diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3c29745fe94193518e0d872140adea40d658dce9..4fc705a53d1b8c266d1c6104982fda5caf33426a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -9,6 +9,7 @@ option(MACE_ENABLE_OPENCL      "whether to enable OpenCL support"           OFF)
 option(MACE_ENABLE_CUDA        "whether to enable CUDA support"             OFF)
 option(MACE_ENABLE_HEXAGON_DSP "whether to enable Hexagon DSP support"      OFF)
 option(MACE_ENABLE_HEXAGON_HTA "whether to enable Hexagon HTA support"      OFF)
+option(MACE_ENABLE_MTK_APU     "whether to enable MTK APU support"          OFF)
 option(MACE_ENABLE_TESTS       "whether to build c++ unit tests"            OFF)
 option(MACE_ENABLE_BENCHMARKS  "whether to build c++ micro benchmarks"      OFF)
 option(MACE_ENABLE_EXAMPLES    "whether to build examples"                  OFF)
@@ -89,6 +90,13 @@ if(MACE_ENABLE_HEXAGON_HTA)
   add_definitions(-DMACE_ENABLE_HEXAGON_HTA)
 endif(MACE_ENABLE_HEXAGON_HTA)
 
+if(MACE_ENABLE_MTK_APU)
+  if(NOT ANDROID)
+    message(FATAL_ERROR "MTK APU is only supported on Android")
+  endif(NOT ANDROID)
+  add_definitions(-DMACE_ENABLE_MTK_APU)
+endif(MACE_ENABLE_MTK_APU)
+
 if(MACE_ENABLE_OBFUSCATE)
   add_definitions(-DMACE_OBFUSCATE_LITERALS)
 endif(MACE_ENABLE_OBFUSCATE)
diff --git a/mace/core/CMakeLists.txt b/mace/core/CMakeLists.txt
index e4b3a0e06d9ea1ebb35b062b113bec91dd1d833c..c9b18826ee5ee409db0da8c76837a2dcea76b7b7 100644
--- a/mace/core/CMakeLists.txt
+++ b/mace/core/CMakeLists.txt
@@ -40,6 +40,11 @@ if(MACE_ENABLE_HEXAGON_HTA)
   set(EXTRA_LINK_LIBS ${EXTRA_LINK_LIBS} hta_controller hta_hexagon_runtime npu)
 endif(MACE_ENABLE_HEXAGON_HTA)
 
+if(MACE_ENABLE_MTK_APU)
+  set(CORE_SRCS ${CORE_SRCS} runtime/apu/apu_wrapper.cc)
+  set(EXTRA_LINK_LIBS ${EXTRA_LINK_LIBS} apu-frontend)
+endif(MACE_ENABLE_MTK_APU)
+
 add_library(core STATIC ${CORE_SRCS})
 target_link_libraries(core PRIVATE
   proto
diff --git a/mace/python/tools/converter_tool/apu_converter.py b/mace/python/tools/converter_tool/apu_converter.py
index 5fbf2da7ce9c993a9fbcc18392a59314913c2a4f..5de7a6fc4b8530c56e3a47841ed78eec4c080787 100644
--- a/mace/python/tools/converter_tool/apu_converter.py
+++ b/mace/python/tools/converter_tool/apu_converter.py
@@ -397,12 +397,13 @@ class ApuConverter(base_converter.ConverterInterface):
             _op.input.extend([tensor.name])
 
     def use_uint8_in_out(self):
+        replace_dict = {}
         for input_info in self._model.input_info:
             if input_info.data_type == mace_pb2.DT_FLOAT:
                 for op in self._model.op:
                     if op.input[0] == input_info.name \
                            and op.type == MaceOp.Quantize.name:
-                        input_info.name = op.output[0]
+                        replace_dict[op.output[0]] = input_info.name
                         input_info.scale = op.quantize_info[0].scale
                         input_info.zero_point = op.quantize_info[0].zero_point
                         break
@@ -412,6 +413,13 @@ class ApuConverter(base_converter.ConverterInterface):
                 for op in self._model.op:
                     if op.output[0] == output_info.name \
                            and op.type == MaceOp.Dequantize.name:
-                        output_info.name = op.input[0]
+                        replace_dict[op.input[0]] = output_info.name
                         break
                 self._model.op.remove(op)
+        for op in self._model.op:
+            for i, op_input in enumerate(op.input):
+                if op_input in replace_dict.keys():
+                    op.input[i] = replace_dict[op_input]
+            for i, op_output in enumerate(op.output):
+                if op_output in replace_dict.keys():
+                    op.output[i] = replace_dict[op_output]
diff --git a/third_party/apu/apu.cmake b/third_party/apu/apu.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..e476e7bec337e402160617c3a0e312115c384d83
--- /dev/null
+++ b/third_party/apu/apu.cmake
@@ -0,0 +1,13 @@
+set(APU_INSTALL_DIR  "${PROJECT_SOURCE_DIR}/third_party/apu")
+set(APU_INCLUDE_DIR  "${APU_INSTALL_DIR}")
+
+include_directories(SYSTEM "${APU_INCLUDE_DIR}")
+
+set(APU-FRONTEND
+  "${APU_INSTALL_DIR}/libapu-frontend.so"
+)
+
+add_library(apu-frontend SHARED IMPORTED GLOBAL)
+set_target_properties(apu-frontend PROPERTIES IMPORTED_LOCATION ${APU-FRONTEND})
+
+install(FILES ${APU-FRONTEND} DESTINATION lib)
diff --git a/third_party/third_party.cmake b/third_party/third_party.cmake
index b25a168977064ba22813d336e3dfdf9f01f30081..4317450c3a9e0e93aed438484548b1b15c6c1cb6 100644
--- a/third_party/third_party.cmake
+++ b/third_party/third_party.cmake
@@ -58,3 +58,7 @@ endif(MACE_ENABLE_HEXAGON_DSP)
 if(MACE_ENABLE_HEXAGON_HTA)
   include(${PROJECT_SOURCE_DIR}/third_party/hta/hta.cmake)
 endif(MACE_ENABLE_HEXAGON_HTA)
+
+if(MACE_ENABLE_MTK_APU)
+  include(${PROJECT_SOURCE_DIR}/third_party/apu/apu.cmake)
+endif(MACE_ENABLE_MTK_APU)
diff --git a/tools/bazel-build-standalone-lib.sh b/tools/bazel-build-standalone-lib.sh
index 2b7116e218d0a27da91cace868bbfc6fe9e1f20f..61ebd9d6f76fc3d89e93aff512e8e5d198c45780 100755
--- a/tools/bazel-build-standalone-lib.sh
+++ b/tools/bazel-build-standalone-lib.sh
@@ -19,6 +19,7 @@ mkdir -p $LIB_DIR/armeabi-v7a/cpu_gpu
 rm -rf $LIB_DIR/arm64-v8a
 mkdir -p $LIB_DIR/arm64-v8a/cpu_gpu_dsp
 mkdir -p $LIB_DIR/arm64-v8a/cpu_gpu
+mkdir -p $LIB_DIR/arm64-v8a/cpu_gpu_apu
 
 rm -rf $LIB_DIR/linux-x86-64
 mkdir -p $LIB_DIR/linux-x86-64
@@ -42,6 +43,11 @@ bazel build --config android --config optimization mace/libmace:libmace_dynamic
 cp bazel-bin/mace/libmace/libmace.so $LIB_DIR/arm64-v8a/cpu_gpu_dsp/
 cp third_party/nnlib/arm64-v8a/*so $LIB_DIR/arm64-v8a/cpu_gpu_dsp/
 
+echo "build shared lib for arm64-v8a + cpu_gpu_apu"
+bazel build --config android --config optimization mace/libmace:libmace_dynamic --define neon=true --define openmp=false --define opencl=true --define apu=true --define quantize=true --cpu=arm64-v8a
+cp bazel-bin/mace/libmace/libmace.so $LIB_DIR/arm64-v8a/cpu_gpu_apu/
+cp third_party/apu/*so $LIB_DIR/arm64-v8a/cpu_gpu_apu/
+
 echo "build shared lib for armeabi-v7a + cpu_gpu"
 bazel build --config android --config optimization mace/libmace:libmace_dynamic --define neon=true --define openmp=false --define opencl=true --define quantize=true --cpu=armeabi-v7a
 cp bazel-bin/mace/libmace/libmace.so $LIB_DIR/armeabi-v7a/cpu_gpu/
@@ -75,6 +81,11 @@ bazel build --config android --config optimization mace/libmace:libmace_static -
 cp bazel-genfiles/mace/libmace/libmace.a $LIB_DIR/arm64-v8a/cpu_gpu_dsp/
 cp third_party/nnlib/arm64-v8a/*so $LIB_DIR/arm64-v8a/cpu_gpu_dsp/
 
+echo "build static lib for arm64-v8a + cpu_gpu_apu"
+bazel build --config android --config optimization mace/libmace:libmace_static --config symbol_hidden --define neon=true --define openmp=false --define opencl=true --define apu=true --define quantize=true --cpu=arm64-v8a
+cp bazel-genfiles/mace/libmace/libmace.a $LIB_DIR/arm64-v8a/cpu_gpu_apu/
+cp third_party/apu/*so $LIB_DIR/arm64-v8a/cpu_gpu_apu/
+
 echo "build static lib for armeabi-v7a + cpu_gpu"
 bazel build --config android --config optimization mace/libmace:libmace_static --config symbol_hidden --define neon=true --define openmp=false --define opencl=true --define quantize=true --cpu=armeabi-v7a
 cp bazel-genfiles/mace/libmace/libmace.a $LIB_DIR/armeabi-v7a/cpu_gpu/
diff --git a/tools/cmake-build-android-arm64-v8a-cpu.sh b/tools/cmake-build-android-arm64-v8a-cpu.sh
index dea3fe5d1c7d9798d69277173618ce6368f8c5a3..57d2a3607c402dc8c6dbde6bb1d79124fe5cc5ae 100755
--- a/tools/cmake-build-android-arm64-v8a-cpu.sh
+++ b/tools/cmake-build-android-arm64-v8a-cpu.sh
@@ -15,6 +15,7 @@ cmake -DANDROID_ABI="arm64-v8a" \
       -DMACE_ENABLE_OPENCL=OFF            \
       -DMACE_ENABLE_HEXAGON_DSP=OFF       \
       -DMACE_ENABLE_HEXAGON_HTA=OFF       \
+      -DMACE_ENABLE_MTK_APU=OFF           \
       -DMACE_ENABLE_OPT_SIZE=ON           \
       -DMACE_ENABLE_OBFUSCATE=ON          \
       -DCMAKE_INSTALL_PREFIX=install      \
diff --git a/tools/cmake-build-android-arm64-v8a-full.sh b/tools/cmake-build-android-arm64-v8a-full.sh
index b51ec2004ed33b179a87771ccf24a388943a16db..3db663553cb13a173c39a350bcecd8ef1bef71b1 100755
--- a/tools/cmake-build-android-arm64-v8a-full.sh
+++ b/tools/cmake-build-android-arm64-v8a-full.sh
@@ -15,6 +15,7 @@ cmake -DANDROID_ABI="arm64-v8a" \
       -DMACE_ENABLE_OPENCL=ON             \
       -DMACE_ENABLE_HEXAGON_DSP=ON        \
       -DMACE_ENABLE_HEXAGON_HTA=ON        \
+      -DMACE_ENABLE_MTK_APU=ON            \
       -DMACE_ENABLE_OPT_SIZE=ON           \
       -DMACE_ENABLE_OBFUSCATE=ON          \
       -DCMAKE_INSTALL_PREFIX=install      \
diff --git a/tools/cmake-build-android-armeabi-v7a-cpu.sh b/tools/cmake-build-android-armeabi-v7a-cpu.sh
index bbea9d4c8eb0f918844ad62360d25f61446e2d31..4018b8d5ba9b00405e79864be4f1822b4e64dd36 100755
--- a/tools/cmake-build-android-armeabi-v7a-cpu.sh
+++ b/tools/cmake-build-android-armeabi-v7a-cpu.sh
@@ -16,6 +16,7 @@ cmake -DANDROID_ABI="armeabi-v7a" \
       -DMACE_ENABLE_OPENCL=OFF            \
       -DMACE_ENABLE_HEXAGON_DSP=OFF       \
       -DMACE_ENABLE_HEXAGON_HTA=OFF       \
+      -DMACE_ENABLE_MTK_APU=OFF           \
       -DMACE_ENABLE_OPT_SIZE=ON           \
       -DMACE_ENABLE_OBFUSCATE=ON          \
       -DCMAKE_INSTALL_PREFIX=install      \
diff --git a/tools/cmake-build-android-armeabi-v7a-full.sh b/tools/cmake-build-android-armeabi-v7a-full.sh
index 09e5c7f896d709b0e1e794e35f087c85be397ea8..6ac7f7f6e2db8a313fb7b1058fce8154bf31cd17 100755
--- a/tools/cmake-build-android-armeabi-v7a-full.sh
+++ b/tools/cmake-build-android-armeabi-v7a-full.sh
@@ -16,6 +16,7 @@ cmake -DANDROID_ABI="armeabi-v7a" \
       -DMACE_ENABLE_OPENCL=ON             \
       -DMACE_ENABLE_HEXAGON_DSP=ON        \
       -DMACE_ENABLE_HEXAGON_HTA=ON        \
+      -DMACE_ENABLE_MTK_APU=OFF           \
       -DMACE_ENABLE_OPT_SIZE=ON           \
       -DMACE_ENABLE_OBFUSCATE=ON          \
       -DCMAKE_INSTALL_PREFIX=install      \
diff --git a/tools/converter.py b/tools/converter.py
index 98fb6bc36b6b88973d16882ffb95ec4b82447f5d..aca556aedcc01c7c6bd9f78acb9c883030071918 100644
--- a/tools/converter.py
+++ b/tools/converter.py
@@ -189,6 +189,19 @@ def get_hta_mode(configs):
     return False
 
 
+def get_apu_mode(configs):
+    runtime_list = []
+    for model_name in configs[YAMLKeyword.models]:
+        model_runtime = \
+            configs[YAMLKeyword.models][model_name].get(
+                YAMLKeyword.runtime, "")
+        runtime_list.append(model_runtime.lower())
+
+    if RuntimeType.apu in runtime_list:
+        return True
+    return False
+
+
 def get_opencl_mode(configs):
     runtime_list = []
     for model_name in configs[YAMLKeyword.models]:
@@ -575,6 +588,7 @@ def format_model_config(flags):
                 DeviceType.GPU: ValidationThreshold.gpu_threshold,
                 DeviceType.HEXAGON: ValidationThreshold.quantize_threshold,
                 DeviceType.HTA: ValidationThreshold.quantize_threshold,
+                DeviceType.APU: ValidationThreshold.quantize_threshold,
                 DeviceType.QUANTIZE: ValidationThreshold.quantize_threshold,
             }
             for k, v in six.iteritems(validation_threshold):
@@ -816,6 +830,7 @@ def build_model_lib(configs, address_sanitizer, debug_mode):
             toolchain=toolchain,
             enable_hexagon=get_hexagon_mode(configs),
             enable_hta=get_hta_mode(configs),
+            enable_apu=get_apu_mode(configs),
             enable_opencl=get_opencl_mode(configs),
             enable_quantize=get_quantize_mode(configs),
             address_sanitizer=address_sanitizer,
@@ -884,6 +899,7 @@ def build_mace_run(configs, target_abi, toolchain, enable_openmp,
         toolchain=toolchain,
         enable_hexagon=get_hexagon_mode(configs),
         enable_hta=get_hta_mode(configs),
+        enable_apu=get_apu_mode(configs),
         enable_openmp=enable_openmp,
         enable_opencl=get_opencl_mode(configs),
         enable_quantize=get_quantize_mode(configs),
@@ -931,6 +947,7 @@ def build_example(configs, target_abi, toolchain, enable_openmp, mace_lib_type,
                             enable_quantize=get_quantize_mode(configs),
                             enable_hexagon=get_hexagon_mode(configs),
                             enable_hta=get_hta_mode(configs),
+                            enable_apu=get_apu_mode(configs),
                             address_sanitizer=flags.address_sanitizer,
                             symbol_hidden=get_symbol_hidden_mode(debug_mode, mace_lib_type),  # noqa
                             debug_mode=debug_mode)
@@ -964,6 +981,7 @@ def build_example(configs, target_abi, toolchain, enable_openmp, mace_lib_type,
                             enable_quantize=get_quantize_mode(configs),
                             enable_hexagon=get_hexagon_mode(configs),
                             enable_hta=get_hta_mode(configs),
+                            enable_apu=get_apu_mode(configs),
                             address_sanitizer=flags.address_sanitizer,
                             debug_mode=debug_mode,
                             extra_args=build_arg)
@@ -1072,6 +1090,7 @@ def build_benchmark_model(configs,
                             enable_quantize=get_quantize_mode(configs),
                             enable_hexagon=get_hexagon_mode(configs),
                             enable_hta=get_hta_mode(configs),
+                            enable_apu=get_apu_mode(configs),
                             symbol_hidden=get_symbol_hidden_mode(debug_mode, mace_lib_type),  # noqa
                             debug_mode=debug_mode,
                             extra_args=build_arg)
diff --git a/tools/device.py b/tools/device.py
index 22648ad9a0806ca914717778865ba9453ef91f8f..ca21c285f0019a842c39f7f47a264d0e6e9445f5 100644
--- a/tools/device.py
+++ b/tools/device.py
@@ -280,6 +280,10 @@ class DeviceWrapper:
                     "third_party/nnlib/%s/libhexagon_controller.so" % abi,
                     self.data_dir)
 
+            if device_type == common.DeviceType.APU:
+                self.push("third_party/apu/libapu-frontend.so",
+                          self.data_dir)
+
             mace_model_phone_path = ""
             if model_graph_format == ModelFormat.file:
                 mace_model_phone_path = "%s/%s.pb" % (self.data_dir,
diff --git a/tools/sh_commands.py b/tools/sh_commands.py
index 172b8d44fb9c17914e719bec9fc811ac9820b3aa..e9d051b3c718e1621e00c7160567944c29b940ae 100644
--- a/tools/sh_commands.py
+++ b/tools/sh_commands.py
@@ -264,6 +264,7 @@ def bazel_build(target,
                 toolchain='android',
                 enable_hexagon=False,
                 enable_hta=False,
+                enable_apu=False,
                 enable_openmp=False,
                 enable_neon=True,
                 enable_opencl=True,
@@ -303,8 +304,9 @@ def bazel_build(target,
             "--define",
             "hexagon=%s" % str(enable_hexagon).lower(),
             "--define",
-            "hta=%s" % str(enable_hta).lower())
-
+            "hta=%s" % str(enable_hta).lower(),
+            "--define",
+            "apu=%s" % str(enable_apu).lower())
     if address_sanitizer:
         bazel_args += ("--config", "asan")
     if debug_mode:
@@ -322,6 +324,7 @@ def bazel_build(target,
     sh.bazel(
         _fg=True,
         *bazel_args)
+    six.print_(bazel_args)
     six.print_("Build done!\n")
 
 
@@ -881,6 +884,7 @@ def build_run_throughput_test(abi,
                               cpu_model_tag,
                               gpu_model_tag,
                               dsp_model_tag,
+                              apu_model_tag,
                               phone_data_dir,
                               strip="always",
                               input_file_name="model_input"):
@@ -896,7 +900,9 @@ def build_run_throughput_test(abi,
     if dsp_model_tag:
         model_tag_build_flag += "--copt=-DMACE_DSP_MODEL_TAG=%s " % \
                                 dsp_model_tag
-
+    if apu_model_tag:
+        model_tag_build_flag += "--copt=-DMACE_APU_MODEL_TAG=%s " % \
+                                apu_model_tag
     sh.cp("-f", merged_lib_file, "mace/benchmark/libmace_merged.a")
     sh.bazel(
         "build",
@@ -948,7 +954,10 @@ def build_run_throughput_test(abi,
     adb_push("third_party/nnlib/%s/libhexagon_controller.so" % abi,
              phone_data_dir,
              serialno)
-
+    if apu_model_tag:
+        adb_push("third_party/apu/libapu-frontend.so",
+                 phone_data_dir,
+                 serialno)
     sh.adb(
         "-s",
         serialno,