Benchmarks with backprop enabled (and removes overhead).

Before:
np.array([[3]])                          took 1.50us (30000 iterations)
Tensor([[3]])                            took 16.30us (30000 iterations)
MatMul [2, 2]: np.dot                         took 0.61us (30000 iterations)
MatMul [2, 2]: tf.matmul                      took 60.53us (30000 iterations)
MatMul [2, 2]: gen_math_ops.mat_mul           took 25.72us (30000 iterations)
MatMul [2, 2]: TFE_Py_Execute                 took 2.82us (30000 iterations)
MatMul [2, 2]: defun(tf.matmul)               took 45.70us (30000 iterations)
MatMul [100, 784]: np.dot                         took 383.32us (1000 iterations)
MatMul [100, 784]: tf.matmul                      took 350.35us (1000 iterations)
MatMul [100, 784]: gen_math_ops.mat_mul           took 315.97us (1000 iterations)
MatMul [100, 784]: TFE_Py_Execute                 took 249.42us (1000 iterations)
MatMul [100, 784]: defun(tf.matmul)               took 280.95us (1000 iterations)

If backprop is enabled:
np.array([[3]])                          took 0.83us (30000 iterations)
Tensor([[3]])                            took 15.21us (30000 iterations)
MatMul [2, 2]: np.dot                         took 0.63us (30000 iterations)
MatMul [2, 2]: tf.matmul                      took 76.31us (30000 iterations)
MatMul [2, 2]: gen_math_ops.mat_mul           took 38.66us (30000 iterations)
MatMul [2, 2]: TFE_Py_Execute                 took 2.31us (30000 iterations)
MatMul [2, 2]: defun(tf.matmul)               took 51.96us (30000 iterations)
MatMul [100, 784]: np.dot                         took 378.34us (1000 iterations)
MatMul [100, 784]: tf.matmul                      took 352.09us (1000 iterations)
MatMul [100, 784]: gen_math_ops.mat_mul           took 364.28us (1000 iterations)
MatMul [100, 784]: TFE_Py_Execute                 took 350.68us (1000 iterations)
MatMul [100, 784]: defun(tf.matmul)               took 377.19us (1000 iterations)

After:
np.array([[3]])                          took 0.86us (30000 iterations)
Tensor([[3]])                            took 15.19us (30000 iterations)
MatMul [2, 2]: np.dot                         took 0.60us (30000 iterations)
MatMul [2, 2]: tf.matmul                      took 64.51us (30000 iterations)
MatMul [2, 2]: gen_math_ops.mat_mul           took 28.34us (30000 iterations)
MatMul [2, 2]: TFE_Py_Execute                 took 2.38us (30000 iterations)
MatMul [2, 2]: defun(tf.matmul)               took 48.50us (30000 iterations)
MatMul [100, 784]: np.dot                         took 475.27us (1000 iterations)
MatMul [100, 784]: tf.matmul                      took 399.50us (1000 iterations)
MatMul [100, 784]: gen_math_ops.mat_mul           took 307.80us (1000 iterations)
MatMul [100, 784]: TFE_Py_Execute                 took 272.83us (1000 iterations)
MatMul [100, 784]: defun(tf.matmul)               took 350.06us (1000 iterations)
PiperOrigin-RevId: 165765641

tensorflow/python/eager/BUILD

@@ -307,6 +307,7 @@ py_test(
     srcs = ["benchmarks_test.py"],
     srcs_version = "PY2AND3",
     deps = [
+        ":backprop",
         ":context",
         ":function",
         ":tensor",

tensorflow/python/eager/backprop.py

@@ -134,6 +134,8 @@ def _record_gradient(op_name, inputs, attrs, results, name):
   Raises:
     An exception on error.
   """
+  if not any(ag_core.isnode(x) for x in inputs):
+    return results
   num_outputs = len(results)
   if num_outputs == 0:
     return results

tensorflow/python/eager/benchmarks_test.py

@@ -33,6 +33,7 @@
 from six.moves import xrange  # pylint: disable=redefined-builtin
 
 from tensorflow.python import pywrap_tensorflow
+from tensorflow.python.eager import backprop  # pylint: disable=unused-import
 from tensorflow.python.eager import context
 from tensorflow.python.eager import function
 from tensorflow.python.eager import tensor