Merged in codeplaysoftware/eigen-upstream-pure/Pointer_type_creation (pull request PR-461)

Creating a pointer type in TensorCustomOp.h

2 files changed, 22 insertions, 13 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
index 47b5a5a5e..cbec5e9b4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h
@@ -88,6 +88,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename PointerType<CoeffReturnType, Device>::Type PointerT;
 
   enum {
     IsAligned = false,
@@ -106,12 +107,12 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(PointerT data) {
     if (data) {
       evalTo(data);
       return false;
     } else {
-      m_result = static_cast<CoeffReturnType*>(
+      m_result = static_cast<PointerT>(
           m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
       evalTo(m_result);
       return true;
@@ -139,23 +140,22 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_result; }
+  EIGEN_DEVICE_FUNC PointerT data() const { return m_result; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
 #endif
 
  protected:
-  EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
-    TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(
-        data, m_dimensions);
+  EIGEN_DEVICE_FUNC void evalTo(PointerT data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(data, m_dimensions);
     m_op.func().eval(m_op.expression(), result, m_device);
   }
 
   Dimensions m_dimensions;
   const ArgType m_op;
   const Device& m_device;
-  CoeffReturnType* m_result;
+  PointerT m_result;
 };
 
 
@@ -250,6 +250,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
   typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename PointerType<CoeffReturnType, Device>::Type PointerT;
 
   enum {
     IsAligned = false,
@@ -268,12 +269,12 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(PointerT data) {
     if (data) {
       evalTo(data);
       return false;
     } else {
-      m_result = static_cast<Scalar *>(m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
+      m_result = static_cast<PointerT>(m_device.allocate_temp(dimensions().TotalSize() * sizeof(CoeffReturnType)));
       evalTo(m_result);
       return true;
     }
@@ -300,22 +301,22 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC typename internal::traits<XprType>::PointerType data() const { return m_result; }
+  EIGEN_DEVICE_FUNC PointerT data() const { return m_result; }
 
 #ifdef EIGEN_USE_SYCL
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
 #endif
 
  protected:
-  EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
-    TensorMap<Tensor<Scalar, NumDims, Layout> > result(data, m_dimensions);
+  EIGEN_DEVICE_FUNC void evalTo(PointerT data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout> > result(data, m_dimensions);
     m_op.func().eval(m_op.lhsExpression(), m_op.rhsExpression(), result, m_device);
   }
 
   Dimensions m_dimensions;
   const XprType m_op;
   const Device& m_device;
-  CoeffReturnType* m_result;
+  PointerT m_result;
 };
 
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 0dd524a30..93a3b0e14 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -24,6 +24,14 @@ template<typename T> struct MakePointer {
   typedef T ScalarType;
 };
 
+// The PointerType class is a container of the device specefic pointer
+// used for refering to a Pointer on TensorEvaluator class. While the TensorExpression
+// is a device-agnostic type and need MakePointer class for type conversion,
+// the TensorEvaluator calss can be specialized for a device, hence it is possible
+// to construct different types of temproray storage memory in TensorEvaluator
+// for different devices by specializing the following PointerType class.
+template<typename T, typename Device> struct PointerType : MakePointer<T>{};
+
 namespace internal{
 template<typename A, typename B> struct Pointer_type_promotion {
   static const bool val=false;