From 5f8c468a4fb896130db7fcc8921f050fd222cc06 Mon Sep 17 00:00:00 2001
From: serge-sans-paille <serge.guelton@telecom-bretagne.eu>
Date: Sun, 25 Feb 2024 17:29:44 +0100
Subject: [PATCH] Use fma and fms instruction when available to speedup complex
 multiply

This leverage the specific layout of xsimd batch of complex.
---
 include/xsimd/types/xsimd_batch.hpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/xsimd/types/xsimd_batch.hpp b/include/xsimd/types/xsimd_batch.hpp
index b4989fc88..d9108823a 100644
--- a/include/xsimd/types/xsimd_batch.hpp
+++ b/include/xsimd/types/xsimd_batch.hpp
@@ -1347,8 +1347,8 @@ namespace xsimd
     template <class T, class A>
     inline batch<std::complex<T>, A>& batch<std::complex<T>, A>::operator*=(batch const& other) noexcept
     {
-        real_batch new_real = real() * other.real() - imag() * other.imag();
-        real_batch new_imag = real() * other.imag() + imag() * other.real();
+        real_batch new_real = fms(real(), other.real(), imag() * other.imag());
+        real_batch new_imag = fma(real(), other.imag(), imag() * other.real());
         m_real = new_real;
         m_imag = new_imag;
         return *this;