Make GRU Operator adapt to sequence2batch

paddle/operators/gru_op.h

@@ -66,7 +66,7 @@ class GRUKernel : public framework::OpKernel<T> {
     bool is_reverse = context.Attr<bool>("is_reverse");
     math::LoDTensor2BatchFunctor<Place, T> to_batch;
     // to_batch(context.device_context(), *input, batch_gate, is_reverse);
-    to_batch(context.device_context(), *input, *batch_gate, is_reverse);
+    to_batch(context.device_context(), *input, *batch_gate, true, is_reverse);
 
     int frame_size = hidden_dims[1];
     int batch_size = hidden_dims[0];
@@ -172,8 +172,8 @@ class GRUGradKernel : public framework::OpKernel<T> {
     batch_hidden_grad.set_lod(batch_hidden->lod());
     // context.ShareLoD(framework::GradVarName("Hidden"),
     //                  framework::GradVarName("Input"));
-    to_batch(context.device_context(), *hidden_grad, batch_hidden_grad,
-             is_reverse, false);
+    to_batch(context.device_context(), *hidden_grad, batch_hidden_grad, false,
+             is_reverse);
 
     math::hl_gru_value<T> gru_value;
     gru_value.gateWeight = const_cast<T*>(weight_data);