 +76,44 @@ namespace xsimd } } + // avgr + template <class A, class T, class = typename std::enable_if<std::is_unsigned<T>::value, void>::type> + inline batch<T, A> avgr(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx2>) noexcept + { + XSIMD_IF_CONSTEXPR(sizeof(T) == 1) + { + return _mm256_avg_epu8(self, other); + } + else XSIMD_IF_CONSTEXPR(sizeof(T) == 2) + { + return _mm256_avg_epu16(self, other); + } + else + { + return avgr(self, other, generic {}); + } + } + + // avg + template <class A, class T, class = typename std::enable_if<std::is_unsigned<T>::value, void>::type> + inline batch<T, A> avg(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx2>) noexcept + { + XSIMD_IF_CONSTEXPR(sizeof(T) == 1) + { + auto adj = ((self ^ other) << 7) >> 7; + return avgr(self, other, A {}) - adj; + } + else XSIMD_IF_CONSTEXPR(sizeof(T) == 2) + { + auto adj = ((self ^ other) << 15) >> 15; + return avgr(self, other, A {}) - adj; + } + else + { + return avg(self, other, generic {}); + } + } + // bitwise_and template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type> inline batch<T, A> bitwise_and(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx2>) noexcept |