TFE_DeleteOp(matmul);
  TFE_DeleteTensorHandle(m);
  TF_Buffer* b = TF_NewBuffer();
  TFE_ContextExportRunMetadata(ctx, b, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  tensorflow::RunMetadata rm;
  EXPECT_TRUE(
      rm.ParseFromString({reinterpret_cast<const char*>(b->data), b->length}));
  TF_DeleteBuffer(b);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  ASSERT_EQ(1, num_retvals);

*   Shape strictness: all fed values must have a shape that is compatible with
    the tensor they are replacing
*   Exposed `tf.while_loop` (deprecated `control_flow_ops.While`)
*   run() now takes RunOptions and RunMetadata, which enable timing stats
*   Fixed lots of potential overflow problems in op kernels
*   Various performance improvements, especially for RNNs and convolutions
*   Many bugfixes