diff --git a/mace/ops/addn_test.cc b/mace/ops/addn_test.cc
index 37f409b61fa10166ab22188fbe4544e8c537385e..471d96ec3a4743821a8ad635205e5bb7b0ed8449 100644
--- a/mace/ops/addn_test.cc
+++ b/mace/ops/addn_test.cc
@@ -42,28 +42,60 @@ template <DeviceType D>
 void SimpleAdd3() {
   // Construct graph
   OpsTestNet net;
-  OpDefBuilder("AddN", "AddNTest")
-      .Input("Input1")
-      .Input("Input2")
-      .Input("Input3")
-      .Output("Output")
-      .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddInputFromArray<D, float>("Input1", {1, 2, 3, 1}, {1, 2, 3, 4, 5, 6});
-  net.AddInputFromArray<D, float>("Input2", {1, 2, 3, 1}, {1, 2, 3, 4, 5, 6});
-  net.AddInputFromArray<D, float>("Input3", {1, 2, 3, 1}, {1, 2, 3, 4, 5, 6});
+  net.AddInputFromArray<D, float>("Input0", {1, 2, 3, 1},
+                                  {-2.06715, 2, 3, 4, 5, 6});
+  net.AddInputFromArray<D, float>("Input1", {1, 2, 3, 1},
+                                  {0.875977, 2, 3, 4, 5, 6});
+  net.AddInputFromArray<D, float>("Input2", {1, 2, 3, 1},
+                                  {1.34866, 2, 3, 4, 5, 6});
+  net.AddInputFromArray<D, float>("Input3", {1, 2, 3, 1},
+                                  {-0.1582, 2, 3, 4, 5, 6});
+
+  const int input_num = 4;
+  if (D == DeviceType::OPENCL) {
+    // run on gpu
+    for (int i = 0; i < input_num; ++i) {
+      BufferToImage<D, half>(&net, MakeString("Input", i),
+                             MakeString("InputImage", i),
+                             kernels::BufferType::IN_OUT_CHANNEL);
+    }
 
-  // Run
-  net.RunOp(D);
+    auto op_def_cl = OpDefBuilder("AddN", "AddNTest");
+    for (int i = 0; i < input_num; ++i) {
+      op_def_cl.Input(MakeString("InputImage", i));
+    }
+    op_def_cl.Output("OutputImage")
+        .AddIntArg("T", static_cast<int>(DataType::DT_HALF))
+        .Finalize(net.NewOperatorDef());
 
-  auto expected = CreateTensor<float>({1, 2, 3, 1}, {3, 6, 9, 12, 15, 18});
+    // Run on device
+    net.RunOp(D);
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
+    ImageToBuffer<D, float>(&net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT_CHANNEL);
+  } else {
+    OpDefBuilder("AddN", "AddNTest")
+        .Input("Input0")
+        .Input("Input1")
+        .Input("Input2")
+        .Input("Input3")
+        .Output("Output")
+        .Finalize(net.NewOperatorDef());
+    // Run
+    net.RunOp(D);
+  }
+
+  auto expected = CreateTensor<float>({1, 2, 3, 1},
+                                      {-0.000713, 8, 12, 16, 20, 24});
+
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-4, 1e-3);
 }
 }  // namespace
 
 TEST_F(AddnOpTest, CPUSimpleAdd3) { SimpleAdd3<DeviceType::CPU>(); }
+TEST_F(AddnOpTest, GPUSimpleAdd3) { SimpleAdd3<DeviceType::OPENCL>(); }
 
 namespace {
 template <DeviceType D>
@@ -118,7 +150,8 @@ void RandomTest() {
     ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
                             kernels::BufferType::IN_OUT_CHANNEL);
 
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.1);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-2, 1e-2);
   }
 }
 }  // namespace
diff --git a/mace/ops/batch_norm_test.cc b/mace/ops/batch_norm_test.cc
index bb546a78524cc4c834ab0adc909f0adf9f7442af..d588eccd746a34acb987cee5b3fa2625253d1900 100644
--- a/mace/ops/batch_norm_test.cc
+++ b/mace/ops/batch_norm_test.cc
@@ -67,10 +67,11 @@ void Simple() {
 
   // Check
   auto expected =
-    CreateTensor<float>({1, 6, 2, 1}, {-3.86, -3.86, -1.51, -1.51, 0.83, 0.83,
-                                       3.17, 3.17, 5.51, 5.51, 7.86, 7.86});
+    CreateTensor<float>({1, 6, 2, 1}, {-3.8543, -3.8543, -1.5125, -1.5125,
+                                       0.8291, 0.8291, 3.1708, 3.1708,
+                                       5.5125, 5.5125, 7.8543, 7.8543});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-2);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-4);
 }
 }  // namespace
 
@@ -81,8 +82,8 @@ TEST_F(BatchNormOpTest, SimpleOPENCL) { Simple<DeviceType::OPENCL>(); }
 TEST_F(BatchNormOpTest, SimpleRandomOPENCL) {
   // generate random input
   static unsigned int seed = time(NULL);
-  index_t batch = 1 + rand_r(&seed) % 10;
-  index_t channels = 3 + rand_r(&seed) % 50;
+  index_t batch = 1 + rand_r(&seed) % 5;
+  index_t channels = 3 + rand_r(&seed) % 25;
   index_t height = 64;
   index_t width = 64;
 
@@ -146,7 +147,7 @@ TEST_F(BatchNormOpTest, SimpleRandomOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5, 1e-4);
 }
 
 TEST_F(BatchNormOpTest, SimpleRandomHalfOPENCL) {
@@ -218,14 +219,14 @@ TEST_F(BatchNormOpTest, SimpleRandomHalfOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2, 1e-2);
 }
 
 TEST_F(BatchNormOpTest, ComplexRandomOPENCL) {
   // generate random input
   static unsigned int seed = time(NULL);
-  index_t batch = 1 + rand_r(&seed) % 10;
-  index_t channels = 3 + rand_r(&seed) % 50;
+  index_t batch = 1 + rand_r(&seed) % 5;
+  index_t channels = 3 + rand_r(&seed) % 25;
   index_t height = 103;
   index_t width = 113;
 
@@ -289,14 +290,14 @@ TEST_F(BatchNormOpTest, ComplexRandomOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5, 1e-4);
 }
 
 TEST_F(BatchNormOpTest, ComplexRandomHalfOPENCL) {
   // generate random input
   static unsigned int seed = time(NULL);
-  index_t batch = 1 + rand_r(&seed) % 10;
-  index_t channels = 3 + rand_r(&seed) % 50;
+  index_t batch = 1 + rand_r(&seed) % 5;
+  index_t channels = 3 + rand_r(&seed) % 25;
   index_t height = 103;
   index_t width = 113;
 
@@ -361,7 +362,7 @@ TEST_F(BatchNormOpTest, ComplexRandomHalfOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2, 1e-2);
 }
 
 TEST_F(BatchNormOpTest, NEONTest) {
@@ -418,7 +419,7 @@ TEST_F(BatchNormOpTest, NEONTest) {
 
   ExpectTensorNear<float>(*net.GetOutput("OutputExptected"),
                           *net.GetOutput("OutputNeon"),
-                          0.001);
+                          1e-5);
 }
 
 }  // namespace test
diff --git a/mace/ops/bias_add_test.cc b/mace/ops/bias_add_test.cc
index ccfcc9451ceb8d6483d5a11652026746fae5c6ec..dc3bca0617e4b4ccf45a2f7c9508a10ce63d0db3 100644
--- a/mace/ops/bias_add_test.cc
+++ b/mace/ops/bias_add_test.cc
@@ -53,7 +53,7 @@ void BiasAddSimple() {
       {1, 6, 2, 1},
       {5.5, 5.5, 7.5, 7.5, 9.5, 9.5, 11.5, 11.5, 13.5, 13.5, 15.5, 15.5});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-2);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -109,7 +109,7 @@ TEST_F(BiasAddOpTest, SimpleRandomOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5);
 }
 
 TEST_F(BiasAddOpTest, ComplexRandomOPENCL) {
@@ -158,7 +158,7 @@ TEST_F(BiasAddOpTest, ComplexRandomOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5);
 }
 
 }  // namespace test
diff --git a/mace/ops/buffer_to_image_test.cc b/mace/ops/buffer_to_image_test.cc
index 151abb8f4ceb4ae96c90c1d2c67258822396e377..679c326e8b5ea7669f1b97c20a50064c785866e3 100644
--- a/mace/ops/buffer_to_image_test.cc
+++ b/mace/ops/buffer_to_image_test.cc
@@ -143,7 +143,7 @@ void TestDiffTypeBidirectionTransform(const int type,
 
   // Check
   ExpectTensorNear<float>(*net.GetOutput("Input"), *net.GetOutput("I2BOutput"),
-                          1e-2);
+                          1e-3, 1e-6);
 }
 }  // namespace
 
@@ -185,7 +185,7 @@ void TestStringHalfBidirectionTransform(const int type,
 
   // Check
   ExpectTensorNear<half>(*net.GetOutput("Input"), *net.GetOutput("I2BOutput"),
-                         1e-2);
+                         1e-3, 1e-6);
 }
 }  // namespace
 
diff --git a/mace/ops/channel_shuffle_test.cc b/mace/ops/channel_shuffle_test.cc
index 610a260a2a98a2f0768830b831ad28e215eae53d..02bbaf2af807b00531e73ce1bf31f9af21df78a0 100644
--- a/mace/ops/channel_shuffle_test.cc
+++ b/mace/ops/channel_shuffle_test.cc
@@ -32,7 +32,7 @@ TEST_F(ChannelShuffleOpTest, C8G4_CPU) {
   auto expected = CreateTensor<float>(
       {1, 1, 2, 8}, {0, 2, 4, 6, 1, 3, 5, 7, 8, 10, 12, 14, 9, 11, 13, 15});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 TEST_F(ChannelShuffleOpTest, C16G4_OPENCL) {
@@ -66,7 +66,7 @@ TEST_F(ChannelShuffleOpTest, C16G4_OPENCL) {
       {0,  4,  8,  12, 1,  5,  9,  13, 2,  6,  10, 14, 3,  7,  11, 15,
        16, 20, 24, 28, 17, 21, 25, 29, 18, 22, 26, 30, 19, 23, 27, 31});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 }  // namespace test
diff --git a/mace/ops/conv_2d_test.cc b/mace/ops/conv_2d_test.cc
index b56f924c255b58bae092620fb76169447365e67a..a9707ab8469174efed16fcc33485b00fa3c2897a 100644
--- a/mace/ops/conv_2d_test.cc
+++ b/mace/ops/conv_2d_test.cc
@@ -68,7 +68,7 @@ void TestNHWCSimple3x3VALID() {
   }
 
   auto expected = CreateTensor<float>({1, 1, 1, 1}, {18.1f});
-  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
+  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 template<DeviceType D, typename T>
@@ -128,7 +128,7 @@ void TestNHWCSimple3x3SAME() {
     {1, 3, 3, 1},
     {8.1f, 12.1f, 8.1f, 12.1f, 18.1f, 12.1f, 8.1f, 12.1f, 8.1f});
 
-  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
+  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -194,7 +194,7 @@ void TestNHWCSimple3x3WithoutBias() {
   // Check
   auto expected = CreateTensor<float>({1, 1, 1, 1}, {18.0f});
 
-  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
+  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -266,7 +266,7 @@ void TestNHWCCombined3x3() {
   auto expected = CreateTensor<float>(
     {1, 3, 3, 2}, {8.1f, 4.2f, 12.1f, 6.2f, 8.1f, 4.2f, 12.1f, 6.2f, 18.1f,
                    9.2f, 12.1f, 6.2f, 8.1f, 4.2f, 12.1f, 6.2f, 8.1f, 4.2f});
-  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
+  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -345,7 +345,7 @@ void TestConv1x1() {
      5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f,
      5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -415,7 +415,8 @@ void TestComplexConvNxNS12(const std::vector<index_t> &shape,
 
     ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
                         kernels::BufferType::IN_OUT_CHANNEL);
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
+    ExpectTensorNear<float>(expected,
+                            *net.GetOutput("OPENCLOutput"), 1e-5, 1e-4);
   };
 
   for (int kernel_size : {1, 3, 7}) {
@@ -516,7 +517,8 @@ void TestHalfComplexConvNxNS12(const std::vector<index_t> &input_shape,
     ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
                             kernels::BufferType::IN_OUT_CHANNEL);
 
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
+    ExpectTensorNear<float>(expected,
+                            *net.GetOutput("OPENCLOutput"), 1e-2, 1e-1);
   };
 
   func(1, 1, VALID);
@@ -641,7 +643,8 @@ void TestDilationConvNxN(const std::vector<index_t> &shape,
 
     ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
                         kernels::BufferType::IN_OUT_CHANNEL);
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5, 1e-4);
   };
 
   for (int kernel_size : {3}) {
@@ -725,7 +728,8 @@ void TestArbitraryPadConvNxN(const std::vector<index_t> &shape,
 
     ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
                         kernels::BufferType::IN_OUT_CHANNEL);
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5, 1e-4);
   };
 
   for (int kernel_size : {3, 5}) {
@@ -807,7 +811,7 @@ static void TestNeonArbitraryPadConvNxN(const std::vector<index_t> &shape,
 
     ExpectTensorNear<float>(*net.GetOutput("OutputExptected"),
                             *net.GetOutput("OutputNeon"),
-                            0.001);
+                            1e-5, 1e-3);
   };
 
   for (int kernel_size : {1, 3, 5}) {
diff --git a/mace/ops/cwise_test.cc b/mace/ops/cwise_test.cc
index 0ae24fd52c0233b2618c6548b90861f74e11af80..56b4359a14c605635d0fd090fb640c228ffa96ab 100644
--- a/mace/ops/cwise_test.cc
+++ b/mace/ops/cwise_test.cc
@@ -54,7 +54,7 @@ void Simple(const kernels::CWiseType type,
 
   auto expected = CreateTensor<float>(shape, output);
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-3);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5, 1e-3);
 }
 }  // namespace
 
@@ -140,10 +140,10 @@ void RandomTest(const kernels::CWiseType type,
 
   if (DataTypeToEnum<T>::value == DT_FLOAT) {
     ExpectTensorNear<float>(*net.GetTensor("Output"),
-                            *net.GetOutput("OPENCLOutput"), 1e-3);
+                            *net.GetOutput("OPENCLOutput"), 1e-5, 1e-4);
   } else {
     ExpectTensorNear<float>(*net.GetTensor("Output"),
-                            *net.GetOutput("OPENCLOutput"), 1e-1);
+                            *net.GetOutput("OPENCLOutput"), 1e-2, 1e-2);
   }
 }
 }  // namespace
diff --git a/mace/ops/depth_to_space_test.cc b/mace/ops/depth_to_space_test.cc
index 93cfac5863da6a142c290d0239fa169c0b959399..f96dfb4d3f1713d6ce21838a211a6db900f65700 100644
--- a/mace/ops/depth_to_space_test.cc
+++ b/mace/ops/depth_to_space_test.cc
@@ -48,7 +48,7 @@ void RunDepthToSpace(const bool d2s,
         kernels::BufferType::IN_OUT_CHANNEL);
   }
   auto expected = CreateTensor<float>(expected_shape, expected_data);
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -194,10 +194,10 @@ void RandomTest(const bool d2s, const int block_size,
 
   if (DataTypeToEnum<T>::value == DT_FLOAT) {
     ExpectTensorNear<float>(*net.GetTensor("Output"),
-                            *net.GetOutput("OPENCLOutput"), 1e-3);
+                            *net.GetOutput("OPENCLOutput"), 1e-5);
   } else {
     ExpectTensorNear<float>(*net.GetTensor("Output"),
-                            *net.GetOutput("OPENCLOutput"), 1e-1);
+                            *net.GetOutput("OPENCLOutput"), 1e-3, 1e-4);
   }
 }
 }  // namespace
diff --git a/mace/ops/depthwise_conv2d_test.cc b/mace/ops/depthwise_conv2d_test.cc
index 73bdceb5605d2e49f5063f6a966eed2c0d83f3a5..50d9a1f7de9ed8c70e44db904e7c331b496d3b82 100644
--- a/mace/ops/depthwise_conv2d_test.cc
+++ b/mace/ops/depthwise_conv2d_test.cc
@@ -46,8 +46,8 @@ void SimpleValidTest() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, T>(&net, "OutputImage", "Output",
-                        kernels::BufferType::IN_OUT_CHANNEL);
+    ImageToBuffer<D, float>(&net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT_CHANNEL);
 
   } else {
     OpDefBuilder("DepthwiseConv2d", "DepthwiseConv2DTest")
@@ -64,11 +64,15 @@ void SimpleValidTest() {
   }
 
   // Check
-  auto expected = CreateTensor<T>(
-    {1, 2, 2, 2}, VectorStaticCast<T>({37.1f, 148.2f, 47.1f, 188.2f, 67.1f,
-                                       268.2f, 77.1f, 308.2f}));
+  auto expected = CreateTensor<float>(
+    {1, 2, 2, 2}, {37.1f, 148.2f, 47.1f, 188.2f,
+                   67.1f, 268.2f, 77.1f, 308.2f});
 
-  ExpectTensorNear<T>(*expected, *net.GetOutput("Output"), 1e-5);
+  if (DataTypeToEnum<T>::value == DT_HALF) {
+    ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-3, 1e-3);
+  } else {
+    ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
+  }
 }
 }  // namespace
 
@@ -189,7 +193,11 @@ void ComplexValidTest() {
        9.13650036, 8.5095005, 8.92500019, 9.34349918, 8.69849968,
        9.12300014, 9.55049992, 4.55220032, 4.80690002, 5.06340027}));
 
-  ExpectTensorNear<T>(*expected, *net.GetOutput("Output"), 0.2);
+  if (DataTypeToEnum<T>::value == DT_FLOAT) {
+    ExpectTensorNear<T>(*expected, *net.GetOutput("Output"), 1e-5);
+  } else {
+    ExpectTensorNear<T>(*expected, *net.GetOutput("Output"), 1e-2);
+  }
 }
 }  // namespace
 
@@ -282,7 +290,13 @@ void TestNxNS12(const index_t height, const index_t width) {
     }
 
     // Check
-    ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"), 0.1);
+    if (DataTypeToEnum<T>::value == DT_FLOAT) {
+      ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"),
+                              1e-5, 1e-4);
+    } else {
+      ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"),
+                              1e-2, 1e-2);
+    }
   };
 
   for (int kernel_size : {2, 3, 4}) {
@@ -303,12 +317,10 @@ TEST_F(DepthwiseConv2dOpTest, OpenCLSimpleNxNS12Half) {
 }
 
 TEST_F(DepthwiseConv2dOpTest, OpenCLAlignedNxNS12) {
-  TestNxNS12<DeviceType::OPENCL, float>(64, 64);
   TestNxNS12<DeviceType::OPENCL, float>(128, 128);
 }
 
 TEST_F(DepthwiseConv2dOpTest, OpenCLAlignedNxNS12Half) {
-  TestNxNS12<DeviceType::OPENCL, half>(64, 64);
   TestNxNS12<DeviceType::OPENCL, half>(128, 128);
 }
 
@@ -380,7 +392,7 @@ void TestNEONNxNS12(const index_t height,
     // Check
     ExpectTensorNear<float>(*net.GetOutput("OutputExptected"),
                             *net.GetOutput("OutputNeon"),
-                            0.001);
+                            1e-5, 1e-3);
   };
 
   for (int kernel_size : {1, 3, 5}) {
diff --git a/mace/ops/eltwise_test.cc b/mace/ops/eltwise_test.cc
index 36dea470f8b9e9f5a783a176cad43bda4494c4f4..a489607730dfe955f73fec36788039ab75ba9efa 100644
--- a/mace/ops/eltwise_test.cc
+++ b/mace/ops/eltwise_test.cc
@@ -60,7 +60,7 @@ void Simple(const kernels::EltwiseType type,
 
   auto expected = CreateTensor<float>(shape, output);
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-3);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -146,10 +146,10 @@ void RandomTest(const kernels::EltwiseType type,
 
   if (DataTypeToEnum<T>::value == DT_FLOAT) {
     ExpectTensorNear<float>(*net.GetTensor("Output"),
-                            *net.GetOutput("OPENCLOutput"), 1e-3);
+                            *net.GetOutput("OPENCLOutput"), 1e-5, 1e-4);
   } else {
     ExpectTensorNear<float>(*net.GetTensor("Output"),
-                            *net.GetOutput("OPENCLOutput"), 1e-1);
+                            *net.GetOutput("OPENCLOutput"), 1e-2, 1e-2);
   }
 }
 }  // namespace
diff --git a/mace/ops/folded_batch_norm_test.cc b/mace/ops/folded_batch_norm_test.cc
index d46e161570a922c8f51f9e06913f713a2e6b8f82..d422bb041c007e05d229c192dbcdd68ad1a82c0d 100644
--- a/mace/ops/folded_batch_norm_test.cc
+++ b/mace/ops/folded_batch_norm_test.cc
@@ -72,10 +72,11 @@ void Simple() {
 
   // Check
   auto expected =
-      CreateTensor<float>({1, 6, 2, 1}, {-3.86, -3.86, -1.51, -1.51, 0.83, 0.83,
-                                         3.17, 3.17, 5.51, 5.51, 7.86, 7.86});
+      CreateTensor<float>({1, 6, 2, 1}, {-3.8543, -3.8543, -1.5125, -1.5125,
+                                         0.8291, 0.8291, 3.1708, 3.1708,
+                                         5.5125, 5.5125, 7.8543, 7.8543});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-2);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-4);
 }
 }  // namespace
 
@@ -226,7 +227,7 @@ TEST_F(FoldedBatchNormOpTest, SimpleRandomOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5, 1e-4);
 }
 
 TEST_F(FoldedBatchNormOpTest, SimpleRandomHalfOPENCL) {
@@ -281,7 +282,7 @@ TEST_F(FoldedBatchNormOpTest, SimpleRandomHalfOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2, 1e-2);
 }
 
 TEST_F(FoldedBatchNormOpTest, ComplexRandomOPENCL) {
@@ -334,7 +335,7 @@ TEST_F(FoldedBatchNormOpTest, ComplexRandomOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5, 1e-4);
 }
 
 TEST_F(FoldedBatchNormOpTest, ComplexRandomHalfOPENCL) {
@@ -388,7 +389,7 @@ TEST_F(FoldedBatchNormOpTest, ComplexRandomHalfOPENCL) {
 
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2, 1e-2);
 }
 
 }  // namespace test
diff --git a/mace/ops/fully_connected_test.cc b/mace/ops/fully_connected_test.cc
index 349acc7f03c30aaa57faa76eaeacc621f1624f32..d4c17fb42fdabfe02709e7c2a8268b6fa138ffcd 100644
--- a/mace/ops/fully_connected_test.cc
+++ b/mace/ops/fully_connected_test.cc
@@ -164,9 +164,11 @@ void Complex(const index_t batch,
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
   if (DataTypeToEnum<T>::value == DataType::DT_HALF) {
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-1, 1e-1);
   } else {
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-3);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5, 1e-4);
   }
 }
 }  // namespace
@@ -247,9 +249,11 @@ void TestWXFormat(const index_t batch,
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_CHANNEL);
   if (DataTypeToEnum<T>::value == DataType::DT_HALF) {
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-1, 1e-1);
   } else {
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-3, 1e-3);
   }
 }
 }  // namespace
@@ -267,7 +271,7 @@ TEST_F(FullyConnectedOpTest, OPENCLWidthFormatMultiBatch) {
 }
 
 TEST_F(FullyConnectedOpTest, OPENCLHalfWidthFormatAligned) {
-  TestWXFormat<float>(1, 2, 2, 512, 2);
+  TestWXFormat<half>(1, 2, 2, 512, 2);
   TestWXFormat<half>(1, 11, 11, 32, 16);
   TestWXFormat<half>(1, 16, 32, 32, 32);
 }
@@ -315,7 +319,7 @@ void FullyConnectedTestNEON(const index_t batch,
 
   ExpectTensorNear<float>(*net.GetOutput("OutputExptected"),
                           *net.GetOutput("OutputNeon"),
-                          0.01);
+                          1e-3, 1e-3);
 }
 }  // namespace
 
diff --git a/mace/ops/fused_conv_2d_test.cc b/mace/ops/fused_conv_2d_test.cc
index e0fb62457254ead8b07dbcc45c74923f33b26525..6ef1298386a8114505bd82aeb30114b1792cb71b 100644
--- a/mace/ops/fused_conv_2d_test.cc
+++ b/mace/ops/fused_conv_2d_test.cc
@@ -18,14 +18,14 @@ template<DeviceType D, typename T>
 void TestNHWCSimple3x3VALID() {
   OpsTestNet net;
   // Add input data
-  net.AddInputFromArray<D, T>(
+  net.AddInputFromArray<D, float>(
     "Input", {1, 3, 3, 2},
     {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1});
-  net.AddInputFromArray<D, T>(
+  net.AddInputFromArray<D, float>(
     "Filter", {3, 3, 1, 2},
     {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
      1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
-  net.AddInputFromArray<D, T>("Bias", {1}, {-0.1f});
+  net.AddInputFromArray<D, float>("Bias", {1}, {-0.1f});
 
   if (D == DeviceType::OPENCL) {
     BufferToImage<D, T>(&net, "Input", "InputImage",
@@ -49,8 +49,8 @@ void TestNHWCSimple3x3VALID() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, T>(&net, "OutputImage", "Output",
-                        kernels::BufferType::IN_OUT_CHANNEL);
+    ImageToBuffer<D, float>(&net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT_CHANNEL);
 
   } else {
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
@@ -69,7 +69,7 @@ void TestNHWCSimple3x3VALID() {
   }
 
   auto expected = CreateTensor<float>({1, 1, 1, 1}, {0.0f});
-  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"));
 }
 
 template<DeviceType D, typename T>
@@ -77,14 +77,14 @@ void TestNHWCSimple3x3SAME() {
   OpsTestNet net;
 
   // Add input data
-  net.AddInputFromArray<D, T>(
+  net.AddInputFromArray<D, float>(
     "Input", {1, 3, 3, 2},
     {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1});
-  net.AddInputFromArray<D, T>(
+  net.AddInputFromArray<D, float>(
     "Filter", {3, 3, 1, 2},
     {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
      1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
-  net.AddInputFromArray<D, T>("Bias", {1}, {-0.1f});
+  net.AddInputFromArray<D, float>("Bias", {1}, {-0.1f});
 
   if (D == DeviceType::OPENCL) {
     BufferToImage<D, T>(&net, "Input", "InputImage",
@@ -108,8 +108,8 @@ void TestNHWCSimple3x3SAME() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, T>(&net, "OutputImage", "Output",
-                        kernels::BufferType::IN_OUT_CHANNEL);
+    ImageToBuffer<D, float>(&net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT_CHANNEL);
 
   } else {
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
@@ -130,7 +130,7 @@ void TestNHWCSimple3x3SAME() {
   auto expected = CreateTensor<float>(
     {1, 3, 3, 1}, {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f});
 
-  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"));
 }
 }  // namespace
 
@@ -150,10 +150,10 @@ void TestNHWCSimple3x3WithoutBias() {
   OpsTestNet net;
 
   // Add input data
-  net.AddInputFromArray<D, T>(
+  net.AddInputFromArray<D, float>(
     "Input", {1, 3, 3, 2},
     {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1});
-  net.AddInputFromArray<D, T>(
+  net.AddInputFromArray<D, float>(
     "Filter", {3, 3, 1, 2},
     {1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
      1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
@@ -177,8 +177,8 @@ void TestNHWCSimple3x3WithoutBias() {
     // Run
     net.RunOp(D);
     // Transfer output
-    ImageToBuffer<D, T>(&net, "OutputImage", "Output",
-                        kernels::BufferType::IN_OUT_CHANNEL);
+    ImageToBuffer<D, float>(&net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT_CHANNEL);
   } else {
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
       .Input("Input")
@@ -198,7 +198,7 @@ void TestNHWCSimple3x3WithoutBias() {
   // Check
   auto expected = CreateTensor<float>({1, 1, 1, 1}, {0.0f});
 
-  ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"));
 }
 }  // namespace
 
@@ -277,7 +277,7 @@ void TestConv1x1() {
      5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f,
      5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f, 5.1f, 10.2f});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"));
 }
 }  // namespace
 
@@ -346,7 +346,8 @@ void TestComplexConvNxNS12(const std::vector<index_t> &shape) {
 
     ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
                         kernels::BufferType::IN_OUT_CHANNEL);
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5, 1e-4);
   };
 
   for (int kernel_size : {1, 3}) {
@@ -434,7 +435,8 @@ void TestHalfComplexConvNxNS12(const std::vector<index_t> &shape) {
     ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
                             kernels::BufferType::IN_OUT_CHANNEL);
 
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.2);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-2, 1e-1);
   };
 
   for (int kernel_size : {1, 3}) {
@@ -513,7 +515,8 @@ void TestGeneralConvNxNS12(const std::vector<index_t> &image_shape,
 
     ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
                         kernels::BufferType::IN_OUT_CHANNEL);
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5, 1e-4);
   };
 
   for (int stride : {1, 2}) {
@@ -594,7 +597,8 @@ void TestAtrousConvNxN(const std::vector<index_t> &shape,
 
     ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
                         kernels::BufferType::IN_OUT_CHANNEL);
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5, 1e-4);
   };
 
   for (int kernel_size : {3}) {
@@ -683,7 +687,8 @@ void TestGeneralHalfAtrousConv(const std::vector<index_t> &image_shape,
 
     ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
                             kernels::BufferType::IN_OUT_CHANNEL);
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.7);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-2, 1e-1);
   };
 
   func(1, 1, VALID);
@@ -763,7 +768,7 @@ void TestNEONGeneralConvNxNS12(
 
     ExpectTensorNear<float>(*net.GetOutput("OutputExptected"),
                             *net.GetOutput("OutputNeon"),
-                            0.001);
+                            1e-5, 1e-4);
   };
 
   for (int stride : {1, 2}) {
diff --git a/mace/ops/global_avg_pooling_test.cc b/mace/ops/global_avg_pooling_test.cc
index 8b82f7ab4e3c9e8a98d0cf5549028a3a4d691a9f..e211268801fded1e632e13aa4ad1dc61e05723ae 100644
--- a/mace/ops/global_avg_pooling_test.cc
+++ b/mace/ops/global_avg_pooling_test.cc
@@ -31,7 +31,7 @@ TEST_F(GlobalAvgPoolingOpTest, 3x7x7_CPU) {
   // Check
   auto expected = CreateTensor<float>({1, 3, 1, 1}, {1, 2, 3});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 }  // namespace test
diff --git a/mace/ops/matmul_test.cc b/mace/ops/matmul_test.cc
index afd92df2333f1e378528ec640e01d5b91d91619c..d5fea8569279a3e7bebe9f4a332325498ecaabdb 100644
--- a/mace/ops/matmul_test.cc
+++ b/mace/ops/matmul_test.cc
@@ -148,9 +148,11 @@ void Complex(const index_t batch,
   ImageToBuffer<DeviceType::OPENCL, float>(&net, "OutputImage", "OPENCLOutput",
                                            kernels::BufferType::IN_OUT_HEIGHT);
   if (DataTypeToEnum<T>::value == DataType::DT_HALF) {
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-1);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-2, 1e-1);
   } else {
-    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-4);
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5, 1e-5);
   }
 }
 }  // namespace
diff --git a/mace/ops/ops_test_util.h b/mace/ops/ops_test_util.h
index 595f8000348ad4a8087fc01787d0d2e5932a3b80..aca7853da53c7cff137921f6d6606b9169165dd4 100644
--- a/mace/ops/ops_test_util.h
+++ b/mace/ops/ops_test_util.h
@@ -439,7 +439,9 @@ struct Expector<EXP_TYPE, RES_TYPE, true> {
     }
   }
 
-  static void Near(const Tensor &x, const Tensor &y, const double abs_err) {
+  static void Near(const Tensor &x, const Tensor &y,
+                   const double rel_err,
+                   const double abs_err) {
     ASSERT_EQ(x.dtype(), DataTypeToEnum<EXP_TYPE>::v());
     ASSERT_EQ(y.dtype(), DataTypeToEnum<RES_TYPE>::v());
     AssertSameDims(x, y);
@@ -452,8 +454,9 @@ struct Expector<EXP_TYPE, RES_TYPE, true> {
         for (int h = 0; h < x.dim(1); ++h) {
           for (int w = 0; w < x.dim(2); ++w) {
             for (int c = 0; c < x.dim(3); ++c) {
-              EXPECT_NEAR(*a, *b, abs_err) << "with index = [" << n << ", " << h
-                                           << ", " << w << ", " << c << "]";
+              const double error = abs_err + rel_err * std::abs(*a);
+              EXPECT_NEAR(*a, *b, error) << "with index = [" << n << ", " << h
+                                         << ", " << w << ", " << c << "]";
               a++;
               b++;
             }
@@ -462,26 +465,31 @@ struct Expector<EXP_TYPE, RES_TYPE, true> {
       }
     } else {
       for (int i = 0; i < x.size(); ++i) {
-        EXPECT_NEAR(a[i], b[i], abs_err) << "a = " << a << " b = " << b
-                                         << " index = " << i;
+        const double error = abs_err + rel_err * std::abs(a[i]);
+        EXPECT_NEAR(a[i], b[i], error) << "a = " << a << " b = " << b
+                                       << " index = " << i;
       }
     }
   }
 };
 
 template<typename T>
-void ExpectTensorNear(const Tensor &x, const Tensor &y, const double abs_err) {
+void ExpectTensorNear(const Tensor &x, const Tensor &y,
+                      const double rel_err = 1e-5,
+                      const double abs_err = 1e-8) {
   static_assert(is_floating_point_type<T>::value,
                 "T is not a floating point type");
-  Expector<T, T>::Near(x, y, abs_err);
+  Expector<T, T>::Near(x, y, rel_err, abs_err);
 }
 
 template<typename EXP_TYPE, typename RES_TYPE>
-void ExpectTensorNear(const Tensor &x, const Tensor &y, const double abs_err) {
+void ExpectTensorNear(const Tensor &x, const Tensor &y,
+                      const double rel_err = 1e-5,
+                      const double abs_err = 1e-8) {
   static_assert(is_floating_point_type<EXP_TYPE>::value &&
                   is_floating_point_type<RES_TYPE>::value,
                 "T is not a floating point type");
-  Expector<EXP_TYPE, RES_TYPE>::Near(x, y, abs_err);
+  Expector<EXP_TYPE, RES_TYPE>::Near(x, y, rel_err, abs_err);
 }
 
 template<DeviceType D, typename T>
diff --git a/mace/ops/pad_test.cc b/mace/ops/pad_test.cc
index 14d30ca5950f84c2c22a995669a05cec10889375..5a87f32a2219f47815c68fe32f835c89adc278be 100644
--- a/mace/ops/pad_test.cc
+++ b/mace/ops/pad_test.cc
@@ -137,7 +137,7 @@ void Complex(const std::vector<index_t> &input_shape,
   auto output = net.GetTensor("OpenCLOutput");
 
   if (DataTypeToEnum<T>::value == DT_HALF) {
-    ExpectTensorNear<float>(expected, *output, 1e-1);
+    ExpectTensorNear<float>(expected, *output, 1e-2, 1e-2);
   } else {
     ExpectTensorNear<float>(expected, *output, 1e-5);
   }
diff --git a/mace/ops/pooling_test.cc b/mace/ops/pooling_test.cc
index d65af7c611b9f4f6acb80c3e4901d0046efb1821..27d4cf891fa8abd44754a6b8f70c5425e5b9103c 100644
--- a/mace/ops/pooling_test.cc
+++ b/mace/ops/pooling_test.cc
@@ -41,7 +41,7 @@ TEST_F(PoolingOpTest, MAX_VALID) {
   auto expected =
     CreateTensor<float>({1, 2, 2, 2}, {5, 21, 7, 23, 13, 29, 15, 31});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 TEST_F(PoolingOpTest, MAX_SAME) {
@@ -67,7 +67,7 @@ TEST_F(PoolingOpTest, MAX_SAME) {
   // Check
   auto expected = CreateTensor<float>({1, 2, 2, 1}, {4, 5, 7, 8});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 TEST_F(PoolingOpTest, MAX_VALID_DILATION) {
@@ -94,7 +94,7 @@ TEST_F(PoolingOpTest, MAX_VALID_DILATION) {
   // Check
   auto expected = CreateTensor<float>({1, 2, 2, 1}, {10, 11, 14, 15});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 TEST_F(PoolingOpTest, MAX_k2x2s2x2) {
@@ -120,7 +120,7 @@ TEST_F(PoolingOpTest, MAX_k2x2s2x2) {
   // Check
   auto expected = CreateTensor<float>({1, 1, 5, 1}, {10, 12, 14, 16, 17});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 namespace {
@@ -167,7 +167,7 @@ void SimpleMaxPooling3S2() {
   // Check
   auto expected = CreateTensor<float>({1, 1, 4, 1}, {20, 22, 24, 26});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -192,11 +192,10 @@ void MaxPooling3S2(const std::vector<index_t> &input_shape,
     .AddIntsArg("strides", strides)
     .AddIntArg("padding", padding)
     .AddIntsArg("dilations", {1, 1})
-    .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
     .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<D, T>("Input", input_shape);
+  net.AddRandomInput<D, float>("Input", input_shape);
 
   // run on cpu
   net.RunOp();
@@ -216,10 +215,15 @@ void MaxPooling3S2(const std::vector<index_t> &input_shape,
     .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
     .Finalize(net.NewOperatorDef());
   net.RunOp(D);
-  ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
+  ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
                       kernels::BufferType::IN_OUT_CHANNEL);
 
-  ExpectTensorNear<T>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
+  if (DataTypeToEnum<T>::value == DT_HALF) {
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-3, 1e-4);
+  } else {
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5);
+  }
 }
 }  // namespace
 
@@ -276,7 +280,7 @@ TEST_F(PoolingOpTest, AVG_VALID) {
   auto expected = CreateTensor<float>(
     {1, 2, 2, 2}, {2.5, 18.5, 4.5, 20.5, 10.5, 26.5, 12.5, 28.5});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 namespace {
@@ -309,7 +313,7 @@ void SimpleAvgPoolingTest() {
   // Check
   auto expected = CreateTensor<float>({1, 1, 4, 1}, {4.5, 6.5, 8.5, 10.5});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -356,10 +360,17 @@ void AvgPoolingTest(const std::vector<index_t> &shape,
     .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
     .Finalize(net.NewOperatorDef());
   net.RunOp(D);
-  ImageToBuffer<D, T>(&net, "OutputImage", "OPENCLOutput",
+  ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
                       kernels::BufferType::IN_OUT_CHANNEL);
 
-  ExpectTensorNear<float, T>(expected, *net.GetOutput("OPENCLOutput"), 0.01);
+
+  if (DataTypeToEnum<T>::value == DT_HALF) {
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-3, 1e-3);
+  } else {
+    ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                            1e-5);
+  }
 }
 }  // namespace
 
@@ -448,7 +459,7 @@ void AvgPoolingNEONTest(const std::vector<index_t> &shape,
 
   ExpectTensorNear<float>(*net.GetOutput("OutputExptected"),
                           *net.GetOutput("OutputNeon"),
-                          0.01);
+                          1e-5, 1e-4);
 }
 }  // namespace
 
diff --git a/mace/ops/resize_bilinear_test.cc b/mace/ops/resize_bilinear_test.cc
index cd1b81d0eec455c8554adaa5a46807501144cf15..3629ed01dcfcf350f84f16c7872c2cc6283ac842 100644
--- a/mace/ops/resize_bilinear_test.cc
+++ b/mace/ops/resize_bilinear_test.cc
@@ -35,7 +35,7 @@ TEST_F(ResizeBilinearTest, CPUResizeBilinearWOAlignCorners) {
   // Check
   auto expected = CreateTensor<float>({1, 1, 2, 3}, {0, 1, 2, 6, 7, 8});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 TEST_F(ResizeBilinearTest, ResizeBilinearWAlignCorners) {
@@ -60,7 +60,7 @@ TEST_F(ResizeBilinearTest, ResizeBilinearWAlignCorners) {
   // Check
   auto expected = CreateTensor<float>({1, 1, 2, 3}, {0, 1, 2, 9, 10, 11});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
 namespace {
@@ -113,7 +113,8 @@ void TestRandomResizeBilinear() {
       // TODO(someone): support NEON
     }
     // Check
-    ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"), 0.001);
+    ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"),
+                            1e-5, 1e-6);
   }
 }
 }  // namespace
diff --git a/mace/ops/softmax_test.cc b/mace/ops/softmax_test.cc
index 156b84d3ef302662fb419d933d9634f5d40f507f..e8a7c48ea42758eb1fd6ffdc1064ea44c453a795 100644
--- a/mace/ops/softmax_test.cc
+++ b/mace/ops/softmax_test.cc
@@ -49,7 +49,7 @@ void Simple() {
     {1, 1, 2, 4},
     {0.25, 0.25, 0.25, 0.25, 0.0320586, 0.08714432, 0.23688282, 0.64391426});
 
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-7);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 }  // namespace
 
@@ -89,7 +89,8 @@ void Complex(const std::vector<index_t> &logits_shape) {
   ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
                           kernels::BufferType::IN_OUT_CHANNEL);
 
-  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5);
+  ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
+                          1e-5);
 }
 }  // namespace
 
@@ -138,7 +139,7 @@ void SoftMaxNEONTest(const std::vector<index_t> &logits_shape) {
 
   ExpectTensorNear<float>(*net.GetOutput("OutputExptected"),
                           *net.GetOutput("OutputNeon"),
-                          0.01);
+                          1e-5, 1e-5);
 }
 }  // namespace
 
diff --git a/mace/ops/space_to_batch_test.cc b/mace/ops/space_to_batch_test.cc
index a62b3493026615a8976b0d0dfcf23d69f5759ae2..a9ae67aca875efc7f98df8c0a744407c77d8ecf1 100644
--- a/mace/ops/space_to_batch_test.cc
+++ b/mace/ops/space_to_batch_test.cc
@@ -36,7 +36,7 @@ void RunSpaceToBatch(const std::vector<index_t> &input_shape,
   ImageToBuffer<D, float>(&net, "OutputImage", "Output",
                           kernels::BufferType::IN_OUT_CHANNEL);
   // Check
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-8);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"));
 }
 
 template <DeviceType D>
@@ -64,7 +64,7 @@ void RunBatchToSpace(const std::vector<index_t> &input_shape,
   ImageToBuffer<D, float>(&net, "OutputImage", "Output",
                           kernels::BufferType::IN_OUT_CHANNEL);
   // Check
-  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-8);
+  ExpectTensorNear<float>(*expected, *net.GetOutput("Output"));
 }
 
 template <typename T>
diff --git a/mace/ops/transpose_test.cc b/mace/ops/transpose_test.cc
index bd61e132b5659cbd865a63d9037dc9f1e05c0cbc..929ac2c0b6d317a9ab40f1a8831ea07d8dbbaf92 100644
--- a/mace/ops/transpose_test.cc
+++ b/mace/ops/transpose_test.cc
@@ -30,8 +30,7 @@ void TransposeNCHWTest(const std::vector<index_t> &input_shape) {
   net.FillNHWCInputToNCHWInput<DeviceType::CPU, float>("InputNCHW", "Input");
 
   ExpectTensorNear<float>(*net.GetOutput("InputNCHW"),
-                          *net.GetOutput("Output"),
-                          0.01);
+                          *net.GetOutput("Output"));
 }
 }  // namespace
 
diff --git a/mace/ops/winograd_convolution_test.cc b/mace/ops/winograd_convolution_test.cc
index ce9bcdca6b27f9ef34cb73130db79b50c1db7cc9..409f9cb3e20e3eb3e16cd269dd53e06b6e0e22f5 100644
--- a/mace/ops/winograd_convolution_test.cc
+++ b/mace/ops/winograd_convolution_test.cc
@@ -127,9 +127,11 @@ void WinogradConvolution(const index_t batch,
   ImageToBuffer<D, float>(&net, "WinoOutputImage", "WinoOutput",
                           kernels::BufferType::IN_OUT_CHANNEL);
   if (DataTypeToEnum<T>::value == DataType::DT_HALF) {
-    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"), 1e-1);
+    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"),
+                            1e-2, 1e-2);
   } else {
-    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"), 1e-4);
+    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"),
+                            1e-5, 1e-4);
   }
 }
 }  // namespace
@@ -246,9 +248,11 @@ void WinogradConvolutionWithPad(const index_t batch,
   ImageToBuffer<D, float>(&net, "WinoOutputImage", "WinoOutput",
                           kernels::BufferType::IN_OUT_CHANNEL);
   if (DataTypeToEnum<T>::value == DataType::DT_HALF) {
-    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"), 1e-1);
+    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"),
+                            1e-2, 1e-2);
   } else {
-    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"), 1e-3);
+    ExpectTensorNear<float>(expected, *net.GetOutput("WinoOutput"),
+                            1e-5, 1e-4);
   }
 }
 }  // namespace