From a19e6251bc7d77014ce73e530c52a8f490828a39 Mon Sep 17 00:00:00 2001
From: ling <lingqiaomin.huawei.com>
Date: Wed, 16 Sep 2020 12:09:23 +0800
Subject: [PATCH] [MSLITE][Develop] modify optimize.so to sdot and fp16 so

---
 cmake/package_lite.cmake                      |  3 +-
 mindspore/lite/nnacl/CMakeLists.txt           | 65 +++++++++----------
 .../assembly/{opt => fp16}/ConvDwFp16Border.S |  0
 .../assembly/{opt => fp16}/ConvDwFp16Center.S |  0
 .../assembly/{opt => fp16}/ConvDwFp16Row.S    |  0
 .../{opt => fp16}/DeconvDwFp16Border.S        |  0
 .../{opt => fp16}/DeconvDwFp16Center.S        |  0
 .../assembly/{opt => fp16}/Float16ToFloat32.S |  0
 .../assembly/{opt => fp16}/Float32ToFloat16.S |  0
 .../{opt => fp16}/IndirectGemmFp16_16x8.S     |  0
 .../nnacl/assembly/{opt => fp16}/MatmulFp16.S |  0
 .../{opt => fp16}/PostFuncBiasReluC8Fp16.S    |  0
 mindspore/lite/nnacl/optimize/CMakeLists.txt  | 28 ++++++++
 mindspore/lite/nnacl/optimized_kernel.h       | 50 ++++++++++++--
 mindspore/lite/src/CMakeLists.txt             | 48 +++++++++-----
 mindspore/lite/src/kernel_registry.cc         |  6 ++
 .../src/runtime/kernel/arm/CMakeLists.txt     | 46 +++++--------
 .../kernel/arm/int8/opt_op_handler.cc}        |  8 +--
 mindspore/lite/test/run_benchmark_nets.sh     |  3 +-
 mindspore/lite/tools/converter/CMakeLists.txt |  1 -
 20 files changed, 165 insertions(+), 93 deletions(-)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/ConvDwFp16Border.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/ConvDwFp16Center.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/ConvDwFp16Row.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/DeconvDwFp16Border.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/DeconvDwFp16Center.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/Float16ToFloat32.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/Float32ToFloat16.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/IndirectGemmFp16_16x8.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/MatmulFp16.S (100%)
 rename mindspore/lite/nnacl/assembly/{opt => fp16}/PostFuncBiasReluC8Fp16.S (100%)
 create mode 100644 mindspore/lite/nnacl/optimize/CMakeLists.txt
 rename mindspore/lite/{nnacl/opt_op_handler.c => src/runtime/kernel/arm/int8/opt_op_handler.cc} (100%)

diff --git a/cmake/package_lite.cmake b/cmake/package_lite.cmake
index 7cfa003af6f..56ce91a28fa 100644
--- a/cmake/package_lite.cmake
+++ b/cmake/package_lite.cmake
@@ -58,7 +58,8 @@ if (PLATFORM_ARM64)
     install(FILES ${TOP_DIR}/mindspore/core/ir/dtype/type_id.h DESTINATION ${INC_DIR}/ir/dtype COMPONENT ${COMPONENT_NAME})
     install(DIRECTORY ${TOP_DIR}/mindspore/lite/include/ DESTINATION ${INC_DIR} COMPONENT ${COMPONENT_NAME} FILES_MATCHING PATTERN "*.h")
     install(DIRECTORY ${TOP_DIR}/mindspore/lite/schema/ DESTINATION ${INC_DIR}/schema COMPONENT ${COMPONENT_NAME} FILES_MATCHING PATTERN "*.h" PATTERN "inner" EXCLUDE)
-    install(FILES ${TOP_DIR}/mindspore/lite/build/nnacl/liboptimize.so DESTINATION ${LIB_DIR} COMPONENT ${COMPONENT_NAME})
+    install(FILES ${TOP_DIR}/mindspore/lite/build/src/libmindspore-lite-optimize.so DESTINATION ${LIB_DIR} COMPONENT ${COMPONENT_NAME})
+    install(FILES ${TOP_DIR}/mindspore/lite/build/src/libmindspore-lite-fp16.so DESTINATION ${LIB_DIR} COMPONENT ${COMPONENT_NAME})
     install(DIRECTORY ${TOP_DIR}/third_party/flatbuffers/include DESTINATION ${FLATBF_DIR} COMPONENT ${COMPONENT_NAME})
 elseif (PLATFORM_ARM32)
     install(FILES ${TOP_DIR}/mindspore/lite/build/src/libmindspore-lite.so DESTINATION ${LIB_DIR} COMPONENT ${COMPONENT_NAME})
diff --git a/mindspore/lite/nnacl/CMakeLists.txt b/mindspore/lite/nnacl/CMakeLists.txt
index 2c8baf93ba1..93d11c3e344 100644
--- a/mindspore/lite/nnacl/CMakeLists.txt
+++ b/mindspore/lite/nnacl/CMakeLists.txt
@@ -1,45 +1,42 @@
 project(nnacl)
 
 set(NNACL_DIR ${CMAKE_CURRENT_SOURCE_DIR})
-set(TOP_DIR ${CMAKE_CURRENT_SOURCE_DIR}/..)
 include_directories(NNACL_DIR)
 
-########################### optimized files ###########################
-file(GLOB OPTIMIZED_ASSEMBLY
-        ${NNACL_DIR}/assembly/opt/*.s
-        ${NNACL_DIR}/assembly/opt/*.S
-        )
+if (PLATFORM_ARM32 OR PLATFORM_ARM64)
+    if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math")
+    endif()
+endif ()
 
-file(GLOB FP16_SRC
-        ${NNACL_DIR}/fp16/*.c
-        ${TOP_DIR}/src/runtime/kernel/arm/fp16/*.cc
-        )
+########################### files ###########################
+file(GLOB KERNEL_SRC
+    ${NNACL_DIR}/*.c
+    ${NNACL_DIR}/fp32/*.c
+    ${NNACL_DIR}/int8/*.c
+    ${NNACL_DIR}/quantization/*.c
+)
 
-########################### share library build ########################
-set(OPTIMIZED_OPS ${NNACL_DIR}/opt_op_handler.c)
-
-set_property(SOURCE ${OPTIMIZED_ASSEMBLY} PROPERTY LANGUAGE C)
-list(APPEND OPTIMIZED_OPS ${OPTIMIZED_ASSEMBLY} ${FP16_SRC})
+if (SUPPORT_TRAIN)
+    file (GLOB TRAIN_SRC ${NNACL_DIR}/fp32_grad/*.c)
+endif()
 
 if (PLATFORM_ARM64)
-    string(REPLACE "-fvisibility=hidden" "-fvisibility=default" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv8.2-a+dotprod+fp16")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8.2-a+dotprod+fp16")
-    add_library(optimize SHARED ${OPTIMIZED_OPS})
-    target_link_libraries(
-            optimize
-            mindspore-lite
-    )
-    set_target_properties(optimize PROPERTIES CLEAN_DIRECT_OUTPUT 1)
+    file(GLOB ASSEMBLY_SRC ${NNACL_DIR}/assembly/arm64/*.S)
+    set_property(SOURCE ${ASSEMBLY_SRC} PROPERTY LANGUAGE C)
+endif()
 
-    if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
-        add_custom_command(TARGET optimize POST_BUILD
-            COMMAND ${ANDROID_NDK}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/aarch64-linux-android/bin/strip
-            ${TOP_DIR}/build/nnacl/liboptimize.so)
-    endif ()
+if (PLATFORM_ARM32)
+    file(GLOB ASSEMBLY_SRC ${NNACL_DIR}/assembly/arm32/*.S)
+    set_property(SOURCE ${ASSEMBLY_SRC} PROPERTY LANGUAGE C)
+endif()
 
-    add_custom_command(TARGET optimize POST_BUILD
-            COMMAND rm -rf ${TOP_DIR}/output/lib/liboptimize.so
-            COMMAND mkdir -pv ${TOP_DIR}/output/lib
-            COMMAND cp ${TOP_DIR}/build/nnacl/liboptimize.so ${TOP_DIR}/output/lib)
-endif ()
+########################### build nnacl static library ########################
+string(REPLACE "-fvisibility=hidden" "-fvisibility=default" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+add_library(nnacl STATIC ${KERNEL_SRC} ${TRAIN_SRC} ${ASSEMBLY_SRC})
+
+########################### arm64 build optimize library ########################
+if (PLATFORM_ARM64)
+    add_subdirectory(${NNACL_DIR}/optimize)
+endif()
\ No newline at end of file
diff --git a/mindspore/lite/nnacl/assembly/opt/ConvDwFp16Border.S b/mindspore/lite/nnacl/assembly/fp16/ConvDwFp16Border.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/ConvDwFp16Border.S
rename to mindspore/lite/nnacl/assembly/fp16/ConvDwFp16Border.S
diff --git a/mindspore/lite/nnacl/assembly/opt/ConvDwFp16Center.S b/mindspore/lite/nnacl/assembly/fp16/ConvDwFp16Center.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/ConvDwFp16Center.S
rename to mindspore/lite/nnacl/assembly/fp16/ConvDwFp16Center.S
diff --git a/mindspore/lite/nnacl/assembly/opt/ConvDwFp16Row.S b/mindspore/lite/nnacl/assembly/fp16/ConvDwFp16Row.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/ConvDwFp16Row.S
rename to mindspore/lite/nnacl/assembly/fp16/ConvDwFp16Row.S
diff --git a/mindspore/lite/nnacl/assembly/opt/DeconvDwFp16Border.S b/mindspore/lite/nnacl/assembly/fp16/DeconvDwFp16Border.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/DeconvDwFp16Border.S
rename to mindspore/lite/nnacl/assembly/fp16/DeconvDwFp16Border.S
diff --git a/mindspore/lite/nnacl/assembly/opt/DeconvDwFp16Center.S b/mindspore/lite/nnacl/assembly/fp16/DeconvDwFp16Center.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/DeconvDwFp16Center.S
rename to mindspore/lite/nnacl/assembly/fp16/DeconvDwFp16Center.S
diff --git a/mindspore/lite/nnacl/assembly/opt/Float16ToFloat32.S b/mindspore/lite/nnacl/assembly/fp16/Float16ToFloat32.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/Float16ToFloat32.S
rename to mindspore/lite/nnacl/assembly/fp16/Float16ToFloat32.S
diff --git a/mindspore/lite/nnacl/assembly/opt/Float32ToFloat16.S b/mindspore/lite/nnacl/assembly/fp16/Float32ToFloat16.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/Float32ToFloat16.S
rename to mindspore/lite/nnacl/assembly/fp16/Float32ToFloat16.S
diff --git a/mindspore/lite/nnacl/assembly/opt/IndirectGemmFp16_16x8.S b/mindspore/lite/nnacl/assembly/fp16/IndirectGemmFp16_16x8.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/IndirectGemmFp16_16x8.S
rename to mindspore/lite/nnacl/assembly/fp16/IndirectGemmFp16_16x8.S
diff --git a/mindspore/lite/nnacl/assembly/opt/MatmulFp16.S b/mindspore/lite/nnacl/assembly/fp16/MatmulFp16.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/MatmulFp16.S
rename to mindspore/lite/nnacl/assembly/fp16/MatmulFp16.S
diff --git a/mindspore/lite/nnacl/assembly/opt/PostFuncBiasReluC8Fp16.S b/mindspore/lite/nnacl/assembly/fp16/PostFuncBiasReluC8Fp16.S
similarity index 100%
rename from mindspore/lite/nnacl/assembly/opt/PostFuncBiasReluC8Fp16.S
rename to mindspore/lite/nnacl/assembly/fp16/PostFuncBiasReluC8Fp16.S
diff --git a/mindspore/lite/nnacl/optimize/CMakeLists.txt b/mindspore/lite/nnacl/optimize/CMakeLists.txt
new file mode 100644
index 00000000000..e6acd66df74
--- /dev/null
+++ b/mindspore/lite/nnacl/optimize/CMakeLists.txt
@@ -0,0 +1,28 @@
+project(optimize)
+
+set(NNACL_DIR ${CMAKE_CURRENT_SOURCE_DIR}/..)
+include_directories(NNACL_DIR)
+
+########################### optimized files ###########################
+file(GLOB SDOT_SRC ${NNACL_DIR}/assembly/opt/*.S)
+file(GLOB FP16_C_SRC ${NNACL_DIR}/fp16/*.c)
+file(GLOB FP16_NEON_SRC ${NNACL_DIR}/assembly/fp16/*.S)
+
+set_property(SOURCE ${SDOT_SRC} PROPERTY LANGUAGE C)
+set_property(SOURCE ${FP16_C_SRC} PROPERTY LANGUAGE C)
+set_property(SOURCE ${FP16_NEON_SRC} PROPERTY LANGUAGE C)
+
+########################### share library build ########################
+list(APPEND SDOT_FILES ${SDOT_SRC})
+list(APPEND FP16_FILES ${FP16_C_SRC})
+list(APPEND FP16_FILES ${FP16_NEON_SRC})
+
+string(REPLACE "-fvisibility=hidden" "-fvisibility=default" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv8.2-a+dotprod+fp16")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8.2-a+dotprod+fp16")
+
+add_library(nnacl_optimize STATIC ${SDOT_FILES})
+target_link_libraries(nnacl_optimize mindspore-lite)
+
+add_library(nnacl_fp16 STATIC ${FP16_FILES})
+target_link_libraries(nnacl_fp16 mindspore-lite)
diff --git a/mindspore/lite/nnacl/optimized_kernel.h b/mindspore/lite/nnacl/optimized_kernel.h
index 216005f4d86..bbe83437573 100644
--- a/mindspore/lite/nnacl/optimized_kernel.h
+++ b/mindspore/lite/nnacl/optimized_kernel.h
@@ -24,14 +24,15 @@
 #include <asm/hwcap.h>
 #include "nnacl/nnacl_utils.h"
 #endif
+#include "utils/log_adapter.h"
 
-#define OPTIMIZE_SHARED_LIBRARY_PATH "liboptimize.so"
+#define OPTIMIZE_SHARED_LIBRARY_PATH "libmindspore-lite-optimize.so"
+#define FLOAT16_SHARED_LIBRARY_PATH "libmindspore-lite-fp16.so"
 
 class OptimizeModule {
  public:
   OptimizeModule() {
     bool support_optimize_ops = false;
-    bool support_fp16 = false;
 #ifdef __ANDROID__
     int hwcap_type = 16;
     uint32_t hwcap = getHwCap(hwcap_type);
@@ -40,8 +41,7 @@ class OptimizeModule {
 #elif defined(__arm__)
     if (hwcap & HWCAP_HALF) {
 #endif
-      MS_LOG(INFO) << "Hw cap support FP16, hwcap: 0x" << hwcap;
-      support_fp16 = true;
+
 #ifdef ENABLE_ARM64
     }
 #elif defined(__arm__)
@@ -57,7 +57,7 @@ class OptimizeModule {
     }
 #endif
 #endif
-    if (!(support_optimize_ops && support_fp16)) {
+    if (support_optimize_ops == false) {
       return;
     }
 #ifndef _WIN32
@@ -77,4 +77,44 @@ class OptimizeModule {
   void *optimized_op_handler_ = nullptr;
 };
 
+class Float16Module {
+ public:
+  Float16Module() {
+    bool support_fp16 = false;
+#ifdef __ANDROID__
+    int hwcap_type = 16;
+    uint32_t hwcap = getHwCap(hwcap_type);
+#ifdef ENABLE_ARM64
+    if (hwcap & HWCAP_FPHP) {
+#elif defined(__arm__)
+    if (hwcap & HWCAP_HALF) {
+#endif
+      MS_LOG(INFO) << "Hw cap support FP16, hwcap: 0x" << hwcap;
+      support_fp16 = true;
+#ifdef ENABLE_ARM64
+    }
+#elif defined(__arm__)
+    }
+#endif
+#endif
+    if (support_fp16 == false) {
+      return;
+    }
+#ifndef _WIN32
+    float16_op_handler_ = dlopen(FLOAT16_SHARED_LIBRARY_PATH, RTLD_LAZY);
+    if (float16_op_handler_ == nullptr) {
+      MS_LOG(INFO) << "Open optimize shared library failed: " << dlerror();
+    }
+#endif
+  }
+
+  ~Float16Module() = default;
+
+  static Float16Module *GetInstance() {
+    static Float16Module fp16_module;
+    return &fp16_module;
+  }
+  void *float16_op_handler_ = nullptr;
+};
+
 #endif  // MINDSPORE_LITE_NNACL_OPTIMIZED_KERNEL_H_
diff --git a/mindspore/lite/src/CMakeLists.txt b/mindspore/lite/src/CMakeLists.txt
index e511cbec73b..88e57da86f2 100644
--- a/mindspore/lite/src/CMakeLists.txt
+++ b/mindspore/lite/src/CMakeLists.txt
@@ -1,7 +1,12 @@
+
+set(LITE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/..)
+include_directories(${LITE_DIR}/nnacl/)
+include_directories(${LITE_DIR}/nnacl/optimize)
+
 if (PLATFORM_ARM32 OR PLATFORM_ARM64)
-  # for performance
+#for performance
   if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
-    #-fno-rtti -fno-exceptions
+#- fno - rtti - fno - exceptions
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math")
   endif()
@@ -65,21 +70,11 @@ set_target_properties(mindspore-lite_static PROPERTIES CLEAN_DIRECT_OUTPUT 1)
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-private-field")
 if (SUPPORT_GPU)
   add_subdirectory(runtime/kernel/opencl)
-  target_link_libraries(mindspore-lite
-      cpu_kernel_mid
-      opencl_kernel_mid
-      )
-  target_link_libraries(mindspore-lite_static
-      cpu_kernel_mid
-      opencl_kernel_mid
-      )
+  target_link_libraries(mindspore-lite cpu_kernel_mid opencl_kernel_mid nnacl)
+  target_link_libraries(mindspore-lite_static cpu_kernel_mid opencl_kernel_mid nnacl)
 else ()
-  target_link_libraries(mindspore-lite
-      cpu_kernel_mid
-      )
-  target_link_libraries(mindspore-lite_static
-      cpu_kernel_mid
-      )
+  target_link_libraries(mindspore-lite cpu_kernel_mid nnacl)
+  target_link_libraries(mindspore-lite_static cpu_kernel_mid nnacl)
 endif ()
 if (PLATFORM_ARM32 OR PLATFORM_ARM64)
   target_link_libraries(mindspore-lite log)
@@ -106,3 +101,24 @@ if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
   endif ()
 endif ()
 
+########################## build optimize and float16 library #################################3
+if (PLATFORM_ARM64)
+  add_library(mindspore-lite-optimize SHARED)
+  target_link_libraries(mindspore-lite-optimize cpu_opt_kernel_mid)
+  target_link_libraries(mindspore-lite-optimize nnacl_optimize)
+
+  add_library(mindspore-lite-fp16 SHARED)
+  target_link_libraries(mindspore-lite-fp16 cpu_fp16_kernel_mid)
+  target_link_libraries(mindspore-lite-fp16 nnacl_fp16)
+endif ()
+
+if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release" AND (PLATFORM_ARM64))
+  add_custom_command(TARGET mindspore-lite-optimize POST_BUILD COMMAND
+    ${ANDROID_NDK}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/aarch64-linux-android/bin/strip
+    ${TOP_DIR}/mindspore/lite/build/src/libmindspore-lite-optimize.so)
+
+  add_custom_command(TARGET mindspore-lite-fp16 POST_BUILD COMMAND
+    ${ANDROID_NDK}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/aarch64-linux-android/bin/strip
+    ${TOP_DIR}/mindspore/lite/build/src/libmindspore-lite-fp16.so)
+endif ()
+
diff --git a/mindspore/lite/src/kernel_registry.cc b/mindspore/lite/src/kernel_registry.cc
index 994867b0768..1207680fb0b 100644
--- a/mindspore/lite/src/kernel_registry.cc
+++ b/mindspore/lite/src/kernel_registry.cc
@@ -43,6 +43,12 @@ int KernelRegistry::Init() {
   } else {
     MS_LOG(INFO) << "load optimize lib failed.";
   }
+  void *float16_op_handler = Float16Module::GetInstance()->float16_op_handler_;
+  if (float16_op_handler != nullptr) {
+    MS_LOG(INFO) << "load float16 lib success.";
+  } else {
+    MS_LOG(INFO) << "load float16 lib failed.";
+  }
 #endif
   return RET_OK;
 }
diff --git a/mindspore/lite/src/runtime/kernel/arm/CMakeLists.txt b/mindspore/lite/src/runtime/kernel/arm/CMakeLists.txt
index 5a136127e83..a108bce633b 100644
--- a/mindspore/lite/src/runtime/kernel/arm/CMakeLists.txt
+++ b/mindspore/lite/src/runtime/kernel/arm/CMakeLists.txt
@@ -1,39 +1,23 @@
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/)
 
 file(GLOB KERNEL_SRC
-        ${CMAKE_CURRENT_SOURCE_DIR}/base/*.cc
-        ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/*.c
-        ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/fp32/*.c
-        ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/int8/*.c
-        ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/quantization/*.c
-        ${CMAKE_CURRENT_SOURCE_DIR}/fp32/*.cc
-        ${CMAKE_CURRENT_SOURCE_DIR}/int8/*.cc
-        )
-list(REMOVE_ITEM KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/opt_op_handler.c)
+    ${CMAKE_CURRENT_SOURCE_DIR}/base/*.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/fp32/*.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/int8/*.cc
+    )
+list(REMOVE_ITEM KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/int8/opt_op_handler.cc)
 
 if (SUPPORT_TRAIN)
-file (GLOB TRAIN_KERNEL_SRC
-        ${CMAKE_CURRENT_SOURCE_DIR}/fp32_grad/*.cc
-        ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/fp32_grad/*.c
-        )
-    set(KERNEL_SRC ${KERNEL_SRC} ${TRAIN_KERNEL_SRC})
-endif()
-
-if (PLATFORM_ARM64)
-    # assembly
-    file(GLOB ASSEMBLY_SRC ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/assembly/arm64/*.s
-            ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/assembly/arm64/*.S)
-    set_property(SOURCE ${ASSEMBLY_SRC} PROPERTY LANGUAGE C)
-    set(KERNEL_SRC ${KERNEL_SRC} ${ASSEMBLY_SRC})
-endif()
-
-if (PLATFORM_ARM32)
-    # assembly
-    file(GLOB ASSEMBLY_SRC ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/assembly/arm32/*.s
-    ${CMAKE_CURRENT_SOURCE_DIR}/../../../../nnacl/assembly/arm32/*.S
-            )
-    set_property(SOURCE ${ASSEMBLY_SRC} PROPERTY LANGUAGE C)
-    set(KERNEL_SRC ${KERNEL_SRC} ${ASSEMBLY_SRC})
+file (GLOB TRAIN_KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/fp32_grad/*.cc)
+set(KERNEL_SRC ${KERNEL_SRC} ${TRAIN_KERNEL_SRC})
 endif()
 
 add_library(cpu_kernel_mid OBJECT ${KERNEL_SRC})
+
+if (PLATFORM_ARM64)
+  file(GLOB FP16_KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/fp16/*.cc)
+  add_library(cpu_fp16_kernel_mid OBJECT ${FP16_KERNEL_SRC})
+  file(GLOB OPT_KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/int8/opt_op_handler.cc)
+  add_library(cpu_opt_kernel_mid OBJECT ${OPT_KERNEL_SRC})
+endif ()
+
diff --git a/mindspore/lite/nnacl/opt_op_handler.c b/mindspore/lite/src/runtime/kernel/arm/int8/opt_op_handler.cc
similarity index 100%
rename from mindspore/lite/nnacl/opt_op_handler.c
rename to mindspore/lite/src/runtime/kernel/arm/int8/opt_op_handler.cc
index 7a16c16990f..9ee46d6c731 100644
--- a/mindspore/lite/nnacl/opt_op_handler.c
+++ b/mindspore/lite/src/runtime/kernel/arm/int8/opt_op_handler.cc
@@ -34,10 +34,6 @@ extern void MatmulInt8DpNeon64(const int8_t *a, const int8_t *b, int8_t *dst, in
                                int *multiplier, int *left_shift, int *right_shift, int row, int col, int stride,
                                size_t peroc);
 
-#ifdef __cplusplus
-}
-#endif
-
 #ifdef ENABLE_ARM64
 void IndirectGemmInt8_optimize_handler(int8_t *dst, const int8_t *src, const int8_t *weight, const int32_t *bias,
                                        size_t ksize, size_t ic4, size_t output_channel, size_t offset,
@@ -61,3 +57,7 @@ void MatMulRInt8_optimize_handler(const int8_t *a, const int8_t *b, int8_t *dst,
                             output_zp, multiplier, left_shift, right_shift, row, col, stride, per_channel);
 }
 #endif
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/mindspore/lite/test/run_benchmark_nets.sh b/mindspore/lite/test/run_benchmark_nets.sh
index 437490de385..0b1377c2316 100644
--- a/mindspore/lite/test/run_benchmark_nets.sh
+++ b/mindspore/lite/test/run_benchmark_nets.sh
@@ -293,7 +293,8 @@ function Run_arm64() {
     fi
 
     cp -a ${arm64_path}/mindspore-lite-${version}-runtime-arm64-${process_unit_arm64}/lib/libmindspore-lite.so ${benchmark_test_path}/libmindspore-lite.so || exit 1
-    cp -a ${arm64_path}/mindspore-lite-${version}-runtime-arm64-${process_unit_arm64}/lib/liboptimize.so ${benchmark_test_path}/liboptimize.so || exit 1
+    cp -a ${arm64_path}/mindspore-lite-${version}-runtime-arm64-${process_unit_arm64}/lib/libmindspore-lite-fp16.so ${benchmark_test_path}/libmindspore-lite-fp16.so || exit 1
+    cp -a ${arm64_path}/mindspore-lite-${version}-runtime-arm64-${process_unit_arm64}/lib/libmindspore-lite-optimize.so ${benchmark_test_path}/libmindspore-lite-optimize.so || exit 1
     cp -a ${arm64_path}/mindspore-lite-${version}-runtime-arm64-${process_unit_arm64}/benchmark/benchmark ${benchmark_test_path}/benchmark || exit 1
 
     # adb push all needed files to the phone
diff --git a/mindspore/lite/tools/converter/CMakeLists.txt b/mindspore/lite/tools/converter/CMakeLists.txt
index 1de0a3896ab..f2c19036184 100644
--- a/mindspore/lite/tools/converter/CMakeLists.txt
+++ b/mindspore/lite/tools/converter/CMakeLists.txt
@@ -106,7 +106,6 @@ file(GLOB KERNEL_SRC
         ${ARM_DIR}/fp32/*.cc
         ${ARM_DIR}/int8/*.cc
         )
-list(REMOVE_ITEM KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/../../nnacl/opt_op_handler.c)
 
 if (PLATFORM_ARM64)
     # assembly