From 7aea350ba1177db2ec7ae3bbb615450b0f74279f Mon Sep 17 00:00:00 2001
From: Chip Kerchner <ckerchner@tenstorrent.com>
Date: Fri, 9 Jan 2026 12:16:28 -0500
Subject: [PATCH] Fix more packetmath issues for RVV

libeigen/eigen!2105
---
 Eigen/src/Core/arch/RVV10/PacketMath.h     | 19 +++++++---
 Eigen/src/Core/arch/RVV10/PacketMathBF16.h | 40 ++++++++++++++--------
 Eigen/src/Core/arch/RVV10/PacketMathFP16.h | 38 ++++++++++++--------
 3 files changed, 64 insertions(+), 33 deletions(-)

diff --git a/Eigen/src/Core/arch/RVV10/PacketMath.h b/Eigen/src/Core/arch/RVV10/PacketMath.h
index 909fd88ce..679d5c1fc 100644
--- a/Eigen/src/Core/arch/RVV10/PacketMath.h
+++ b/Eigen/src/Core/arch/RVV10/PacketMath.h
@@ -1665,11 +1665,20 @@ EIGEN_STRONG_INLINE Packet1Xd plset<Packet1Xd>(const double& a) {
 template <>
 EIGEN_STRONG_INLINE void pbroadcast4<Packet1Xd>(const double* a, Packet1Xd& a0, Packet1Xd& a1, Packet1Xd& a2,
                                                 Packet1Xd& a3) {
-  vfloat64m1_t aa = __riscv_vle64_v_f64m1(a, 4);
-  a0 = __riscv_vrgather_vx_f64m1(aa, 0, unpacket_traits<Packet1Xd>::size);
-  a1 = __riscv_vrgather_vx_f64m1(aa, 1, unpacket_traits<Packet1Xd>::size);
-  a2 = __riscv_vrgather_vx_f64m1(aa, 2, unpacket_traits<Packet1Xd>::size);
-  a3 = __riscv_vrgather_vx_f64m1(aa, 3, unpacket_traits<Packet1Xd>::size);
+  if (EIGEN_RISCV64_RVV_VL >= 256) {
+    vfloat64m1_t aa = __riscv_vle64_v_f64m1(a, 4);
+    a0 = __riscv_vrgather_vx_f64m1(aa, 0, unpacket_traits<Packet1Xd>::size);
+    a1 = __riscv_vrgather_vx_f64m1(aa, 1, unpacket_traits<Packet1Xd>::size);
+    a2 = __riscv_vrgather_vx_f64m1(aa, 2, unpacket_traits<Packet1Xd>::size);
+    a3 = __riscv_vrgather_vx_f64m1(aa, 3, unpacket_traits<Packet1Xd>::size);
+  } else {
+    vfloat64m1_t aa0 = __riscv_vle64_v_f64m1(a + 0, 2);
+    vfloat64m1_t aa1 = __riscv_vle64_v_f64m1(a + 2, 2);
+    a0 = __riscv_vrgather_vx_f64m1(aa0, 0, unpacket_traits<Packet1Xd>::size);
+    a1 = __riscv_vrgather_vx_f64m1(aa0, 1, unpacket_traits<Packet1Xd>::size);
+    a2 = __riscv_vrgather_vx_f64m1(aa1, 0, unpacket_traits<Packet1Xd>::size);
+    a3 = __riscv_vrgather_vx_f64m1(aa1, 1, unpacket_traits<Packet1Xd>::size);
+  }
 }
 
 template <>
diff --git a/Eigen/src/Core/arch/RVV10/PacketMathBF16.h b/Eigen/src/Core/arch/RVV10/PacketMathBF16.h
index ec0e42b2a..2522efd99 100644
--- a/Eigen/src/Core/arch/RVV10/PacketMathBF16.h
+++ b/Eigen/src/Core/arch/RVV10/PacketMathBF16.h
@@ -164,11 +164,6 @@ EIGEN_STRONG_INLINE Packet1Xbf pabs(const Packet1Xbf& a) {
                                                                    unpacket_traits<Packet1Xs>::size));
 }
 
-template <>
-EIGEN_STRONG_INLINE Packet1Xbf pabsdiff(const Packet1Xbf& a, const Packet1Xbf& b) {
-  return F32ToBf16(pabsdiff<Packet2Xf>(Bf16ToF32(a), Bf16ToF32(b)));
-}
-
 template <>
 EIGEN_STRONG_INLINE Packet1Xbf pset1<Packet1Xbf>(const bfloat16& from) {
   return __riscv_vreinterpret_bf16m1(
@@ -197,12 +192,23 @@ EIGEN_STRONG_INLINE void pbroadcast4<Packet1Xbf>(const bfloat16* a, Packet1Xbf&
 
 template <>
 EIGEN_STRONG_INLINE Packet1Xbf padd<Packet1Xbf>(const Packet1Xbf& a, const Packet1Xbf& b) {
-  return F32ToBf16(padd<Packet2Xf>(Bf16ToF32(a), Bf16ToF32(b)));
+  // b + (1 * a)
+  return F32ToBf16(__riscv_vfwmaccbf16_vf_f32m2(Bf16ToF32(b),
+                                                numext::bit_cast<__bf16>(static_cast<numext::int16_t>(0x3f80u)), a,
+                                                unpacket_traits<Packet1Xbf>::size));
 }
 
 template <>
 EIGEN_STRONG_INLINE Packet1Xbf psub<Packet1Xbf>(const Packet1Xbf& a, const Packet1Xbf& b) {
-  return F32ToBf16(psub<Packet2Xf>(Bf16ToF32(a), Bf16ToF32(b)));
+  // a + (-1 * b)
+  return F32ToBf16(__riscv_vfwmaccbf16_vf_f32m2(Bf16ToF32(a),
+                                                numext::bit_cast<__bf16>(static_cast<numext::int16_t>(0xbf80u)), b,
+                                                unpacket_traits<Packet1Xbf>::size));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1Xbf pabsdiff(const Packet1Xbf& a, const Packet1Xbf& b) {
+  return pabs<Packet1Xbf>(psub<Packet1Xbf>(a, b));
 }
 
 template <>
@@ -493,11 +499,6 @@ EIGEN_STRONG_INLINE Packet2Xbf pabs(const Packet2Xbf& a) {
                                                                    unpacket_traits<Packet2Xs>::size));
 }
 
-template <>
-EIGEN_STRONG_INLINE Packet2Xbf pabsdiff(const Packet2Xbf& a, const Packet2Xbf& b) {
-  return F32ToBf16(pabsdiff<Packet4Xf>(Bf16ToF32(a), Bf16ToF32(b)));
-}
-
 template <>
 EIGEN_STRONG_INLINE Packet2Xbf pset1<Packet2Xbf>(const bfloat16& from) {
   return __riscv_vreinterpret_bf16m2(
@@ -526,12 +527,23 @@ EIGEN_STRONG_INLINE void pbroadcast4<Packet2Xbf>(const bfloat16* a, Packet2Xbf&
 
 template <>
 EIGEN_STRONG_INLINE Packet2Xbf padd<Packet2Xbf>(const Packet2Xbf& a, const Packet2Xbf& b) {
-  return F32ToBf16(padd<Packet4Xf>(Bf16ToF32(a), Bf16ToF32(b)));
+  // b + (1 * a)
+  return F32ToBf16(__riscv_vfwmaccbf16_vf_f32m4(Bf16ToF32(b),
+                                                numext::bit_cast<__bf16>(static_cast<numext::int16_t>(0x3f80u)), a,
+                                                unpacket_traits<Packet2Xbf>::size));
 }
 
 template <>
 EIGEN_STRONG_INLINE Packet2Xbf psub<Packet2Xbf>(const Packet2Xbf& a, const Packet2Xbf& b) {
-  return F32ToBf16(psub<Packet4Xf>(Bf16ToF32(a), Bf16ToF32(b)));
+  // a + (-1 * b)
+  return F32ToBf16(__riscv_vfwmaccbf16_vf_f32m4(Bf16ToF32(a),
+                                                numext::bit_cast<__bf16>(static_cast<numext::int16_t>(0xbf80u)), b,
+                                                unpacket_traits<Packet2Xbf>::size));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2Xbf pabsdiff(const Packet2Xbf& a, const Packet2Xbf& b) {
+  return pabs<Packet2Xbf>(psub<Packet2Xbf>(a, b));
 }
 
 template <>
diff --git a/Eigen/src/Core/arch/RVV10/PacketMathFP16.h b/Eigen/src/Core/arch/RVV10/PacketMathFP16.h
index 517b703e3..73118a31c 100644
--- a/Eigen/src/Core/arch/RVV10/PacketMathFP16.h
+++ b/Eigen/src/Core/arch/RVV10/PacketMathFP16.h
@@ -480,17 +480,21 @@ EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet1Xh>(const Packet1Xh& a) {
 template <>
 EIGEN_STRONG_INLINE Eigen::half predux_min<Packet1Xh>(const Packet1Xh& a) {
   const Eigen::half max = (std::numeric_limits<Eigen::half>::max)();
-  return static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmin_vs_f16m1_f16m1(
-      a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(max), unpacket_traits<Packet1Xh>::size),
-      unpacket_traits<Packet1Xh>::size)));
+  const Eigen::half nan = (std::numeric_limits<Eigen::half>::quiet_NaN)();
+  return (std::min)(static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmin_vs_f16m1_f16m1(
+                        a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(nan), unpacket_traits<Packet1Xh>::size),
+                        unpacket_traits<Packet1Xh>::size))),
+                    max);
 }
 
 template <>
 EIGEN_STRONG_INLINE Eigen::half predux_max<Packet1Xh>(const Packet1Xh& a) {
-  const Eigen::half min = (std::numeric_limits<Eigen::half>::min)();
-  return static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmax_vs_f16m1_f16m1(
-      a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(min), unpacket_traits<Packet1Xh>::size),
-      unpacket_traits<Packet1Xh>::size)));
+  const Eigen::half min = -(std::numeric_limits<Eigen::half>::max)();
+  const Eigen::half nan = (std::numeric_limits<Eigen::half>::quiet_NaN)();
+  return (std::max)(static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmax_vs_f16m1_f16m1(
+                        a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(nan), unpacket_traits<Packet1Xh>::size),
+                        unpacket_traits<Packet1Xh>::size))),
+                    min);
 }
 
 template <int N>
@@ -837,17 +841,23 @@ EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet2Xh>(const Packet2Xh& a) {
 template <>
 EIGEN_STRONG_INLINE Eigen::half predux_min<Packet2Xh>(const Packet2Xh& a) {
   const Eigen::half max = (std::numeric_limits<Eigen::half>::max)();
-  return static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmin_vs_f16m2_f16m1(
-      a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(max), unpacket_traits<Packet2Xh>::size / 2),
-      unpacket_traits<Packet2Xh>::size)));
+  const Eigen::half nan = (std::numeric_limits<Eigen::half>::quiet_NaN)();
+  return (std::min)(
+      static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmin_vs_f16m2_f16m1(
+          a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(nan), unpacket_traits<Packet2Xh>::size / 2),
+          unpacket_traits<Packet2Xh>::size))),
+      max);
 }
 
 template <>
 EIGEN_STRONG_INLINE Eigen::half predux_max<Packet2Xh>(const Packet2Xh& a) {
-  const Eigen::half min = (std::numeric_limits<Eigen::half>::min)();
-  return static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmax_vs_f16m2_f16m1(
-      a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(min), unpacket_traits<Packet2Xh>::size / 2),
-      unpacket_traits<Packet2Xh>::size)));
+  const Eigen::half min = -(std::numeric_limits<Eigen::half>::max)();
+  const Eigen::half nan = (std::numeric_limits<Eigen::half>::quiet_NaN)();
+  return (std::max)(
+      static_cast<Eigen::half>(__riscv_vfmv_f(__riscv_vfredmax_vs_f16m2_f16m1(
+          a, __riscv_vfmv_v_f_f16m1(numext::bit_cast<_Float16>(nan), unpacket_traits<Packet2Xh>::size / 2),
+          unpacket_traits<Packet2Xh>::size))),
+      min);
 }
 
 template <int N>