Don't crash when attempting to reduce empty tensors.

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index bf6e10a..c3edae4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h

@@ -238,7 +238,7 @@
                            device.maxCudaThreadsPerMultiProcessor() / block_size;
     const Index size = array_prod(evaluator.dimensions());
     // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0.
-    const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, (size + block_size - 1) / block_size), 1);
+    const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1);
 
     LAUNCH_CUDA_KERNEL(
         (EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, Index>),

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
index 6af2d45..cd04716 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h

@@ -24,9 +24,17 @@
   return second;
 }
 
-template <typename T> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+
+template <typename T, typename X, typename Y>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T divup(const X x, const Y y) {
+  return static_cast<T>((x + y - 1) / y);
+}
+
+template <typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 T divup(const T x, const T y) {
-  return (x + y - 1) / y;
+  return static_cast<T>((x + y - 1) / y);
 }
 
 template <size_t n> struct max_n_1 {

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index afa5a25..fd2587d 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h

@@ -134,9 +134,14 @@
     typedef typename Self::Index Index;
 
     const Index num_coeffs = array_prod(self.m_impl.dimensions());
+    // Don't crash when we're called with an input tensor of size 0.
+    if (num_coeffs == 0) {
+      return;
+    }
+
     const int block_size = 256;
     const int num_per_thread = 128;
-    const int num_blocks = numext::ceil(static_cast<float>(num_coeffs) / (block_size * num_per_thread));
+    const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread);
 
     if (num_blocks > 1) {
       // We initialize the outputs outside the reduction kernel when we can't be sure that there