Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fluid_api

Merge branch develop

paddle/fluid/operators/lrn_op.cc

@@ -214,7 +214,10 @@ class LRNOpMaker : public framework::OpProtoAndCheckerMaker {
         "Defaults to \"NHWC\". Specify the data format of the output data, "
         "the input will be transformed automatically. ")
         .SetDefault("AnyLayout");
-    AddAttr<bool>("is_test", "").SetDefault(false);
+    AddAttr<bool>("is_test",
+                  "Turns on memory optimization that optimizes away "
+                  "unnecessary memory allocations. Used by MKLDNN.")
+        .SetDefault(false);
 
     AddComment(R"DOC(
 Local Response Normalization Operator.

paddle/fluid/operators/lrn_op.h

@@ -121,6 +121,10 @@ class LRNGradKernel : public framework::OpKernel<T> {
     T alpha = ctx.Attr<T>("alpha");
     T beta = ctx.Attr<T>("beta");
 
+    PADDLE_ENFORCE(
+        !ctx.Attr<bool>("is_test"),
+        "is_test attribute should be set to False in training phase.");
+
     LRNGradFunctor<DeviceContext, T> f;
     f(ctx, x, out, mid, x_g, out_g, N, C, H, W, n, alpha, beta);
   }

paddle/fluid/platform/profiler_test.cc

@@ -13,7 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/platform/profiler.h"
+#ifdef PADDLE_WITH_CUDA
 #include "cuda_runtime.h"
+#endif
 #include "gtest/gtest.h"
 
 TEST(Event, CpuElapsedTime) {
@@ -159,6 +161,7 @@ TEST(RecordEvent, RecordEvent) {
   DisableProfiler(EventSortingKey::kTotal, "/tmp/profiler");
 }
 
+#ifdef PADDLE_WITH_CUDA
 TEST(TMP, stream_wait) {
   cudaStream_t stream;
   cudaStreamCreate(&stream);
@@ -166,3 +169,4 @@ TEST(TMP, stream_wait) {
   cudaStreamSynchronize(stream);
   cudaStreamSynchronize(stream);
 }
+#endif