From 802568322625be9ab162c3529d525b960ed96a0c Mon Sep 17 00:00:00 2001
From: Alexander Grund <github@grundis.de>
Date: Sat, 3 Aug 2024 03:49:51 +0000
Subject: [PATCH] Fix conversion of `Eigen::half` to `_Float16` in AVX512 code

---
 Eigen/src/Core/arch/AVX512/PacketMathFP16.h | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/arch/AVX512/PacketMathFP16.h b/Eigen/src/Core/arch/AVX512/PacketMathFP16.h
index d4a5816ab..93ae95beb 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMathFP16.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMathFP16.h
@@ -109,7 +109,13 @@ struct unpacket_traits<Packet8h> {
 
 template <>
 EIGEN_STRONG_INLINE Packet32h pset1<Packet32h>(const Eigen::half& from) {
-  return _mm512_set1_ph(static_cast<_Float16>(from));
+  // half/half_raw is bit compatible
+  return _mm512_set1_ph(numext::bit_cast<_Float16>(from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet32h pzero(const Packet32h& /*a*/) {
+  return _mm512_setzero_ph();
 }
 
 // pset1frombits
@@ -209,10 +215,8 @@ EIGEN_STRONG_INLINE Packet32h pmax<Packet32h>(const Packet32h& a, const Packet32
 // plset
 template <>
 EIGEN_STRONG_INLINE Packet32h plset<Packet32h>(const half& a) {
-  return _mm512_add_ph(_mm512_set1_ph(a),
-                       _mm512_set_ph(31.0f, 30.0f, 29.0f, 28.0f, 27.0f, 26.0f, 25.0f, 24.0f, 23.0f, 22.0f, 21.0f, 20.0f,
-                                     19.0f, 18.0f, 17.0f, 16.0f, 15.0f, 14.0f, 13.0f, 12.0f, 11.0f, 10.0f, 9.0f, 8.0f,
-                                     7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f, 0.0f));
+  return _mm512_add_ph(pset1<Packet32h>(a), _mm512_set_ph(31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17,
+                                                          16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0));
 }
 
 // por
@@ -510,7 +514,7 @@ EIGEN_STRONG_INLINE Packet8h pnmsub(const Packet8h& a, const Packet8h& b, const
 
 template <>
 EIGEN_STRONG_INLINE Packet32h pnegate<Packet32h>(const Packet32h& a) {
-  return _mm512_sub_ph(_mm512_set1_ph(0.0), a);
+  return psub(pzero(a), a);
 }
 
 // pconj