Update Eigen to commit 1724f8760da8.

(1) Revert gebp_neon.patch.oss.

PiperOrigin-RevId: 232725552

tensorflow/opensource_only.files

@@ -120,8 +120,8 @@ tensorflow/third_party/eigen3/unsupported/Eigen/CXX11/src/FixedPoint/MatMatProdu
 tensorflow/third_party/eigen3/unsupported/Eigen/CXX11/ThreadPool
 tensorflow/third_party/eigen3/unsupported/Eigen/SpecialFunctions
 tensorflow/third_party/eigen3/unsupported/Eigen/MatrixFunctions
+tensorflow/third_party/eigen3/gpu_packet_math.patch
 tensorflow/third_party/eigen3/LICENSE
-tensorflow/third_party/eigen3/gebp_neon.patch
 tensorflow/third_party/eigen3/BUILD
 tensorflow/third_party/systemlibs/build_defs.bzl.tpl
 tensorflow/third_party/systemlibs/absl_py.BUILD

tensorflow/workspace.bzl

@@ -140,12 +140,12 @@ def tf_workspace(path_prefix = "", tf_repo_name = ""):
     tf_http_archive(
         name = "eigen_archive",
         build_file = clean_dep("//third_party:eigen.BUILD"),
-        patch_file = clean_dep("//third_party/eigen3:gebp_neon.patch"),
+        patch_file = clean_dep("//third_party/eigen3:gpu_packet_math.patch"),
-        sha256 = "48678550a32665331d729be87076e576f2502fff325f5b6c2c78ebf7b1b22c7b",
+        sha256 = "d1d2ac19b8ef386ad70b91932a90bfbc3014b801d14723b9c8373239128bd2dd",
-        strip_prefix = "eigen-eigen-bcc817c0ba98",
+        strip_prefix = "eigen-eigen-1724f8760da8",
         urls = [
-            "https://mirror.bazel.build/bitbucket.org/eigen/eigen/get/bcc817c0ba98.tar.gz",
+            "https://mirror.bazel.build/bitbucket.org/eigen/eigen/get/1724f8760da8.tar.gz",
-            "https://bitbucket.org/eigen/eigen/get/bcc817c0ba98.tar.gz",
+            "https://bitbucket.org/eigen/eigen/get/1724f8760da8.tar.gz",
         ],
     )
 

third_party/eigen3/gebp_neon.patch

@@ -1,11 +0,0 @@
---- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h	2019-01-22 20:46:51.000000000 -0800
-+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h	2019-01-25 13:48:49.000000000 -0800
-@@ -1031,7 +1031,7 @@
- 
-   EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const
-   {
--    c += a * b;
-+    c = vfmaq_n_f32(c, a, b);
-   }
- 
-   EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const

third_party/eigen3/gpu_packet_math.patch

@@ -0,0 +1,18 @@
+--- a/Eigen/src/Core/arch/GPU/PacketMath.h
++++ b/Eigen/src/Core/arch/GPU/PacketMath.h
+@@ -100,6 +100,7 @@
+   return make_double2(from, from);
+ }
+ 
++#if defined(EIGEN_CUDA_ARCH)
+ namespace {
+ 
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float bitwise_and(const float& a,
+@@ -211,6 +212,7 @@
+ pcmp_eq<double2>(const double2& a, const double2& b) {
+   return make_double2(eq_mask(a.x, b.x), eq_mask(a.y, b.y));
+ }
++#endif  // EIGEN_CUDA_ARCH
+ 
+ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 plset<float4>(const float& a) {
+   return make_float4(a, a+1, a+2, a+3);

third_party/eigen3/unsupported/Eigen/CXX11/src/FixedPoint/PacketMathAVX2.h

@@ -178,25 +178,37 @@ template <>
 struct unpacket_traits<Packet32q8i> {
   typedef QInt8 type;
   typedef Packet16q8i half;
-  enum { size = 32, alignment = Aligned32 };
+  enum { size = 32, alignment = Aligned32, vectorizable = true };
+};
+template <>
+struct unpacket_traits<Packet16q8i> {
+  typedef QInt8 type;
+  typedef Packet16q8i half;
+  enum { size = 16, alignment = Aligned32, vectorizable = true };
 };
 template <>
 struct unpacket_traits<Packet16q16i> {
   typedef QInt16 type;
   typedef Packet8q16i half;
-  enum { size = 16, alignment = Aligned32 };
+  enum { size = 16, alignment = Aligned32, vectorizable = true };
+};
+template <>
+struct unpacket_traits<Packet8q16i> {
+  typedef QInt16 type;
+  typedef Packet8q16i half;
+  enum { size = 8, alignment = Aligned32, vectorizable = true };
 };
 template <>
 struct unpacket_traits<Packet32q8u> {
   typedef QUInt8 type;
   typedef Packet16q8u half;
-  enum { size = 32, alignment = Aligned32 };
+  enum { size = 32, alignment = Aligned32, vectorizable = true };
 };
 template <>
 struct unpacket_traits<Packet8q32i> {
   typedef QInt32 type;
   typedef Packet4q32i half;
-  enum { size = 8, alignment = Aligned32 };
+  enum { size = 8, alignment = Aligned32, vectorizable = true };
 };
 
 // Unaligned load
@@ -206,6 +218,11 @@ EIGEN_STRONG_INLINE Packet32q8i ploadu<Packet32q8i>(const QInt8* from) {
       reinterpret_cast<const __m256i*>(from));
 }
 template <>
+EIGEN_STRONG_INLINE Packet16q8i ploadu<Packet16q8i>(const QInt8* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(
+      reinterpret_cast<const __m128i*>(from));
+}
+template <>
 EIGEN_STRONG_INLINE Packet32q8u ploadu<Packet32q8u>(const QUInt8* from) {
   EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(
       reinterpret_cast<const __m256i*>(from));
@@ -215,6 +232,11 @@ EIGEN_STRONG_INLINE Packet16q16i ploadu<Packet16q16i>(const QInt16* from) {
   EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(
       reinterpret_cast<const __m256i*>(from));
 }
+template<>
+EIGEN_STRONG_INLINE Packet8q16i ploadu<Packet8q16i>(const QInt16* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(
+      reinterpret_cast<const __m128i*>(from));
+}
 template <>
 EIGEN_STRONG_INLINE Packet8q32i ploadu<Packet8q32i>(const QInt32* from) {
   EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(
@@ -228,6 +250,11 @@ EIGEN_STRONG_INLINE Packet32q8i pload<Packet32q8i>(const QInt8* from) {
       reinterpret_cast<const __m256i*>(from));
 }
 template <>
+EIGEN_STRONG_INLINE Packet16q8i pload<Packet16q8i>(const QInt8* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(
+      reinterpret_cast<const __m128i*>(from));
+}
+template <>
 EIGEN_STRONG_INLINE Packet32q8u pload<Packet32q8u>(const QUInt8* from) {
   EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(
       reinterpret_cast<const __m256i*>(from));
@@ -238,6 +265,11 @@ EIGEN_STRONG_INLINE Packet16q16i pload<Packet16q16i>(const QInt16* from) {
       reinterpret_cast<const __m256i*>(from));
 }
 template <>
+EIGEN_STRONG_INLINE Packet8q16i pload<Packet8q16i>(const QInt16* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(
+      reinterpret_cast<const __m128i*>(from));
+}
+template <>
 EIGEN_STRONG_INLINE Packet8q32i pload<Packet8q32i>(const QInt32* from) {
   EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(
       reinterpret_cast<const __m256i*>(from));
@@ -250,6 +282,11 @@ EIGEN_STRONG_INLINE void pstoreu<QInt8>(QInt8* to, const Packet32q8i& from) {
       reinterpret_cast<__m256i*>(to), from.val);
 }
 template <>
+EIGEN_STRONG_INLINE void pstoreu<QInt8>(QInt8* to, const Packet16q8i& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(
+      reinterpret_cast<__m128i*>(to), from.val);
+}
+template <>
 EIGEN_STRONG_INLINE void pstoreu<QUInt8>(QUInt8* to, const Packet32q8u& from) {
   EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(
       reinterpret_cast<__m256i*>(to), from.val);
@@ -260,6 +297,11 @@ EIGEN_STRONG_INLINE void pstoreu<QInt16>(QInt16* to, const Packet16q16i& from) {
       reinterpret_cast<__m256i*>(to), from.val);
 }
 template <>
+EIGEN_STRONG_INLINE void pstoreu<QInt16>(QInt16* to, const Packet8q16i& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(
+      reinterpret_cast<__m128i*>(to), from.val);
+}
+template <>
 EIGEN_STRONG_INLINE void pstoreu<QInt32>(QInt32* to, const Packet8q32i& from) {
   EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(
       reinterpret_cast<__m256i*>(to), from.val);
@@ -277,6 +319,11 @@ EIGEN_STRONG_INLINE void pstore<QInt16>(QInt16* to, const Packet16q16i& from) {
                                                from.val);
 }
 template <>
+EIGEN_STRONG_INLINE void pstore<QInt16>(QInt16* to, const Packet8q16i& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to),
+                                            from.val);
+}
+template <>
 EIGEN_STRONG_INLINE void pstore<QUInt8>(QUInt8* to, const Packet32q8u& from) {
   EIGEN_DEBUG_ALIGNED_STORE _mm256_store_si256(reinterpret_cast<__m256i*>(to),
                                                from.val);
@@ -286,6 +333,11 @@ EIGEN_STRONG_INLINE void pstore<QInt8>(QInt8* to, const Packet32q8i& from) {
   EIGEN_DEBUG_ALIGNED_STORE _mm256_store_si256(reinterpret_cast<__m256i*>(to),
                                                from.val);
 }
+template <>
+EIGEN_STRONG_INLINE void pstore<QInt8>(QInt8* to, const Packet16q8i& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to),
+                                            from.val);
+}
 
 // Extract first element.
 template <>