diff --git a/lite/api/mobilenetv1_test.cc b/lite/api/mobilenetv1_test.cc
index 085f7f3ad7101a59b8035ac3a8ad8a1e602fb102..bcc9644f81542ab6fb8a0badf8ecaea89fc8dedb 100644
--- a/lite/api/mobilenetv1_test.cc
+++ b/lite/api/mobilenetv1_test.cc
@@ -23,6 +23,10 @@
 #include "lite/core/op_registry.h"
 
 DEFINE_string(optimized_model, "", "optimized_model");
+DEFINE_int32(N, 1, "input_batch");
+DEFINE_int32(C, 3, "input_channel");
+DEFINE_int32(H, 224, "input_height");
+DEFINE_int32(W, 224, "input_width");
 
 namespace paddle {
 namespace lite {
@@ -37,7 +41,8 @@ void TestModel(const std::vector<Place>& valid_places,
   predictor.Build(model_dir, "", "", valid_places);
 
   auto* input_tensor = predictor.GetInput(0);
-  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
+  input_tensor->Resize(DDim(
+      std::vector<DDim::value_type>({FLAGS_N, FLAGS_C, FLAGS_H, FLAGS_W})));
   auto* data = input_tensor->mutable_data<float>();
   auto item_size = input_tensor->dims().production();
   for (int i = 0; i < item_size; i++) {
@@ -58,6 +63,8 @@ void TestModel(const std::vector<Place>& valid_places,
     predictor.SaveModel(FLAGS_optimized_model);
   }
 
+  LOG(INFO) << "input shape(NCHW):" << FLAGS_N << " " << FLAGS_C << " "
+            << FLAGS_H << " " << FLAGS_W;
   LOG(INFO) << "================== Speed Report ===================";
   LOG(INFO) << "Model: " << model_dir << ", threads num " << FLAGS_threads
             << ", warmup: " << FLAGS_warmup << ", repeats: " << FLAGS_repeats
diff --git a/lite/api/mobilenetv2_test.cc b/lite/api/mobilenetv2_test.cc
index 84bd27e352f549d619cfa51f9127f973023e6d45..012d6d48d9e6d3747f83a7f1089944bbaf359f71 100644
--- a/lite/api/mobilenetv2_test.cc
+++ b/lite/api/mobilenetv2_test.cc
@@ -23,6 +23,10 @@
 #include "lite/core/op_registry.h"
 
 DEFINE_string(optimized_model, "", "optimized_model");
+DEFINE_int32(N, 1, "input_batch");
+DEFINE_int32(C, 3, "input_channel");
+DEFINE_int32(H, 224, "input_height");
+DEFINE_int32(W, 224, "input_width");
 
 namespace paddle {
 namespace lite {
@@ -38,7 +42,8 @@ void TestModel(const std::vector<Place>& valid_places,
   predictor.Build(model_dir, "", "", valid_places);
 
   auto* input_tensor = predictor.GetInput(0);
-  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
+  input_tensor->Resize(DDim(
+      std::vector<DDim::value_type>({FLAGS_N, FLAGS_C, FLAGS_H, FLAGS_W})));
   auto* data = input_tensor->mutable_data<float>();
   auto item_size = input_tensor->dims().production();
   for (int i = 0; i < item_size; i++) {
@@ -59,6 +64,8 @@ void TestModel(const std::vector<Place>& valid_places,
     predictor.SaveModel(FLAGS_optimized_model);
   }
 
+  LOG(INFO) << "input shape(NCHW):" << FLAGS_N << " " << FLAGS_C << " "
+            << FLAGS_H << " " << FLAGS_W;
   LOG(INFO) << "================== Speed Report ===================";
   LOG(INFO) << "Model: " << model_dir << ", threads num " << FLAGS_threads
             << ", warmup: " << FLAGS_warmup << ", repeats: " << FLAGS_repeats
@@ -123,8 +130,11 @@ TEST(MobileNetV2, test_arm) {
 #ifdef LITE_WITH_OPENCL
 TEST(MobileNetV2, test_opencl) {
   std::vector<Place> valid_places({
-      Place{TARGET(kOpenCL), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kFloat)},
+      Place{TARGET(kOpenCL), PRECISION(kFloat), DATALAYOUT(kImageDefault)},
+      Place{TARGET(kOpenCL), PRECISION(kFloat), DATALAYOUT(kNCHW)},
+      Place{TARGET(kOpenCL), PRECISION(kAny), DATALAYOUT(kImageDefault)},
+      Place{TARGET(kOpenCL), PRECISION(kAny), DATALAYOUT(kNCHW)},
+      TARGET(kARM),  // enable kARM CPU kernel when no opencl kernel
   });
 
   TestModel(valid_places);
diff --git a/lite/kernels/opencl/conv_compute.cc b/lite/kernels/opencl/conv_compute.cc
index a9cfb32aa5b141b4e3b3c7d28d2d3694524fa34c..d00101552d4376bc4ac2a176016c1a9a449c35a7 100644
--- a/lite/kernels/opencl/conv_compute.cc
+++ b/lite/kernels/opencl/conv_compute.cc
@@ -362,6 +362,9 @@ void ConvImageCompute::PrepareForRun() {
         filter_image_dims[0], filter_image_dims[1], filter_image_v.data());
 
     impl_ = &ConvImageCompute::Conv2d1x1;
+#if 1  // TODO(ysh329): enable general dwconv
+  } else if (filter_dims[1] == 1 && x_dims[1] == output_dims[1]) {
+#else  // TODO(ysh329): remove dwconv3x3s1 and dwconv3x3 temporarily, need fix
   } else if (filter_dims[1] == 1 && x_dims[1] == output_dims[1] &&
              kernel_h == 3 && kernel_w == 3 && groups > 1) {
     // depth_conv2d_3x3s1, depth_conv2d_3x3
@@ -374,7 +377,7 @@ void ConvImageCompute::PrepareForRun() {
     }
     kernel_func_paths_.push_back("image/depthwise_conv2d_kernel.cl");
 
-    CLImageConverterDWBlock converter;
+    CLImageConverterNWBlock converter;
     const DDim& filter_image_dims = converter.InitImageDimInfoWith(filter_dims);
     std::vector<float> filter_image_v(filter_image_dims[0] *
                                       filter_image_dims[1] * 4);  // 4 : RGBA
@@ -382,12 +385,13 @@ void ConvImageCompute::PrepareForRun() {
     filter_gpu_image_.mutable_data<float, cl::Image2D>(
         filter_image_dims[0], filter_image_dims[1], filter_image_v.data());
   } else if (filter_dims[1] == 1 && x_dims[1] == output_dims[1] &&
-             kernel_h != 3 && groups > 1) {
+             kernel_h != 3) {
+#endif
     // depth_conv2d
     kernel_func_names_.push_back("depth_conv2d");
     kernel_func_paths_.push_back("image/depthwise_conv2d_basic_kernel.cl");
 
-    CLImageConverterDWBlock converter;
+    CLImageConverterNWBlock converter;
     const DDim& filter_image_dims = converter.InitImageDimInfoWith(filter_dims);
     std::vector<float> filter_image_v(filter_image_dims[0] *
                                       filter_image_dims[1] * 4);  // 4 : RGBA