Set the default value for the use_gpu-param to True in test utilities

Also, fix a bug in IsFunctionCallOp found by this CL.

Contrary to what it sounds like, `use_gpu` does not force the test to run on
GPUs, it merely *allows* the test to run on GPUs (there is a separate
`force_gpu` option for forcing).  This means setting `use_gpu` to `True` means
that the test will run on GPUs if one is available.

Given that setting `use_gpu` to `True` by default makes sense, and there should
be a good reason for a test to set it to `False` (which disallows GPU use, even
when one is available).

For this reason, this CL changes the default value of `use_gpu`.  As you can
see, this has already found a few real bugs.

In a later CL I will remove instances that pass use_gpu=True explicitly as those
should no longer be necessary.

PiperOrigin-RevId: 356906251
Change-Id: Ibd0f785af0d2b1290dc40e84f928ff4291a58fe7

tensorflow/core/framework/memory_types.cc

@@ -64,7 +64,8 @@ void MemoryTypesHelper(const NameRangeMap& name_map,
 
 bool IsFunctionCallOp(const string& op_type) {
   return op_type == "SymbolicGradient" || op_type == "PartitionedCall" ||
-         op_type == "StatefulPartitionedCall" || op_type == "While";
+         op_type == "StatefulPartitionedCall" || op_type == "While" ||
+         op_type == "StatelessWhile";
 }
 
 }  // namespace

tensorflow/python/framework/function_test.py

@@ -422,7 +422,9 @@ class FunctionTest(test.TestCase):
       with ops.control_dependencies([z]):
         return x * 2
 
-    with ops.Graph().as_default(), self.cached_session():
+    # @function.Defun creates a non-partitioned function.  If we place this on
+    # the GPU then the inner `Print` op cannot be run.
+    with ops.Graph().as_default(), self.cached_session(use_gpu=False):
       z = Foo(constant_op.constant(3.0))
       self.assertAllEqual(z, 6.0)
 

tensorflow/python/framework/test_util.py

@@ -2273,7 +2273,7 @@ class TensorFlowTestCase(googletest.TestCase):
 
   # pylint: disable=g-doc-return-or-yield
   @contextlib.contextmanager
-  def session(self, graph=None, config=None, use_gpu=False, force_gpu=False):
+  def session(self, graph=None, config=None, use_gpu=True, force_gpu=False):
     """A context manager for a TensorFlow Session for use in executing tests.
 
     Note that this will set this session and the graph as global defaults.
@@ -2320,7 +2320,7 @@ class TensorFlowTestCase(googletest.TestCase):
   def cached_session(self,
                      graph=None,
                      config=None,
-                     use_gpu=False,
+                     use_gpu=True,
                      force_gpu=False):
     """Returns a TensorFlow Session for use in executing tests.
 
@@ -2374,7 +2374,7 @@ class TensorFlowTestCase(googletest.TestCase):
   def test_session(self,
                    graph=None,
                    config=None,
-                   use_gpu=False,
+                   use_gpu=True,
                    force_gpu=False):
     """Use cached_session instead."""
     if self.id().endswith(".test_session"):

tensorflow/python/kernel_tests/atrous_convolution_test.py

@@ -265,6 +265,7 @@ class AtrousConvolutionTest(test.TestCase):
     self.assertLess(err, err_tolerance)
 
   @test_util.run_v1_only("b/120545219")
+  @test_util.disable_xla("b/178665095")
   def testGradient(self):
     with self.cached_session():
       for padding in ["SAME", "VALID"]:

tensorflow/python/kernel_tests/betainc_op_test.py

@@ -135,6 +135,7 @@ class BetaincTest(test.TestCase):
     self._testBetaInc(a_s, b_s, x_s, dtypes.float64)
 
   @test_util.run_deprecated_v1
+  @test_util.disable_xla("b/178338235")
   def testBetaIncDoubleVerySmallValues(self):
     a_s = np.abs(np.random.randn(10, 10) * 1e-16)  # in (0, infty)
     b_s = np.abs(np.random.randn(10, 10) * 1e-16)  # in (0, infty)
@@ -142,6 +143,7 @@ class BetaincTest(test.TestCase):
     self._testBetaInc(a_s, b_s, x_s, dtypes.float64)
 
   @test_util.run_deprecated_v1
+  @test_util.disable_xla("b/178338235")
   def testBetaIncFloatVerySmallValues(self):
     a_s = np.abs(np.random.randn(10, 10) * 1e-8)  # in (0, infty)
     b_s = np.abs(np.random.randn(10, 10) * 1e-8)  # in (0, infty)

tensorflow/python/kernel_tests/cond_v2_test.py

@@ -1130,7 +1130,10 @@ class CondV2Test(test.TestCase):
 
   @test_util.run_deprecated_v1
   def testLoweringDisabledWithSingleThreadedExecutorContext(self):
-    with self.session(graph=ops.Graph()) as sess:
+    # Single threaded executor does not support partitioned graphs, so we can't
+    # run on GPUs (running on GPU requires a mixed CPU/GPU graph).
+    with self.session(graph=ops.Graph(), use_gpu=False) as sess:
+
       @function.defun
       def _add_cond(x):
         return cond_v2.cond_v2(

tensorflow/python/kernel_tests/conv2d_backprop_filter_grad_test.py

@@ -68,8 +68,11 @@ class Conv2DBackpropFilterGradTest(test.TestCase):
               [in_val, out_backprop_val], [in_shape, out_backprop_shape],
               output, filter_shape)
           print("conv2d_backprop_filter gradient err = %g " % err)
-          err_tolerance = 2e-3
-          self.assertLess(err, err_tolerance)
+          err_tolerance = 3e-2 if test.is_gpu_available() else 2e-3
+          self.assertLess(
+              err,
+              err_tolerance,
+              msg="padding={0},stride={1},".format(str(padding), stride))
 
   @test_util.run_deprecated_v1
   def testGradientDilatedConv(self):

tensorflow/python/kernel_tests/conv3d_transpose_test.py

@@ -220,7 +220,7 @@ class Conv3DTransposeTest(test.TestCase):
       err = gradient_checker.compute_gradient_error([x, f], [x_shape, f_shape],
                                                     output, y_shape)
     print("conv3d_transpose gradient err = %g " % err)
-    err_tolerance = 0.0005
+    err_tolerance = 0.00055
     self.assertLess(err, err_tolerance)
 
 

tensorflow/python/kernel_tests/dynamic_partition_op_test.py

@@ -306,7 +306,8 @@ class DynamicPartitionTest(test.TestCase):
 
   @test_util.run_deprecated_v1
   def testScalarIndexOutOfRange(self):
-    with self.cached_session() as sess:
+    # GPU kernels don't throw exceptions.
+    with self.cached_session(use_gpu=False):
       bad = 17
       data = np.zeros(5)
       partitions = data_flow_ops.dynamic_partition(data, bad, num_partitions=7)
@@ -315,7 +316,8 @@ class DynamicPartitionTest(test.TestCase):
 
   @test_util.run_deprecated_v1
   def testHigherRankIndexOutOfRange(self):
-    with self.cached_session() as sess:
+    # GPU kernels don't throw exceptions.
+    with self.cached_session(use_gpu=False) as sess:
       shape = (2, 3)
       indices = array_ops.placeholder(shape=shape, dtype=np.int32)
       data = np.zeros(shape + (5,))

tensorflow/python/kernel_tests/matrix_triangular_solve_op_test.py

@@ -218,9 +218,10 @@ class MatrixTriangularSolveOpTest(test.TestCase):
   def testNotInvertible(self):
     # The input should be invertible.
     # The matrix is singular because it has a zero on the diagonal.
-    # FIXME(rmlarsen): The GPU kernel does not check for singularity.
     singular_matrix = np.array([[1., 0., -1.], [-1., 0., 1.], [0., -1., 1.]])
-    with self.cached_session():
+
+    # FIXME(rmlarsen): The GPU kernel does not check for singularity.
+    with self.cached_session(use_gpu=False):
       with self.assertRaisesOpError("Input matrix is not invertible."):
         self._verifySolve(singular_matrix, singular_matrix)
       with self.assertRaisesOpError("Input matrix is not invertible."):

tensorflow/python/ops/nn_test.py

@@ -1015,7 +1015,8 @@ class ReluTest(test_lib.TestCase):
     # Test that relu(nan) = nan for various sizes.
     for i in range(18):
       x = np.zeros(i) + np.nan
-      with self.cached_session():
+      # TODO(b/178335491): This is broken on GPU today.
+      with self.cached_session(use_gpu=False):
         z = nn_ops.relu(constant_op.constant(x)).eval()
         self.assertTrue(np.isnan(z).all())