From 09e75d408bc98a6d6830ac458472019a94655e44 Mon Sep 17 00:00:00 2001
From: "David G. Andersen" <dga@google.com>
Date: Fri, 27 May 2016 11:44:28 -0800
Subject: [PATCH] Very Large Tensors:  Force 64 bit indexing on GPU if total
 size of tensor is too large for 32 bit.  This change is barely relevant - the
 only way to bump into it today is using half-precision floats, and even
 that's a stretch - but might as well anticipate the future. Change: 123445306

---
 tensorflow/core/kernels/slice_op.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/tensorflow/core/kernels/slice_op.h b/tensorflow/core/kernels/slice_op.h
index a906724be95..18047396256 100644
--- a/tensorflow/core/kernels/slice_op.h
+++ b/tensorflow/core/kernels/slice_op.h
@@ -30,7 +30,9 @@ struct Slice {
                   typename TTypes<T, NDIMS>::ConstTensor input,
                   const Eigen::DSizes<Eigen::DenseIndex, NDIMS>& slice_indices,
                   const Eigen::DSizes<Eigen::DenseIndex, NDIMS>& slice_sizes) {
-    if (Eigen::internal::is_same<Device, Eigen::GpuDevice>::value) {
+    bool use_64bit = (input.size() > Eigen::NumTraits<int>::highest());
+    if (!use_64bit &&
+        Eigen::internal::is_same<Device, Eigen::GpuDevice>::value) {
       Eigen::DSizes<int, NDIMS> indices;
       for (int i = 0; i < NDIMS; ++i) {
         indices[i] = slice_indices[i];