Add tiled evaluation for TensorForcedEvalOp

1 files changed, 20 insertions, 5 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index 78068be35..74b905329 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -90,14 +90,21 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
 
   enum {
-    IsAligned = true,
-    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
-    BlockAccess = false,
+    IsAligned         = true,
+    PacketAccess      = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess       = internal::is_arithmetic<CoeffReturnType>::value,
     PreferBlockAccess = false,
-    Layout = TensorEvaluator<ArgType, Device>::Layout,
-    RawAccess = true
+    Layout            = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess         = true
   };
 
+  typedef typename internal::TensorBlock<
+      CoeffReturnType, Index, internal::traits<ArgType>::NumDimensions, Layout>
+      TensorBlock;
+  typedef typename internal::TensorBlockReader<
+      CoeffReturnType, Index, internal::traits<ArgType>::NumDimensions, Layout>
+      TensorBlockReader;
+
   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device)
   /// op_ is used for sycl
       : m_impl(op.expression(), device), m_op(op.expression()), m_device(device), m_buffer(NULL)
@@ -139,6 +146,14 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
     return internal::ploadt<PacketReturnType, LoadMode>(m_buffer + index);
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void getResourceRequirements(
+      std::vector<internal::TensorOpResourceRequirements>*) const {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(TensorBlock* block) const {
+    assert(m_buffer != NULL);
+    TensorBlockReader::Run(block, m_buffer);
+  }
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
   }