diff --git a/mace/ops/BUILD b/mace/ops/BUILD
index a35e9cc3da4b9ac2a10cea7d971213e76b3dc9c7..ed251b51884016e01f8cc0dd56821bed3005a3ec 100644
--- a/mace/ops/BUILD
+++ b/mace/ops/BUILD
@@ -59,6 +59,7 @@ cc_test(
     srcs = glob(["*_benchmark.cc"]),
     deps = [
         ":ops",
+        ":test",
         "//mace/core:core",
         "//mace/core:test_benchmark_main",
     ],
diff --git a/mace/ops/conv_2d_benchmark.cc b/mace/ops/conv_2d_benchmark.cc
index 49c80b67c9b969c0e9f3f3f733d63af9a0611395..8148c02df5800933b9c45be8ef5f827313c01e95 100644
--- a/mace/ops/conv_2d_benchmark.cc
+++ b/mace/ops/conv_2d_benchmark.cc
@@ -4,12 +4,12 @@
 
 #include <algorithm>
 
+#include "mace/core/operator.h"
 #include "mace/core/testing/test_benchmark.h"
-#include "mace/kernels/conv_2d.h"
-#include "mace/kernels/conv_pool_2d_util.h"
+#include "mace/ops/conv_2d.h"
+#include "mace/ops/ops_test_util.h"
 
 namespace mace {
-namespace kernels {
 
 template <DeviceType D, typename T>
 static void Conv2d(int iters, int batch, int channels, int height, int width,
@@ -17,8 +17,32 @@ static void Conv2d(int iters, int batch, int channels, int height, int width,
                    Padding padding, int output_channels) {
   mace::testing::StopTiming();
 
+  OpsTestNet net;
+  OpDefBuilder("Conv2d", "Conv2dTest")
+        .Input("Input")
+        .Input("Filter")
+        .Input("Bias")
+        .Output("Output")
+        .Finalize(net.operator_def());
+
+  // Add args
+  net.AddIntsArg("strides", {stride, stride});
+  net.AddIntArg("padding", padding);
+  net.AddIntsArg("dilations", {1, 1});
+
+  // Add input data
+  net.AddRandomInput<float>("Input", {batch, channels, height, width});
+  net.AddRandomInput<float>("Filter", {output_channels, channels, kernel_h, kernel_w});
+  net.AddRandomInput<float>("Bias", {output_channels});
+
+  // Worm-up
+  for (int i = 0; i < 5; ++i) {
+    net.RunOp(D);
+  }
+
   mace::testing::StartTiming();
   while(iters--) {
+    net.RunOp(D);
   }
 }
 
@@ -36,7 +60,7 @@ static void Conv2d(int iters, int batch, int channels, int height, int width,
   BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, TYPE, CPU);  \
   BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, TYPE, NEON);
 
+BM_CONV_2D(1, 3, 256, 256, 1, 1, 1, VALID, 64, float);
 BM_CONV_2D(1, 64, 32, 32, 1, 1, 1, VALID, 128, float);
 
-} //  namespace kernels
 } //  namespace mace
diff --git a/mace/ops/conv_2d_test.cc b/mace/ops/conv_2d_test.cc
index 1aec07f84bbd2ddf2c8488c4e5dcabf6ab91016c..b8af88ae1f1e8deb79d8e3058cea45a702a1d34d 100644
--- a/mace/ops/conv_2d_test.cc
+++ b/mace/ops/conv_2d_test.cc
@@ -3,8 +3,8 @@
 //
 
 #include "mace/core/operator.h"
-#include "mace/ops/ops_test_util.h"
 #include "mace/ops/conv_2d.h"
+#include "mace/ops/ops_test_util.h"
 
 using namespace mace;
 
diff --git a/mace/ops/ops_test_util.h b/mace/ops/ops_test_util.h
index 14ec485ce583447b1f91a22544b616621dfba7c7..dbec589db8fb54cee3b19fc935232be7a603d0c0 100644
--- a/mace/ops/ops_test_util.h
+++ b/mace/ops/ops_test_util.h
@@ -37,17 +37,33 @@ class OpDefBuilder {
 
 class OpsTestNet {
   public:
+    OpsTestNet() {}
+
     template <typename T>
     void AddInputFromArray(const char* name,
                            const std::vector<index_t>& shape,
                            const std::vector<T>& data) {
       Tensor* input = ws_.CreateTensor(name, cpu_allocator(), DataTypeToEnum<T>::v());
       input->Resize(shape);
-      float* input_data = input->mutable_data<float>();
-      // TODO check the dims
+      T* input_data = input->mutable_data<T>();
+      MACE_CHECK(input->size() == data.size());
       memcpy(input_data, data.data(), data.size() * sizeof(T));
     }
 
+    template <typename T>
+    void AddRandomInput(const char* name, const std::vector<index_t>& shape) {
+      Tensor* input = ws_.CreateTensor(name, cpu_allocator(), DataTypeToEnum<T>::v());
+      input->Resize(shape);
+      float* input_data = input->mutable_data<T>();
+
+      std::random_device rd;
+      std::mt19937 gen(rd());
+      std::normal_distribution<T> nd(0, 1);
+
+      std::generate(input_data, input_data + input->size(),
+                    [&gen, &nd]{ return nd(gen); });
+    }
+
     void AddIntArg(const char* name, const int value) {
       auto arg = op_def_.add_arg();
       arg->set_name(name);
@@ -98,7 +114,7 @@ class OpsTestNet {
       if (!net_) {
         NetDef net_def;
         net_def.add_op()->CopyFrom(op_def_);
-        VLOG(0) << net_def.DebugString();
+        VLOG(3) << net_def.DebugString();
         net_ = CreateNet(net_def, &ws_, device);
       }
       return net_->Run();