diff --git a/tensorflow/core/common_runtime/eager/kernel_and_device.cc b/tensorflow/core/common_runtime/eager/kernel_and_device.cc
index 3c586e8188a..bf7c083f24b 100644
--- a/tensorflow/core/common_runtime/eager/kernel_and_device.cc
+++ b/tensorflow/core/common_runtime/eager/kernel_and_device.cc
@@ -35,8 +35,10 @@ limitations under the License.
 #include "tensorflow/core/lib/gtl/cleanup.h"
 #include "tensorflow/core/lib/gtl/map_util.h"
 #include "tensorflow/core/lib/random/random.h"
+#include "tensorflow/core/platform/denormal.h"
 #include "tensorflow/core/platform/errors.h"
 #include "tensorflow/core/platform/fingerprint.h"
+#include "tensorflow/core/platform/setround.h"
 #include "tensorflow/core/profiler/lib/annotated_traceme.h"
 #include "tensorflow/core/profiler/lib/traceme.h"
 #include "tensorflow/core/public/version.h"
@@ -281,6 +283,8 @@ Status KernelAndDeviceOp::Run(
   OpKernelContext context(&params);
 
   {
+    port::ScopedFlushDenormal flush;
+    port::ScopedSetRound round(FE_TONEAREST);
     // 'AnnotatedTraceMe' will trace both scheduling time on host and execution
     // time on device of the OpKernel.
     profiler::AnnotatedTraceMe activity(
diff --git a/tensorflow/python/kernel_tests/denormal_test.py b/tensorflow/python/kernel_tests/denormal_test.py
index d824e95f213..6e073f0d526 100644
--- a/tensorflow/python/kernel_tests/denormal_test.py
+++ b/tensorflow/python/kernel_tests/denormal_test.py
@@ -23,7 +23,6 @@ import platform
 
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import test_util
-from tensorflow.python.ops import array_ops
 from tensorflow.python.platform import test
 
 
@@ -35,32 +34,30 @@ class DenormalTest(test.TestCase):
       tiny = np.finfo(dtype).tiny
       self.assertEqual(tiny, tiny / 16 * 16)
 
-  def _flushDenormalsTest(self, use_gpu, dtypes):
-    if platform.machine() == "ppc64le" or platform.machine(
-    ) == "s390x" or platform.machine() == "aarch64":
+  def _flushDenormalsTest(self, dtypes):
+    if (platform.machine() == "ppc64le" or platform.machine() == "s390x" or
+        platform.machine() == "aarch64"):
       # Disabled denormal_test on power/s390x/aarch64 platform
       # Check relevant discussion - https://github.com/tensorflow/tensorflow/issues/11902
       return
-    with self.cached_session(use_gpu=use_gpu):
-      array_ops.identity(7).eval()
-      for dtype in dtypes:
-        tiny = np.finfo(dtype).tiny
-        # Small shape to test main thread, large shape to test thread pool
-        for shape in (), (1 << 20,):
-          flush = 0.1 * constant_op.constant(tiny, shape=shape)
-          self.assertAllEqual(flush.eval(), np.zeros(shape))
-          # Make sure the flags don't leak out
-          self.testPythonHasDenormals()
+    for dtype in dtypes:
+      tiny = np.finfo(dtype).tiny
+      # Small shape to test main thread, large shape to test thread pool
+      for shape in (), (1 << 20,):
+        flush = 0.1 * constant_op.constant(tiny, shape=shape)
+        self.assertAllEqual(self.evaluate(flush), np.zeros(shape))
+        # Make sure the flags don't leak out
+        self.testPythonHasDenormals()
 
-  @test_util.run_deprecated_v1
+  @test_util.run_in_graph_and_eager_modes(use_gpu=False)
   def testFlushDenormalsCPU(self):
     # On CPUs, the processor flags flush for both single and double precision.
-    self._flushDenormalsTest(use_gpu=False, dtypes=(np.float32, np.float64))
+    self._flushDenormalsTest(dtypes=(np.float32, np.float64))
 
-  @test_util.run_deprecated_v1
+  @test_util.run_in_graph_and_eager_modes(use_gpu=True)
   def testFlushDenormalsGPU(self):
     # On GPUs, only single precision can flush to zero.
-    self._flushDenormalsTest(use_gpu=True, dtypes=(np.float32,))
+    self._flushDenormalsTest(dtypes=(np.float32,))
 
 
 if __name__ == "__main__":