Allow (safe) in-place computation in TensorFlow C++ ops. When at least one input tensor has the same size and type as the output, and the underlying buffer is owned by the op, i.e. when its refcount is 1 at the time the op's Compute method executes, the computation can be performed in place and allocation of the output buffer avoided.

I updated the following ops to perform in-place computation automatically when possible:
   * All standard coefficient-wise unary and binary operators (including with broadcasting) inheriting from base classes in kernels/cwise_ops_common.h.
   * unary and binary operators inheriting from base classes in framework/numeric_op.h. This is mostly old code for the Relu family and associated gradients.
   * All linear algebra ops inheriting from linalg_common.
   * Misc individual files/ops: softmax, select, bias, aggregate ops, batch_norm & fused_batch_norm, adjust_hue, constant, depthwise_conv_grad, fractional_avg_pool, misc. pooling ops, matrix_set_diag, xent & sparse_xent, unique_op.
Change: 148166936

1 files changed, 9 insertions, 0 deletions

diff --git a/tensorflow/python/kernel_tests/slice_op_test.py b/tensorflow/python/kernel_tests/slice_op_test.py
index 29f76a2182..c11f78b77e 100644
--- a/tensorflow/python/kernel_tests/slice_op_test.py
+++ b/tensorflow/python/kernel_tests/slice_op_test.py
@@ -269,6 +269,15 @@ class SliceTest(test.TestCase):
     c = array_ops.slice(a, [begin, 0], [-1, 2])
     self.assertEqual([None, 2], c.get_shape().as_list())
 
+  def testSliceOfSlice(self):
+    with self.test_session(use_gpu=True):
+      a = constant_op.constant([[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]])
+      b = a[1:, :]
+      c = b[:-1, :]
+      d = c[1, :]
+      res = 2 * d - c[1, :] + a[2, :] - 2 * b[-2, :]
+      self.assertAllEqual([0, 0, 0], res.eval())
+
 
 if __name__ == "__main__":
   test.main()