1 files changed, 17 insertions, 7 deletions
diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index 195d40fb4..636230944 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -183,12 +183,22 @@ template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d&
 }
 #endif
 
-template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_min_pd(a,b); }
-
-template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_max_ps(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_max_pd(a,b); }
-
+template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::min.
+  return _mm256_min_ps(b,a);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::min.
+  return _mm256_min_pd(b,a);
+}
+template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::max.
+  return _mm256_max_ps(b,a);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) {
+  // Arguments are swapped to match NaN propagation behavior of std::max.
+  return _mm256_max_pd(b,a);
+}
 template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); }
 template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); }
 
@@ -225,7 +235,7 @@ template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from)
 //   Packet8f tmp  = _mm256_castps128_ps256(_mm_loadu_ps(from));
 //   tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1);
 //   return _mm256_unpacklo_ps(tmp,tmp);
-  
+
   // _mm256_insertf128_ps is very slow on Haswell, thus:
   Packet8f tmp = _mm256_broadcast_ps((const __m128*)(const void*)from);
   // mimic an "inplace" permutation of the lower 128bits using a blend