diff --git a/mace/ops/conv_2d_benchmark.cc b/mace/ops/conv_2d_benchmark.cc
index 98575c819ada9747c4bb8e316051086e53c92398..d54ad7095044f2af2c268f67c19e0e443bea48f7 100644
--- a/mace/ops/conv_2d_benchmark.cc
+++ b/mace/ops/conv_2d_benchmark.cc
@@ -27,15 +27,15 @@ static void Conv2d(int iters,
   OpsTestNet net;
 
   // Add input data
-  net.AddRandomInput<D, T>("Input", {batch, height, width, channels});
-  net.AddRandomInput<D, T>("Filter",
+  net.AddRandomInput<D, half>("Input", {batch, height, width, channels});
+  net.AddRandomInput<D, half>("Filter",
                                {kernel_h, kernel_w, channels, output_channels});
-  net.AddRandomInput<D, T>("Bias", {output_channels});
+  net.AddRandomInput<D, half>("Bias", {output_channels});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, half>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, half>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
+    BufferToImage<D, half>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("InputImage")
         .Input("FilterImage")
@@ -55,6 +55,7 @@ static void Conv2d(int iters,
         .AddIntsArg("strides", {stride, stride})
         .AddIntArg("padding", padding)
         .AddIntsArg("dilations", {1, 1})
+        .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
         .Finalize(net.NewOperatorDef());
   }
 
@@ -92,15 +93,15 @@ static void Conv2d(int iters,
   BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, TYPE, OPENCL);
 
 // ICNet
-BM_CONV_2D(1, 512, 15, 15, 1, 1, 1, VALID, 1024, half);
+BM_CONV_2D(1, 512, 15, 15, 1, 1, 1, VALID, 1024, float);
 // SNPE GPU ExecutionDuration = 448us, % ALU Utilization = 105
-BM_CONV_2D(1, 64, 60, 60, 1, 1, 1, VALID, 128, half);
+BM_CONV_2D(1, 64, 60, 60, 1, 1, 1, VALID, 128, float);
 // SNPE GPU ExecutionDuration = 258us, % ALU Utilization = 108
-BM_CONV_2D(1, 32, 60, 60, 1, 1, 1, VALID, 128, half);
+BM_CONV_2D(1, 32, 60, 60, 1, 1, 1, VALID, 128, float);
 
-BM_CONV_2D(1, 128, 60, 60, 3, 3, 1, VALID, 128, half);
+BM_CONV_2D(1, 128, 60, 60, 3, 3, 1, VALID, 128, float);
 // SNPE GPU ExecutionDuration = 506us, % ALU Utilization = 106.8
-BM_CONV_2D(1, 32, 60, 60, 3, 3, 1, SAME, 32, half);
+BM_CONV_2D(1, 32, 60, 60, 3, 3, 1, SAME, 32, float);
 
 // Test RGB <-> YUV
 //BM_CONV_2D(1, 3, 2160, 1080, 1, 1, 1, VALID, 3, float);
diff --git a/mace/ops/conv_2d_test.cc b/mace/ops/conv_2d_test.cc
index 1cda801747df6e57284673f00c8e5266723743dd..faaf508c5d2f576da8632c800cadfbfdd4cae389 100644
--- a/mace/ops/conv_2d_test.cc
+++ b/mace/ops/conv_2d_test.cc
@@ -634,9 +634,9 @@ static void TestHalfComplexConvNxNS12(const std::vector<index_t> &shape) {
     // Run on device
     net.RunOp(D);
 
-    ImageToBuffer<D, half>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
 
-    ExpectTensorNear<float, half>(expected, *net.GetOutput("OPENCLOutput"), 0.2);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.2);
   };
 
   for (int kernel_size : {1, 3}) {