Fix bug.

paddle/operators/math/CMakeLists.txt

@@ -10,7 +10,7 @@ if(WITH_GPU)
     nv_library(pooling SRCS pooling.cc pooling.cu DEPS device_context)
     nv_library(sequence_pooling SRCS sequence_pooling.cc sequence_pooling.cu DEPS device_context math_function)
     nv_library(vol2col SRCS vol2col.cc vol2col.cu DEPS device_context)
-    nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context)
+    nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context math_function)
     nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context)
     nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions)
     nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
@@ -22,7 +22,7 @@ else()
     cc_library(pooling SRCS pooling.cc DEPS device_context)
     cc_library(sequence_pooling SRCS sequence_pooling.cc DEPS device_context math_function)
     cc_library(vol2col SRCS vol2col.cc DEPS device_context)
-    cc_library(context_project SRCS context_project.cc DEPS device_context)
+    cc_library(context_project SRCS context_project.cc DEPS device_context math_function)
     cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context)
     cc_library(lstm_compute SRCS lstm_compute.cc DEPS device_context activation_functions)
     cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function)

paddle/operators/math/context_project.h

@@ -91,9 +91,6 @@ class ContextProjectFunctor {
     auto lod_level_0 = in.lod()[0];
 
     math::Im2ColFunctor<math::ColFormat::kOCF, Place, float> im2col_ocf;
-    if (platform::is_gpu_place(context.GetPlace())) {
-      LOG(INFO) << "========= gpu ==========";
-    }
 
     int input_row_begin, input_row_end;
     int sequence_height, sequence_width;

paddle/operators/math/math_function_impl.h

@@ -23,8 +23,7 @@ template <typename Place, typename T>
 void SetConstant<Place, T>::operator()(const platform::DeviceContext& context,
                                        framework::Tensor* tensor, T num) {
   auto t = framework::EigenVector<T>::Flatten(*tensor);
-  t.device(*context.GetEigenDevice<platform::CPUPlace>()) =
-      t.constant(static_cast<T>(num));
+  t.device(*context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(num));
 }
 
 template <typename Place, typename T, int Rank>

paddle/operators/softmax_op.h

@@ -27,9 +27,6 @@ class SoftmaxKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* X = context.Input<Tensor>("X");
     auto* Y = context.Output<Tensor>("Y");
-    if (platform::is_gpu_place(context.GetPlace())) {
-      LOG(INFO) << "==========gpu=========";
-    }
 
     // allocate memory on device.
     Y->mutable_data<T>(context.GetPlace());