From: A. Unique TensorFlower <gardener@tensorflow.org>
Date: Thu, 15 Mar 2018 21:40:12 +0000 (-0700)
Subject: Adds a GPU kernel registration for PlaceholderWithDefault, so we can avoid the issue
X-Git-Tag: tflite-v0.1.7~167^2^2~16
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=a325b0ae208b3295dca574f93c214f7c924cea11;p=platform%2Fupstream%2Ftensorflow.git

Adds a GPU kernel registration for PlaceholderWithDefault, so we can avoid the issue
of using it with a registered GPU device without soft placement.

PiperOrigin-RevId: 189248024
---

diff --git a/tensorflow/core/kernels/identity_op.cc b/tensorflow/core/kernels/identity_op.cc
index a18a72c..b5603fe 100644
--- a/tensorflow/core/kernels/identity_op.cc
+++ b/tensorflow/core/kernels/identity_op.cc
@@ -101,6 +101,10 @@ REGISTER_SYCL_HOST_KERNEL(bool);
   REGISTER_KERNEL_BUILDER(Name("DebugGradientIdentity")                     \
                               .Device(DEVICE_GPU)                           \
                               .TypeConstraint<type>("T"),                   \
+                          IdentityOp);                                      \
+  REGISTER_KERNEL_BUILDER(Name("PlaceholderWithDefault")                    \
+                              .Device(DEVICE_GPU)                           \
+                              .TypeConstraint<type>("dtype"),               \
                           IdentityOp)
 
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_KERNEL);
@@ -112,18 +116,24 @@ REGISTER_GPU_KERNEL(Variant);
 // A special GPU kernel for int32 and bool.
 // TODO(b/25387198): Also enable int32 in device memory. This kernel
 // registration requires all int32 inputs and outputs to be in host memory.
-#define REGISTER_GPU_HOST_KERNEL(type)                    \
-  REGISTER_KERNEL_BUILDER(Name("Identity")                \
-                              .Device(DEVICE_GPU)         \
-                              .HostMemory("input")        \
-                              .HostMemory("output")       \
-                              .TypeConstraint<type>("T"), \
-                          IdentityOp);                    \
-  REGISTER_KERNEL_BUILDER(Name("RefIdentity")             \
-                              .Device(DEVICE_GPU)         \
-                              .HostMemory("input")        \
-                              .HostMemory("output")       \
-                              .TypeConstraint<type>("T"), \
+#define REGISTER_GPU_HOST_KERNEL(type)                        \
+  REGISTER_KERNEL_BUILDER(Name("Identity")                    \
+                              .Device(DEVICE_GPU)             \
+                              .HostMemory("input")            \
+                              .HostMemory("output")           \
+                              .TypeConstraint<type>("T"),     \
+                          IdentityOp);                        \
+  REGISTER_KERNEL_BUILDER(Name("RefIdentity")                 \
+                              .Device(DEVICE_GPU)             \
+                              .HostMemory("input")            \
+                              .HostMemory("output")           \
+                              .TypeConstraint<type>("T"),     \
+                          IdentityOp);                        \
+  REGISTER_KERNEL_BUILDER(Name("PlaceholderWithDefault")      \
+                              .Device(DEVICE_GPU)             \
+                              .HostMemory("input")            \
+                              .HostMemory("output")           \
+                              .TypeConstraint<type>("dtype"), \
                           IdentityOp)
 
 REGISTER_GPU_HOST_KERNEL(int32);
diff --git a/tensorflow/python/kernel_tests/constant_op_test.py b/tensorflow/python/kernel_tests/constant_op_test.py
index ffbdb0e..18796f7 100644
--- a/tensorflow/python/kernel_tests/constant_op_test.py
+++ b/tensorflow/python/kernel_tests/constant_op_test.py
@@ -881,7 +881,7 @@ versions {
 class PlaceholderWithDefaultTest(test.TestCase):
 
   def testFullShape(self):
-    with self.test_session():
+    with self.test_session(force_gpu=test_util.is_gpu_available()):
       p = array_ops.placeholder_with_default([[2, 2], [2, 2]], shape=[2, 2])
       a = array_ops.identity(p)
       self.assertAllEqual([[2, 2], [2, 2]], a.eval())
@@ -892,7 +892,7 @@ class PlaceholderWithDefaultTest(test.TestCase):
         a.eval(feed_dict={p: [[6, 6, 6], [6, 6, 6]]})
 
   def testPartialShape(self):
-    with self.test_session():
+    with self.test_session(force_gpu=test_util.is_gpu_available()):
       p = array_ops.placeholder_with_default([1, 2, 3], shape=[None])
       a = array_ops.identity(p)
       self.assertAllEqual([1, 2, 3], a.eval())
@@ -902,7 +902,7 @@ class PlaceholderWithDefaultTest(test.TestCase):
         a.eval(feed_dict={p: [[2, 2], [2, 2]]})
 
   def testNoShape(self):
-    with self.test_session():
+    with self.test_session(force_gpu=test_util.is_gpu_available()):
       p = array_ops.placeholder_with_default([17], shape=None)
       a = array_ops.identity(p)
       self.assertAllEqual([17], a.eval())
@@ -911,11 +911,12 @@ class PlaceholderWithDefaultTest(test.TestCase):
           [[3, 3], [3, 3]], a.eval(feed_dict={p: [[3, 3], [3, 3]]}))
 
   def testGradient(self):
-    with self.test_session():
+    with self.test_session(force_gpu=test_util.is_gpu_available()):
       x = array_ops.placeholder(dtypes_lib.float32, [5, 7])
       y = array_ops.placeholder_with_default(x, None)
       err = gradient_checker.compute_gradient_error(x, [5, 7], y, [5, 7])
       self.assertLess(err, 1e-3)
 
+
 if __name__ == "__main__":
   test.main()