diff --git a/src/framework/load_ops.h b/src/framework/load_ops.h
index dd46776219565243f23955e279f57f76f94df24a..03fdd8d433cd40aa7ba4786f02221bd24bd3a050 100644
--- a/src/framework/load_ops.h
+++ b/src/framework/load_ops.h
@@ -199,6 +199,12 @@ LOAD_OP3(pool2d, CPU, MALI_GPU, FPGA);
 #ifdef MULTICLASSNMS_OP
 LOAD_OP1(multiclass_nms, CPU);
 #endif
+#ifdef SUM_OP
+LOAD_OP1(sum, CPU);
+#endif
+#ifdef ELEMENTWISEMUL_OP
+LOAD_OP1(elementwise_mul, CPU);
+#endif
 #ifdef SLICE_OP
 LOAD_OP2(slice, CPU, MALI_GPU);
 #endif
diff --git a/src/operators/kernel/central-arm-func/sum_arm_func.h b/src/operators/kernel/central-arm-func/sum_arm_func.h
index 0319f2b23418f36670ca51993e97726879f12ec1..25c1c51c7abd62a900665197ab4e221b76a3fa04 100644
--- a/src/operators/kernel/central-arm-func/sum_arm_func.h
+++ b/src/operators/kernel/central-arm-func/sum_arm_func.h
@@ -27,13 +27,11 @@ void SumCompute(const SumParam<CPU> &param) {
   auto *outvar = param.OutVar();
 
   bool in_place = outvar == inputsvars[0];
-  DLOG << "11:";
   if (outvar->IsType<framework::LoDTensor>()) {
     auto *out = outvar->GetMutable<LoDTensor>();
     if (!in_place) {
       out->mutable_data<float>();
     }
-    DLOG << "1:";
     auto *outptr = out->data<float>();
     // auto result = Flatten(*out);
 
@@ -62,7 +60,6 @@ void SumCompute(const SumParam<CPU> &param) {
     }
 
   } else if (outvar->IsType<framework::SelectedRows>()) {
-    DLOG << "2:";
     std::unique_ptr<framework::SelectedRows> in0;
     if (in_place) {
       // If is in_place, we store the input[0] to in0
@@ -119,12 +116,12 @@ void SumCompute(const SumParam<CPU> &param) {
       if (sel_row.rows().size() == 0) {
         continue;
       }
-      PADDLE_MOBILE_ENFORCE(out->height() == sel_row.height());
+      PADDLE_MOBILE_ENFORCE(out->height() == sel_row.height(),
+                            "seletrows height != outheight");
       functor(sel_row, offset, out);
       offset += sel_row.value().numel();
     }
   } else if (outvar->IsType<LoDTensorArray>()) {
-    DLOG << "3:";
     auto &out_array = *outvar->GetMutable<LoDTensorArray>();
     for (size_t i = in_place ? 1 : 0; i < inputsvars.size(); ++i) {
       PADDLE_MOBILE_ENFORCE(inputsvars[i]->IsType<LoDTensorArray>(),
@@ -140,7 +137,8 @@ void SumCompute(const SumParam<CPU> &param) {
             framework::TensorCopy((*in_array)[i], &out_array[i]);
             out_array[i].set_lod((*in_array)[i].lod());
           } else {
-            PADDLE_MOBILE_ENFORCE(out_array[i].lod() == (*in_array)[i].lod());
+            PADDLE_MOBILE_ENFORCE(out_array[i].lod() == (*in_array)[i].lod(),
+                                  "outLod != inLod");
             auto *inptr = (*in_array)[i].data<float>();
             auto *outptr = out_array[i].data<float>();
 
@@ -152,9 +150,7 @@ void SumCompute(const SumParam<CPU> &param) {
       }
     }
   } else {
-    DLOG << "2:";
     if (outvar->IsType<framework::Tensor>()) {
-      DLOG << "3: ";
     }
     PADDLE_MOBILE_THROW_EXCEPTION(
         "Unexpected branch, output variable type is %s", outvar->Type().name());
diff --git a/src/operators/math/selected_rows_functor.h b/src/operators/math/selected_rows_functor.h
index 8cf1f5ca395d111ecca90f802773703ecb3286c9..f8b5521e4d19fd3199e7b05a902c98b731c9fbd0 100644
--- a/src/operators/math/selected_rows_functor.h
+++ b/src/operators/math/selected_rows_functor.h
@@ -47,7 +47,7 @@ struct SelectedRowsAddTo {
                   const int64_t input2_offset,
                   framework::SelectedRows* input2) {
     auto in1_height = input1.height();
-    PADDLE_MOBILE_ENFORCE(in1_height == input2->height());
+    PADDLE_MOBILE_ENFORCE(in1_height == input2->height(), "height error");
 
     auto& in1_rows = input1.rows();
     auto& in2_rows = *(input2->mutable_rows());
@@ -77,13 +77,14 @@ struct SelectedRowsAddToTensor {
                   framework::Tensor* input2) {
     auto in1_height = input1.height();
     auto in2_dims = input2->dims();
-    PADDLE_MOBILE_ENFORCE(in1_height == in2_dims[0]);
+    PADDLE_MOBILE_ENFORCE(in1_height == in2_dims[0], "height != dims[0]");
 
     auto& in1_value = input1.value();
     auto& in1_rows = input1.rows();
 
     int64_t in1_row_numel = in1_value.numel() / in1_rows.size();
-    PADDLE_MOBILE_ENFORCE(in1_row_numel == input2->numel() / in1_height);
+    PADDLE_MOBILE_ENFORCE(in1_row_numel == input2->numel() / in1_height,
+                          "row_numel error");
 
     auto* in1_data = in1_value.data<T>();
     auto* input2_data = input2->data<T>();
diff --git a/src/operators/op_param.h b/src/operators/op_param.h
index d1795ae3475b5dccb1d50cfe37bde356e70468b1..70562da8f8961daed9c0057f3ebc8e1a1a6e340e 100644
--- a/src/operators/op_param.h
+++ b/src/operators/op_param.h
@@ -471,15 +471,6 @@ class ElementwiseMulParam : OpParam {
   GType *input_y_;
   GType *out_;
   int axis_;
-#ifdef PADDLE_MOBILE_FPGA
-
- private:
-  fpga::EWMulArgs fpga_EW_mul_args;
-
- public:
-  const fpga::EWMulArgs &FpgaArgs() const { return fpga_EW_mul_args; }
-  void SetFpgaArgs(const fpga::EWMulArgs &args) { fpga_EW_mul_args = args; }
-#endif
 };
 #endif
 
@@ -628,15 +619,6 @@ class SumParam : public OpParam {
   Variable *out_var_;
   vector<GType *> inputs_;
   GType *out_;
-#ifdef PADDLE_MOBILE_FPGA
-
- private:
-  fpga::SumArgs fpga_sum_args;
-
- public:
-  const fpga::SumArgs &FpgaArgs() const { return fpga_sum_args; }
-  void SetFpgaArgs(const fpga::SumArgs &args) { fpga_sum_args = args; }
-#endif
 };
 #endif