From 95f4837fd8830f20cb193fdc7950d353c6a63984 Mon Sep 17 00:00:00 2001
From: Andreas Krebbel <krebbel@linux.ibm.com>
Date: Fri, 24 Apr 2026 10:49:22 +0200
Subject: [PATCH] Fix VSX rounding

---
 include/xsimd/arch/xsimd_vsx.hpp | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/include/xsimd/arch/xsimd_vsx.hpp b/include/xsimd/arch/xsimd_vsx.hpp
index bf4c08308..122b916c7 100644
--- a/include/xsimd/arch/xsimd_vsx.hpp
+++ b/include/xsimd/arch/xsimd_vsx.hpp
@@ -605,8 +605,18 @@ namespace xsimd
         }
 
         // round
-        template <class A, class T, class = std::enable_if_t<std::is_floating_point<T>::value>>
-        XSIMD_INLINE batch<T, A> round(batch<T, A> const& self, requires_arch<vsx>) noexcept
+
+        // vec_round exists also for float vectors but is mapped to vrfin instruction which uses the wrong rounding mode
+#if defined __has_builtin && __has_builtin(__builtin_vsx_xvrspi)
+        template <class A>
+        XSIMD_INLINE batch<float, A> round(batch<float, A> const& self, requires_arch<vsx>) noexcept
+        {
+            return __builtin_vsx_xvrspi(self.data);
+        }
+#endif
+        // For double vectors vec_round uses xvrdpi which does the right thing
+        template <class A>
+        XSIMD_INLINE batch<double, A> round(batch<double, A> const& self, requires_arch<vsx>) noexcept
         {
             return vec_round(self.data);
         }