Add async evaluation support to TensorSlicingOp.

Device::memcpy is not async-safe and might lead to deadlocks. Always evaluate slice expression in async mode.

1 files changed, 8 insertions, 0 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 879a67ea4..d04dcae17 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -559,6 +559,14 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
     }
     return true;
   }
+  
+#ifdef EIGEN_USE_THREADS
+  template <typename EvalSubExprsCallback>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
+      EvaluatorPointerType data, EvalSubExprsCallback done) {
+    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
+  }
+#endif  // EIGEN_USE_THREADS  
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
     m_impl.cleanup();