3 files changed, 7 insertions, 5 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 261dbffa8..82c9f0ae3 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -14,7 +14,7 @@
 // clang is incompatible with the CUDA syntax wrt making a kernel a class friend,
 // so we'll use a macro to make clang happy.
 #ifndef KERNEL_FRIEND
-#if defined(__clang__) && defined(__CUDA__)
+#if defined(__clang__) && (defined(__CUDA__) || defined(__HIP__))
 #define KERNEL_FRIEND friend __global__
 #else
 #define KERNEL_FRIEND friend
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
index 22c952ae1..2cac2d0f1 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/EventCount.h
@@ -169,9 +169,7 @@ class EventCount {
 
   class Waiter {
     friend class EventCount;
-    // Align to 128 byte boundary to prevent false sharing with other Waiter
-    // objects in the same vector.
-    EIGEN_ALIGN_TO_BOUNDARY(128) std::atomic<Waiter*> next;
+    std::atomic<Waiter*> next;
     std::mutex mu;
     std::condition_variable cv;
     uint64_t epoch;
@@ -181,6 +179,9 @@ class EventCount {
       kWaiting,
       kSignaled,
     };
+    // Pad past 128 byte boundary to prevent false sharing with other Waiter
+    // objects in the same vector.
+    char pad_[128];
   };
 
  private:
diff --git a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
index a41731c34..7229839ac 100644
--- a/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
+++ b/unsupported/Eigen/CXX11/src/ThreadPool/ThreadLocal.h
@@ -12,7 +12,8 @@
 
 #if EIGEN_MAX_CPP_VER >= 11 &&                         \
     ((EIGEN_COMP_GNUC && EIGEN_GNUC_AT_LEAST(4, 8)) || \
-     __has_feature(cxx_thread_local))
+     __has_feature(cxx_thread_local)                || \
+     (EIGEN_COMP_MSVC >= 1900) )
 #define EIGEN_THREAD_LOCAL static thread_local
 #endif