fix windows compile (#13147)

paddle/fluid/operators/activation_op.h

@@ -865,8 +865,8 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     auto temp1 = static_cast<T>(1) /
                  (static_cast<T>(1) + (static_cast<T>(-beta) * x).exp());
-    auto temp2 = temp1 * (static_cast<T>(1) - (beta * out));
-    dx.device(d) = dout * ((beta * out) + temp2);
+    auto temp2 = temp1 * (static_cast<T>(1) - (static_cast<T>(beta) * out));
+    dx.device(d) = dout * ((static_cast<T>(beta) * out) + temp2);
   }
 };
 

paddle/fluid/operators/attention_lstm_op.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/attention_lstm_op.h"
-#include <sys/time.h>
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/cpu_vec.h"

paddle/fluid/operators/gru_unit_op.h

@@ -92,12 +92,12 @@ class GRUUnitKernel : public framework::OpKernel<T> {
               gate_data, frame_size * 3);
 
     // calculate activited gate
-    Eigen::array<int, 2> extents({{batch_size, frame_size}});
-    Eigen::array<int, 2> u_offsets({{0, 0}});
+    Eigen::array<int, 2> extents = {batch_size, frame_size};
+    Eigen::array<int, 2> u_offsets = {0, 0};
     ActCompute(context.Attr<int>("gate_activation"), place,
                g.slice(u_offsets, extents), g.slice(u_offsets, extents));
     auto u = g.slice(u_offsets, extents);  // update gate
-    Eigen::array<int, 2> r_offsets({{0, frame_size}});
+    Eigen::array<int, 2> r_offsets = {0, frame_size};
     ActCompute(context.Attr<int>("gate_activation"), place,
                g.slice(r_offsets, extents), g.slice(r_offsets, extents));
     auto r = g.slice(r_offsets, extents);  // reset gate
@@ -107,7 +107,7 @@ class GRUUnitKernel : public framework::OpKernel<T> {
               weight_data + frame_size * frame_size * 2, frame_size, 1,
               gate_data + frame_size * 2, frame_size * 3);
 
-    Eigen::array<int, 2> c_offsets({{0, frame_size * 2}});
+    Eigen::array<int, 2> c_offsets = {0, frame_size * 2};
     ActCompute(context.Attr<int>("activation"), place,
                g.slice(c_offsets, extents), g.slice(c_offsets, extents));
     auto c = g.slice(c_offsets, extents);  // output candidate
@@ -171,12 +171,12 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
     int batch_size = input->dims()[0];
     int frame_size = hidden_prev->dims()[1];
 
-    Eigen::array<int, 2> extents({{batch_size, frame_size}});
-    Eigen::array<int, 2> u_offsets({{0, 0}});
+    Eigen::array<int, 2> extents = {batch_size, frame_size};
+    Eigen::array<int, 2> u_offsets = {0, 0};
     auto u = g.slice(u_offsets, extents);  // update gate
-    Eigen::array<int, 2> r_offsets({{0, frame_size}});
+    Eigen::array<int, 2> r_offsets = {0, frame_size};
     auto r = g.slice(r_offsets, extents);  // reset gate
-    Eigen::array<int, 2> c_offsets({{0, frame_size * 2}});
+    Eigen::array<int, 2> c_offsets = {0, frame_size * 2};
     auto c = g.slice(c_offsets, extents);  // output candidate
 
     // backward for unactivated update gate

paddle/fluid/operators/label_smooth_op.h

@@ -38,7 +38,8 @@ class LabelSmoothKernel : public framework::OpKernel<T> {
       auto dist = framework::EigenVector<T>::Flatten(*dist_t);
       out.device(dev) =
           static_cast<T>(1 - epsilon) * in +
-          epsilon * dist.broadcast(Eigen::DSizes<int, 1>(in_t->numel()));
+          static_cast<T>(epsilon) *
+              dist.broadcast(Eigen::DSizes<int, 1>(in_t->numel()));
     } else {
       out.device(dev) = static_cast<T>(1 - epsilon) * in +
                         static_cast<T>(epsilon / label_dim);