diff --git a/lite/CMakeLists.txt b/lite/CMakeLists.txt
index ebb761601273edd2f83f684c74852ff3dea6811c..df6b7d3648409e13d88c049ec86173905f8b3cb6 100644
--- a/lite/CMakeLists.txt
+++ b/lite/CMakeLists.txt
@@ -179,8 +179,10 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
                 add_dependencies(tiny_publish_cxx_lib paddle_light_api_shared)
                 add_dependencies(tiny_publish_cxx_lib bundle_light_api)
                 add_dependencies(publish_inference tiny_publish_cxx_lib)
-                add_custom_command(TARGET tiny_publish_cxx_lib POST_BUILD
-                            COMMAND ${CMAKE_STRIP} "-s" ${INFER_LITE_PUBLISH_ROOT}/cxx/lib/libpaddle_light_api_shared.so)
+                if(NOT "${CMAKE_BUILD_TYPE}" STREQUAL "Debug")
+                    add_custom_command(TARGET tiny_publish_cxx_lib POST_BUILD
+                                COMMAND ${CMAKE_STRIP} "-s" ${INFER_LITE_PUBLISH_ROOT}/cxx/lib/libpaddle_light_api_shared.so)
+                endif()
             endif()
         endif()
     endif()
diff --git a/lite/api/paddle_place.cc b/lite/api/paddle_place.cc
index 894d839185ea9e1b6b47b87c398f249f044c2b51..6d12df67ac70d5d922680fc76763123117045175 100644
--- a/lite/api/paddle_place.cc
+++ b/lite/api/paddle_place.cc
@@ -77,7 +77,8 @@ const std::string& PrecisionToStr(PrecisionType precision) {
 }
 
 const std::string& DataLayoutToStr(DataLayoutType layout) {
-  static const std::string datalayout2string[] = {"unk", "NCHW", "any", "NHWC"};
+  static const std::string datalayout2string[] = {
+      "unk", "NCHW", "any", "NHWC", "ImageDefault", "ImageFolder", "ImageNW"};
   auto x = static_cast<int>(layout);
   CHECK_LT(x, static_cast<int>(DATALAYOUT(NUM)));
   return datalayout2string[x];
@@ -115,8 +116,13 @@ const std::string& PrecisionRepr(PrecisionType precision) {
 }
 
 const std::string& DataLayoutRepr(DataLayoutType layout) {
-  static const std::string datalayout2string[] = {
-      "kUnk", "kNCHW", "kAny", "kNHWC"};
+  static const std::string datalayout2string[] = {"kUnk",
+                                                  "kNCHW",
+                                                  "kAny",
+                                                  "kNHWC",
+                                                  "kImageDefault",
+                                                  "kImageFolder",
+                                                  "kImageNW"};
   auto x = static_cast<int>(layout);
   CHECK_LT(x, static_cast<int>(DATALAYOUT(NUM)));
   return datalayout2string[x];
@@ -146,8 +152,12 @@ std::set<PrecisionType> ExpandValidPrecisions(PrecisionType precision) {
 }
 
 std::set<DataLayoutType> ExpandValidLayouts(DataLayoutType layout) {
-  static const std::set<DataLayoutType> valid_set(
-      {DATALAYOUT(kNCHW), DATALAYOUT(kAny), DATALAYOUT(kNHWC)});
+  static const std::set<DataLayoutType> valid_set({DATALAYOUT(kNCHW),
+                                                   DATALAYOUT(kAny),
+                                                   DATALAYOUT(kNHWC),
+                                                   DATALAYOUT(kImageDefault),
+                                                   DATALAYOUT(kImageFolder),
+                                                   DATALAYOUT(kImageNW)});
   if (layout == DATALAYOUT(kAny)) {
     return valid_set;
   }
diff --git a/lite/api/paddle_place.h b/lite/api/paddle_place.h
index 07284be095c05e5dfa069b0973d5982cf1f07c8a..1aa41522352e9c2832e3c9919249887480e871a3 100644
--- a/lite/api/paddle_place.h
+++ b/lite/api/paddle_place.h
@@ -71,8 +71,11 @@ enum class DataLayoutType : int {
   kUnk = 0,
   kNCHW = 1,
   kNHWC = 3,
-  kAny = 2,  // any data layout
-  NUM = 4,   // number of fields.
+  kImageDefault = 4,  // for opencl image2d
+  kImageFolder = 5,   // for opencl image2d
+  kImageNW = 6,       // for opencl image2d
+  kAny = 2,           // any data layout
+  NUM = 7,            // number of fields.
 };
 
 typedef enum {
diff --git a/lite/api/python/pybind/pybind.cc b/lite/api/python/pybind/pybind.cc
index 2df2e8f8f8aa56bb71b0e1cb293df2ecbbafd0bb..7d4ed4e98701a5328b0f05387dc73ad8b93dfe18 100644
--- a/lite/api/python/pybind/pybind.cc
+++ b/lite/api/python/pybind/pybind.cc
@@ -165,6 +165,9 @@ void BindLitePlace(py::module *m) {
   py::enum_<DataLayoutType>(*m, "DataLayoutType")
       .value("NCHW", DataLayoutType::kNCHW)
       .value("NHWC", DataLayoutType::kNHWC)
+      .value("ImageDefault", DataLayoutType::kImageDefault)
+      .value("ImageFolder", DataLayoutType::kImageFolder)
+      .value("ImageNW", DataLayoutType::kImageNW)
       .value("Any", DataLayoutType::kAny);
 
   // Place
diff --git a/lite/core/op_registry.cc b/lite/core/op_registry.cc
index 490a1e116d2d1596adf2ed8a3cd1ca6a9b2847f4..716ce9d6a82b07270b5029f4cddf6a6b808c6c21 100644
--- a/lite/core/op_registry.cc
+++ b/lite/core/op_registry.cc
@@ -40,6 +40,18 @@ std::list<std::unique_ptr<KernelBase>> KernelRegistry::Create(
       return Create<TARGET(target__),                                        \
                     PRECISION(precision__),                                  \
                     DATALAYOUT(kNHWC)>(op_type);                             \
+    case DATALAYOUT(kImageDefault):                                          \
+      return Create<TARGET(target__),                                        \
+                    PRECISION(precision__),                                  \
+                    DATALAYOUT(kImageDefault)>(op_type);                     \
+    case DATALAYOUT(kImageFolder):                                           \
+      return Create<TARGET(target__),                                        \
+                    PRECISION(precision__),                                  \
+                    DATALAYOUT(kImageFolder)>(op_type);                      \
+    case DATALAYOUT(kImageNW):                                               \
+      return Create<TARGET(target__),                                        \
+                    PRECISION(precision__),                                  \
+                    DATALAYOUT(kImageNW)>(op_type);                          \
     default:                                                                 \
       LOG(FATAL) << "unsupported kernel layout " << DataLayoutToStr(layout); \
   }
@@ -147,6 +159,17 @@ KernelRegistry::KernelRegistry()
   INIT_FOR(kOpenCL, kFloat, kAny);
   INIT_FOR(kOpenCL, kInt8, kNCHW);
   INIT_FOR(kOpenCL, kAny, kAny);
+  INIT_FOR(kOpenCL, kFP16, kNCHW);
+  INIT_FOR(kOpenCL, kFP16, kNHWC);
+  INIT_FOR(kOpenCL, kFP16, kImageDefault);
+  INIT_FOR(kOpenCL, kFP16, kImageFolder);
+  INIT_FOR(kOpenCL, kFP16, kImageNW);
+  INIT_FOR(kOpenCL, kFloat, kImageDefault);
+  INIT_FOR(kOpenCL, kFloat, kImageFolder);
+  INIT_FOR(kOpenCL, kFloat, kImageNW);
+  INIT_FOR(kOpenCL, kAny, kImageDefault);
+  INIT_FOR(kOpenCL, kAny, kImageFolder);
+  INIT_FOR(kOpenCL, kAny, kImageNW);
 
   INIT_FOR(kNPU, kFloat, kNCHW);
   INIT_FOR(kNPU, kInt8, kNCHW);
diff --git a/lite/core/op_registry.h b/lite/core/op_registry.h
index 81dba38d934ac6a3d21cf2c3a7e154f0e5cb91ce..0df5cb41ecc4c631e8540f9595c3182122b99f5f 100644
--- a/lite/core/op_registry.h
+++ b/lite/core/op_registry.h
@@ -176,6 +176,39 @@ class KernelRegistry final {
               KernelRegistryForTarget<TARGET(kOpenCL),
                                       PRECISION(kAny),
                                       DATALAYOUT(kAny)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFP16),
+                                      DATALAYOUT(kNCHW)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFP16),
+                                      DATALAYOUT(kNHWC)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFP16),
+                                      DATALAYOUT(kImageDefault)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFP16),
+                                      DATALAYOUT(kImageFolder)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFP16),
+                                      DATALAYOUT(kImageNW)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFloat),
+                                      DATALAYOUT(kImageDefault)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFloat),
+                                      DATALAYOUT(kImageFolder)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kFloat),
+                                      DATALAYOUT(kImageNW)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kAny),
+                                      DATALAYOUT(kImageDefault)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kAny),
+                                      DATALAYOUT(kImageFolder)> *,  //
+              KernelRegistryForTarget<TARGET(kOpenCL),
+                                      PRECISION(kAny),
+                                      DATALAYOUT(kImageNW)> *,  //
 
               KernelRegistryForTarget<TARGET(kNPU),
                                       PRECISION(kAny),