From dcceb9afec3be00dda399a622a5f6c2e520a3906 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Antonio=20S=C3=A1nchez?= <cantonios@google.com>
Date: Fri, 26 Apr 2024 15:28:03 +0000
Subject: [PATCH] Unbork avx512 preduce_mul on MSVC.

---
 Eigen/src/Core/arch/AVX512/PacketMath.h | 18 ++++++++++++++++++
 test/packetmath.cpp                     |  9 +++++++++
 2 files changed, 27 insertions(+)

diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index aa338d112..8f7662f47 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -1562,10 +1562,28 @@ template <>
 EIGEN_STRONG_INLINE int predux_mul<Packet16i>(const Packet16i& a) {
   return _mm512_reduce_mul_epi32(a);
 }
+
+#if EIGEN_COMP_MSVC
+// MSVC's _mm512_reduce_mul_epi64 is borked, at least up to and including 1939.
+//    alignas(64) int64_t data[] = { 1,1,-1,-1,1,-1,-1,-1 };
+//    int64_t out = _mm512_reduce_mul_epi64(_mm512_load_epi64(data));
+// produces garbage: 4294967295.  It seems to happen whenever the output is supposed to be negative.
+// Fall back to a manual approach:
+template <>
+EIGEN_STRONG_INLINE int64_t predux_mul<Packet8l>(const Packet8l& a) {
+  Packet4l lane0 = _mm512_extracti64x4_epi64(a, 0);
+  Packet4l lane1 = _mm512_extracti64x4_epi64(a, 1);
+  Packet4l res = pmul(lane0, lane1);
+  res = pmul(res, Packet4l(_mm256_permute2x128_si256(res, res, 1)));
+  res = pmul(res, Packet4l(_mm256_shuffle_epi32(res, 0xE)));
+  return pfirst(res);
+}
+#else
 template <>
 EIGEN_STRONG_INLINE int64_t predux_mul<Packet8l>(const Packet8l& a) {
   return _mm512_reduce_mul_epi64(a);
 }
+#endif
 
 template <>
 EIGEN_STRONG_INLINE float predux_min<Packet16f>(const Packet16f& a) {
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 8bfa32196..b9a06b869 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -623,6 +623,15 @@ void packetmath() {
     VERIFY(test::areApprox(ref, data2, HalfPacketSize) && "internal::predux_half_dowto4");
   }
 
+  // Avoid overflows.
+  if (NumTraits<Scalar>::IsInteger && NumTraits<Scalar>::IsSigned) {
+    Scalar limit =
+        static_cast<Scalar>(std::pow(static_cast<double>(numext::real(NumTraits<Scalar>::highest())),
+                                     1.0 / static_cast<double>(Eigen::internal::unpacket_traits<Packet>::size)));
+    for (int i = 0; i < PacketSize; ++i) {
+      data1[i] = internal::random<Scalar>(-limit, limit);
+    }
+  }
   ref[0] = Scalar(1);
   for (int i = 0; i < PacketSize; ++i) ref[0] = REF_MUL(ref[0], data1[i]);
   VERIFY(internal::isApprox(ref[0], internal::predux_mul(internal::pload<Packet>(data1))) && "internal::predux_mul");