Rollback or PR-746 and partial rollback of https://bitbucket.org/eigen/eigen/commits/668ab3fc474e54c7919eda4fbaf11f3a99246494

.

std::array is still not supported in CUDA device code on Windows.

2 files changed, 28 insertions, 3 deletions

diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h
index e8a7f65ed..a8bb8a624 100644
--- a/Eigen/src/Core/DenseStorage.h
+++ b/Eigen/src/Core/DenseStorage.h
@@ -369,7 +369,7 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
       , m_cols(other.m_cols)
     {
       EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*m_cols)
-      std::copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data);
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data);
     }
     EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
@@ -452,7 +452,7 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
       , m_cols(other.m_cols)
     {
       EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_cols*_Rows)
-      std::copy(other.m_data, other.m_data+_Rows*m_cols, m_data);
+      internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data);
     }
     EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
@@ -528,7 +528,7 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
       , m_rows(other.m_rows)
     {
       EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*_Cols)
-      std::copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data);
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data);
     }
     EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h
index d0622d0b4..1b12544d2 100644
--- a/Eigen/src/Core/util/Memory.h
+++ b/Eigen/src/Core/util/Memory.h
@@ -507,6 +507,31 @@ inline Index first_multiple(Index size, Index base)
   return ((size+base-1)/base)*base;
 }
 
+// std::copy is much slower than memcpy, so let's introduce a smart_copy which
+// use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
+template<typename T, bool UseMemcpy> struct smart_copy_helper;
+
+template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target)
+{
+  smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
+}
+
+template<typename T> struct smart_copy_helper<T,true> {
+  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
+  {
+    IntPtr size = IntPtr(end)-IntPtr(start);
+    if(size==0) return;
+    eigen_internal_assert(start!=0 && end!=0 && target!=0);
+    EIGEN_USING_STD(memcpy)
+    memcpy(target, start, size);
+  }
+};
+
+template<typename T> struct smart_copy_helper<T,false> {
+  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
+  { std::copy(start, end, target); }
+};
+
 // intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
 template<typename T, bool UseMemmove> struct smart_memmove_helper;