hadd tweaks

Author: DiamonDinoia

include/xsimd/arch/xsimd_avx.hpp

@@ -1067,14 +1067,10 @@ namespace xsimd
         template <class A>
         XSIMD_INLINE double reduce_add(batch<double, A> const& rhs, requires_arch<avx>) noexcept
         {
-            // rhs = (x0, x1, x2, x3)
-            // tmp = (x2, x3, x0, x1)
-            __m256d tmp = _mm256_permute2f128_pd(rhs, rhs, 1);
-            // tmp = (x2+x0, x3+x1, -, -)
-            tmp = _mm256_add_pd(rhs, tmp);
-            // tmp = (x2+x0+x3+x1, -, -, -)
-            tmp = _mm256_hadd_pd(tmp, tmp);
-            return _mm_cvtsd_f64(_mm256_extractf128_pd(tmp, 0));
+            __m128d low, high;
+            detail::split_avx(rhs, low, high);
+            batch<double, sse4_2> blow(low), bhigh(high);
+            return reduce_add(blow+bhigh, sse4_2 {});
         }
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
         XSIMD_INLINE T reduce_add(batch<T, A> const& self, requires_arch<avx>) noexcept

include/xsimd/arch/xsimd_avx512f.hpp

@@ -1410,10 +1410,14 @@ namespace xsimd
         template <class A>
         XSIMD_INLINE double reduce_add(batch<double, A> const& rhs, requires_arch<avx512f>) noexcept
         {
+#if defined(__INTEL_COMPILER)
+            return _mm512_reduce_add_pd(rhs);
+#else
             __m256d tmp1 = _mm512_extractf64x4_pd(rhs, 1);
             __m256d tmp2 = _mm512_extractf64x4_pd(rhs, 0);
             __m256d res1 = _mm256_add_pd(tmp1, tmp2);
             return reduce_add(batch<double, avx2>(res1), avx2 {});
+#endif
         }
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
         XSIMD_INLINE T reduce_add(batch<T, A> const& self, requires_arch<avx512f>) noexcept

include/xsimd/arch/xsimd_sse3.hpp

@@ -53,8 +53,9 @@ namespace xsimd
         template <class A>
         XSIMD_INLINE double reduce_add(batch<double, A> const& self, requires_arch<sse3>) noexcept
         {
-            __m128d tmp0 = _mm_hadd_pd(self, self);
-            return _mm_cvtsd_f64(tmp0);
+            double low = _mm_cvtsd_f64(self); // get lower element
+            double high = _mm_cvtsd_f64(_mm_shuffle_pd(self, self, 0x1)); // get upper element
+            return low + high;
         }
 
     }