Merge pull request #431 from MegEngine/try-import

4 years ago · bd3c4a0527
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -29,6 +29,7 @@ jobs:
        uses: actions/checkout@v2
      - name: Checkout submodules
        run: |
          apt update&&apt install ninja-build
          ./third_party/prepare.sh
          ./third_party/install-mkl.sh
      - name: Build MegEngine
@@ -57,6 +58,7 @@ jobs:
        uses: actions/checkout@v2
      - name: Checkout submodules
        run: |
          apt update&&apt install ninja-build
          ./third_party/prepare.sh
          ./third_party/install-mkl.sh
      - name: Build MegEngine
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
--- a/ci/cmake.sh
+++ b/ci/cmake.sh
@@ -27,7 +27,8 @@ function build() {
            -DMGE_WITH_DISTRIBUTED=${DMGE_WITH_DISTRIBUTED} \
            -DMGE_WITH_CUDA=${DMGE_WITH_CUDA} \
            -DMGE_WITH_TEST=ON \
            -DCMAKE_BUILD_TYPE=RelWithDebInfo
            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
 	    -DMGE_WITH_CUSTOM_OP=ON
        make -j$(($(nproc) * 2)) -I ${build_dir}
        make develop
    popd >/dev/null
--- a/cmake/BuildFlatBuffers.cmake
+++ b/cmake/BuildFlatBuffers.cmake
@@ -1,59 +1,56 @@
 # Copyright 2015 Google Inc. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 # Licensed under the Apache License, Version 2.0 (the "License"); you may not use this
 # file except in compliance with the License. You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 # http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # Unless required by applicable law or agreed to in writing, software distributed under
 # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
 # ANY KIND, either express or implied. See the License for the specific language
 # governing permissions and limitations under the License.
 # General function to create FlatBuffer build rules for the given list of
 # schemas.
 # General function to create FlatBuffer build rules for the given list of schemas.
 #
 # flatbuffers_schemas: A list of flatbuffer schema files to process.
 #
 # schema_include_dirs: A list of schema file include directories, which will be
 # passed to flatc via the -I parameter.
 # schema_include_dirs: A list of schema file include directories, which will be passed
 # to flatc via the -I parameter.
 #
 # custom_target_name: The generated files will be added as dependencies for a
 # new custom target with this name. You should add that target as a dependency
 # for your main target to ensure these files are built. You can also retrieve
 # various properties from this target, such as GENERATED_INCLUDES_DIR,
 # BINARY_SCHEMAS_DIR, and COPY_TEXT_SCHEMAS_DIR.
 # custom_target_name: The generated files will be added as dependencies for a new custom
 # target with this name. You should add that target as a dependency for your main target
 # to ensure these files are built. You can also retrieve various properties from this
 # target, such as GENERATED_INCLUDES_DIR, BINARY_SCHEMAS_DIR, and COPY_TEXT_SCHEMAS_DIR.
 #
 # additional_dependencies: A list of additional dependencies that you'd like
 # all generated files to depend on. Pass in a blank string if you have none.
 # additional_dependencies: A list of additional dependencies that you'd like all
 # generated files to depend on. Pass in a blank string if you have none.
 #
 # generated_includes_dir: Where to generate the C++ header files for these
 # schemas. The generated includes directory will automatically be added to
 # CMake's include directories, and will be where generated header files are
 # placed. This parameter is optional; pass in empty string if you don't want to
 # generate include files for these schemas.
 # generated_includes_dir: Where to generate the C++ header files for these schemas. The
 # generated includes directory will automatically be added to CMake's include
 # directories, and will be where generated header files are placed. This parameter is
 # optional; pass in empty string if you don't want to generate include files for these
 # schemas.
 #
 # binary_schemas_dir: If you specify an optional binary schema directory, binary
 # schemas will be generated for these schemas as well, and placed into the given
 # directory.
 # binary_schemas_dir: If you specify an optional binary schema directory, binary schemas
 # will be generated for these schemas as well, and placed into the given directory.
 #
 # copy_text_schemas_dir: If you want all text schemas (including schemas from
 # all schema include directories) copied into a directory (for example, if you
 # need them within your project to build JSON files), you can specify that
 # folder here. All text schemas will be copied to that folder.
 # copy_text_schemas_dir: If you want all text schemas (including schemas from all schema
 # include directories) copied into a directory (for example, if you need them within
 # your project to build JSON files), you can specify that folder here. All text schemas
 # will be copied to that folder.
 #
 # IMPORTANT: Make sure you quote all list arguments you pass to this function!
 # Otherwise CMake will only pass in the first element.
 # Example: build_flatbuffers("${fb_files}" "${include_dirs}" target_name ...)
 function(build_flatbuffers flatbuffers_schemas
                           schema_include_dirs
                           custom_target_name
                           additional_dependencies
                           generated_includes_dir
                           binary_schemas_dir
                           copy_text_schemas_dir)
 # IMPORTANT: Make sure you quote all list arguments you pass to this function! Otherwise
 # CMake will only pass in the first element. Example: build_flatbuffers("${fb_files}"
 # "${include_dirs}" target_name ...)
 function(
  build_flatbuffers
  flatbuffers_schemas
  schema_include_dirs
  custom_target_name
  additional_dependencies
  generated_includes_dir
  binary_schemas_dir
  copy_text_schemas_dir)
  # Test if including from FindFlatBuffers
  if(FLATBUFFERS_FLATC_EXECUTABLE)
@@ -65,10 +62,7 @@ function(build_flatbuffers flatbuffers_schemas
  endif()
  set(FLATC_SCHEMA_ARGS --gen-mutable)
  if(FLATBUFFERS_FLATC_SCHEMA_EXTRA_ARGS)
    set(FLATC_SCHEMA_ARGS
      ${FLATBUFFERS_FLATC_SCHEMA_EXTRA_ARGS}
      ${FLATC_SCHEMA_ARGS}
      )
    set(FLATC_SCHEMA_ARGS ${FLATBUFFERS_FLATC_SCHEMA_EXTRA_ARGS} ${FLATC_SCHEMA_ARGS})
  endif()
  set(working_dir "${CMAKE_CURRENT_SOURCE_DIR}")
@@ -77,12 +71,12 @@ function(build_flatbuffers flatbuffers_schemas
  # Generate the include files parameters.
  set(include_params "")
  set(all_generated_files "")
  foreach (include_dir ${schema_include_dirs})
  foreach(include_dir ${schema_include_dirs})
    set(include_params -I ${include_dir} ${include_params})
    if (NOT ${copy_text_schemas_dir} STREQUAL "")
    if(NOT ${copy_text_schemas_dir} STREQUAL "")
      # Copy text schemas from dependent folders.
      file(GLOB_RECURSE dependent_schemas ${include_dir}/${schema_glob})
      foreach (dependent_schema ${dependent_schemas})
      foreach(dependent_schema ${dependent_schemas})
        file(COPY ${dependent_schema} DESTINATION ${copy_text_schemas_dir})
      endforeach()
    endif()
@@ -91,62 +85,54 @@ function(build_flatbuffers flatbuffers_schemas
  foreach(schema ${flatbuffers_schemas})
    get_filename_component(filename ${schema} NAME_WE)
    # For each schema, do the things we requested.
    if (NOT ${generated_includes_dir} STREQUAL "")
    if(NOT ${generated_includes_dir} STREQUAL "")
      set(generated_include ${generated_includes_dir}/${filename}_generated.h)
      add_custom_command(
        OUTPUT ${generated_include}
        COMMAND ${FLATC} ${FLATC_SCHEMA_ARGS}
        -o ${generated_includes_dir}
        ${include_params}
        -c ${schema}
        COMMAND ${FLATC} ${FLATC_SCHEMA_ARGS} -o ${generated_includes_dir}
                ${include_params} -c ${schema}
        DEPENDS ${FLATC_TARGET} ${schema} ${additional_dependencies}
        WORKING_DIRECTORY "${working_dir}")
      list(APPEND all_generated_files ${generated_include})
    endif()
    if (NOT ${binary_schemas_dir} STREQUAL "")
    if(NOT ${binary_schemas_dir} STREQUAL "")
      set(binary_schema ${binary_schemas_dir}/${filename}.bfbs)
      add_custom_command(
        OUTPUT ${binary_schema}
        COMMAND ${FLATC} -b --schema
        -o ${binary_schemas_dir}
        ${include_params}
        ${schema}
        COMMAND ${FLATC} -b --schema -o ${binary_schemas_dir} ${include_params}
                ${schema}
        DEPENDS ${FLATC_TARGET} ${schema} ${additional_dependencies}
        WORKING_DIRECTORY "${working_dir}")
      list(APPEND all_generated_files ${binary_schema})
    endif()
    if (NOT ${copy_text_schemas_dir} STREQUAL "")
    if(NOT ${copy_text_schemas_dir} STREQUAL "")
      file(COPY ${schema} DESTINATION ${copy_text_schemas_dir})
    endif()
  endforeach()
  # Create a custom target that depends on all the generated files.
  # This is the target that you can depend on to trigger all these
  # to be built.
  add_custom_target(${custom_target_name}
                    DEPENDS ${all_generated_files} ${additional_dependencies})
  # Create a custom target that depends on all the generated files. This is the target
  # that you can depend on to trigger all these to be built.
  add_custom_target(${custom_target_name} DEPENDS ${all_generated_files}
                                                  ${additional_dependencies})
  # Register the include directory we are using.
  if (NOT ${generated_includes_dir} STREQUAL "")
  if(NOT ${generated_includes_dir} STREQUAL "")
    include_directories(${generated_includes_dir})
    set_property(TARGET ${custom_target_name}
      PROPERTY GENERATED_INCLUDES_DIR
      ${generated_includes_dir})
    set_property(TARGET ${custom_target_name} PROPERTY GENERATED_INCLUDES_DIR
                                                       ${generated_includes_dir})
  endif()
  # Register the binary schemas dir we are using.
  if (NOT ${binary_schemas_dir} STREQUAL "")
    set_property(TARGET ${custom_target_name}
      PROPERTY BINARY_SCHEMAS_DIR
      ${binary_schemas_dir})
  if(NOT ${binary_schemas_dir} STREQUAL "")
    set_property(TARGET ${custom_target_name} PROPERTY BINARY_SCHEMAS_DIR
                                                       ${binary_schemas_dir})
  endif()
  # Register the text schema copy dir we are using.
  if (NOT ${copy_text_schemas_dir} STREQUAL "")
    set_property(TARGET ${custom_target_name}
      PROPERTY COPY_TEXT_SCHEMAS_DIR
      ${copy_text_schemas_dir})
  if(NOT ${copy_text_schemas_dir} STREQUAL "")
    set_property(TARGET ${custom_target_name} PROPERTY COPY_TEXT_SCHEMAS_DIR
                                                       ${copy_text_schemas_dir})
  endif()
 endfunction()
--- a/cmake/FetchMegBrainVersion.cmake
+++ b/cmake/FetchMegBrainVersion.cmake
@@ -1,49 +1,45 @@
 # Parses the version set in src/core/include/megbrain/version.h
 # Exports the following variables:
 # MGB_VER_MAJOR: Major version
 # MGB_VER_MINOR: Minor version
 # MGB_VER_PATCH: Patch version
 # MGB_IS_DEV: Is development version
 # MGB_VER_STRING: Version string
 # Parses the version set in src/core/include/megbrain/version.h Exports the following
 # variables: MGB_VER_MAJOR: Major version MGB_VER_MINOR: Minor version MGB_VER_PATCH:
 # Patch version MGB_IS_DEV: Is development version MGB_VER_STRING: Version string
 option(MGB_FORCE_DEV_VERSION "Force -dev tag in version stamp" OFF)
 file (READ "${CMAKE_CURRENT_SOURCE_DIR}/src/core/include/megbrain/version.h" content)
 file(READ "${CMAKE_CURRENT_SOURCE_DIR}/src/core/include/megbrain/version.h" content)
 string (REGEX MATCH "MGB_MAJOR +([0-9]+)" _ ${content})
 set (MGB_VER_MAJOR ${CMAKE_MATCH_1})
 string(REGEX MATCH "MGB_MAJOR +([0-9]+)" _ ${content})
 set(MGB_VER_MAJOR ${CMAKE_MATCH_1})
 string (REGEX MATCH "MGB_MINOR +([0-9]+)" _ ${content})
 set (MGB_VER_MINOR ${CMAKE_MATCH_1})
 string(REGEX MATCH "MGB_MINOR +([0-9]+)" _ ${content})
 set(MGB_VER_MINOR ${CMAKE_MATCH_1})
 string (REGEX MATCH "MGB_PATCH *([0-9]+)" _ ${content})
 set (MGB_VER_PATCH ${CMAKE_MATCH_1})
 string(REGEX MATCH "MGB_PATCH *([0-9]+)" _ ${content})
 set(MGB_VER_PATCH ${CMAKE_MATCH_1})
 string (REGEX MATCH "MGE_MAJOR +([0-9]+)" _ ${content})
 set (MGE_VER_MAJOR ${CMAKE_MATCH_1})
 string(REGEX MATCH "MGE_MAJOR +([0-9]+)" _ ${content})
 set(MGE_VER_MAJOR ${CMAKE_MATCH_1})
 string (REGEX MATCH "MGE_MINOR +([0-9]+)" _ ${content})
 set (MGE_VER_MINOR ${CMAKE_MATCH_1})
 string(REGEX MATCH "MGE_MINOR +([0-9]+)" _ ${content})
 set(MGE_VER_MINOR ${CMAKE_MATCH_1})
 string (REGEX MATCH "MGE_PATCH *([0-9]+)" _ ${content})
 set (MGE_VER_PATCH ${CMAKE_MATCH_1})
 string(REGEX MATCH "MGE_PATCH *([0-9]+)" _ ${content})
 set(MGE_VER_PATCH ${CMAKE_MATCH_1})
 string (REGEX MATCH "MGE_EXTRA_NAME *\"(.*)\"" _ ${content})
 set (MGE_EXTRA_NAME ${CMAKE_MATCH_1})
 string(REGEX MATCH "MGE_EXTRA_NAME *\"(.*)\"" _ ${content})
 set(MGE_EXTRA_NAME ${CMAKE_MATCH_1})
 if (MGB_FORCE_DEV_VERSION)
    set (MGB_IS_DEV 1)
 if(MGB_FORCE_DEV_VERSION)
  set(MGB_IS_DEV 1)
 else()
    string (REGEX MATCH "MGB_IS_DEV +([01])" _ ${content})
    set (MGB_IS_DEV ${CMAKE_MATCH_1})
  string(REGEX MATCH "MGB_IS_DEV +([01])" _ ${content})
  set(MGB_IS_DEV ${CMAKE_MATCH_1})
 endif()
 if (DEFINED MGB_VER_MAJOR)
    set (MGB_VER_STRING "${MGB_VER_MAJOR}.${MGB_VER_MINOR}.${MGB_VER_PATCH}")
 if(DEFINED MGB_VER_MAJOR)
  set(MGB_VER_STRING "${MGB_VER_MAJOR}.${MGB_VER_MINOR}.${MGB_VER_PATCH}")
 else()
    set (MGB_VER_STRING "${MGE_VER_MAJOR}.${MGE_VER_MINOR}.${MGE_VER_PATCH}")
  set(MGB_VER_STRING "${MGE_VER_MAJOR}.${MGE_VER_MINOR}.${MGE_VER_PATCH}")
 endif(DEFINED MGB_VER_MAJOR)
 if (MGB_IS_DEV)
    set (MGB_VER_STRING "${MGB_VER_STRING}-dev")
 if(MGB_IS_DEV)
  set(MGB_VER_STRING "${MGB_VER_STRING}-dev")
 endif()
 message(STATUS "Building MegBrain ${MGB_VER_STRING}")
--- a/cmake/Halide.cmake
+++ b/cmake/Halide.cmake
@@ -2,31 +2,40 @@
 include(ExternalProject)
 find_package(LLVM 6.0 REQUIRED CONFIG)
 STRING(REPLACE "." ";" LLVM_VERSION_LIST ${LLVM_PACKAGE_VERSION})
 string(REPLACE "." ";" LLVM_VERSION_LIST ${LLVM_PACKAGE_VERSION})
 list(GET LLVM_VERSION_LIST 0 LLVM_VERSION_MAJOR)
 list(GET LLVM_VERSION_LIST 1 LLVM_VERSION_MINOR)
 set(HALIDE_DIR "${PROJECT_SOURCE_DIR}/third_party/Halide" CACHE STRING "halide directory")
 set(HALIDE_DIR
    "${PROJECT_SOURCE_DIR}/third_party/Halide"
    CACHE STRING "halide directory")
 set(HALIDE_BUILD_DIR ${PROJECT_BINARY_DIR}/third_party/Halide)
 set(HALIDE_LIB ${HALIDE_BUILD_DIR}/lib/libHalide.a)
 ExternalProject_add(
    halide
    SOURCE_DIR ${HALIDE_DIR}
    PREFIX ${HALIDE_BUILD_DIR}
    CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER} -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER} -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} -DCMAKE_INSTALL_PREFIX=${HALIDE_BUILD_DIR} -DWITH_APPS=OFF -DWITH_TESTS=OFF -DWITH_TUTORIALS=OFF -DHALIDE_SHARED_LIBRARY=OFF -DHALIDE_REQUIRE_LLVM_VERSION=${LLVM_VERSION_MAJOR}${LLVM_VERSION_MINOR} -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DTARGET_MIPS=OFF -DTARGET_POWERPC=OFF
    BUILD_BYPRODUCTS ${HALIDE_LIB}
 )
 ExternalProject_Add(
  halide
  SOURCE_DIR ${HALIDE_DIR}
  PREFIX ${HALIDE_BUILD_DIR}
  CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER}
             -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER}
             -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
             -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE}
             -DCMAKE_INSTALL_PREFIX=${HALIDE_BUILD_DIR}
             -DWITH_APPS=OFF
             -DWITH_TESTS=OFF
             -DWITH_TUTORIALS=OFF
             -DHALIDE_SHARED_LIBRARY=OFF
             -DHALIDE_REQUIRE_LLVM_VERSION=${LLVM_VERSION_MAJOR}${LLVM_VERSION_MINOR}
             -DCMAKE_POSITION_INDEPENDENT_CODE=ON
             -DTARGET_MIPS=OFF
             -DTARGET_POWERPC=OFF
  BUILD_BYPRODUCTS ${HALIDE_LIB})
 set(HALIDE_INC ${HALIDE_BUILD_DIR}/include)
 file(MAKE_DIRECTORY ${HALIDE_INC})
 add_library(libhalide STATIC IMPORTED GLOBAL)
 add_dependencies(libhalide halide)
 set_target_properties(
    libhalide PROPERTIES
    IMPORTED_LOCATION ${HALIDE_LIB}
    INTERFACE_INCLUDE_DIRECTORIES ${HALIDE_INC}
 )
 set_target_properties(libhalide PROPERTIES IMPORTED_LOCATION ${HALIDE_LIB}
                                           INTERFACE_INCLUDE_DIRECTORIES ${HALIDE_INC})
 set(LLVM_COMPONENTS mcjit;bitwriter;linker;passes;X86;ARM;AArch64;Hexagon;NVPTX;AMDGPU)
 llvm_map_components_to_libnames(HALIDE_LLVM_LIBS ${LLVM_COMPONENTS})
--- a/cmake/MKL_DNN.cmake
+++ b/cmake/MKL_DNN.cmake
@@ -1,25 +1,31 @@
 if (MGE_USE_SYSTEM_LIB)
    find_package(dnnl)
    if (dnnl_FOUND)
        message(STATUS "Using system provided MKL-DNN.")
        set (MGE_USE_SYSTEM_MKLDNN ON)
        return()
    endif()
 if(MGE_USE_SYSTEM_LIB)
  find_package(dnnl)
  if(dnnl_FOUND)
    message(STATUS "Using system provided MKL-DNN.")
    set(MGE_USE_SYSTEM_MKLDNN ON)
    return()
  endif()
 endif()
 option(DNNL_BUILD_TESTS "" OFF)
 option(DNNL_BUILD_EXAMPLES "" OFF)
 # we do not want to use OMP now, so config to CPU mode
 # if set to OMP, some dnnl algo will be more fast
 set(DNNL_CPU_RUNTIME "SEQ" CACHE STRING "config dnnl to DNNL_RUNTIME_SEQ")
 # we do not want to use OMP now, so config to CPU mode if set to OMP, some dnnl algo
 # will be more fast
 set(DNNL_CPU_RUNTIME
    "SEQ"
    CACHE STRING "config dnnl to DNNL_RUNTIME_SEQ")
 if(MGE_BLAS STREQUAL "MKL")
    option(_DNNL_USE_MKL "" ON)
    set(MKLROOT ${MKL_ROOT_DIR} CACHE STRING "MKL ROOT FOR DNNL")
    set(MKLLIB libmkl)
  option(_DNNL_USE_MKL "" ON)
  set(MKLROOT
      ${MKL_ROOT_DIR}
      CACHE STRING "MKL ROOT FOR DNNL")
  set(MKLLIB libmkl)
 else()
    option(_DNNL_USE_MKL "" OFF)
  option(_DNNL_USE_MKL "" OFF)
 endif()
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-unused-parameter -Wno-extra")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-parameter -Wno-extra")
 set(DNNL_LIBRARY_TYPE STATIC CACHE STRING "config dnnl to STATIC")
 set(DNNL_LIBRARY_TYPE
    STATIC
    CACHE STRING "config dnnl to STATIC")
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/intel-mkl-dnn)
--- a/cmake/Modules/FindNumPy.cmake
+++ b/cmake/Modules/FindNumPy.cmake
@@ -1,30 +1,28 @@
 # - Find the NumPy libraries
 # This module finds if NumPy is installed, and sets the following variables
 # indicating where it is.
 # * Find the NumPy libraries This module finds if NumPy is installed, and sets the
 #   following variables indicating where it is.
 #
 # TODO: Update to provide the libraries and paths for linking npymath lib.
 #
 #  NUMPY_FOUND               - was NumPy found
 #  NUMPY_VERSION             - the version of NumPy found as a string
 #  NUMPY_VERSION_MAJOR       - the major version number of NumPy
 #  NUMPY_VERSION_MINOR       - the minor version number of NumPy
 #  NUMPY_VERSION_PATCH       - the patch version number of NumPy
 #  NUMPY_VERSION_DECIMAL     - e.g. version 1.6.1 is 10601
 #  NUMPY_INCLUDE_DIR         - path to the NumPy include files
 # NUMPY_FOUND               - was NumPy found NUMPY_VERSION             - the version of
 # NumPy found as a string NUMPY_VERSION_MAJOR       - the major version number of NumPy
 # NUMPY_VERSION_MINOR       - the minor version number of NumPy NUMPY_VERSION_PATCH -
 # the patch version number of NumPy NUMPY_VERSION_DECIMAL     - e.g. version 1.6.1 is
 # 10601 NUMPY_INCLUDE_DIR         - path to the NumPy include files
 unset(NUMPY_VERSION)
 unset(NUMPY_INCLUDE_DIR)
 if(PYTHONINTERP_FOUND)
  execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
    "import numpy as n; print(n.__version__); print(n.get_include());"
  execute_process(
    COMMAND "${PYTHON_EXECUTABLE}" "-c"
            "import numpy as n; print(n.__version__); print(n.get_include());"
    RESULT_VARIABLE __result
    OUTPUT_VARIABLE __output
    OUTPUT_STRIP_TRAILING_WHITESPACE)
  if(__result MATCHES 0)
    string(REGEX REPLACE ";" "\\\\;" __values ${__output})
    string(REGEX REPLACE "\r?\n" ";"    __values ${__values})
    string(REGEX REPLACE "\r?\n" ";" __values ${__values})
    list(GET __values 0 NUMPY_VERSION)
    list(GET __values 1 NUMPY_INCLUDE_DIR)
@@ -33,13 +31,18 @@ if(PYTHONINTERP_FOUND)
      set(NUMPY_VERSION_MAJOR ${CMAKE_MATCH_1})
      set(NUMPY_VERSION_MINOR ${CMAKE_MATCH_2})
      set(NUMPY_VERSION_PATCH ${CMAKE_MATCH_3})
      math(EXPR NUMPY_VERSION_DECIMAL
        "(${NUMPY_VERSION_MAJOR} * 10000) + (${NUMPY_VERSION_MINOR} * 100) + ${NUMPY_VERSION_PATCH}")
      string(REGEX REPLACE "\\\\" "/"  NUMPY_INCLUDE_DIR ${NUMPY_INCLUDE_DIR})
      math(
        EXPR
        NUMPY_VERSION_DECIMAL
        "(${NUMPY_VERSION_MAJOR} * 10000) + (${NUMPY_VERSION_MINOR} * 100) + ${NUMPY_VERSION_PATCH}"
      )
      string(REGEX REPLACE "\\\\" "/" NUMPY_INCLUDE_DIR ${NUMPY_INCLUDE_DIR})
    else()
     unset(NUMPY_VERSION)
     unset(NUMPY_INCLUDE_DIR)
     message(STATUS "Requested NumPy version and include path, but got instead:\n${__output}\n")
      unset(NUMPY_VERSION)
      unset(NUMPY_INCLUDE_DIR)
      message(
        STATUS
          "Requested NumPy version and include path, but got instead:\n${__output}\n")
    endif()
  endif()
 else()
@@ -47,8 +50,10 @@ else()
 endif()
 include(FindPackageHandleStandardArgs)
 find_package_handle_standard_args(NumPy REQUIRED_VARS NUMPY_INCLUDE_DIR NUMPY_VERSION
                                        VERSION_VAR   NUMPY_VERSION)
 find_package_handle_standard_args(
  NumPy
  REQUIRED_VARS NUMPY_INCLUDE_DIR NUMPY_VERSION
  VERSION_VAR NUMPY_VERSION)
 if(NUMPY_FOUND)
  message(STATUS "NumPy ver. ${NUMPY_VERSION} found (include: ${NUMPY_INCLUDE_DIR})")
--- a/cmake/OpenBLAS.cmake
+++ b/cmake/OpenBLAS.cmake
@@ -1,48 +1,50 @@
 if (MGE_USE_SYSTEM_LIB)
    find_package(OpenBLAS)
    set (MGE_USE_SYSTEM_OPENBLAS ON)
    message(STATUS "Using system provided OpenBLAS ${OpenBLAS_VERSION}")
    add_library(libopenblas IMPORTED GLOBAL)
    set_target_properties(
        libopenblas PROPERTIES
        IMPORTED_LOCATION ${OpenBLAS_LIBRARIES}
        INTERFACE_INCLUDE_DIRECTORIES ${OpenBLAS_INCLUDE_DIRS}
    )
    return()
 if(MGE_USE_SYSTEM_LIB)
  find_package(OpenBLAS)
  set(MGE_USE_SYSTEM_OPENBLAS ON)
  message(STATUS "Using system provided OpenBLAS ${OpenBLAS_VERSION}")
  add_library(libopenblas IMPORTED GLOBAL)
  set_target_properties(
    libopenblas PROPERTIES IMPORTED_LOCATION ${OpenBLAS_LIBRARIES}
                           INTERFACE_INCLUDE_DIRECTORIES ${OpenBLAS_INCLUDE_DIRS})
  return()
 endif()
 include(ExternalProject)
 include(GNUInstallDirs)
 set(OPENBLAS_DIR "${PROJECT_SOURCE_DIR}/third_party/OpenBLAS" CACHE STRING "OpenBLAS directory")
 set(OPENBLAS_DIR
    "${PROJECT_SOURCE_DIR}/third_party/OpenBLAS"
    CACHE STRING "OpenBLAS directory")
 set(OPENBLAS_BUILD_DIR ${PROJECT_BINARY_DIR}/third_party/OpenBLAS)
 set(OPENBLAS_INC ${OPENBLAS_BUILD_DIR}/include)
 set(OPENBLAS_LIB ${OPENBLAS_BUILD_DIR}/${CMAKE_INSTALL_LIBDIR}/libopenblas.a)
 if(${CMAKE_GENERATOR} STREQUAL "Ninja")
    set(MAKE_COMMAND make)
  set(MAKE_COMMAND make)
 else()
    set(MAKE_COMMAND "$(MAKE)")
  set(MAKE_COMMAND "$(MAKE)")
 endif()
 ExternalProject_add(
    openblas
    SOURCE_DIR ${OPENBLAS_DIR}
    PREFIX ${OPENBLAS_BUILD_DIR}
    CMAKE_GENERATOR "Unix Makefiles"
    CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER} -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER} -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} -DCMAKE_INSTALL_PREFIX=${OPENBLAS_BUILD_DIR} -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} -DCMAKE_POSITION_INDEPENDENT_CODE=ON
    BUILD_COMMAND ${MAKE_COMMAND}
    BUILD_BYPRODUCTS ${OPENBLAS_LIB} ${OPENBLAS_PROTOC_EXECUTABLE}
 )
 ExternalProject_Add(
  openblas
  SOURCE_DIR ${OPENBLAS_DIR}
  PREFIX ${OPENBLAS_BUILD_DIR}
  CMAKE_GENERATOR "Unix Makefiles"
  CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER}
             -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER}
             -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
             -DCMAKE_INSTALL_PREFIX=${OPENBLAS_BUILD_DIR}
             -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE}
             -DCMAKE_POSITION_INDEPENDENT_CODE=ON
  BUILD_COMMAND ${MAKE_COMMAND}
  BUILD_BYPRODUCTS ${OPENBLAS_LIB} ${OPENBLAS_PROTOC_EXECUTABLE})
 file(MAKE_DIRECTORY ${OPENBLAS_INC})
 add_library(libopenblas STATIC IMPORTED GLOBAL)
 add_dependencies(libopenblas openblas)
 set_target_properties(
    libopenblas PROPERTIES
    IMPORTED_LOCATION ${OPENBLAS_LIB}
    INTERFACE_INCLUDE_DIRECTORIES ${OPENBLAS_BUILD_DIR}/include
 )
  libopenblas PROPERTIES IMPORTED_LOCATION ${OPENBLAS_LIB}
                         INTERFACE_INCLUDE_DIRECTORIES ${OPENBLAS_BUILD_DIR}/include)
--- a/cmake/aclrt.cmake
+++ b/cmake/aclrt.cmake
@@ -1,31 +1,31 @@
 find_library(ACLRT_LIBRARY
    NAMES libascendcl.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{ACLRT_HOME}/lib64/stub" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES stub
    DOC "ACL library." )
 find_library(
  ACLRT_LIBRARY
  NAMES libascendcl.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{ACLRT_HOME}/lib64/stub" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES stub
  DOC "ACL library.")
 if(ACLRT_LIBRARY STREQUAL "ACLRT_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find ACLRT Library")
  message(FATAL_ERROR "Can not find ACLRT Library")
 endif()
 get_filename_component(__found_aclrt_root "${ACLRT_LIBRARY}/../../../" REALPATH)
 find_path(ACLRT_INCLUDE_DIR
    NAMES acl/acl.h
    HINTS "$ENV{ACLRT_HOME}/include" ${__found_aclrt_root}
    PATH_SUFFIXES include
    DOC "Path to ACLRT include directory." )
 find_path(
  ACLRT_INCLUDE_DIR
  NAMES acl/acl.h
  HINTS "$ENV{ACLRT_HOME}/include" ${__found_aclrt_root}
  PATH_SUFFIXES include
  DOC "Path to ACLRT include directory.")
 if(ACLRT_INCLUDE_DIR STREQUAL "ACLRT_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find ACLRT Library")
  message(FATAL_ERROR "Can not find ACLRT Library")
 endif()
 add_library(libascendcl SHARED IMPORTED)
 set_target_properties(libascendcl PROPERTIES
    IMPORTED_LOCATION ${ACLRT_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${ACLRT_INCLUDE_DIR}
 )
 set_target_properties(
  libascendcl PROPERTIES IMPORTED_LOCATION ${ACLRT_LIBRARY}
                         INTERFACE_INCLUDE_DIRECTORIES ${ACLRT_INCLUDE_DIR})
 message(STATUS "Found ACLRT: ${__found_aclrt_root}")
--- a/cmake/cndev.cmake
+++ b/cmake/cndev.cmake
@@ -1,44 +1,57 @@
 find_library(CNDEV_LIBRARY 
    NAMES libcndev.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CNDEV library." )
 find_library(
  CNDEV_LIBRARY
  NAMES libcndev.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES lib lib64
  DOC "CNDEV library.")
 if(CNDEV_LIBRARY STREQUAL "CNDEV_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find CNDEV Library")
  message(FATAL_ERROR "Can not find CNDEV Library")
 endif()
 get_filename_component(__found_cndev_root ${CNDEV_LIBRARY}/../.. REALPATH)
 find_path(CNDEV_INCLUDE_DIR 
    NAMES cndev.h
    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cndev_root}
    PATH_SUFFIXES include 
    DOC "Path to CNDEV include directory." )
 find_path(
  CNDEV_INCLUDE_DIR
  NAMES cndev.h
  HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cndev_root}
  PATH_SUFFIXES include
  DOC "Path to CNDEV include directory.")
 if(CNDEV_INCLUDE_DIR STREQUAL "CNDEV_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find CNDEV Library")
  message(FATAL_ERROR "Can not find CNDEV Library")
 endif()
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_1 REGEX "^#define CNDEV_VERSION_1 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_2 REGEX "^#define CNDEV_VERSION_2 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_3 REGEX "^#define CNDEV_VERSION_3 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_4 REGEX "^#define CNDEV_VERSION_4 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_5 REGEX "^#define CNDEV_VERSION_5 [0-9]+.*$")
 string(REGEX REPLACE "^#define CNDEV_VERSION_1 ([0-9]+).*$" "\\1" CNDEV_VERSION_1 "${CNDEV_1}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_2 ([0-9]+).*$" "\\1" CNDEV_VERSION_2 "${CNDEV_2}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_3 ([0-9]+).*$" "\\1" CNDEV_VERSION_3 "${CNDEV_3}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_4 ([0-9]+).*$" "\\1" CNDEV_VERSION_4 "${CNDEV_4}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_5 ([0-9]+).*$" "\\1" CNDEV_VERSION_5 "${CNDEV_5}")
 set(CNDEV_VERSION_STRING "${CNDEV_VERSION_1}.${CNDEV_VERSION_2}.${CNDEV_VERSION_3}.${CNDEV_VERSION_4}.${CNDEV_VERSION_5}")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_1
     REGEX "^#define CNDEV_VERSION_1 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_2
     REGEX "^#define CNDEV_VERSION_2 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_3
     REGEX "^#define CNDEV_VERSION_3 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_4
     REGEX "^#define CNDEV_VERSION_4 [0-9]+.*$")
 file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_5
     REGEX "^#define CNDEV_VERSION_5 [0-9]+.*$")
 string(REGEX REPLACE "^#define CNDEV_VERSION_1 ([0-9]+).*$" "\\1" CNDEV_VERSION_1
                     "${CNDEV_1}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_2 ([0-9]+).*$" "\\1" CNDEV_VERSION_2
                     "${CNDEV_2}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_3 ([0-9]+).*$" "\\1" CNDEV_VERSION_3
                     "${CNDEV_3}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_4 ([0-9]+).*$" "\\1" CNDEV_VERSION_4
                     "${CNDEV_4}")
 string(REGEX REPLACE "^#define CNDEV_VERSION_5 ([0-9]+).*$" "\\1" CNDEV_VERSION_5
                     "${CNDEV_5}")
 set(CNDEV_VERSION_STRING
    "${CNDEV_VERSION_1}.${CNDEV_VERSION_2}.${CNDEV_VERSION_3}.${CNDEV_VERSION_4}.${CNDEV_VERSION_5}"
 )
 add_library(libcndev SHARED IMPORTED)
 set_target_properties(libcndev PROPERTIES
    IMPORTED_LOCATION ${CNDEV_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${CNDEV_INCLUDE_DIR}
 )
 message(STATUS "Found CNDEV: ${__found_cndev_root} (found version: ${CNDEV_VERSION_STRING})")
 set_target_properties(
  libcndev PROPERTIES IMPORTED_LOCATION ${CNDEV_LIBRARY} INTERFACE_INCLUDE_DIRECTORIES
                                                         ${CNDEV_INCLUDE_DIR})
 message(
  STATUS "Found CNDEV: ${__found_cndev_root} (found version: ${CNDEV_VERSION_STRING})")
--- a/cmake/cnlight.cmake
+++ b/cmake/cnlight.cmake
@@ -1,40 +1,49 @@
 find_library(CNLIGHT_LIBRARY 
    NAMES libcnlight.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CNLIGHT library." )
 find_library(
  CNLIGHT_LIBRARY
  NAMES libcnlight.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES lib lib64
  DOC "CNLIGHT library.")
 if(CNLIGHT_LIBRARY STREQUAL "CNLIGHT_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find CNLIGHT Library")
  message(FATAL_ERROR "Can not find CNLIGHT Library")
 endif()
 get_filename_component(__found_cnlight_root "${CNLIGHT_LIBRARY}/../.." REALPATH)
 find_path(CNLIGHT_INCLUDE_DIR 
    NAMES cnlight.h
    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnlight_root}
    PATH_SUFFIXES include 
    DOC "Path to CNLIGHT include directory." )
 find_path(
  CNLIGHT_INCLUDE_DIR
  NAMES cnlight.h
  HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnlight_root}
  PATH_SUFFIXES include
  DOC "Path to CNLIGHT include directory.")
 if(CNLIGHT_INCLUDE_DIR STREQUAL "CNLIGHT_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find CNLIGHT Library")
  message(FATAL_ERROR "Can not find CNLIGHT Library")
 endif()
 file(STRINGS "${CNLIGHT_INCLUDE_DIR}/cnlight.h" CNLIGHT_MAJOR REGEX "^#define CNLIGHT_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNLIGHT_INCLUDE_DIR}/cnlight.h" CNLIGHT_MINOR REGEX "^#define CNLIGHT_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNLIGHT_INCLUDE_DIR}/cnlight.h" CNLIGHT_PATCH REGEX "^#define CNLIGHT_PATCH_VERSION [0-9]+.*$")
 string(REGEX REPLACE "^#define CNLIGHT_MAJOR_VERSION ([0-9]+).*$" "\\1" CNLIGHT_VERSION_MAJOR "${CNLIGHT_MAJOR}")
 string(REGEX REPLACE "^#define CNLIGHT_MINOR_VERSION ([0-9]+).*$" "\\1" CNLIGHT_VERSION_MINOR "${CNLIGHT_MINOR}")
 string(REGEX REPLACE "^#define CNLIGHT_PATCH_VERSION ([0-9]+).*$" "\\1" CNLIGHT_VERSION_PATCH "${CNLIGHT_PATCH}")
 set(CNLIGHT_VERSION_STRING "${CNLIGHT_VERSION_MAJOR}.${CNLIGHT_VERSION_MINOR}.${CNLIGHT_VERSION_PATCH}")
 file(STRINGS "${CNLIGHT_INCLUDE_DIR}/cnlight.h" CNLIGHT_MAJOR
     REGEX "^#define CNLIGHT_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNLIGHT_INCLUDE_DIR}/cnlight.h" CNLIGHT_MINOR
     REGEX "^#define CNLIGHT_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNLIGHT_INCLUDE_DIR}/cnlight.h" CNLIGHT_PATCH
     REGEX "^#define CNLIGHT_PATCH_VERSION [0-9]+.*$")
 string(REGEX REPLACE "^#define CNLIGHT_MAJOR_VERSION ([0-9]+).*$" "\\1"
                     CNLIGHT_VERSION_MAJOR "${CNLIGHT_MAJOR}")
 string(REGEX REPLACE "^#define CNLIGHT_MINOR_VERSION ([0-9]+).*$" "\\1"
                     CNLIGHT_VERSION_MINOR "${CNLIGHT_MINOR}")
 string(REGEX REPLACE "^#define CNLIGHT_PATCH_VERSION ([0-9]+).*$" "\\1"
                     CNLIGHT_VERSION_PATCH "${CNLIGHT_PATCH}")
 set(CNLIGHT_VERSION_STRING
    "${CNLIGHT_VERSION_MAJOR}.${CNLIGHT_VERSION_MINOR}.${CNLIGHT_VERSION_PATCH}")
 add_library(libcnlight SHARED IMPORTED)
 set_target_properties(libcnlight PROPERTIES
    IMPORTED_LOCATION ${CNLIGHT_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${CNLIGHT_INCLUDE_DIR}
 )
 message(STATUS "Found CNLIGHT: ${__found_cnlight_root} (found version: ${CNLIGHT_VERSION_STRING})")
 set_target_properties(
  libcnlight PROPERTIES IMPORTED_LOCATION ${CNLIGHT_LIBRARY}
                        INTERFACE_INCLUDE_DIRECTORIES ${CNLIGHT_INCLUDE_DIR})
 message(
  STATUS
    "Found CNLIGHT: ${__found_cnlight_root} (found version: ${CNLIGHT_VERSION_STRING})")
--- a/cmake/cnml.cmake
+++ b/cmake/cnml.cmake
@@ -1,40 +1,48 @@
 find_library(CNML_LIBRARY 
    NAMES libcnml.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CNML library." )
 find_library(
  CNML_LIBRARY
  NAMES libcnml.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES lib lib64
  DOC "CNML library.")
 if(CNML_LIBRARY STREQUAL "CNML_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find CNML Library")
  message(FATAL_ERROR "Can not find CNML Library")
 endif()
 get_filename_component(__found_cnml_root "${CNML_LIBRARY}/../.." REALPATH)
 find_path(CNML_INCLUDE_DIR 
    NAMES cnml.h
    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnml_root}
    PATH_SUFFIXES include 
    DOC "Path to CNML include directory." )
 find_path(
  CNML_INCLUDE_DIR
  NAMES cnml.h
  HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnml_root}
  PATH_SUFFIXES include
  DOC "Path to CNML include directory.")
 if(CNML_INCLUDE_DIR STREQUAL "CNML_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find CNML Library")
  message(FATAL_ERROR "Can not find CNML Library")
 endif()
 file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_MAJOR REGEX "^#define CNML_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_MINOR REGEX "^#define CNML_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_PATCH REGEX "^#define CNML_PATCH_VERSION [0-9]+.*$")
 string(REGEX REPLACE "^#define CNML_MAJOR_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_MAJOR "${CNML_MAJOR}")
 string(REGEX REPLACE "^#define CNML_MINOR_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_MINOR "${CNML_MINOR}")
 string(REGEX REPLACE "^#define CNML_PATCH_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_PATCH "${CNML_PATCH}")
 set(CNML_VERSION_STRING "${CNML_VERSION_MAJOR}.${CNML_VERSION_MINOR}.${CNML_VERSION_PATCH}")
 file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_MAJOR
     REGEX "^#define CNML_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_MINOR
     REGEX "^#define CNML_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_PATCH
     REGEX "^#define CNML_PATCH_VERSION [0-9]+.*$")
 string(REGEX REPLACE "^#define CNML_MAJOR_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_MAJOR
                     "${CNML_MAJOR}")
 string(REGEX REPLACE "^#define CNML_MINOR_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_MINOR
                     "${CNML_MINOR}")
 string(REGEX REPLACE "^#define CNML_PATCH_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_PATCH
                     "${CNML_PATCH}")
 set(CNML_VERSION_STRING
    "${CNML_VERSION_MAJOR}.${CNML_VERSION_MINOR}.${CNML_VERSION_PATCH}")
 add_library(libcnml SHARED IMPORTED)
 set_target_properties(libcnml PROPERTIES
    IMPORTED_LOCATION ${CNML_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${CNML_INCLUDE_DIR}
 )
 message(STATUS "Found CNML: ${__found_cnml_root} (found version: ${CNML_VERSION_STRING})")
 set_target_properties(
  libcnml PROPERTIES IMPORTED_LOCATION ${CNML_LIBRARY} INTERFACE_INCLUDE_DIRECTORIES
                                                       ${CNML_INCLUDE_DIR})
 message(
  STATUS "Found CNML: ${__found_cnml_root} (found version: ${CNML_VERSION_STRING})")
--- a/cmake/cnnl.cmake
+++ b/cmake/cnnl.cmake
@@ -1,80 +1,100 @@
 find_library(CNNL_LIBRARY 
    NAMES libcnnl.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CNNL library." )
 find_library(
  CNNL_LIBRARY
  NAMES libcnnl.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES lib lib64
  DOC "CNNL library.")
 if(CNNL_LIBRARY STREQUAL "CNNL_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find CNNL Library")
  message(FATAL_ERROR "Can not find CNNL Library")
 endif()
 get_filename_component(__found_cnnl_root "${CNNL_LIBRARY}/../.." REALPATH)
 find_path(CNNL_INCLUDE_DIR 
    NAMES cnnl.h
    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnnl_root}
    PATH_SUFFIXES include 
    DOC "Path to CNNL include directory." )
 find_path(
  CNNL_INCLUDE_DIR
  NAMES cnnl.h
  HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnnl_root}
  PATH_SUFFIXES include
  DOC "Path to CNNL include directory.")
 if(CNNL_INCLUDE_DIR STREQUAL "CNNL_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find CNNL Library")
  message(FATAL_ERROR "Can not find CNNL Library")
 endif()
 file(STRINGS "${CNNL_INCLUDE_DIR}/cnnl.h" CNNL_MAJOR REGEX "^#define CNNL_MAJOR [0-9]+.*$")
 file(STRINGS "${CNNL_INCLUDE_DIR}/cnnl.h" CNNL_MINOR REGEX "^#define CNNL_MINOR [0-9]+.*$")
 file(STRINGS "${CNNL_INCLUDE_DIR}/cnnl.h" CNNL_PATCH REGEX "^#define CNNL_PATCHLEVEL [0-9]+.*$")
 string(REGEX REPLACE "^#define CNNL_MAJOR ([0-9]+).*$" "\\1" CNNL_VERSION_MAJOR "${CNNL_MAJOR}")
 string(REGEX REPLACE "^#define CNNL_MINOR ([0-9]+).*$" "\\1" CNNL_VERSION_MINOR "${CNNL_MINOR}")
 string(REGEX REPLACE "^#define CNNL_PATCHLEVEL ([0-9]+).*$" "\\1" CNNL_VERSION_PATCH "${CNNL_PATCH}")
 set(CNNL_VERSION_STRING "${CNNL_VERSION_MAJOR}.${CNNL_VERSION_MINOR}.${CNNL_VERSION_PATCH}")
 file(STRINGS "${CNNL_INCLUDE_DIR}/cnnl.h" CNNL_MAJOR
     REGEX "^#define CNNL_MAJOR [0-9]+.*$")
 file(STRINGS "${CNNL_INCLUDE_DIR}/cnnl.h" CNNL_MINOR
     REGEX "^#define CNNL_MINOR [0-9]+.*$")
 file(STRINGS "${CNNL_INCLUDE_DIR}/cnnl.h" CNNL_PATCH
     REGEX "^#define CNNL_PATCHLEVEL [0-9]+.*$")
 string(REGEX REPLACE "^#define CNNL_MAJOR ([0-9]+).*$" "\\1" CNNL_VERSION_MAJOR
                     "${CNNL_MAJOR}")
 string(REGEX REPLACE "^#define CNNL_MINOR ([0-9]+).*$" "\\1" CNNL_VERSION_MINOR
                     "${CNNL_MINOR}")
 string(REGEX REPLACE "^#define CNNL_PATCHLEVEL ([0-9]+).*$" "\\1" CNNL_VERSION_PATCH
                     "${CNNL_PATCH}")
 set(CNNL_VERSION_STRING
    "${CNNL_VERSION_MAJOR}.${CNNL_VERSION_MINOR}.${CNNL_VERSION_PATCH}")
 add_library(libcnnl SHARED IMPORTED)
 set_target_properties(libcnnl PROPERTIES
    IMPORTED_LOCATION ${CNNL_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${CNNL_INCLUDE_DIR}
 )
 set_target_properties(
  libcnnl PROPERTIES IMPORTED_LOCATION ${CNNL_LIBRARY} INTERFACE_INCLUDE_DIRECTORIES
                                                       ${CNNL_INCLUDE_DIR})
 message(STATUS "Found CNNL: ${__found_cnnl_root} (found version: ${CNNL_VERSION_STRING})")
 message(
  STATUS "Found CNNL: ${__found_cnnl_root} (found version: ${CNNL_VERSION_STRING})")
 find_library(CNNL_EXTRA_LIBRARY 
    NAMES libcnnl_extra.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CNNL_EXTRA library." )
 find_library(
  CNNL_EXTRA_LIBRARY
  NAMES libcnnl_extra.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES lib lib64
  DOC "CNNL_EXTRA library.")
 if(CNNL_EXTRA_LIBRARY STREQUAL "CNNL_EXTRA_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find CNNL_EXTRA Library")
  message(FATAL_ERROR "Can not find CNNL_EXTRA Library")
 endif()
 get_filename_component(__found_cnnl_extra_root "${CNNL_EXTRA_LIBRARY}/../.." REALPATH)
 find_path(CNNL_EXTRA_INCLUDE_DIR 
    NAMES cnnl_extra.h
    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnnl_extra_root}
    PATH_SUFFIXES include 
    DOC "Path to CNNL_EXTRA include directory." )
 find_path(
  CNNL_EXTRA_INCLUDE_DIR
  NAMES cnnl_extra.h
  HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnnl_extra_root}
  PATH_SUFFIXES include
  DOC "Path to CNNL_EXTRA include directory.")
 if(CNNL_EXTRA_INCLUDE_DIR STREQUAL "CNNL_EXTRA_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find CNNL_EXTRA Library")
  message(FATAL_ERROR "Can not find CNNL_EXTRA Library")
 endif()
 file(STRINGS "${CNNL_EXTRA_INCLUDE_DIR}/cnnl_extra.h" CNNL_EXTRA_MAJOR REGEX "^#define CNNL_EXTRA_MAJOR [0-9]+.*$")
 file(STRINGS "${CNNL_EXTRA_INCLUDE_DIR}/cnnl_extra.h" CNNL_EXTRA_MINOR REGEX "^#define CNNL_EXTRA_MINOR [0-9]+.*$")
 file(STRINGS "${CNNL_EXTRA_INCLUDE_DIR}/cnnl_extra.h" CNNL_EXTRA_PATCH REGEX "^#define CNNL_EXTRA_PATCHLEVEL [0-9]+.*$")
 string(REGEX REPLACE "^#define CNNL_EXTRA_MAJOR ([0-9]+).*$" "\\1" CNNL_EXTRA_VERSION_MAJOR "${CNNL_EXTRA_MAJOR}")
 string(REGEX REPLACE "^#define CNNL_EXTRA_MINOR ([0-9]+).*$" "\\1" CNNL_EXTRA_VERSION_MINOR "${CNNL_EXTRA_MINOR}")
 string(REGEX REPLACE "^#define CNNL_EXTRA_PATCHLEVEL ([0-9]+).*$" "\\1" CNNL_EXTRA_VERSION_PATCH "${CNNL_EXTRA_PATCH}")
 set(CNNL_EXTRA_VERSION_STRING "${CNNL_EXTRA_VERSION_MAJOR}.${CNNL_EXTRA_VERSION_MINOR}.${CNNL_EXTRA_VERSION_PATCH}")
 file(STRINGS "${CNNL_EXTRA_INCLUDE_DIR}/cnnl_extra.h" CNNL_EXTRA_MAJOR
     REGEX "^#define CNNL_EXTRA_MAJOR [0-9]+.*$")
 file(STRINGS "${CNNL_EXTRA_INCLUDE_DIR}/cnnl_extra.h" CNNL_EXTRA_MINOR
     REGEX "^#define CNNL_EXTRA_MINOR [0-9]+.*$")
 file(STRINGS "${CNNL_EXTRA_INCLUDE_DIR}/cnnl_extra.h" CNNL_EXTRA_PATCH
     REGEX "^#define CNNL_EXTRA_PATCHLEVEL [0-9]+.*$")
 string(REGEX REPLACE "^#define CNNL_EXTRA_MAJOR ([0-9]+).*$" "\\1"
                     CNNL_EXTRA_VERSION_MAJOR "${CNNL_EXTRA_MAJOR}")
 string(REGEX REPLACE "^#define CNNL_EXTRA_MINOR ([0-9]+).*$" "\\1"
                     CNNL_EXTRA_VERSION_MINOR "${CNNL_EXTRA_MINOR}")
 string(REGEX REPLACE "^#define CNNL_EXTRA_PATCHLEVEL ([0-9]+).*$" "\\1"
                     CNNL_EXTRA_VERSION_PATCH "${CNNL_EXTRA_PATCH}")
 set(CNNL_EXTRA_VERSION_STRING
    "${CNNL_EXTRA_VERSION_MAJOR}.${CNNL_EXTRA_VERSION_MINOR}.${CNNL_EXTRA_VERSION_PATCH}"
 )
 add_library(libcnnl_extra SHARED IMPORTED)
 set_target_properties(libcnnl_extra PROPERTIES
    IMPORTED_LOCATION ${CNNL_EXTRA_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${CNNL_EXTRA_INCLUDE_DIR}
 )
 message(STATUS "Found CNNL_EXTRA: ${__found_cnnl_extra_root} (found version: ${CNNL_EXTRA_VERSION_STRING})")
 set_target_properties(
  libcnnl_extra PROPERTIES IMPORTED_LOCATION ${CNNL_EXTRA_LIBRARY}
                           INTERFACE_INCLUDE_DIRECTORIES ${CNNL_EXTRA_INCLUDE_DIR})
 message(
  STATUS
    "Found CNNL_EXTRA: ${__found_cnnl_extra_root} (found version: ${CNNL_EXTRA_VERSION_STRING})"
 )
--- a/cmake/cnrt.cmake
+++ b/cmake/cnrt.cmake
@@ -1,40 +1,48 @@
 find_library(CNRT_LIBRARY 
    NAMES libcnrt.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CNRT library." )
 find_library(
  CNRT_LIBRARY
  NAMES libcnrt.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES lib lib64
  DOC "CNRT library.")
 if(CNRT_LIBRARY STREQUAL "CNRT_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find CNRT Library")
  message(FATAL_ERROR "Can not find CNRT Library")
 endif()
 get_filename_component(__found_cnrt_root ${CNRT_LIBRARY}/../../ REALPATH)
 find_path(CNRT_INCLUDE_DIR 
    NAMES cnrt.h
    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnrt_root}
    PATH_SUFFIXES include 
    DOC "Path to CNRT include directory." )
 find_path(
  CNRT_INCLUDE_DIR
  NAMES cnrt.h
  HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnrt_root}
  PATH_SUFFIXES include
  DOC "Path to CNRT include directory.")
 if(CNRT_INCLUDE_DIR STREQUAL "CNRT_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find CNRT Library")
  message(FATAL_ERROR "Can not find CNRT Library")
 endif()
 file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_MAJOR REGEX "^#define CNRT_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_MINOR REGEX "^#define CNRT_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_PATCH REGEX "^#define CNRT_PATCH_VERSION [0-9]+.*$")
 string(REGEX REPLACE "^#define CNRT_MAJOR_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_MAJOR "${CNRT_MAJOR}")
 string(REGEX REPLACE "^#define CNRT_MINOR_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_MINOR "${CNRT_MINOR}")
 string(REGEX REPLACE "^#define CNRT_PATCH_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_PATCH "${CNRT_PATCH}")
 set(CNRT_VERSION_STRING "${CNRT_VERSION_MAJOR}.${CNRT_VERSION_MINOR}.${CNRT_VERSION_PATCH}")
 file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_MAJOR
     REGEX "^#define CNRT_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_MINOR
     REGEX "^#define CNRT_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_PATCH
     REGEX "^#define CNRT_PATCH_VERSION [0-9]+.*$")
 string(REGEX REPLACE "^#define CNRT_MAJOR_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_MAJOR
                     "${CNRT_MAJOR}")
 string(REGEX REPLACE "^#define CNRT_MINOR_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_MINOR
                     "${CNRT_MINOR}")
 string(REGEX REPLACE "^#define CNRT_PATCH_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_PATCH
                     "${CNRT_PATCH}")
 set(CNRT_VERSION_STRING
    "${CNRT_VERSION_MAJOR}.${CNRT_VERSION_MINOR}.${CNRT_VERSION_PATCH}")
 add_library(libcnrt SHARED IMPORTED)
 set_target_properties(libcnrt PROPERTIES
    IMPORTED_LOCATION ${CNRT_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${CNRT_INCLUDE_DIR}
 )
 message(STATUS "Found CNRT: ${__found_cnrt_root} (found version: ${CNRT_VERSION_STRING})")
 set_target_properties(
  libcnrt PROPERTIES IMPORTED_LOCATION ${CNRT_LIBRARY} INTERFACE_INCLUDE_DIRECTORIES
                                                       ${CNRT_INCLUDE_DIR})
 message(
  STATUS "Found CNRT: ${__found_cnrt_root} (found version: ${CNRT_VERSION_STRING})")
--- a/cmake/cpp_redis.cmake
+++ b/cmake/cpp_redis.cmake
@@ -1,2 +1,5 @@
 file(GLOB_RECURSE CPP_REDIS_SRCS ${PROJECT_SOURCE_DIR}/third_party/cpp_redis/sources/*.cpp ${PROJECT_SOURCE_DIR}/third_party/tacopie/sources/*.cpp)
 set(CPP_REDIS_INCLUDES ${PROJECT_SOURCE_DIR}/third_party/cpp_redis/includes ${PROJECT_SOURCE_DIR}/third_party/tacopie/includes)
 file(GLOB_RECURSE CPP_REDIS_SRCS
     ${PROJECT_SOURCE_DIR}/third_party/cpp_redis/sources/*.cpp
     ${PROJECT_SOURCE_DIR}/third_party/tacopie/sources/*.cpp)
 set(CPP_REDIS_INCLUDES ${PROJECT_SOURCE_DIR}/third_party/cpp_redis/includes
                       ${PROJECT_SOURCE_DIR}/third_party/tacopie/includes)
--- a/cmake/cpuinfo.cmake
+++ b/cmake/cpuinfo.cmake
@@ -1,20 +1,20 @@
 if (MGE_USE_SYSTEM_LIB)
    find_package(Cpuinfo)
    message(STATUS "Using system provided cpuinfo ${cpuinfo_VERSION}")
    add_library(libcpuinfo IMPORTED GLOBAL)
    set_target_properties(
        libcpuinfo PROPERTIES
        IMPORTED_LOCATION ${cpuinfo_LIBRARIES}
        INTERFACE_INCLUDE_DIRECTORIES ${cpuinfo_INCLUDE_DIRS}
        )
    return()
 if(MGE_USE_SYSTEM_LIB)
  find_package(Cpuinfo)
  message(STATUS "Using system provided cpuinfo ${cpuinfo_VERSION}")
  add_library(libcpuinfo IMPORTED GLOBAL)
  set_target_properties(
    libcpuinfo PROPERTIES IMPORTED_LOCATION ${cpuinfo_LIBRARIES}
                          INTERFACE_INCLUDE_DIRECTORIES ${cpuinfo_INCLUDE_DIRS})
  return()
 endif()
 SET(CPUINFO_LIBRARY_TYPE "static" CACHE STRING "Type of cpuinfo library (shared, static, or default) to build")
 OPTION(CPUINFO_BUILD_TOOLS "Build command-line tools" OFF)
 OPTION(CPUINFO_BUILD_UNIT_TESTS "Build cpuinfo unit tests" OFF)
 OPTION(CPUINFO_BUILD_MOCK_TESTS "Build cpuinfo mock tests" OFF)
 OPTION(CPUINFO_BUILD_BENCHMARKS "Build cpuinfo micro-benchmarks" OFF)
 set(CPUINFO_LIBRARY_TYPE
    "static"
    CACHE STRING "Type of cpuinfo library (shared, static, or default) to build")
 option(CPUINFO_BUILD_TOOLS "Build command-line tools" OFF)
 option(CPUINFO_BUILD_UNIT_TESTS "Build cpuinfo unit tests" OFF)
 option(CPUINFO_BUILD_MOCK_TESTS "Build cpuinfo mock tests" OFF)
 option(CPUINFO_BUILD_BENCHMARKS "Build cpuinfo micro-benchmarks" OFF)
 include_directories("${PROJECT_SOURCE_DIR}/third_party/cpuinfo/include")
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/cpuinfo ${CMAKE_CURRENT_BINARY_DIR}/cpuinfo EXCLUDE_FROM_ALL)
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/cpuinfo
                 ${CMAKE_CURRENT_BINARY_DIR}/cpuinfo EXCLUDE_FROM_ALL)
--- a/cmake/cudnn.cmake
+++ b/cmake/cudnn.cmake
@@ -1,73 +1,83 @@
 find_package(PkgConfig)
 if(${PkgConfig_FOUND})
    pkg_check_modules(PC_CUDNN QUIET CUDNN)
  pkg_check_modules(PC_CUDNN QUIET CUDNN)
 endif()
 if("${CUDNN_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{CUDNN_ROOT_DIR}"  STREQUAL "")
    set(CUDNN_ROOT_DIR $ENV{CUDNN_ROOT_DIR})
 if("${CUDNN_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{CUDNN_ROOT_DIR}" STREQUAL "")
  set(CUDNN_ROOT_DIR $ENV{CUDNN_ROOT_DIR})
 endif()
 if(MGE_CUDA_USE_STATIC AND NOT MGE_WITH_CUDNN_SHARED)
    find_library(CUDNN_LIBRARY 
        NAMES libcudnn_static.a cudnn.lib
        PATHS ${ALTER_LD_LIBRARY_PATHS} ${CUDNN_ROOT_DIR} ${PC_CUDNN_LIBRARY_DIRS} ${CMAKE_INSTALL_PREFIX}
        HINTS ${ALTER_LIBRARY_PATHS}
        PATH_SUFFIXES lib lib64
        DOC "CUDNN library." )
  find_library(
    CUDNN_LIBRARY
    NAMES libcudnn_static.a cudnn.lib
    PATHS ${ALTER_LD_LIBRARY_PATHS} ${CUDNN_ROOT_DIR} ${PC_CUDNN_LIBRARY_DIRS}
          ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CUDNN library.")
 else()
    find_library(CUDNN_LIBRARY 
        NAMES libcudnn.so libcudnn.dylib cudnn64.dll
        PATHS ${ALTER_LD_LIBRARY_PATHS} ${CUDNN_ROOT_DIR} ${PC_CUDNN_LIBRARY_DIRS} ${CMAKE_INSTALL_PREFIX}
        HINTS ${ALTER_LIBRARY_PATHS}
        PATH_SUFFIXES lib lib64
        DOC "CUDNN library." )
  find_library(
    CUDNN_LIBRARY
    NAMES libcudnn.so libcudnn.dylib cudnn64.dll
    PATHS ${ALTER_LD_LIBRARY_PATHS} ${CUDNN_ROOT_DIR} ${PC_CUDNN_LIBRARY_DIRS}
          ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "CUDNN library.")
 endif()
 if(CUDNN_LIBRARY STREQUAL "CUDNN_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find CuDNN Library, please refer to scripts/cmake-build/BUILD_README.md to init CUDNN env")
  message(
    FATAL_ERROR
      "Can not find CuDNN Library, please refer to scripts/cmake-build/BUILD_README.md to init CUDNN env"
  )
 endif()
 get_filename_component(__found_cudnn_root ${CUDNN_LIBRARY}/../.. REALPATH)
 find_path(CUDNN_INCLUDE_DIR 
    NAMES cudnn.h
    HINTS $ENV{PC_CUDNN_INCLUDE_DIRS} ${CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_INCLUDE} ${__found_cudnn_root}
    PATH_SUFFIXES include 
    DOC "Path to CUDNN include directory." )
 find_path(
  CUDNN_INCLUDE_DIR
  NAMES cudnn.h
  HINTS $ENV{PC_CUDNN_INCLUDE_DIRS} ${CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_INCLUDE}
        ${__found_cudnn_root}
  PATH_SUFFIXES include
  DOC "Path to CUDNN include directory.")
 if(CUDNN_INCLUDE_DIR STREQUAL "CUDNN_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find CuDNN INCLUDE, please refer to scripts/cmake-build/BUILD_README.md to init CUDNN env")
  message(
    FATAL_ERROR
      "Can not find CuDNN INCLUDE, please refer to scripts/cmake-build/BUILD_README.md to init CUDNN env"
  )
 endif()
 if(EXISTS ${CUDNN_INCLUDE_DIR}/cudnn_version.h)
    file(READ ${CUDNN_INCLUDE_DIR}/cudnn_version.h CUDNN_VERSION_FILE_CONTENTS)
  file(READ ${CUDNN_INCLUDE_DIR}/cudnn_version.h CUDNN_VERSION_FILE_CONTENTS)
 else()
    file(READ ${CUDNN_INCLUDE_DIR}/cudnn.h CUDNN_VERSION_FILE_CONTENTS)
  file(READ ${CUDNN_INCLUDE_DIR}/cudnn.h CUDNN_VERSION_FILE_CONTENTS)
 endif()
 string(REGEX MATCH "define CUDNN_MAJOR * +([0-9]+)"
    CUDNN_MAJOR_VERSION "${CUDNN_VERSION_FILE_CONTENTS}")
 string(REGEX REPLACE "define CUDNN_MAJOR * +([0-9]+)" "\\1"
    CUDNN_MAJOR_VERSION "${CUDNN_MAJOR_VERSION}")
 string(REGEX MATCH "define CUDNN_MINOR * +([0-9]+)"
    CUDNN_MINOR_VERSION "${CUDNN_VERSION_FILE_CONTENTS}")
 string(REGEX REPLACE "define CUDNN_MINOR * +([0-9]+)" "\\1"
    CUDNN_MINOR_VERSION "${CUDNN_MINOR_VERSION}")
 string(REGEX MATCH "define CUDNN_PATCHLEVEL * +([0-9]+)"
    CUDNN_PATCH_VERSION "${CUDNN_VERSION_FILE_CONTENTS}")
 string(REGEX REPLACE "define CUDNN_PATCHLEVEL * +([0-9]+)" "\\1"
    CUDNN_PATCH_VERSION "${CUDNN_PATCH_VERSION}")  
 string(REGEX MATCH "define CUDNN_MAJOR * +([0-9]+)" CUDNN_MAJOR_VERSION
             "${CUDNN_VERSION_FILE_CONTENTS}")
 string(REGEX REPLACE "define CUDNN_MAJOR * +([0-9]+)" "\\1" CUDNN_MAJOR_VERSION
                     "${CUDNN_MAJOR_VERSION}")
 string(REGEX MATCH "define CUDNN_MINOR * +([0-9]+)" CUDNN_MINOR_VERSION
             "${CUDNN_VERSION_FILE_CONTENTS}")
 string(REGEX REPLACE "define CUDNN_MINOR * +([0-9]+)" "\\1" CUDNN_MINOR_VERSION
                     "${CUDNN_MINOR_VERSION}")
 string(REGEX MATCH "define CUDNN_PATCHLEVEL * +([0-9]+)" CUDNN_PATCH_VERSION
             "${CUDNN_VERSION_FILE_CONTENTS}")
 string(REGEX REPLACE "define CUDNN_PATCHLEVEL * +([0-9]+)" "\\1" CUDNN_PATCH_VERSION
                     "${CUDNN_PATCH_VERSION}")
 set(CUDNN_VERSION ${CUDNN_MAJOR_VERSION}.${CUDNN_MINOR_VERSION}.${CUDNN_PATCH_VERSION})
 if(MGE_CUDA_USE_STATIC)
    add_library(libcudnn STATIC IMPORTED)
  add_library(libcudnn STATIC IMPORTED)
 else()
    add_library(libcudnn SHARED IMPORTED)
  add_library(libcudnn SHARED IMPORTED)
 endif()
 set_target_properties(libcudnn PROPERTIES
    IMPORTED_LOCATION ${CUDNN_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${CUDNN_INCLUDE_DIR})
 set_target_properties(
  libcudnn PROPERTIES IMPORTED_LOCATION ${CUDNN_LIBRARY} INTERFACE_INCLUDE_DIRECTORIES
                                                         ${CUDNN_INCLUDE_DIR})
 message(STATUS "Found CuDNN: ${__found_cudnn_root} (found version: ${CUDNN_VERSION})")
--- a/cmake/flatbuffers.cmake
+++ b/cmake/flatbuffers.cmake
@@ -1,27 +1,47 @@
 if (MGE_USE_SYSTEM_LIB)
    find_package(Flatbuffers REQUIRED)
    message(STATUS "Using system provided Flatbuffers ${Flatbuffers_VERSION}")
    include(cmake/BuildFlatBuffers.cmake)
    return()
 if(MGE_USE_SYSTEM_LIB)
  find_package(Flatbuffers REQUIRED)
  message(STATUS "Using system provided Flatbuffers ${Flatbuffers_VERSION}")
  include(cmake/BuildFlatBuffers.cmake)
  return()
 endif()
 if(MSVC OR WIN32)
    message(DEBUG "add flags flatc for clang-cl build")
    set(FLATC_FLAGS "")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=unknown-argument -Wno-error=c++98-compat -Wno-error=reserved-id-macro")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=sign-conversion -Wno-error=exceptions -Wno-error=argument-outside-range")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=delete-non-virtual-dtor -Wno-error=ignored-attributes -Wno-error=format")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=sign-compare -Wno-error=unused-private-field -Wno-error=braced-scalar-init")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=return-type-c-linkage -Wno-error=invalid-noreturn -Wno-error=c++98-compat-pedantic")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=extra-semi-stmt -Wno-error=missing-prototypes -Wno-error=documentation-unknown-command")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=missing-variable-declarations  -Wno-error=nonportable-system-include-path")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=exit-time-destructors -Wno-error=unused-macros -Wno-error=global-constructors")
    set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=switch-enum -Wno-error=missing-noreturn -Wno-error=float-equal")
    if (${CMAKE_CXX_COMPILER_VERSION} VERSION_GREATER_EQUAL "11.0.0")
        set(FLATC_FLAGS "${FLATC_FLAGS} -Wno-error=suggest-override -Wno-error=suggest-destructor-override")
    endif()
  message(DEBUG "add flags flatc for clang-cl build")
  set(FLATC_FLAGS "")
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=unknown-argument -Wno-error=c++98-compat -Wno-error=reserved-id-macro"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=sign-conversion -Wno-error=exceptions -Wno-error=argument-outside-range"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=delete-non-virtual-dtor -Wno-error=ignored-attributes -Wno-error=format"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=sign-compare -Wno-error=unused-private-field -Wno-error=braced-scalar-init"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=return-type-c-linkage -Wno-error=invalid-noreturn -Wno-error=c++98-compat-pedantic"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=extra-semi-stmt -Wno-error=missing-prototypes -Wno-error=documentation-unknown-command"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=missing-variable-declarations  -Wno-error=nonportable-system-include-path"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=exit-time-destructors -Wno-error=unused-macros -Wno-error=global-constructors"
  )
  set(FLATC_FLAGS
      "${FLATC_FLAGS} -Wno-error=switch-enum -Wno-error=missing-noreturn -Wno-error=float-equal"
  )
  if(${CMAKE_CXX_COMPILER_VERSION} VERSION_GREATER_EQUAL "11.0.0")
    set(FLATC_FLAGS
        "${FLATC_FLAGS} -Wno-error=suggest-override -Wno-error=suggest-destructor-override"
    )
  endif()
    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${FLATC_FLAGS}")
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${FLATC_FLAGS}")
  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${FLATC_FLAGS}")
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${FLATC_FLAGS}")
 endif()
 option(FLATBUFFERS_BUILD_TESTS "" OFF)
--- a/cmake/gflags.cmake
+++ b/cmake/gflags.cmake
@@ -1 +1,2 @@
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/gflags ${CMAKE_CURRENT_BINARY_DIR}/gflags)
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/gflags
                 ${CMAKE_CURRENT_BINARY_DIR}/gflags)
--- a/cmake/gtest.cmake
+++ b/cmake/gtest.cmake
@@ -1,2 +1,2 @@
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/gtest ${CMAKE_CURRENT_BINARY_DIR}/gtest EXCLUDE_FROM_ALL)
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/gtest
                 ${CMAKE_CURRENT_BINARY_DIR}/gtest EXCLUDE_FROM_ALL)
--- a/cmake/llvm-project.cmake
+++ b/cmake/llvm-project.cmake
@@ -1,88 +1,136 @@
 # - Find the llvm/mlir libraries
 # This module finds if llvm/mlir is installed, or build llvm/mlir from source.
 # This module sets the following variables.
 # * Find the llvm/mlir libraries This module finds if llvm/mlir is installed, or build
 #   llvm/mlir from source. This module sets the following variables.
 #
 #  MLIR_LLVM_INCLUDE_DIR     - path to the LLVM/MLIR include files
 #  MLIR_LLVM_LIBS            - path to the LLVM/MLIR libraries
 # MLIR_LLVM_INCLUDE_DIR     - path to the LLVM/MLIR include files MLIR_LLVM_LIBS - path
 # to the LLVM/MLIR libraries
 #
 # This module define the following functions.
 #
 # external_tablegen_library  - created interface library which depends on tablegen outputs
 # external_tablegen_library  - created interface library which depends on tablegen
 # outputs
 include(CMakeParseArguments)
 function(external_tablegen_library)
    cmake_parse_arguments(
        _RULE
        "TESTONLY"
        "NAME;TBLGEN"
        "SRCS;INCLUDES;OUTS"
        ${ARGN}
        )
  cmake_parse_arguments(_RULE "TESTONLY" "NAME;TBLGEN" "SRCS;INCLUDES;OUTS" ${ARGN})
    if(_RULE_TESTONLY AND NOT MGE_WITH_TEST)
        return()
    endif()
  if(_RULE_TESTONLY AND NOT MGE_WITH_TEST)
    return()
  endif()
    set(_NAME ${_RULE_NAME})
  set(_NAME ${_RULE_NAME})
    set(LLVM_TARGET_DEFINITIONS ${_RULE_SRCS})
    set(_INCLUDE_DIRS ${_RULE_INCLUDES})
    list(TRANSFORM _INCLUDE_DIRS PREPEND "-I")
    set(_OUTPUTS)
    while(_RULE_OUTS)
        list(GET _RULE_OUTS 0 _COMMAND)
        list(REMOVE_AT _RULE_OUTS 0)
        list(GET _RULE_OUTS 0 _FILE)
        list(REMOVE_AT _RULE_OUTS 0)
        tablegen(${_RULE_TBLGEN} ${_FILE} ${_COMMAND} ${_INCLUDE_DIRS})
        list(APPEND _OUTPUTS ${CMAKE_CURRENT_BINARY_DIR}/${_FILE})
    endwhile()
    add_custom_target(${_NAME}_target DEPENDS ${_OUTPUTS})
  set(LLVM_TARGET_DEFINITIONS ${_RULE_SRCS})
  set(_INCLUDE_DIRS ${_RULE_INCLUDES})
  list(TRANSFORM _INCLUDE_DIRS PREPEND "-I")
  set(_OUTPUTS)
  while(_RULE_OUTS)
    list(GET _RULE_OUTS 0 _COMMAND)
    list(REMOVE_AT _RULE_OUTS 0)
    list(GET _RULE_OUTS 0 _FILE)
    list(REMOVE_AT _RULE_OUTS 0)
    tablegen(${_RULE_TBLGEN} ${_FILE} ${_COMMAND} ${_INCLUDE_DIRS})
    list(APPEND _OUTPUTS ${CMAKE_CURRENT_BINARY_DIR}/${_FILE})
  endwhile()
  add_custom_target(${_NAME}_target DEPENDS ${_OUTPUTS})
    add_library(${_NAME} INTERFACE)
    add_dependencies(${_NAME} ${_NAME}_target)
  add_library(${_NAME} INTERFACE)
  add_dependencies(${_NAME} ${_NAME}_target)
    target_include_directories(${_NAME} INTERFACE
        "$<BUILD_INTERFACE:${_RULE_INCLUDES}>")
  target_include_directories(${_NAME} INTERFACE "$<BUILD_INTERFACE:${_RULE_INCLUDES}>")
    install(TARGETS ${_NAME} EXPORT ${MGE_EXPORT_TARGETS})
  install(TARGETS ${_NAME} EXPORT ${MGE_EXPORT_TARGETS})
 endfunction()
 set(LLVM_LIBS LLVMCore LLVMSupport LLVMX86CodeGen LLVMOrcJIT LLVMNVPTXCodeGen LLVMNVPTXDesc LLVMNVPTXInfo)
 set(MLIR_CORE_LIBS MLIRAnalysis MLIRExecutionEngine MLIRIR MLIRParser MLIRPass MLIRSideEffectInterfaces MLIRTransforms)
 set(MLIR_DIALECT_LIBS MLIRAsync MLIRAVX512 MLIRGPU MLIRLLVMAVX512 MLIRNVVMIR MLIROpenACC MLIRPDL MLIRPDLInterp MLIRQuant MLIRROCDLIR MLIRSDBM MLIRShape MLIRSPIRV MLIRStandardOpsTransforms MLIRTosa)
 set(MLIR_CONVERSION_LIBS MLIRAffineToStandard MLIRAVX512ToLLVM MLIRGPUToGPURuntimeTransforms MLIRGPUToNVVMTransforms MLIRSCFToStandard)
 set(LLVM_LIBS
    LLVMCore
    LLVMSupport
    LLVMX86CodeGen
    LLVMOrcJIT
    LLVMNVPTXCodeGen
    LLVMNVPTXDesc
    LLVMNVPTXInfo)
 set(MLIR_CORE_LIBS
    MLIRAnalysis
    MLIRExecutionEngine
    MLIRIR
    MLIRParser
    MLIRPass
    MLIRSideEffectInterfaces
    MLIRTransforms)
 set(MLIR_DIALECT_LIBS
    MLIRAsync
    MLIRAVX512
    MLIRGPU
    MLIRLLVMAVX512
    MLIRNVVMIR
    MLIROpenACC
    MLIRPDL
    MLIRPDLInterp
    MLIRQuant
    MLIRROCDLIR
    MLIRSDBM
    MLIRShape
    MLIRSPIRV
    MLIRStandardOpsTransforms
    MLIRTosa)
 set(MLIR_CONVERSION_LIBS
    MLIRAffineToStandard MLIRAVX512ToLLVM MLIRGPUToGPURuntimeTransforms
    MLIRGPUToNVVMTransforms MLIRSCFToStandard)
 set(MLIR_TRANSLATION_LIBS MLIRTargetLLVMIR MLIRTargetNVVMIR)
 set(MLIR_LIBS ${MLIR_CORE_LIBS} ${MLIR_DIALECT_LIBS} ${MLIR_CONVERSION_LIBS} ${MLIR_TRANSLATION_LIBS})
 set(MLIR_LIBS ${MLIR_CORE_LIBS} ${MLIR_DIALECT_LIBS} ${MLIR_CONVERSION_LIBS}
              ${MLIR_TRANSLATION_LIBS})
 set(MLIR_LLVM_LIBS ${LLVM_LIBS} ${MLIR_LIBS})
 function(add_mge_mlir_src_dep llvm_monorepo_path)
    set(_CMAKE_BUILD_TYPE "${CMAKE_BUILD_TYPE}")
    string(TOUPPER "${CMAKE_BUILD_TYPE}" uppercase_CMAKE_BUILD_TYPE)
    if(NOT uppercase_CMAKE_BUILD_TYPE MATCHES "^(DEBUG|RELEASE|RELWITHDEBINFO|MINSIZEREL)$")
        set(CMAKE_BUILD_TYPE "Debug")
    endif()
    set(_CMAKE_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
    set(BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
  set(_CMAKE_BUILD_TYPE "${CMAKE_BUILD_TYPE}")
  string(TOUPPER "${CMAKE_BUILD_TYPE}" uppercase_CMAKE_BUILD_TYPE)
  if(NOT uppercase_CMAKE_BUILD_TYPE MATCHES
     "^(DEBUG|RELEASE|RELWITHDEBINFO|MINSIZEREL)$")
    set(CMAKE_BUILD_TYPE "Debug")
  endif()
  set(_CMAKE_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
  set(BUILD_SHARED_LIBS
      OFF
      CACHE BOOL "" FORCE)
    add_subdirectory("${llvm_monorepo_path}/llvm" ${LLVM_BUILD_DIR} EXCLUDE_FROM_ALL)
  add_subdirectory("${llvm_monorepo_path}/llvm" ${LLVM_BUILD_DIR} EXCLUDE_FROM_ALL)
    # Reset CMAKE_BUILD_TYPE to its previous setting
    set(CMAKE_BUILD_TYPE "${_CMAKE_BUILD_TYPE}" CACHE STRING "Build type" FORCE)
    # Reset BUILD_SHARED_LIBS to its previous setting
    set(BUILD_SHARED_LIBS ${_CMAKE_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libraries" FORCE)
  # Reset CMAKE_BUILD_TYPE to its previous setting
  set(CMAKE_BUILD_TYPE
      "${_CMAKE_BUILD_TYPE}"
      CACHE STRING "Build type" FORCE)
  # Reset BUILD_SHARED_LIBS to its previous setting
  set(BUILD_SHARED_LIBS
      ${_CMAKE_BUILD_SHARED_LIBS}
      CACHE BOOL "Build shared libraries" FORCE)
 endfunction()
 # llvm build options
 set(LLVM_INCLUDE_EXAMPLES OFF CACHE BOOL "" FORCE)
 set(LLVM_INCLUDE_TESTS OFF CACHE BOOL "" FORCE)
 set(LLVM_INCLUDE_DOCS OFF CACHE BOOL "" FORCE)
 set(LLVM_ENABLE_BINDINGS OFF CACHE BOOL "" FORCE)
 set(LLVM_INCLUDE_BENCHMARKS OFF CACHE BOOL "" FORCE)
 set(LLVM_ENABLE_RTTI ${MGE_ENABLE_RTTI} CACHE BOOL "" FORCE)
 set(LLVM_TARGETS_TO_BUILD "X86;NVPTX;AArch64;ARM" CACHE STRING "" FORCE)
 set(LLVM_ENABLE_PROJECTS "mlir" CACHE STRING "" FORCE)
 set(LLVM_INCLUDE_EXAMPLES
    OFF
    CACHE BOOL "" FORCE)
 set(LLVM_INCLUDE_TESTS
    OFF
    CACHE BOOL "" FORCE)
 set(LLVM_INCLUDE_DOCS
    OFF
    CACHE BOOL "" FORCE)
 set(LLVM_ENABLE_BINDINGS
    OFF
    CACHE BOOL "" FORCE)
 set(LLVM_INCLUDE_BENCHMARKS
    OFF
    CACHE BOOL "" FORCE)
 set(LLVM_ENABLE_RTTI
    ${MGE_ENABLE_RTTI}
    CACHE BOOL "" FORCE)
 set(LLVM_TARGETS_TO_BUILD
    "X86;NVPTX;AArch64;ARM"
    CACHE STRING "" FORCE)
 set(LLVM_ENABLE_PROJECTS
    "mlir"
    CACHE STRING "" FORCE)
 set(LLVM_BUILD_DIR ${PROJECT_BINARY_DIR}/third_party/llvm-project/llvm)
 add_mge_mlir_src_dep("third_party/llvm-project")
@@ -91,6 +139,5 @@ set(MLIR_LLVM_INCLUDE_DIR
    ${PROJECT_SOURCE_DIR}/third_party/llvm-project/llvm/include
    ${PROJECT_BINARY_DIR}/third_party/llvm-project/llvm/include
    ${PROJECT_SOURCE_DIR}/third_party/llvm-project/mlir/include
    ${PROJECT_BINARY_DIR}/third_party/llvm-project/llvm/tools/mlir/include
    )
    ${PROJECT_BINARY_DIR}/third_party/llvm-project/llvm/tools/mlir/include)
 set(MLIR_TABLEGEN_EXE mlir-tblgen)
--- a/cmake/magicmind.cmake
+++ b/cmake/magicmind.cmake
@@ -1,54 +1,64 @@
 find_library(MAGICMIND_LIBRARY 
    NAMES libmagicmind.so
    PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "MAGICMIND library." )
 find_library(
  MAGICMIND_LIBRARY
  NAMES libmagicmind.so
  PATHS ${ALTER_LD_LIBRARY_PATHS} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
  HINTS ${ALTER_LIBRARY_PATHS}
  PATH_SUFFIXES lib lib64
  DOC "MAGICMIND library.")
 if(MAGICMIND_LIBRARY STREQUAL "MAGICMIND_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find MAGICMIND Library")
  message(FATAL_ERROR "Can not find MAGICMIND Library")
 endif()
 get_filename_component(__found_magicmind_root "${MAGICMIND_LIBRARY}/../../" REALPATH)
 find_path(MAGICMIND_INCLUDE_DIR 
    NAMES common.h
    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_magicmind_root}
    PATH_SUFFIXES include 
    DOC "Path to MAGICMIND include directory." )
 find_path(
  MAGICMIND_INCLUDE_DIR
  NAMES common.h
  HINTS "$ENV{NEUWARE_HOME}/include" ${__found_magicmind_root}
  PATH_SUFFIXES include
  DOC "Path to MAGICMIND include directory.")
 if(MAGICMIND_INCLUDE_DIR STREQUAL "MAGICMIND_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find MAGICMIND Library")
  message(FATAL_ERROR "Can not find MAGICMIND Library")
 endif()
 file(STRINGS "${MAGICMIND_INCLUDE_DIR}/common.h" MAGICMIND_MAJOR REGEX "^#define MM_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${MAGICMIND_INCLUDE_DIR}/common.h" MAGICMIND_MINOR REGEX "^#define MM_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${MAGICMIND_INCLUDE_DIR}/common.h" MAGICMIND_PATCH REGEX "^#define MM_PATCH_VERSION [0-9]+.*$")
 file(STRINGS "${MAGICMIND_INCLUDE_DIR}/common.h" MAGICMIND_MAJOR
     REGEX "^#define MM_MAJOR_VERSION [0-9]+.*$")
 file(STRINGS "${MAGICMIND_INCLUDE_DIR}/common.h" MAGICMIND_MINOR
     REGEX "^#define MM_MINOR_VERSION [0-9]+.*$")
 file(STRINGS "${MAGICMIND_INCLUDE_DIR}/common.h" MAGICMIND_PATCH
     REGEX "^#define MM_PATCH_VERSION [0-9]+.*$")
 string(REGEX REPLACE "^#define MM_MAJOR_VERSION ([0-9]+).*$" "\\1" MAGICMIND_VERSION_MAJOR "${MAGICMIND_MAJOR}")
 string(REGEX REPLACE "^#define MM_MINOR_VERSION ([0-9]+).*$" "\\1" MAGICMIND_VERSION_MINOR "${MAGICMIND_MINOR}")
 string(REGEX REPLACE "^#define MM_PATCH_VERSION ([0-9]+).*$" "\\1" MAGICMIND_VERSION_PATCH "${MAGICMIND_PATCH}")
 set(MAGICMIND_VERSION_STRING "${MAGICMIND_VERSION_MAJOR}.${MAGICMIND_VERSION_MINOR}.${MAGICMIND_VERSION_PATCH}")
 string(REGEX REPLACE "^#define MM_MAJOR_VERSION ([0-9]+).*$" "\\1"
                     MAGICMIND_VERSION_MAJOR "${MAGICMIND_MAJOR}")
 string(REGEX REPLACE "^#define MM_MINOR_VERSION ([0-9]+).*$" "\\1"
                     MAGICMIND_VERSION_MINOR "${MAGICMIND_MINOR}")
 string(REGEX REPLACE "^#define MM_PATCH_VERSION ([0-9]+).*$" "\\1"
                     MAGICMIND_VERSION_PATCH "${MAGICMIND_PATCH}")
 set(MAGICMIND_VERSION_STRING
    "${MAGICMIND_VERSION_MAJOR}.${MAGICMIND_VERSION_MINOR}.${MAGICMIND_VERSION_PATCH}")
 add_library(libmagicmind SHARED IMPORTED)
 set_target_properties(libmagicmind PROPERTIES
    IMPORTED_LOCATION ${MAGICMIND_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${MAGICMIND_INCLUDE_DIR}
 )
 set_target_properties(
  libmagicmind PROPERTIES IMPORTED_LOCATION ${MAGICMIND_LIBRARY}
                          INTERFACE_INCLUDE_DIRECTORIES ${MAGICMIND_INCLUDE_DIR})
 message(STATUS "Found MAGICMIND: ${__found_magicmind_root} (found version: ${MAGICMIND_VERSION_STRING})")
 message(
  STATUS
    "Found MAGICMIND: ${__found_magicmind_root} (found version: ${MAGICMIND_VERSION_STRING})"
 )
 find_library(MAGICMIND_RUNTIME_LIBRARY 
    NAMES libmagicmind_runtime.so
    PATHS "${__found_magicmind_root}/lib64"
    )
 find_library(
  MAGICMIND_RUNTIME_LIBRARY
  NAMES libmagicmind_runtime.so
  PATHS "${__found_magicmind_root}/lib64")
 if(MAGICMIND_RUNTIME_LIBRARY STREQUAL "MAGICMIND_RUNTIME_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find MAGICMIND_RUNTIME Library")
  message(FATAL_ERROR "Can not find MAGICMIND_RUNTIME Library")
 else()
    message(STATUS "Found MAGICMIND_RUNTIME: ${MAGICMIND_RUNTIME_LIBRARY}")
  message(STATUS "Found MAGICMIND_RUNTIME: ${MAGICMIND_RUNTIME_LIBRARY}")
 endif()
 add_library(libmagicmind_runtime SHARED IMPORTED)
 set_target_properties(libmagicmind_runtime PROPERTIES
    IMPORTED_LOCATION ${MAGICMIND_RUNTIME_LIBRARY}
 )
 set_target_properties(libmagicmind_runtime PROPERTIES IMPORTED_LOCATION
                                                      ${MAGICMIND_RUNTIME_LIBRARY})
--- a/cmake/mkl.cmake
+++ b/cmake/mkl.cmake
@@ -1,77 +1,83 @@
 find_path(MKL_ROOT_DIR
    include/mkl_cblas.h
    PATHS
    ${PROJECT_SOURCE_DIR}/third_party/mkl/${MGE_ARCH}
    ${PROJECT_SOURCE_DIR}/third_party/mkl/${MGE_ARCH}/Library
    ${PROJECT_SOURCE_DIR}/third_party/mkl/x86_32/Library
    ${PROJECT_SOURCE_DIR}/third_party/mkl/x86_32
    $ENV{MKLDIR}
    /opt/intel/mkl/*/
    /opt/intel/cmkl/*/
    /Library/Frameworks/Intel_MKL.framework/Versions/Current/lib/universal
 )
 find_path(
  MKL_ROOT_DIR include/mkl_cblas.h
  PATHS ${PROJECT_SOURCE_DIR}/third_party/mkl/${MGE_ARCH}
        ${PROJECT_SOURCE_DIR}/third_party/mkl/${MGE_ARCH}/Library
        ${PROJECT_SOURCE_DIR}/third_party/mkl/x86_32/Library
        ${PROJECT_SOURCE_DIR}/third_party/mkl/x86_32
        $ENV{MKLDIR}
        /opt/intel/mkl/*/
        /opt/intel/cmkl/*/
        /Library/Frameworks/Intel_MKL.framework/Versions/Current/lib/universal)
 if(${MKL_ROOT_DIR} STREQUAL "MKL_ROOT_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find MKL")
  message(FATAL_ERROR "Can not find MKL")
 endif()
 message(STATUS "Build with MKL in ${MKL_ROOT_DIR}")
 find_path(MKL_INCLUDE_DIR
    mkl_cblas.h
    PATHS
    ${MKL_ROOT_DIR}/include
    ${INCLUDE_INSTALL_DIR}
 )
 find_path(MKL_INCLUDE_DIR mkl_cblas.h PATHS ${MKL_ROOT_DIR}/include
                                            ${INCLUDE_INSTALL_DIR})
 option(MGE_MKL_USE_STATIC "Build MegEngine with static MKL" ON)
 if(MGE_MKL_USE_STATIC)
    find_library(MKL_CORE_LIBRARY
        NAMES libmkl_core.a mkl_core.lib
        PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  find_library(
    MKL_CORE_LIBRARY
    NAMES libmkl_core.a mkl_core.lib
    PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    find_library(MKL_SEQUENTIAL_LIBRARY
        NAMES libmkl_sequential.a mkl_sequential.lib
        PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  find_library(
    MKL_SEQUENTIAL_LIBRARY
    NAMES libmkl_sequential.a mkl_sequential.lib
    PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    if(${MGE_ARCH} STREQUAL "x86_64")
        find_library(MKL_IPL_LIBRARY
            NAMES libmkl_intel_ilp64.a mkl_intel_ilp64.lib
            PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    elseif(${MGE_ARCH} STREQUAL "i386")
        find_library(MKL_IPL_LIBRARY
            NAMES libmkl_intel_32.a mkl_intel_32.lib mkl_intel_c.lib
            PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    endif()
  if(${MGE_ARCH} STREQUAL "x86_64")
    find_library(
      MKL_IPL_LIBRARY
      NAMES libmkl_intel_ilp64.a mkl_intel_ilp64.lib
      PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  elseif(${MGE_ARCH} STREQUAL "i386")
    find_library(
      MKL_IPL_LIBRARY
      NAMES libmkl_intel_32.a mkl_intel_32.lib mkl_intel_c.lib
      PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  endif()
    add_library(libmkl INTERFACE IMPORTED)
    if(UNIX AND NOT APPLE)
        target_link_libraries(libmkl INTERFACE -Wl,--start-group ${MKL_CORE_LIBRARY} ${MKL_SEQUENTIAL_LIBRARY} ${MKL_IPL_LIBRARY} -Wl,--end-group)
    else()
        target_link_libraries(libmkl INTERFACE ${MKL_CORE_LIBRARY} ${MKL_SEQUENTIAL_LIBRARY} ${MKL_IPL_LIBRARY})
    endif()
    target_include_directories(libmkl INTERFACE ${MKL_INCLUDE_DIR})
  add_library(libmkl INTERFACE IMPORTED)
  if(UNIX AND NOT APPLE)
    target_link_libraries(
      libmkl INTERFACE -Wl,--start-group ${MKL_CORE_LIBRARY} ${MKL_SEQUENTIAL_LIBRARY}
                       ${MKL_IPL_LIBRARY} -Wl,--end-group)
  else()
    target_link_libraries(libmkl INTERFACE ${MKL_CORE_LIBRARY}
                                           ${MKL_SEQUENTIAL_LIBRARY} ${MKL_IPL_LIBRARY})
  endif()
  target_include_directories(libmkl INTERFACE ${MKL_INCLUDE_DIR})
 else()
    find_library(MKL_CORE_LIBRARY
        NAMES libmkl_core.so libmkl_core.dylib
        PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  find_library(
    MKL_CORE_LIBRARY
    NAMES libmkl_core.so libmkl_core.dylib
    PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    find_library(MKL_SEQUENTIAL_LIBRARY
        NAMES libmkl_sequential.so libmkl_sequential.dylib
        PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  find_library(
    MKL_SEQUENTIAL_LIBRARY
    NAMES libmkl_sequential.so libmkl_sequential.dylib
    PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    if(${MGE_ARCH} STREQUAL "x86_64")
        find_library(MKL_IPL_LIBRARY
            NAMES libmkl_intel_ilp64.so libmkl_intel_ilp64.dylib
            PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    elseif(${MGE_ARCH} STREQUAL "x86_32")
        find_library(MKL_IPL_LIBRARY
            NAMES libmkl_intel_32.so libmkl_intel_32.dylib
            PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
    endif()
    target_link_libraries(libmkl INTERFACE ${MKL_CORE_LIBRARY} ${MKL_SEQUENTIAL_LIBRARY} ${MKL_IPL_LIBRARY})
    target_include_directories(libmkl INTERFACE ${MKL_INCLUDE_DIR})
  if(${MGE_ARCH} STREQUAL "x86_64")
    find_library(
      MKL_IPL_LIBRARY
      NAMES libmkl_intel_ilp64.so libmkl_intel_ilp64.dylib
      PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  elseif(${MGE_ARCH} STREQUAL "x86_32")
    find_library(
      MKL_IPL_LIBRARY
      NAMES libmkl_intel_32.so libmkl_intel_32.dylib
      PATHS ${MKL_ROOT_DIR}/lib/${MKL_ARCH_DIR} ${MKL_ROOT_DIR}/lib/)
  endif()
  target_link_libraries(libmkl INTERFACE ${MKL_CORE_LIBRARY} ${MKL_SEQUENTIAL_LIBRARY}
                                         ${MKL_IPL_LIBRARY})
  target_include_directories(libmkl INTERFACE ${MKL_INCLUDE_DIR})
 endif()
 if(${MGE_ARCH} STREQUAL "x86_64")
    target_compile_definitions(libmkl INTERFACE -DMKL_ILP64)
  target_compile_definitions(libmkl INTERFACE -DMKL_ILP64)
 endif()
--- a/cmake/protobuf.cmake
+++ b/cmake/protobuf.cmake
@@ -1,70 +1,83 @@
 function(PROTOBUF_GENERATE_CPP_WITH_ROOT SRCS HDRS ROOT_DIR)
    if(NOT ARGN)
        message(SEND_ERROR "Error: PROTOBUF_GENERATE_CPP_WITH_ROOT() called without any proto files")
        return()
    endif()
  if(NOT ARGN)
    message(
      SEND_ERROR
        "Error: PROTOBUF_GENERATE_CPP_WITH_ROOT() called without any proto files")
    return()
  endif()
    set(${SRCS})
    set(${HDRS})
    foreach(FIL ${ARGN})
        set(ABS_FIL ${ROOT_DIR}/${FIL})
        get_filename_component(FIL_WE ${FIL} NAME_WE)
        get_filename_component(FIL_DIR ${ABS_FIL} PATH)
        file(RELATIVE_PATH REL_DIR ${ROOT_DIR} ${FIL_DIR})
  set(${SRCS})
  set(${HDRS})
  foreach(FIL ${ARGN})
    set(ABS_FIL ${ROOT_DIR}/${FIL})
    get_filename_component(FIL_WE ${FIL} NAME_WE)
    get_filename_component(FIL_DIR ${ABS_FIL} PATH)
    file(RELATIVE_PATH REL_DIR ${ROOT_DIR} ${FIL_DIR})
        list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.cc")
        list(APPEND ${HDRS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.h")
    list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.cc")
    list(APPEND ${HDRS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.h")
        add_custom_command(
            OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.cc"
                   "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.h"
            COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
            ARGS --cpp_out ${CMAKE_CURRENT_BINARY_DIR} -I ${FIL_DIR} ${ABS_FIL} -I ${PROTOBUF_INCLUDE_DIRS}
            DEPENDS ${ABS_FIL} libprotobuf
            COMMENT "Running C++ protocol buffer compiler on ${FIL}"
            VERBATIM)
    endforeach()
    add_custom_command(
      OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.cc"
             "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.h"
      COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} ARGS --cpp_out ${CMAKE_CURRENT_BINARY_DIR}
              -I ${FIL_DIR} ${ABS_FIL} -I ${PROTOBUF_INCLUDE_DIRS}
      DEPENDS ${ABS_FIL} libprotobuf
      COMMENT "Running C++ protocol buffer compiler on ${FIL}"
      VERBATIM)
  endforeach()
    set_source_files_properties(${${SRCS}} ${${HDRS}} PROPERTIES GENERATED TRUE)
    set(${SRCS} ${${SRCS}} PARENT_SCOPE)
    set(${HDRS} ${${HDRS}} PARENT_SCOPE)
  set_source_files_properties(${${SRCS}} ${${HDRS}} PROPERTIES GENERATED TRUE)
  set(${SRCS}
      ${${SRCS}}
      PARENT_SCOPE)
  set(${HDRS}
      ${${HDRS}}
      PARENT_SCOPE)
 endfunction()
 if(MGE_USE_SYSTEM_LIB)
    find_package(Protobuf)
    if(Protobuf_FOUND)
        add_library(libprotobuf INTERFACE)
        target_link_libraries(libprotobuf INTERFACE ${Protobuf_LIBRARIES})
        target_include_directories(libprotobuf INTERFACE ${Protobuf_INCLUDE_DIRS})
        get_filename_component(Protobuf_ROOT ${Protobuf_INCLUDE_DIR} DIRECTORY)
        set(PROTOBUF_ROOT ${Protobuf_ROOT})
        set(PROTOBUF_PROTOC_EXECUTABLE ${Protobuf_PROTOC_EXECUTABLE})
        set(PROTOBUF_INCLUDE_DIRS ${Protobuf_INCLUDE_DIRS})
        return()
    endif()
  find_package(Protobuf)
  if(Protobuf_FOUND)
    add_library(libprotobuf INTERFACE)
    target_link_libraries(libprotobuf INTERFACE ${Protobuf_LIBRARIES})
    target_include_directories(libprotobuf INTERFACE ${Protobuf_INCLUDE_DIRS})
    get_filename_component(Protobuf_ROOT ${Protobuf_INCLUDE_DIR} DIRECTORY)
    set(PROTOBUF_ROOT ${Protobuf_ROOT})
    set(PROTOBUF_PROTOC_EXECUTABLE ${Protobuf_PROTOC_EXECUTABLE})
    set(PROTOBUF_INCLUDE_DIRS ${Protobuf_INCLUDE_DIRS})
    return()
  endif()
 endif()
 include(ExternalProject)
 include(GNUInstallDirs)
 set(PROTOBUF_DIR "${PROJECT_SOURCE_DIR}/third_party/protobuf" CACHE STRING "protobuf directory")
 set(PROTOBUF_DIR
    "${PROJECT_SOURCE_DIR}/third_party/protobuf"
    CACHE STRING "protobuf directory")
 set(PROTOBUF_BUILD_DIR ${PROJECT_BINARY_DIR}/third_party/protobuf)
 if(${CMAKE_BUILD_TYPE} STREQUAL "Debug")
    set(PROTOBUF_LIB ${PROTOBUF_BUILD_DIR}/${CMAKE_INSTALL_LIBDIR}/libprotobufd.a)
  set(PROTOBUF_LIB ${PROTOBUF_BUILD_DIR}/${CMAKE_INSTALL_LIBDIR}/libprotobufd.a)
 else()
    set(PROTOBUF_LIB ${PROTOBUF_BUILD_DIR}/${CMAKE_INSTALL_LIBDIR}/libprotobuf.a)
  set(PROTOBUF_LIB ${PROTOBUF_BUILD_DIR}/${CMAKE_INSTALL_LIBDIR}/libprotobuf.a)
 endif()
 set(PROTOBUF_PROTOC_EXECUTABLE ${PROTOBUF_BUILD_DIR}/bin/protoc)
 ExternalProject_add(
    protobuf
    SOURCE_DIR ${PROTOBUF_DIR}/cmake
    PREFIX ${PROTOBUF_BUILD_DIR}
    CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER} -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER} -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} -DCMAKE_INSTALL_PREFIX=${PROTOBUF_BUILD_DIR} -Dprotobuf_BUILD_EXAMPLES=OFF -Dprotobuf_BUILD_TESTS=OFF -DBUILD_SHARED_LIBS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON
    BUILD_BYPRODUCTS ${PROTOBUF_LIB} ${PROTOBUF_PROTOC_EXECUTABLE}
 )
 ExternalProject_Add(
  protobuf
  SOURCE_DIR ${PROTOBUF_DIR}/cmake
  PREFIX ${PROTOBUF_BUILD_DIR}
  CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER}
             -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER}
             -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
             -DCMAKE_INSTALL_PREFIX=${PROTOBUF_BUILD_DIR}
             -Dprotobuf_BUILD_EXAMPLES=OFF
             -Dprotobuf_BUILD_TESTS=OFF
             -DBUILD_SHARED_LIBS=OFF
             -DCMAKE_POSITION_INDEPENDENT_CODE=ON
  BUILD_BYPRODUCTS ${PROTOBUF_LIB} ${PROTOBUF_PROTOC_EXECUTABLE})
 set(PROTOBUF_INC ${PROTOBUF_BUILD_DIR}/include)
 file(MAKE_DIRECTORY ${PROTOBUF_INC})
@@ -72,19 +85,14 @@ file(MAKE_DIRECTORY ${PROTOBUF_INC})
 add_library(libprotobuf STATIC IMPORTED GLOBAL)
 add_dependencies(libprotobuf protobuf)
 set_target_properties(
    libprotobuf PROPERTIES
    IMPORTED_LOCATION ${PROTOBUF_LIB}
    INTERFACE_INCLUDE_DIRECTORIES ${PROTOBUF_BUILD_DIR}/include
 )
  libprotobuf PROPERTIES IMPORTED_LOCATION ${PROTOBUF_LIB}
                         INTERFACE_INCLUDE_DIRECTORIES ${PROTOBUF_BUILD_DIR}/include)
 add_executable(protoc IMPORTED GLOBAL)
 add_dependencies(protoc protobuf)
 set_target_properties(
    protoc PROPERTIES
    IMPORTED_LOCATION ${PROTOBUF_BUILD_DIR}/bin/protoc
 )
 set_target_properties(protoc PROPERTIES IMPORTED_LOCATION
                                        ${PROTOBUF_BUILD_DIR}/bin/protoc)
 set(PROTOBUF_ROOT ${PROTOBUF_BUILD_DIR})
 set(PROTOBUF_PROTOC_EXECUTABLE protoc)
 set(PROTOBUF_INCLUDE_DIRS ${PROTOBUF_BUILD_DIR}/include)
--- a/cmake/rocm.cmake
+++ b/cmake/rocm.cmake
@@ -1,28 +1,34 @@
 if(NOT DEFINED HIP_PATH)
    if(NOT DEFINED ENV{HIP_PATH})
        set(HIP_PATH "/opt/rocm/hip" CACHE PATH "Path to which HIP has been installed")
    else()
        set(HIP_PATH $ENV{HIP_PATH} CACHE PATH "Path to which HIP has been installed")
    endif()
  if(NOT DEFINED ENV{HIP_PATH})
    set(HIP_PATH
        "/opt/rocm/hip"
        CACHE PATH "Path to which HIP has been installed")
  else()
    set(HIP_PATH
        $ENV{HIP_PATH}
        CACHE PATH "Path to which HIP has been installed")
  endif()
 endif()
 set(CMAKE_MODULE_PATH "${HIP_PATH}/cmake" ${CMAKE_MODULE_PATH})
 find_package(HIP QUIET)
 if (HIP_FOUND)
    message(STATUS "Found HIP: " ${HIP_VERSION})
 if(HIP_FOUND)
  message(STATUS "Found HIP: " ${HIP_VERSION})
 else()
    message(FATAL_ERROR "Could not find HIP. Ensure that HIP is either installed in /opt/rocm/hip or the variable HIP_PATH is set to point to the right location.")
  message(
    FATAL_ERROR
      "Could not find HIP. Ensure that HIP is either installed in /opt/rocm/hip or the variable HIP_PATH is set to point to the right location."
  )
 endif()
 if (${HIP_VERSION} VERSION_LESS 3.0)
    message(FATAL_ERROR "ROCM version needed 3. Please update ROCM.")
 if(${HIP_VERSION} VERSION_LESS 3.0)
  message(FATAL_ERROR "ROCM version needed 3. Please update ROCM.")
 endif()
 macro(hipconfig_get_option variable option)
    if(NOT DEFINED ${variable})
        execute_process(
            COMMAND ${HIP_HIPCONFIG_EXECUTABLE} ${option}
            OUTPUT_VARIABLE ${variable})
    endif()
  if(NOT DEFINED ${variable})
    execute_process(COMMAND ${HIP_HIPCONFIG_EXECUTABLE} ${option}
                    OUTPUT_VARIABLE ${variable})
  endif()
 endmacro()
 hipconfig_get_option(HIP_COMPILER "--compiler")
@@ -31,30 +37,33 @@ hipconfig_get_option(HIP_CPP_CONFIG "--cpp_config")
 separate_arguments(HIP_CPP_CONFIG)
 foreach(hip_config_item ${HIP_CPP_CONFIG})
    foreach(macro_name "__HIP_PLATFORM_HCC__" "__HIP_ROCclr__")
        if(${hip_config_item} STREQUAL "-D${macro_name}=")
            set(HIP_CPP_DEFINE "${HIP_CPP_DEFINE}#define ${macro_name}\n")
            set(HIP_CPP_UNDEFINE "${HIP_CPP_UNDEFINE}\
  foreach(macro_name "__HIP_PLATFORM_HCC__" "__HIP_ROCclr__")
    if(${hip_config_item} STREQUAL "-D${macro_name}=")
      set(HIP_CPP_DEFINE "${HIP_CPP_DEFINE}#define ${macro_name}\n")
      set(HIP_CPP_UNDEFINE
          "${HIP_CPP_UNDEFINE}\
                    #ifdef ${macro_name}\n#undef ${macro_name}\n\
                    #else\n#error\n\
                    #endif\n")
        elseif(${hip_config_item} STREQUAL "-D${macro_name}")
            set(HIP_CPP_DEFINE "${HIP_CPP_DEFINE}#define ${macro_name} 1\n")
            set(HIP_CPP_UNDEFINE "${HIP_CPP_UNDEFINE}\
    elseif(${hip_config_item} STREQUAL "-D${macro_name}")
      set(HIP_CPP_DEFINE "${HIP_CPP_DEFINE}#define ${macro_name} 1\n")
      set(HIP_CPP_UNDEFINE
          "${HIP_CPP_UNDEFINE}\
                    #ifdef ${macro_name}\n#undef ${macro_name}\n\
                    #else\n#error\n\
                    #endif\n")
        endif()
    endforeach()
    endif()
  endforeach()
 endforeach()
 message(STATUS "Using HIP compiler ${HIP_COMPILER}")
 if(${HIP_COMPILER} STREQUAL "hcc")
    set(MGE_ROCM_LIBS hip_hcc)
    message(WARNING "hcc is not well supported, please modify link.txt to link with hipcc")
 elseif (${HIP_COMPILER} STREQUAL "clang")
    set(MGE_ROCM_LIBS amdhip64)
  set(MGE_ROCM_LIBS hip_hcc)
  message(
    WARNING "hcc is not well supported, please modify link.txt to link with hipcc")
 elseif(${HIP_COMPILER} STREQUAL "clang")
  set(MGE_ROCM_LIBS amdhip64)
 endif()
 list(APPEND MGE_ROCM_LIBS amdocl64 MIOpen rocblas rocrand)
@@ -63,26 +72,28 @@ set(HIP_INCLUDE_DIR ${HIP_ROOT_DIR}/../include)
 set(HIP_LIBRARY_DIR ${HIP_ROOT_DIR}/../lib)
 function(find_rocm_library name dirname include library)
    find_path(${name}_LIBRARY_DIR
        NAMES ${library}
        HINTS "${${name}_ROOT_DIR}" "${HIP_ROOT_DIR}/../${dirname}"
        PATH_SUFFIXES lib lib/x86_64
        DOC "Path to ${name} library directory")
  find_path(
    ${name}_LIBRARY_DIR
    NAMES ${library}
    HINTS "${${name}_ROOT_DIR}" "${HIP_ROOT_DIR}/../${dirname}"
    PATH_SUFFIXES lib lib/x86_64
    DOC "Path to ${name} library directory")
    if(${${name}_LIBRARY_DIR} MATCHES "NOTFOUND$")
        message(FATAL_ERROR "Can not find ${name} library")
    endif()
  if(${${name}_LIBRARY_DIR} MATCHES "NOTFOUND$")
    message(FATAL_ERROR "Can not find ${name} library")
  endif()
    find_path(${name}_INCLUDE_DIR
        NAMES ${include}
        HINTS "${${name}_ROOT_DIR}" "${HIP_ROOT_DIR}/../${dirname}"
        PATH_SUFFIXES include
        DOC "Path to ${name} include directory")
  find_path(
    ${name}_INCLUDE_DIR
    NAMES ${include}
    HINTS "${${name}_ROOT_DIR}" "${HIP_ROOT_DIR}/../${dirname}"
    PATH_SUFFIXES include
    DOC "Path to ${name} include directory")
    if(${name}_INCLUDE_DIR MATCHES "NOTFOUND$")
        message(FATAL_ERROR "Can not find ${name} include")
    endif()
    message(DEBUG "Found lib ${${name}_LIBRARY_DIR}, include ${${name}_INCLUDE_DIR}")
  if(${name}_INCLUDE_DIR MATCHES "NOTFOUND$")
    message(FATAL_ERROR "Can not find ${name} include")
  endif()
  message(DEBUG "Found lib ${${name}_LIBRARY_DIR}, include ${${name}_INCLUDE_DIR}")
 endfunction()
 find_rocm_library(MIOPEN miopen miopen libMIOpen.so)
--- a/cmake/tensorrt.cmake
+++ b/cmake/tensorrt.cmake
@@ -1,166 +1,189 @@
 if("${TRT_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{TRT_ROOT_DIR}"  STREQUAL "")
    set(TRT_ROOT_DIR $ENV{TRT_ROOT_DIR})
 if("${TRT_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{TRT_ROOT_DIR}" STREQUAL "")
  set(TRT_ROOT_DIR $ENV{TRT_ROOT_DIR})
 endif()
 if(MGE_CUDA_USE_STATIC)
    find_library(TRT_LIBRARY 
        NAMES libnvinfer_static.a nvinfer.lib
        PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
        HINTS ${ALTER_LIBRARY_PATHS}
        PATH_SUFFIXES lib lib64
        DOC "TRT library." )
    find_library(TRT_PLUGIN_LIBRARY
        NAMES libnvinfer_plugin_static.a nvinfer_plugin.lib
        PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
        HINTS ${ALTER_LIBRARY_PATHS}
        PATH_SUFFIXES lib lib64
        DOC "TRT plugin library." )
  find_library(
    TRT_LIBRARY
    NAMES libnvinfer_static.a nvinfer.lib
    PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "TRT library.")
  find_library(
    TRT_PLUGIN_LIBRARY
    NAMES libnvinfer_plugin_static.a nvinfer_plugin.lib
    PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "TRT plugin library.")
 else()
    find_library(TRT_LIBRARY 
        NAMES libnvinfer.so libnvinfer.dylib nvinfer.dll
        PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
        HINTS ${ALTER_LIBRARY_PATHS}
        PATH_SUFFIXES lib lib64
        DOC "TRT library." )
    find_library(TRT_PLUGIN_LIBRARY
        NAMES libnvinfer_plugin.so libnvinfer_plugin.dylib nvinfer_plugin.dll
        PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
        HINTS ${ALTER_LIBRARY_PATHS}
        PATH_SUFFIXES lib lib64
        DOC "TRT plugin library." )
  find_library(
    TRT_LIBRARY
    NAMES libnvinfer.so libnvinfer.dylib nvinfer.dll
    PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "TRT library.")
  find_library(
    TRT_PLUGIN_LIBRARY
    NAMES libnvinfer_plugin.so libnvinfer_plugin.dylib nvinfer_plugin.dll
    PATHS ${ALTER_LD_LIBRARY_PATHS} ${TRT_ROOT_DIR} ${CMAKE_INSTALL_PREFIX}
    HINTS ${ALTER_LIBRARY_PATHS}
    PATH_SUFFIXES lib lib64
    DOC "TRT plugin library.")
 endif()
 if(TRT_LIBRARY STREQUAL "TRT_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find TensorRT Library, please refer to scripts/cmake-build/BUILD_README.md to init TRT env")
  message(
    FATAL_ERROR
      "Can not find TensorRT Library, please refer to scripts/cmake-build/BUILD_README.md to init TRT env"
  )
 endif()
 if(TRT_PLUGIN_LIBRARY STREQUAL "TRT_PLUGIN_LIBRARY-NOTFOUND")
    message(FATAL_ERROR "Can not find TensorRT Plugin Library, please refer to scripts/cmake-build/BUILD_README.md to init TRT env")
  message(
    FATAL_ERROR
      "Can not find TensorRT Plugin Library, please refer to scripts/cmake-build/BUILD_README.md to init TRT env"
  )
 endif()
 get_filename_component(__found_trt_root ${TRT_LIBRARY}/../.. REALPATH)
 find_path(TRT_INCLUDE_DIR 
    NAMES NvInfer.h
    HINTS ${TRT_ROOT_DIR} ${CUDA_TOOLKIT_INCLUDE} ${__found_trt_root}
    PATH_SUFFIXES include 
    DOC "Path to TRT include directory." )
 find_path(TRT_PLUGIN_INCLUDE_DIR
    NAMES NvInferPlugin.h
    HINTS ${TRT_ROOT_DIR} ${CUDA_TOOLKIT_INCLUDE} ${__found_trt_root}
    PATH_SUFFIXES include
    DOC "Path to TRT plugin include directory." )
 find_path(
  TRT_INCLUDE_DIR
  NAMES NvInfer.h
  HINTS ${TRT_ROOT_DIR} ${CUDA_TOOLKIT_INCLUDE} ${__found_trt_root}
  PATH_SUFFIXES include
  DOC "Path to TRT include directory.")
 find_path(
  TRT_PLUGIN_INCLUDE_DIR
  NAMES NvInferPlugin.h
  HINTS ${TRT_ROOT_DIR} ${CUDA_TOOLKIT_INCLUDE} ${__found_trt_root}
  PATH_SUFFIXES include
  DOC "Path to TRT plugin include directory.")
 if(TRT_INCLUDE_DIR STREQUAL "TRT_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find TensorRT INCLUDE, please refer to scripts/cmake-build/BUILD_README.md to init TRT env")
  message(
    FATAL_ERROR
      "Can not find TensorRT INCLUDE, please refer to scripts/cmake-build/BUILD_README.md to init TRT env"
  )
 endif()
 if(TRT_PLUGIN_INCLUDE_DIR STREQUAL "TRT_PLUGIN_INCLUDE_DIR-NOTFOUND")
    message(FATAL_ERROR "Can not find TensorRT Plugin INCLUDE, please refer to scripts/cmake-build/BUILD_README.md to init TRT env")
  message(
    FATAL_ERROR
      "Can not find TensorRT Plugin INCLUDE, please refer to scripts/cmake-build/BUILD_README.md to init TRT env"
  )
 endif()
 file(STRINGS "${TRT_INCLUDE_DIR}/NvInfer.h" TensorRT_MAJOR REGEX "^#define NV_TENSORRT_MAJOR [0-9]+.*$")
 file(STRINGS "${TRT_INCLUDE_DIR}/NvInfer.h" TensorRT_MINOR REGEX "^#define NV_TENSORRT_MINOR [0-9]+.*$")
 file(STRINGS "${TRT_INCLUDE_DIR}/NvInfer.h" TensorRT_PATCH REGEX "^#define NV_TENSORRT_PATCH [0-9]+.*$")
 file(STRINGS "${TRT_INCLUDE_DIR}/NvInfer.h" TensorRT_MAJOR
     REGEX "^#define NV_TENSORRT_MAJOR [0-9]+.*$")
 file(STRINGS "${TRT_INCLUDE_DIR}/NvInfer.h" TensorRT_MINOR
     REGEX "^#define NV_TENSORRT_MINOR [0-9]+.*$")
 file(STRINGS "${TRT_INCLUDE_DIR}/NvInfer.h" TensorRT_PATCH
     REGEX "^#define NV_TENSORRT_PATCH [0-9]+.*$")
 if (TensorRT_MAJOR STREQUAL "")
    file(STRINGS "${TRT_INCLUDE_DIR}/NvInferVersion.h" TensorRT_MAJOR REGEX "^#define NV_TENSORRT_MAJOR [0-9]+.*$")
    file(STRINGS "${TRT_INCLUDE_DIR}/NvInferVersion.h" TensorRT_MINOR REGEX "^#define NV_TENSORRT_MINOR [0-9]+.*$")
    file(STRINGS "${TRT_INCLUDE_DIR}/NvInferVersion.h" TensorRT_PATCH REGEX "^#define NV_TENSORRT_PATCH [0-9]+.*$")
 if(TensorRT_MAJOR STREQUAL "")
  file(STRINGS "${TRT_INCLUDE_DIR}/NvInferVersion.h" TensorRT_MAJOR
       REGEX "^#define NV_TENSORRT_MAJOR [0-9]+.*$")
  file(STRINGS "${TRT_INCLUDE_DIR}/NvInferVersion.h" TensorRT_MINOR
       REGEX "^#define NV_TENSORRT_MINOR [0-9]+.*$")
  file(STRINGS "${TRT_INCLUDE_DIR}/NvInferVersion.h" TensorRT_PATCH
       REGEX "^#define NV_TENSORRT_PATCH [0-9]+.*$")
 endif()
 string(REGEX REPLACE "^#define NV_TENSORRT_MAJOR ([0-9]+).*$" "\\1" TensorRT_VERSION_MAJOR "${TensorRT_MAJOR}")
 string(REGEX REPLACE "^#define NV_TENSORRT_MINOR ([0-9]+).*$" "\\1" TensorRT_VERSION_MINOR "${TensorRT_MINOR}")
 string(REGEX REPLACE "^#define NV_TENSORRT_PATCH ([0-9]+).*$" "\\1" TensorRT_VERSION_PATCH "${TensorRT_PATCH}")
 set(TRT_VERSION_STRING "${TensorRT_VERSION_MAJOR}.${TensorRT_VERSION_MINOR}.${TensorRT_VERSION_PATCH}")
 string(REGEX REPLACE "^#define NV_TENSORRT_MAJOR ([0-9]+).*$" "\\1"
                     TensorRT_VERSION_MAJOR "${TensorRT_MAJOR}")
 string(REGEX REPLACE "^#define NV_TENSORRT_MINOR ([0-9]+).*$" "\\1"
                     TensorRT_VERSION_MINOR "${TensorRT_MINOR}")
 string(REGEX REPLACE "^#define NV_TENSORRT_PATCH ([0-9]+).*$" "\\1"
                     TensorRT_VERSION_PATCH "${TensorRT_PATCH}")
 set(TRT_VERSION_STRING
    "${TensorRT_VERSION_MAJOR}.${TensorRT_VERSION_MINOR}.${TensorRT_VERSION_PATCH}")
 if(MGE_CUDA_USE_STATIC)
    add_library(libnvinfer STATIC IMPORTED)
    add_library(libnvinfer_plugin STATIC IMPORTED)
  add_library(libnvinfer STATIC IMPORTED)
  add_library(libnvinfer_plugin STATIC IMPORTED)
 else()
    add_library(libnvinfer SHARED IMPORTED)
    add_library(libnvinfer_plugin SHARED IMPORTED)
  add_library(libnvinfer SHARED IMPORTED)
  add_library(libnvinfer_plugin SHARED IMPORTED)
 endif()
 set_target_properties(libnvinfer PROPERTIES
    IMPORTED_LOCATION ${TRT_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${TRT_INCLUDE_DIR}
 )
 set_target_properties(libnvinfer_plugin PROPERTIES
    IMPORTED_LOCATION ${TRT_PLUGIN_LIBRARY}
    INTERFACE_INCLUDE_DIRECTORIES ${TRT_PLUGIN_INCLUDE_DIR}
 )
 set_target_properties(
  libnvinfer PROPERTIES IMPORTED_LOCATION ${TRT_LIBRARY} INTERFACE_INCLUDE_DIRECTORIES
                                                         ${TRT_INCLUDE_DIR})
 set_target_properties(
  libnvinfer_plugin PROPERTIES IMPORTED_LOCATION ${TRT_PLUGIN_LIBRARY}
                               INTERFACE_INCLUDE_DIRECTORIES ${TRT_PLUGIN_INCLUDE_DIR})
 message(STATUS "Found TensorRT: ${__found_trt_root} (found version: ${TRT_VERSION_STRING})")
 message(
  STATUS "Found TensorRT: ${__found_trt_root} (found version: ${TRT_VERSION_STRING})")
 if(TensorRT_VERSION_MAJOR GREATER_EQUAL 7)
    if(MGE_CUDA_USE_STATIC)
        find_library(LIBMYELIN_COMPILER
            NAMES libmyelin_compiler_static.a myelin_compiler_static.lib
            PATHS ${__found_trt_root}/lib
            )
        if(LIBMYELIN_COMPILER STREQUAL "LIBMYELIN_COMPILER-NOTFOUND")
            message(FATAL_ERROR "Can not find LIBMYELIN_COMPILER Library")
        else()
            message(STATUS "Found TensorRT myelin_compiler: ${LIBMYELIN_COMPILER}")
        endif()
        add_library(libmyelin_compiler STATIC IMPORTED)
        set_target_properties(libmyelin_compiler PROPERTIES
            IMPORTED_LOCATION ${LIBMYELIN_COMPILER}
            )
  if(MGE_CUDA_USE_STATIC)
    find_library(
      LIBMYELIN_COMPILER
      NAMES libmyelin_compiler_static.a myelin_compiler_static.lib
      PATHS ${__found_trt_root}/lib)
    if(LIBMYELIN_COMPILER STREQUAL "LIBMYELIN_COMPILER-NOTFOUND")
      message(FATAL_ERROR "Can not find LIBMYELIN_COMPILER Library")
    else()
      message(STATUS "Found TensorRT myelin_compiler: ${LIBMYELIN_COMPILER}")
    endif()
    add_library(libmyelin_compiler STATIC IMPORTED)
    set_target_properties(libmyelin_compiler PROPERTIES IMPORTED_LOCATION
                                                        ${LIBMYELIN_COMPILER})
        find_library(LIBMYELIN_EXECUTOR
            NAMES libmyelin_executor_static.a myelin_executor_static.lib
            PATHS ${__found_trt_root}/lib
            )
        if(LIBMYELIN_EXECUTOR STREQUAL "LIBMYELIN_EXECUTOR-NOTFOUND")
            message(FATAL_ERROR "Can not find LIBMYELIN_EXECUTOR Library")
        else()
            message(STATUS "Found TensorRT libmyelin_executor: ${LIBMYELIN_EXECUTOR}")
        endif()
        add_library(libmyelin_executor STATIC IMPORTED)
        set_target_properties(libmyelin_executor PROPERTIES
            IMPORTED_LOCATION ${LIBMYELIN_EXECUTOR}
            )
    find_library(
      LIBMYELIN_EXECUTOR
      NAMES libmyelin_executor_static.a myelin_executor_static.lib
      PATHS ${__found_trt_root}/lib)
    if(LIBMYELIN_EXECUTOR STREQUAL "LIBMYELIN_EXECUTOR-NOTFOUND")
      message(FATAL_ERROR "Can not find LIBMYELIN_EXECUTOR Library")
    else()
      message(STATUS "Found TensorRT libmyelin_executor: ${LIBMYELIN_EXECUTOR}")
    endif()
    add_library(libmyelin_executor STATIC IMPORTED)
    set_target_properties(libmyelin_executor PROPERTIES IMPORTED_LOCATION
                                                        ${LIBMYELIN_EXECUTOR})
        find_library(LIBMYELIN_PATTERN_RUNTIME
            NAMES libmyelin_pattern_runtime_static.a myelin_pattern_runtime_static.lib
            PATHS ${__found_trt_root}/lib
            )
        if(LIBMYELIN_PATTERN_RUNTIME STREQUAL "LIBMYELIN_PATTERN_RUNTIME-NOTFOUND")
            message(FATAL_ERROR "Can not find LIBMYELIN_PATTERN_RUNTIME Library")
        else()
            message(STATUS "Found TensorRT libmyelin_pattern_runtime: ${LIBMYELIN_PATTERN_RUNTIME}")
        endif()
        add_library(libmyelin_pattern_runtime STATIC IMPORTED)
        set_target_properties(libmyelin_pattern_runtime PROPERTIES
            IMPORTED_LOCATION ${LIBMYELIN_PATTERN_RUNTIME}
            )
    find_library(
      LIBMYELIN_PATTERN_RUNTIME
      NAMES libmyelin_pattern_runtime_static.a myelin_pattern_runtime_static.lib
      PATHS ${__found_trt_root}/lib)
    if(LIBMYELIN_PATTERN_RUNTIME STREQUAL "LIBMYELIN_PATTERN_RUNTIME-NOTFOUND")
      message(FATAL_ERROR "Can not find LIBMYELIN_PATTERN_RUNTIME Library")
    else()
      message(
        STATUS "Found TensorRT libmyelin_pattern_runtime: ${LIBMYELIN_PATTERN_RUNTIME}")
    endif()
    add_library(libmyelin_pattern_runtime STATIC IMPORTED)
    set_target_properties(libmyelin_pattern_runtime
                          PROPERTIES IMPORTED_LOCATION ${LIBMYELIN_PATTERN_RUNTIME})
        find_library(LIBMYELIN_PATTERN_LIBRARY
            NAMES libmyelin_pattern_library_static.a myelin_pattern_library_static.lib
            PATHS ${__found_trt_root}/lib
            )
        if(LIBMYELIN_PATTERN_LIBRARY STREQUAL "LIBMYELIN_PATTERN_LIBRARY-NOTFOUND")
            message(FATAL_ERROR "Can not find LIBMYELIN_PATTERN_LIBRARY Library")
        else()
            message(STATUS "Found TensorRT libmyelin_pattern_library: ${LIBMYELIN_PATTERN_LIBRARY}")
        endif()
        add_library(libmyelin_pattern_library STATIC IMPORTED)
        set_target_properties(libmyelin_pattern_library PROPERTIES
            IMPORTED_LOCATION ${LIBMYELIN_PATTERN_LIBRARY}
            )
    find_library(
      LIBMYELIN_PATTERN_LIBRARY
      NAMES libmyelin_pattern_library_static.a myelin_pattern_library_static.lib
      PATHS ${__found_trt_root}/lib)
    if(LIBMYELIN_PATTERN_LIBRARY STREQUAL "LIBMYELIN_PATTERN_LIBRARY-NOTFOUND")
      message(FATAL_ERROR "Can not find LIBMYELIN_PATTERN_LIBRARY Library")
    else()
        find_library(LIBMYELIN_SHARED
            NAMES libmyelin.so myelin.dll
            PATHS ${__found_trt_root}/lib
            )
      message(
        STATUS "Found TensorRT libmyelin_pattern_library: ${LIBMYELIN_PATTERN_LIBRARY}")
    endif()
    add_library(libmyelin_pattern_library STATIC IMPORTED)
    set_target_properties(libmyelin_pattern_library
                          PROPERTIES IMPORTED_LOCATION ${LIBMYELIN_PATTERN_LIBRARY})
  else()
    find_library(
      LIBMYELIN_SHARED
      NAMES libmyelin.so myelin.dll
      PATHS ${__found_trt_root}/lib)
        if(LIBMYELIN_SHARED STREQUAL "LIBMYELIN_SHARED-NOTFOUND")
            message(FATAL_ERROR "Can not find LIBMYELIN_SHARED Library")
        else()
            message(STATUS "Found TensorRT libmyelin_shared: ${LIBMYELIN_SHARED}")
        endif()
        add_library(libmyelin SHARED IMPORTED)
        set_target_properties(libmyelin PROPERTIES
            IMPORTED_LOCATION ${LIBMYELIN_SHARED}
            )
    if(LIBMYELIN_SHARED STREQUAL "LIBMYELIN_SHARED-NOTFOUND")
      message(FATAL_ERROR "Can not find LIBMYELIN_SHARED Library")
    else()
      message(STATUS "Found TensorRT libmyelin_shared: ${LIBMYELIN_SHARED}")
    endif()
    add_library(libmyelin SHARED IMPORTED)
    set_target_properties(libmyelin PROPERTIES IMPORTED_LOCATION ${LIBMYELIN_SHARED})
  endif()
 endif()
--- a/cmake/zmq.cmake
+++ b/cmake/zmq.cmake
@@ -1,17 +1,26 @@
 include(ExternalProject)
 include(GNUInstallDirs)
 set(ZMQ_DIR ${PROJECT_SOURCE_DIR}/third_party/libzmq CACHE STRING "ZMQ directory")
 set(ZMQ_DIR
    ${PROJECT_SOURCE_DIR}/third_party/libzmq
    CACHE STRING "ZMQ directory")
 set(ZMQ_BUILD_DIR ${PROJECT_BINARY_DIR}/third_party/libzmq)
 set(ZMQ_LIB ${ZMQ_BUILD_DIR}/${CMAKE_INSTALL_LIBDIR}/libzmq.a)
 ExternalProject_add(
    zmq
    SOURCE_DIR ${ZMQ_DIR}
    PREFIX ${ZMQ_BUILD_DIR}
    CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER} -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER} -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} -DCMAKE_INSTALL_PREFIX=${ZMQ_BUILD_DIR} -DWITH_PERF_TOOL=OFF -DZMQ_BUILD_TESTS=OFF -DENABLE_CPACK=OFF -DENABLE_CURVE=OFF
    BUILD_BYPRODUCTS ${ZMQ_LIB}
 )
 ExternalProject_Add(
  zmq
  SOURCE_DIR ${ZMQ_DIR}
  PREFIX ${ZMQ_BUILD_DIR}
  CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER}
             -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER}
             -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
             -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE}
             -DCMAKE_INSTALL_PREFIX=${ZMQ_BUILD_DIR}
             -DWITH_PERF_TOOL=OFF
             -DZMQ_BUILD_TESTS=OFF
             -DENABLE_CPACK=OFF
             -DENABLE_CURVE=OFF
  BUILD_BYPRODUCTS ${ZMQ_LIB})
 set(ZMQ_INC ${ZMQ_BUILD_DIR}/include)
 include_directories(${ZMQ_INC})
@@ -19,8 +28,5 @@ file(MAKE_DIRECTORY ${ZMQ_INC})
 add_library(libzmq STATIC IMPORTED GLOBAL)
 add_dependencies(libzmq zmq)
 set_target_properties(
    libzmq PROPERTIES
    IMPORTED_LOCATION ${ZMQ_LIB}
    INTERFACE_INCLUDE_DIRECTORIES ${ZMQ_INC}
 )
 set_target_properties(libzmq PROPERTIES IMPORTED_LOCATION ${ZMQ_LIB}
                                        INTERFACE_INCLUDE_DIRECTORIES ${ZMQ_INC})
--- a/dnn/CMakeLists.txt
+++ b/dnn/CMakeLists.txt
@@ -4,66 +4,61 @@ set(OPR_PARAM_DEFS_SCRIPT ${CMAKE_CURRENT_SOURCE_DIR}/scripts/gen_param_defs.py)
 set(OPR_PARAM_DEFS_OUT_DIR ${CMAKE_CURRENT_BINARY_DIR}/include/)
 file(MAKE_DIRECTORY ${OPR_PARAM_DEFS_OUT_DIR}/megdnn)
 add_custom_command(
    OUTPUT
        ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_defs.h
        ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_json.h
        COMMAND ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${OPR_PARAM_DEFS_SCRIPT} ${OPR_PARAM_DEFS_SRCS}
        ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_defs.h
        COMMAND ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${OPR_PARAM_DEFS_SCRIPT} ${OPR_PARAM_DEFS_SRCS}
       tmp_unuse.log --write-cppjson ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_json.h
    DEPENDS ${OPR_PARAM_DEFS_SRCS} ${OPR_PARAM_DEFS_SCRIPT}
    VERBATIM
 )
 list(APPEND OPR_PARAM_DEFS_OUTS
    ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_defs.h
  OUTPUT ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_defs.h
         ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_json.h
  COMMAND ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${OPR_PARAM_DEFS_SCRIPT}
          ${OPR_PARAM_DEFS_SRCS} ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_defs.h
  COMMAND
    ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${OPR_PARAM_DEFS_SCRIPT}
    ${OPR_PARAM_DEFS_SRCS} tmp_unuse.log --write-cppjson
    ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_json.h
 )
  DEPENDS ${OPR_PARAM_DEFS_SRCS} ${OPR_PARAM_DEFS_SCRIPT}
  VERBATIM)
 list(APPEND OPR_PARAM_DEFS_OUTS ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_defs.h
     ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_json.h)
 list(APPEND OPR_PARAM_DEFS_INC ${OPR_PARAM_DEFS_OUT_DIR})
 set(OPR_PARAM_DEFS_OUT_DIR ${CMAKE_CURRENT_BINARY_DIR})
 file(MAKE_DIRECTORY ${OPR_PARAM_DEFS_OUT_DIR}/src/common)
 add_custom_command(
    OUTPUT
        ${OPR_PARAM_DEFS_OUT_DIR}/src/common/opr_param_defs_enumv.cuh
        COMMAND ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${OPR_PARAM_DEFS_SCRIPT}
        --enumv ${OPR_PARAM_DEFS_SRCS}
        ${OPR_PARAM_DEFS_OUT_DIR}/src/common/opr_param_defs_enumv.cuh
    DEPENDS ${OPR_PARAM_DEFS_SRCS} ${OPR_PARAM_DEFS_SCRIPT}
    VERBATIM
 )
  OUTPUT ${OPR_PARAM_DEFS_OUT_DIR}/src/common/opr_param_defs_enumv.cuh
  COMMAND
    ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${OPR_PARAM_DEFS_SCRIPT} --enumv
    ${OPR_PARAM_DEFS_SRCS} ${OPR_PARAM_DEFS_OUT_DIR}/src/common/opr_param_defs_enumv.cuh
  DEPENDS ${OPR_PARAM_DEFS_SRCS} ${OPR_PARAM_DEFS_SCRIPT}
  VERBATIM)
 list(APPEND OPR_PARAM_DEFS_OUTS
    ${OPR_PARAM_DEFS_OUT_DIR}/src/common/opr_param_defs_enumv.cuh
 )
     ${OPR_PARAM_DEFS_OUT_DIR}/src/common/opr_param_defs_enumv.cuh)
 list(APPEND OPR_PARAM_DEFS_INC ${OPR_PARAM_DEFS_OUT_DIR})
 install(DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/include/megdnn DESTINATION ${CMAKE_INSTALL_INCLUDEDIR} FILES_MATCHING PATTERN "*.h")
 install(
  DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/include/megdnn
  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
  FILES_MATCHING
  PATTERN "*.h")
 add_custom_target(_opr_param_defs DEPENDS ${OPR_PARAM_DEFS_OUTS})
 add_library(opr_param_defs INTERFACE)
 target_include_directories(opr_param_defs
    INTERFACE
        $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
 )
 foreach (INCPATH IN LISTS OPR_PARAM_DEFS_INC)
    target_include_directories(opr_param_defs
        INTERFACE $<BUILD_INTERFACE:${INCPATH}>
    )
                           INTERFACE $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
 foreach(INCPATH IN LISTS OPR_PARAM_DEFS_INC)
  target_include_directories(opr_param_defs INTERFACE $<BUILD_INTERFACE:${INCPATH}>)
 endforeach()
 add_dependencies(opr_param_defs _opr_param_defs)
 install(TARGETS opr_param_defs EXPORT ${MGE_EXPORT_TARGETS})
 if(MGE_WITH_CUDA)
    add_library(cutlass INTERFACE)
    target_include_directories(cutlass 
        INTERFACE 
        $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/third_party/cutlass/include>)
  add_library(cutlass INTERFACE)
  target_include_directories(
    cutlass
    INTERFACE $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/third_party/cutlass/include>)
 endif()
 if(MGE_WITH_TEST)
    add_subdirectory(test)
  add_subdirectory(test)
 endif()
 add_subdirectory(src)
--- a/dnn/atlas-stub/CMakeLists.txt
+++ b/dnn/atlas-stub/CMakeLists.txt
@@ -1,6 +1,8 @@
 add_library(atlas-stub STATIC src/libatlas-wrap.cpp)
 target_include_directories(atlas-stub PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
 target_include_directories(
  atlas-stub PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
 install(TARGETS atlas-stub EXPORT ${MGE_EXPORT_TARGETS})
 add_library(acl-cblas STATIC src/libacl_cblas-wrap.cpp)
 target_include_directories(acl-cblas PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
 target_include_directories(
  acl-cblas PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
--- a/dnn/cuda-stub/CMakeLists.txt
+++ b/dnn/cuda-stub/CMakeLists.txt
@@ -1,26 +1,27 @@
 file (GLOB_RECURSE CUDA_STUB src/libcuda.cpp)
 file (GLOB_RECURSE NVRTC_STUB src/libnvrtc.cpp)
 file(GLOB_RECURSE CUDA_STUB src/libcuda.cpp)
 file(GLOB_RECURSE NVRTC_STUB src/libnvrtc.cpp)
 if(MGE_WITH_CUDA_STUB)
    list(APPEND STUB_SRC ${CUDA_STUB})
  list(APPEND STUB_SRC ${CUDA_STUB})
 endif()
 if(MGE_WITH_NVRTC_STUB)
    list(APPEND STUB_SRC ${NVRTC_STUB})
  list(APPEND STUB_SRC ${NVRTC_STUB})
 endif()
 if(MSVC OR WIN32)
    add_library (cuda-stub STATIC ${STUB_SRC})
  add_library(cuda-stub STATIC ${STUB_SRC})
 else()
    add_library (cuda-stub SHARED ${STUB_SRC})
  add_library(cuda-stub SHARED ${STUB_SRC})
 endif()
 set_target_properties(cuda-stub PROPERTIES OUTPUT_NAME cuda_stub)
 target_compile_definitions(cuda-stub PRIVATE __CUDA_API_VERSION_INTERNAL)
 if (MSVC OR WIN32)
    target_link_libraries(cuda-stub PRIVATE -Wl,--no-undefined)
 if(MSVC OR WIN32)
  target_link_libraries(cuda-stub PRIVATE -Wl,--no-undefined)
 else()
    target_link_libraries(cuda-stub PRIVATE dl -Wl,--no-undefined)
  target_link_libraries(cuda-stub PRIVATE dl -Wl,--no-undefined)
 endif()
 target_include_directories(cuda-stub PRIVATE $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/genfiles>)
 install (TARGETS cuda-stub EXPORT ${MGE_EXPORT_TARGETS})
 target_include_directories(cuda-stub
                           PRIVATE $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/genfiles>)
 install(TARGETS cuda-stub EXPORT ${MGE_EXPORT_TARGETS})
--- a/dnn/include/megdnn/common.h
+++ b/dnn/include/megdnn/common.h
@@ -12,6 +12,7 @@
 #pragma once
 #include "megbrain_build_config.h"
 #include "megdnn/oprs/base.h"
 #if MGB_ENABLE_GETENV
 #define MGB_GETENV ::std::getenv
@@ -36,6 +37,11 @@ bool has_available_algo(Opr* opr, Args&&... args) {
    return !all_algos.empty();
 }
 template <class Opr, typename... Args>
 bool has_no_naive_heuristic_algo(Opr* opr, Args&&... args) {
    auto&& algo = opr->get_algorithm_info_heuristic(std::forward<Args>(args)...);
    return !static_cast<bool>(algo.attribute & detail::Algorithm::Attribute::NAIVE);
 }
 }  // namespace megdnn
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
--- a/dnn/include/megdnn/oprs/nn.h
+++ b/dnn/include/megdnn/oprs/nn.h
@@ -1936,6 +1936,119 @@ protected:
            const TensorLayout& grad_s, size_t workspace_in_bytes);
 };
 class LayerNormBase : public OperatorBase {
    DEF_OPR_IMPL_CTOR(LayerNormBase, OperatorBase);
    DEF_OPR_PARAM(LayerNorm);
 protected:
    void deduce_layout_fwd(
            const TensorLayout& data, const TensorLayout& weight,
            const TensorLayout& bias, TensorLayout& dst, TensorLayout& mean,
            TensorLayout& rstd);
    void check_layout_fwd(
            const TensorLayout& data, const TensorLayout& weight,
            const TensorLayout& bias, const TensorLayout& dst, const TensorLayout& mean,
            const TensorLayout& rstd);
 };
 class LayerNormForward : public LayerNormBase {
    DEF_OPR_IMPL(LayerNormForward, LayerNormBase, 3, 3);
 public:
    virtual void exec(
            _megdnn_tensor_in data, _megdnn_tensor_in weight, _megdnn_tensor_in bias,
            _megdnn_tensor_out dst, _megdnn_tensor_out mean, _megdnn_tensor_out rstd,
            _megdnn_workspace workspace) = 0;
    void deduce_layout(
            const TensorLayout& data, const TensorLayout& weight,
            const TensorLayout& bias, TensorLayout& dst, TensorLayout& mean,
            TensorLayout& rstd);
    virtual size_t get_workspace_in_bytes(
            const TensorLayout& data, const TensorLayout& weight,
            const TensorLayout& bias, const TensorLayout& dst, const TensorLayout& mean,
            const TensorLayout& rstd) = 0;
 protected:
    void check_exec(
            const TensorLayout& data, const TensorLayout& weight,
            const TensorLayout& bias, const TensorLayout& dst, const TensorLayout& mean,
            const TensorLayout& rstd, size_t workspace_in_bytes);
 };
 using LayerNorm = LayerNormForward;
 class LayerNormBackward : public LayerNormBase {
    DEF_OPR_IMPL(LayerNormBackward, LayerNormBase, 5, 3);
 public:
    virtual void exec(
            _megdnn_tensor_in diff, _megdnn_tensor_in data, _megdnn_tensor_in weight,
            _megdnn_tensor_in mean, _megdnn_tensor_in rstd, _megdnn_tensor_out ddata,
            _megdnn_tensor_out dweight, _megdnn_tensor_out dbias,
            _megdnn_workspace workspace) = 0;
    void deduce_layout(
            const TensorLayout& diff, const TensorLayout& data,
            const TensorLayout& weight, const TensorLayout& mean,
            const TensorLayout& rstd, TensorLayout& ddata, TensorLayout& dweight,
            TensorLayout& dbias);
    virtual size_t get_workspace_in_bytes(
            const TensorLayout& diff, const TensorLayout& data,
            const TensorLayout& weight, const TensorLayout& mean,
            const TensorLayout& rstd, const TensorLayout& ddata,
            const TensorLayout& dweight, const TensorLayout& dbias) = 0;
 protected:
    void check_exec(
            const TensorLayout& diff, const TensorLayout& data,
            const TensorLayout& weight, const TensorLayout& mean,
            const TensorLayout& rstd, const TensorLayout& ddata,
            const TensorLayout& dweight, const TensorLayout& dbias,
            size_t workspace_in_bytes);
 };
 class DropoutBase : public OperatorBase {
    DEF_OPR_IMPL_CTOR(DropoutBase, OperatorBase);
    DEF_OPR_PARAM(Dropout);
 };
 class DropoutForward : public DropoutBase {
    DEF_OPR_IMPL(DropoutForward, DropoutBase, 1, 2);
 public:
    void deduce_layout(const TensorLayout& inp, TensorLayout& oup, TensorLayout& mask);
    virtual void exec(
            _megdnn_tensor_in inp, _megdnn_tensor_out oup, _megdnn_tensor_out mask,
            _megdnn_workspace workspace) = 0;
    virtual size_t get_workspace_in_bytes(
            const TensorLayout& inp, const TensorLayout& oup,
            const TensorLayout& mask) = 0;
    virtual size_t get_mask_size_in_bytes(const TensorLayout& inp) = 0;
 protected:
    void check_exec(
            const TensorLayout& inp, const TensorLayout& oup, const TensorLayout& mask,
            size_t workspace_in_bytes);
 };
 using Dropout = DropoutForward;
 class DropoutBackward : public DropoutBase {
    DEF_OPR_IMPL(DropoutBackward, DropoutBase, 2, 1);
 public:
    void deduce_layout(
            const TensorLayout& doup, const TensorLayout& mask, TensorLayout& dinp);
    virtual void exec(
            _megdnn_tensor_in doup, _megdnn_tensor_in mask, _megdnn_tensor_out dinp,
            _megdnn_workspace workspace) = 0;
    virtual size_t get_workspace_in_bytes(
            const TensorLayout& doup, const TensorLayout& mask,
            const TensorLayout& dinp) = 0;
 protected:
    void check_exec(
            const TensorLayout& doup, const TensorLayout& mask,
            const TensorLayout& dinp, size_t workspace_in_bytes);
 };
 }  // namespace megdnn
 #include "megdnn/internal/opr_header_epilogue.h"
--- a/dnn/scripts/opr_param_defs.py
+++ b/dnn/scripts/opr_param_defs.py
@@ -1212,3 +1212,15 @@ PADDING_MODES = [Doc('REPLICATE = 0', 'aaaaaa|abcdefgh|hhhhhhh'),
          member_alias=[(i, 'PADDING_{}'.format(i)) for i in PADDING_MODES]
          )
 )
 (pdef('LayerNorm')
 .add_fields('bool', 'affine', 'true')
 .add_fields('float32', 'eps', '1e-5f')
 .add_fields('uint64', 'normalized_dim', '1')
 .add_fields('uint64', 'normalized_size', '1')
 )
 (pdef('Dropout')
 .add_fields('float32', 'drop_prob', '0')
 .add_fields('uint64', 'seed', '0')
 )
--- a/dnn/src/CMakeLists.txt
+++ b/dnn/src/CMakeLists.txt
@@ -5,168 +5,190 @@ file(GLOB_RECURSE SOURCES common/*.cpp naive/*.cpp)
 list(APPEND SOURCES ${PROJECT_BINARY_DIR}/genfiles/megbrain_build_config.h)
 if(NOT ${MGE_ARCH} STREQUAL "naive")
    file(GLOB_RECURSE SOURCES_ fallback/*.cpp)
  file(GLOB_RECURSE SOURCES_ fallback/*.cpp)
  list(APPEND SOURCES ${SOURCES_})
  if(${MGE_ARCH} STREQUAL "fallback")
    message(WARNING "build only with fallback")
  elseif(${MGE_ARCH} STREQUAL "x86_64" OR ${MGE_ARCH} STREQUAL "i386")
    file(GLOB_RECURSE SOURCES_ x86/*.cpp)
    list(APPEND SOURCES ${SOURCES_})
    if(${MGE_ARCH} STREQUAL "fallback")
        message(WARNING "build only with fallback")
    elseif(${MGE_ARCH} STREQUAL "x86_64" OR ${MGE_ARCH} STREQUAL "i386")
        file(GLOB_RECURSE SOURCES_ x86/*.cpp)
        list(APPEND SOURCES ${SOURCES_})
        if(NOT MSVC)
            file(GLOB_RECURSE SOURCES_ x86/*.S)
            set_source_files_properties(${SOURCES_} PROPERTIES LANGUAGE C)
            list(APPEND SOURCES ${SOURCES_})
        endif()
    elseif(${MGE_ARCH} STREQUAL "armv7")
        file(GLOB_RECURSE SOURCES_ armv7/*.cpp)
        list(APPEND SOURCES ${SOURCES_})
        file(GLOB_RECURSE SOURCES_ arm_common/*.cpp)
        list(APPEND SOURCES ${SOURCES_})
        file(GLOB_RECURSE SOURCES_ armv7/*.S)
        set_source_files_properties(${SOURCES_} PROPERTIES LANGUAGE C)
        list(APPEND SOURCES ${SOURCES_})
    elseif(${MGE_ARCH} STREQUAL "aarch64")
        file(GLOB_RECURSE SOURCES_ aarch64/*.cpp)
        list(APPEND SOURCES ${SOURCES_})
        file(GLOB_RECURSE SOURCES_ arm_common/*.cpp)
        list(APPEND SOURCES ${SOURCES_})
        file(GLOB_RECURSE SOURCES_ aarch64/*.S)
        set_source_files_properties(${SOURCES_} PROPERTIES LANGUAGE C)
        list(APPEND SOURCES ${SOURCES_})
    if(NOT MSVC)
      file(GLOB_RECURSE SOURCES_ x86/*.S)
      set_source_files_properties(${SOURCES_} PROPERTIES LANGUAGE C)
      list(APPEND SOURCES ${SOURCES_})
    endif()
  elseif(${MGE_ARCH} STREQUAL "armv7")
    file(GLOB_RECURSE SOURCES_ armv7/*.cpp)
    list(APPEND SOURCES ${SOURCES_})
    file(GLOB_RECURSE SOURCES_ arm_common/*.cpp)
    list(APPEND SOURCES ${SOURCES_})
    file(GLOB_RECURSE SOURCES_ armv7/*.S)
    set_source_files_properties(${SOURCES_} PROPERTIES LANGUAGE C)
    list(APPEND SOURCES ${SOURCES_})
  elseif(${MGE_ARCH} STREQUAL "aarch64")
    file(GLOB_RECURSE SOURCES_ aarch64/*.cpp)
    list(APPEND SOURCES ${SOURCES_})
    file(GLOB_RECURSE SOURCES_ arm_common/*.cpp)
    list(APPEND SOURCES ${SOURCES_})
    file(GLOB_RECURSE SOURCES_ aarch64/*.S)
    set_source_files_properties(${SOURCES_} PROPERTIES LANGUAGE C)
    list(APPEND SOURCES ${SOURCES_})
  endif()
 endif()
 if(MGE_WITH_MIDOUT_PROFILE)
    list(APPEND SOURCES ${PROJECT_SOURCE_DIR}/third_party/midout/src/midout.cpp)
  list(APPEND SOURCES ${PROJECT_SOURCE_DIR}/third_party/midout/src/midout.cpp)
 endif()
 ###############################################################################
 # ######################################################################################
 # HIP_COMPILE
 ###############################################################################
 macro (HIP_COMPILE _hip_target _hip_objs)
    # Separate the sources from the options
    HIP_GET_SOURCES_AND_OPTIONS(_sources
        _cmake_options
        _hipcc_options
        _hcc_options
        _nvcc_options
        ${ARGN})
    HIP_PREPARE_TARGET_COMMANDS(${_hip_target}
        OBJ _generated_files _source_files ${_sources} ${_cmake_options}
        HIPCC_OPTIONS ${_hipcc_options}
        HCC_OPTIONS ${_hcc_options}
        NVCC_OPTIONS ${_nvcc_options})
    if(_source_files)
        list(REMOVE_ITEM _sources ${_source_files})
    endif()
 # ######################################################################################
 macro(HIP_COMPILE _hip_target _hip_objs)
  # Separate the sources from the options
  hip_get_sources_and_options(_sources _cmake_options _hipcc_options _hcc_options
                              _nvcc_options ${ARGN})
  hip_prepare_target_commands(
    ${_hip_target}
    OBJ
    _generated_files
    _source_files
    ${_sources}
    ${_cmake_options}
    HIPCC_OPTIONS
    ${_hipcc_options}
    HCC_OPTIONS
    ${_hcc_options}
    NVCC_OPTIONS
    ${_nvcc_options})
  if(_source_files)
    list(REMOVE_ITEM _sources ${_source_files})
  endif()
    add_custom_target(${_hip_target})
  add_custom_target(${_hip_target})
    # set return value
    set(${_hip_objs} ${_generated_files})
  # set return value
  set(${_hip_objs} ${_generated_files})
 endmacro()
 if (MGE_WITH_ROCM)
    file (GLOB_RECURSE SOURCES_ rocm/*.cpp)
    list (APPEND SOURCES ${SOURCES_})
    # FIXME rocm may lost the first hip file, so currently we just create an
    # empty file to bypass this error.
    file(GLOB start.cpp.hip "" )
    list(APPEND HIP_SOURCES start.cpp.hip)
    configure_file(
        ${PROJECT_SOURCE_DIR}/dnn/include/hcc_detail/hcc_defs_prologue.h.in
        ${PROJECT_BINARY_DIR}/dnn/include/hcc_detail/hcc_defs_prologue.h)
    configure_file(
        ${PROJECT_SOURCE_DIR}/dnn/include/hcc_detail/hcc_defs_epilogue.h.in
        ${PROJECT_BINARY_DIR}/dnn/include/hcc_detail/hcc_defs_epilogue.h)
    file(GLOB_RECURSE HIP_SOURCES_ rocm/*.cpp.hip)
    set(HIP_TARGET_NAME megdnn_hip_kernel)
    set(_HIPCC_OPTIONS "-fPIC")
    set(_HCC_OPTIONS "-fPIC")
    set(_NVCC_OPTIONS "-fPIC")
    list(APPEND HIP_SOURCES ${HIP_SOURCES_})
    set_source_files_properties(${HIP_SOURCES} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
    HIP_INCLUDE_DIRECTORIES(${PROJECT_SOURCE_DIR}/dnn
        ${PROJECT_SOURCE_DIR}/dnn/include
        ${PROJECT_BINARY_DIR}/dnn
        ${PROJECT_BINARY_DIR}/genfiles
        ${PROJECT_BINARY_DIR}/dnn/include
        ${HIP_INCLUDE_DIR}
        ${MIOPEN_INCLUDE_DIR}
        ${ROCBLAS_INCLUDE_DIR}
        ${ROCRAND_INCLUDE_DIR}
        ${AMDOCL_INCLUDE_DIR})
    hip_compile(
        ${HIP_TARGET_NAME} HIPOBJS ${HIP_SOURCES}
        HIPCC_OPTIONS ${_HIPCC_OPTIONS}
        HCC_OPTIONS ${_HCC_OPTIONS}
        NVCC_OPTIONS ${_NVCC_OPTIONS})
    list(APPEND SOURCES ${HIPOBJS})
 endif ()
 if(MGE_WITH_ROCM)
  file(GLOB_RECURSE SOURCES_ rocm/*.cpp)
  list(APPEND SOURCES ${SOURCES_})
  # FIXME rocm may lost the first hip file, so currently we just create an empty file to
  # bypass this error.
  file(GLOB start.cpp.hip "")
  list(APPEND HIP_SOURCES start.cpp.hip)
  configure_file(${PROJECT_SOURCE_DIR}/dnn/include/hcc_detail/hcc_defs_prologue.h.in
                 ${PROJECT_BINARY_DIR}/dnn/include/hcc_detail/hcc_defs_prologue.h)
  configure_file(${PROJECT_SOURCE_DIR}/dnn/include/hcc_detail/hcc_defs_epilogue.h.in
                 ${PROJECT_BINARY_DIR}/dnn/include/hcc_detail/hcc_defs_epilogue.h)
  file(GLOB_RECURSE HIP_SOURCES_ rocm/*.cpp.hip)
  set(HIP_TARGET_NAME megdnn_hip_kernel)
  set(_HIPCC_OPTIONS "-fPIC")
  set(_HCC_OPTIONS "-fPIC")
  set(_NVCC_OPTIONS "-fPIC")
  list(APPEND HIP_SOURCES ${HIP_SOURCES_})
  set_source_files_properties(${HIP_SOURCES} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
  hip_include_directories(
    ${PROJECT_SOURCE_DIR}/dnn
    ${PROJECT_SOURCE_DIR}/dnn/include
    ${PROJECT_BINARY_DIR}/dnn
    ${PROJECT_BINARY_DIR}/genfiles
    ${PROJECT_BINARY_DIR}/dnn/include
    ${HIP_INCLUDE_DIR}
    ${MIOPEN_INCLUDE_DIR}
    ${ROCBLAS_INCLUDE_DIR}
    ${ROCRAND_INCLUDE_DIR}
    ${AMDOCL_INCLUDE_DIR})
  hip_compile(
    ${HIP_TARGET_NAME}
    HIPOBJS
    ${HIP_SOURCES}
    HIPCC_OPTIONS
    ${_HIPCC_OPTIONS}
    HCC_OPTIONS
    ${_HCC_OPTIONS}
    NVCC_OPTIONS
    ${_NVCC_OPTIONS})
  list(APPEND SOURCES ${HIPOBJS})
 endif()
 if(MGE_WITH_CUDA)
    file(GLOB_RECURSE SOURCES_ cuda/*.cpp)
    list(APPEND SOURCES ${SOURCES_})
  file(GLOB_RECURSE SOURCES_ cuda/*.cpp)
  list(APPEND SOURCES ${SOURCES_})
    file(GLOB_RECURSE CUSOURCES cuda/*.cu)
    set(CUTLASS_GEN_SCRIPT ${CMAKE_CURRENT_SOURCE_DIR}/../scripts/cutlass_generator/generator.py)
    set(CUTLASS_GEN_DIR ${CMAKE_CURRENT_BINARY_DIR}/cuda/cutlass/generated)
    set(CUTLASS_SOURCES "")
    function(gen_cutlass_kimpl op type gen_files)
        set(CURRENT_CUTLASS_STAGE_DIR ${CUTLASS_GEN_DIR}/${op}_${type}.stage)
        set(CURRENT_CUTLASS_GEN_DIR ${CUTLASS_GEN_DIR}/${op}_${type})
        set_directory_properties(PROPERTIES CMAKE_CONFIGURE_DEPENDS ${CUTLASS_GEN_SCRIPT})
        file(REMOVE_RECURSE ${CURRENT_CUTLASS_STAGE_DIR})
        file(MAKE_DIRECTORY ${CURRENT_CUTLASS_STAGE_DIR})
        file(MAKE_DIRECTORY ${CURRENT_CUTLASS_GEN_DIR})
        execute_process(
            COMMAND ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${CUTLASS_GEN_SCRIPT} --operations ${op} --type ${type} ${CURRENT_CUTLASS_STAGE_DIR}
            RESULT_VARIABLE gen_cutlass_result
            OUTPUT_FILE ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log
            ERROR_FILE ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log
        )
        if (NOT gen_cutlass_result EQUAL 0)
            message(FATAL_ERROR "Error generating library instances. See ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log")
        endif()
        file(GLOB CUTLASS_GEN_FILES RELATIVE "${CURRENT_CUTLASS_GEN_DIR}/" "${CURRENT_CUTLASS_GEN_DIR}/*.cu")
        foreach(FILE ${CUTLASS_GEN_FILES}) 
            if (NOT EXISTS "${CURRENT_CUTLASS_STAGE_DIR}/${FILE}")
                file(REMOVE "${CURRENT_CUTLASS_GEN_DIR}/${FILE}")
            endif()
        endforeach()
        file(GLOB CUTLASS_GEN_FILES RELATIVE "${CURRENT_CUTLASS_STAGE_DIR}" "${CURRENT_CUTLASS_STAGE_DIR}/*.cu")
        foreach(FILE ${CUTLASS_GEN_FILES})
            execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CURRENT_CUTLASS_STAGE_DIR}/${FILE}" "${CURRENT_CUTLASS_GEN_DIR}")
        endforeach()
        file(REMOVE_RECURSE ${CURRENT_CUTLASS_STAGE_DIR})
        file(GLOB_RECURSE CUTLASS_GEN_FILES "${CURRENT_CUTLASS_GEN_DIR}/*.cu")
        list(APPEND ${gen_files} ${CUTLASS_GEN_FILES})
        set(${gen_files} "${${gen_files}}" PARENT_SCOPE)
    endfunction()
    gen_cutlass_kimpl(gemm simt CUTLASS_SOURCES)
    gen_cutlass_kimpl(gemm tensorop884 CUTLASS_SOURCES)
    gen_cutlass_kimpl(gemm tensorop1688 CUTLASS_SOURCES)
    gen_cutlass_kimpl(gemv simt CUTLASS_SOURCES)
    gen_cutlass_kimpl(deconv simt CUTLASS_SOURCES)
    gen_cutlass_kimpl(deconv tensorop8816 CUTLASS_SOURCES)
    gen_cutlass_kimpl(conv2d simt CUTLASS_SOURCES)
    gen_cutlass_kimpl(conv2d tensorop8816 CUTLASS_SOURCES)
    gen_cutlass_kimpl(conv2d tensorop8832 CUTLASS_SOURCES)
    list(APPEND SOURCES ${CUTLASS_SOURCES})
    list(APPEND SOURCES ${CUSOURCES})
  file(GLOB_RECURSE CUSOURCES cuda/*.cu)
  set(CUTLASS_GEN_SCRIPT
      ${CMAKE_CURRENT_SOURCE_DIR}/../scripts/cutlass_generator/generator.py)
  set(CUTLASS_GEN_DIR ${CMAKE_CURRENT_BINARY_DIR}/cuda/cutlass/generated)
  set(CUTLASS_SOURCES "")
  function(gen_cutlass_kimpl op type gen_files)
    set(CURRENT_CUTLASS_STAGE_DIR ${CUTLASS_GEN_DIR}/${op}_${type}.stage)
    set(CURRENT_CUTLASS_GEN_DIR ${CUTLASS_GEN_DIR}/${op}_${type})
    set_directory_properties(PROPERTIES CMAKE_CONFIGURE_DEPENDS ${CUTLASS_GEN_SCRIPT})
    file(REMOVE_RECURSE ${CURRENT_CUTLASS_STAGE_DIR})
    file(MAKE_DIRECTORY ${CURRENT_CUTLASS_STAGE_DIR})
    file(MAKE_DIRECTORY ${CURRENT_CUTLASS_GEN_DIR})
    execute_process(
      COMMAND ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${CUTLASS_GEN_SCRIPT} --operations
              ${op} --type ${type} ${CURRENT_CUTLASS_STAGE_DIR}
      RESULT_VARIABLE gen_cutlass_result
      OUTPUT_FILE ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log
      ERROR_FILE ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log)
    if(NOT gen_cutlass_result EQUAL 0)
      message(
        FATAL_ERROR
          "Error generating library instances. See ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log"
      )
    endif()
    file(
      GLOB CUTLASS_GEN_FILES
      RELATIVE "${CURRENT_CUTLASS_GEN_DIR}/"
      "${CURRENT_CUTLASS_GEN_DIR}/*.cu")
    foreach(FILE ${CUTLASS_GEN_FILES})
      if(NOT EXISTS "${CURRENT_CUTLASS_STAGE_DIR}/${FILE}")
        file(REMOVE "${CURRENT_CUTLASS_GEN_DIR}/${FILE}")
      endif()
    endforeach()
    file(
      GLOB CUTLASS_GEN_FILES
      RELATIVE "${CURRENT_CUTLASS_STAGE_DIR}"
      "${CURRENT_CUTLASS_STAGE_DIR}/*.cu")
    foreach(FILE ${CUTLASS_GEN_FILES})
      execute_process(
        COMMAND ${CMAKE_COMMAND} -E copy_if_different
                "${CURRENT_CUTLASS_STAGE_DIR}/${FILE}" "${CURRENT_CUTLASS_GEN_DIR}")
    endforeach()
    file(REMOVE_RECURSE ${CURRENT_CUTLASS_STAGE_DIR})
    file(GLOB_RECURSE CUTLASS_GEN_FILES "${CURRENT_CUTLASS_GEN_DIR}/*.cu")
    list(APPEND ${gen_files} ${CUTLASS_GEN_FILES})
    set(${gen_files}
        "${${gen_files}}"
        PARENT_SCOPE)
  endfunction()
  gen_cutlass_kimpl(gemm simt CUTLASS_SOURCES)
  gen_cutlass_kimpl(gemm tensorop884 CUTLASS_SOURCES)
  gen_cutlass_kimpl(gemm tensorop1688 CUTLASS_SOURCES)
  gen_cutlass_kimpl(gemv simt CUTLASS_SOURCES)
  gen_cutlass_kimpl(deconv simt CUTLASS_SOURCES)
  gen_cutlass_kimpl(deconv tensorop8816 CUTLASS_SOURCES)
  gen_cutlass_kimpl(conv2d simt CUTLASS_SOURCES)
  gen_cutlass_kimpl(conv2d tensorop8816 CUTLASS_SOURCES)
  gen_cutlass_kimpl(conv2d tensorop8832 CUTLASS_SOURCES)
  list(APPEND SOURCES ${CUTLASS_SOURCES})
  list(APPEND SOURCES ${CUSOURCES})
 endif()
 if(MGE_WITH_ATLAS)
    file(GLOB_RECURSE SOURCES_ atlas/*.cpp)
    list(APPEND SOURCES ${SOURCES_})
    list(APPEND LIBMEGDNN_DEF -DMEGDNN_WITH_ATLAS=1)
  file(GLOB_RECURSE SOURCES_ atlas/*.cpp)
  list(APPEND SOURCES ${SOURCES_})
  list(APPEND LIBMEGDNN_DEF -DMEGDNN_WITH_ATLAS=1)
 endif()
 add_definitions(${LIBMEGDNN_DEF})
@@ -174,81 +196,85 @@ add_library(megdnn EXCLUDE_FROM_ALL OBJECT ${SOURCES})
 target_link_libraries(megdnn PUBLIC opr_param_defs)
 if(MGE_WITH_CUDA)
    target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:cutlass>)
    target_include_directories(megdnn PRIVATE ${CUDNN_INCLUDE_DIR})
  target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:cutlass>)
  target_include_directories(megdnn PRIVATE ${CUDNN_INCLUDE_DIR})
 endif()
 if(MGE_WITH_ROCM)
    target_include_directories(megdnn PUBLIC
        ${HIP_INCLUDE_DIR}
        ${MIOPEN_INCLUDE_DIR}
        ${ROCBLAS_INCLUDE_DIR}
        ${ROCRAND_INCLUDE_DIR}
        ${AMDOCL_INCLUDE_DIR})
    target_link_directories(megdnn PUBLIC
        ${HIP_LIBRARY_DIR}
        ${MIOPEN_LIBRARY_DIR}
        ${ROCBLAS_LIBRARY_DIR}
        ${ROCRAND_LIBRARY_DIR}
        ${AMDOCL_LIBRARY_DIR})
  target_include_directories(
    megdnn PUBLIC ${HIP_INCLUDE_DIR} ${MIOPEN_INCLUDE_DIR} ${ROCBLAS_INCLUDE_DIR}
                  ${ROCRAND_INCLUDE_DIR} ${AMDOCL_INCLUDE_DIR})
  target_link_directories(
    megdnn
    PUBLIC
    ${HIP_LIBRARY_DIR}
    ${MIOPEN_LIBRARY_DIR}
    ${ROCBLAS_LIBRARY_DIR}
    ${ROCRAND_LIBRARY_DIR}
    ${AMDOCL_LIBRARY_DIR})
 endif()
 if(${MGE_ARCH} STREQUAL "x86_64" OR ${MGE_ARCH} STREQUAL "i386" OR ${MGE_ARCH} STREQUAL "armv7" OR ${MGE_ARCH} STREQUAL "aarch64")
    if(MGE_ENABLE_CPUINFO)
        target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:cpuinfo>)
    endif()
 if(${MGE_ARCH} STREQUAL "x86_64"
   OR ${MGE_ARCH} STREQUAL "i386"
   OR ${MGE_ARCH} STREQUAL "armv7"
   OR ${MGE_ARCH} STREQUAL "aarch64")
  if(MGE_ENABLE_CPUINFO)
    target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:cpuinfo>)
  endif()
 endif()
 target_include_directories(megdnn
    PUBLIC
        $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/genfiles>
        $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/dnn/include>
        $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
    PRIVATE
        ${PROJECT_SOURCE_DIR}/dnn
        ${PROJECT_SOURCE_DIR}/third_party/midout/src
 )
 target_include_directories(
  megdnn
  PUBLIC $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/genfiles>
         $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/dnn/include>
         $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
  PRIVATE ${PROJECT_SOURCE_DIR}/dnn ${PROJECT_SOURCE_DIR}/third_party/midout/src)
 install(DIRECTORY ${PROJECT_SOURCE_DIR}/dnn/include DESTINATION . FILES_MATCHING PATTERN "*.h*")
 install(
  DIRECTORY ${PROJECT_SOURCE_DIR}/dnn/include
  DESTINATION .
  FILES_MATCHING
  PATTERN "*.h*")
 if(CXX_SUPPORT_WCLASS_MEMACCESS)
    if(MGE_WITH_CUDA)
        target_compile_options(megdnn PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-Wno-class-memaccess>"
            "$<$<NOT:$<COMPILE_LANGUAGE:CUDA>>:-Wno-class-memaccess>")
    else()
        target_compile_options(megdnn PRIVATE "-Wno-class-memaccess")
    endif()
  if(MGE_WITH_CUDA)
    target_compile_options(
      megdnn PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-Wno-class-memaccess>"
                     "$<$<NOT:$<COMPILE_LANGUAGE:CUDA>>:-Wno-class-memaccess>")
  else()
    target_compile_options(megdnn PRIVATE "-Wno-class-memaccess")
  endif()
 endif()
 target_compile_definitions(megdnn INTERFACE ${LIBMEGDNN_DEF})
 if(MGE_WITH_MKLDNN AND ${MGE_ARCH} STREQUAL "x86_64")
    if (BUILD_SHARED_LIBS)
        target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:dnnl>)
    else()
        target_link_libraries(megdnn PRIVATE dnnl)
    endif()
  if(BUILD_SHARED_LIBS)
    target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:dnnl>)
  else()
    target_link_libraries(megdnn PRIVATE dnnl)
  endif()
 endif()
 if (BUILD_SHARED_LIBS)
    target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:${MGE_BLAS_LIBS}>)
 if(BUILD_SHARED_LIBS)
  target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:${MGE_BLAS_LIBS}>)
 else()
    target_link_libraries(megdnn PRIVATE ${MGE_BLAS_LIBS})
  target_link_libraries(megdnn PRIVATE ${MGE_BLAS_LIBS})
 endif()
 if (MGE_WITH_ROCM)
    target_link_libraries(megdnn PRIVATE ${HIPOBJS} ${MGE_ROCM_LIBS})
 endif ()
 if(MGE_WITH_ROCM)
  target_link_libraries(megdnn PRIVATE ${HIPOBJS} ${MGE_ROCM_LIBS})
 endif()
 if(MGE_WITH_ATLAS)
    if (BUILD_SHARED_LIBS)
        target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:${MGE_ATLAS_LIBS}>)
    else()
        target_link_libraries(megdnn PRIVATE ${MGE_ATLAS_LIBS})
    endif()
  if(BUILD_SHARED_LIBS)
    target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:${MGE_ATLAS_LIBS}>)
  else()
    target_link_libraries(megdnn PRIVATE ${MGE_ATLAS_LIBS})
  endif()
 endif()
 if(CMAKE_THREAD_LIBS_INIT)
    target_link_libraries(megdnn PRIVATE Threads::Threads)
  target_link_libraries(megdnn PRIVATE Threads::Threads)
 endif()
 install(TARGETS megdnn EXPORT ${MGE_EXPORT_TARGETS})
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1(2);
 INSTANTIATION_CONV_S1_BIAS(2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_BROADCAST_CHANNEL_BIAS(2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s1_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_NO_BIAS(2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2(5);
 INSTANTIATION_CONV_S2_BIAS(2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_BROADCAST_CHANNEL_BIAS(2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_NO_BIAS(2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1(5);
 INSTANTIATION_CONV_S1_BIAS(3);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_BROADCAST_CHANNEL_BIAS(3);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_NO_BIAS(3);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_2x2s2.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2(2);
 INSTANTIATION_CONV_S2_BIAS(3);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_BROADCAST_CHANNEL_BIAS(3);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_NO_BIAS(3);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s1.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1(3);
 INSTANTIATION_CONV_S1_BIAS(5);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_BROADCAST_CHANNEL_BIAS(5);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s1_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_NO_BIAS(5);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2(7);
 INSTANTIATION_CONV_S2_BIAS(5);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_BROADCAST_CHANNEL_BIAS(5);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_5x5s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_NO_BIAS(5);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1(7);
 INSTANTIATION_CONV_S1_BIAS(7);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_BROADCAST_CHANNEL_BIAS(7);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s1_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h"
 INSTANTIATION_CONV_S1_NO_BIAS(7);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_3x3s2.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2(3);
 INSTANTIATION_CONV_S2_BIAS(7);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_BROADCAST_CHANNEL_BIAS(7);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_7x7s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h"
 INSTANTIATION_CONV_S2_NO_BIAS(7);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s1.h
@@ -469,9 +469,12 @@ void conv_bias::conv_direct_fp32_nchw44(
    INSTANTIATION(filter_size, bias, HSwishOp<dt_float32>) \
    INSTANTIATION(filter_size, bias, SigmoidOp<dt_float32>)
 #define INSTANTIATION_CONV_S1(filter_size)                \
    FOR_OP(filter_size, BiasMode::NO_BIAS)                \
    FOR_OP(filter_size, BiasMode::BROADCAST_CHANNEL_BIAS) \
    FOR_OP(filter_size, BiasMode::BIAS)
 #define INSTANTIATION_CONV_S1_NO_BIAS(filter_size) \
    FOR_OP(filter_size, BiasMode::NO_BIAS)
 // vim: syntax=cpp.doxygen
 #define INSTANTIATION_CONV_S1_BROADCAST_CHANNEL_BIAS(filter_size) \
    FOR_OP(filter_size, BiasMode::BROADCAST_CHANNEL_BIAS)
 #define INSTANTIATION_CONV_S1_BIAS(filter_size) FOR_OP(filter_size, BiasMode::BIAS)
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw44_kern_common_s2.h
@@ -550,9 +550,12 @@ void conv_bias::conv_direct_fp32_nchw44(
    INSTANTIATION(filter_size, bias, HSwishOp<dt_float32>) \
    INSTANTIATION(filter_size, bias, SigmoidOp<dt_float32>)
 #define INSTANTIATION_CONV_S2(filter_size)                \
    FOR_OP(filter_size, BiasMode::NO_BIAS)                \
    FOR_OP(filter_size, BiasMode::BROADCAST_CHANNEL_BIAS) \
    FOR_OP(filter_size, BiasMode::BIAS)
 #define INSTANTIATION_CONV_S2_NO_BIAS(filter_size) \
    FOR_OP(filter_size, BiasMode::NO_BIAS)
 // vim: syntax=cpp.doxygen
 #define INSTANTIATION_CONV_S2_BROADCAST_CHANNEL_BIAS(filter_size) \
    FOR_OP(filter_size, BiasMode::BROADCAST_CHANNEL_BIAS)
 #define INSTANTIATION_CONV_S2_BIAS(filter_size) FOR_OP(filter_size, BiasMode::BIAS)
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV(2, 1);
 INSTANCE_CONV_BIAS(2, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(2, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s1_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(2, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV(2, 2);
 INSTANCE_CONV_BIAS(2, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(2, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_2x2s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(2, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV(3, 1);
 INSTANCE_CONV_BIAS(3, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(3, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s1_no_bias
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(3, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_bias.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2.cpp
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -11,4 +11,5 @@
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV(3, 2);
 INSTANCE_CONV_BIAS(3, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(3, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_3x3s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(3, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BIAS(5, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(5, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s1_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(5, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BIAS(5, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(5, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_5x5s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(5, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BIAS(7, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(7, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s1_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(7, 1);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BIAS(7, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_broadcast_channel_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_broadcast_channel_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_broadcast_channel_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(7, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_no_bias.cpp
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_no_bias.cpp
@@ -0,0 +1,15 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_7x7s2_no_bias.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h"
 INSTANCE_CONV_NO_BIAS(7, 2);
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h
+++ b/dnn/src/arm_common/conv_bias/fp32/direct_kernels/f32_direct_nchw_nchw44_kern_common.h
@@ -928,9 +928,11 @@ void fp32_direct_nchw_nchw44::conv_direct_fp32_nchw_nchw44(
    INSTANTIATION(stride, filter, bias, ReluOp<dt_float32>) \
    INSTANTIATION(stride, filter, bias, HSwishOp<dt_float32>)
 #define INSTANCE_CONV(filter, stride)                        \
    FOR_OP(stride, filter, BiasMode::NO_BIAS)                \
    FOR_OP(stride, filter, BiasMode::BROADCAST_CHANNEL_BIAS) \
    FOR_OP(stride, filter, BiasMode::BIAS)
 #define INSTANCE_CONV_NO_BIAS(filter, stride) FOR_OP(stride, filter, BiasMode::NO_BIAS)
 #define INSTANCE_CONV_BROADCAST_CHANNEL_BIAS(filter, stride) \
    FOR_OP(stride, filter, BiasMode::BROADCAST_CHANNEL_BIAS)
 #define INSTANCE_CONV_BIAS(filter, stride) FOR_OP(stride, filter, BiasMode::BIAS)
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.cpp
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.h
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -265,7 +265,8 @@ void conv_direct_sdot_int8_nchw44(
 #define INSTANTIATION(dst_type, stride, filter_size, bias_mode, Op)                \
    template void                                                                  \
    conv_direct_sdot_int8_nchw44<dst_type, stride, bias_mode, Op, filter_size>(    \
    megdnn::arm_common::direct_dotprod_nchw44::conv_direct_sdot_int8_nchw44<       \
            dst_type, stride, bias_mode, Op, filter_size>(                         \
            dst_type * dst, const int oh, const int ow, const int8_t* src,         \
            const int ih, const int iw, const int8_t* weight, const int32_t* bias, \
            const int oh_size, const int oc, const int ic, const Op& op);
@@ -284,22 +285,6 @@ void conv_direct_sdot_int8_nchw44(
    FOR_OP(stride, i, BiasMode::NO_BIAS) \
    FOR_OP(stride, i, BiasMode::BROADCAST_CHANNEL_BIAS)
 #define FOR_FILTER(stride) \
    FOR_BIAS(stride, 2)    \
    FOR_BIAS(stride, 3)    \
    FOR_BIAS(stride, 5)    \
    FOR_BIAS(stride, 7)
 FOR_FILTER(1)
 #undef FOR_STRIDE
 #undef FOR_FILTER
 #undef FOR_IC
 #undef FOR_BIAS
 #undef FOR_NONLINEAR
 #undef FOR_REMAIN
 #undef INSTANTIATION
 }  // namespace direct_dotprod_nchw44
 }  // namespace arm_common
 }  // namespace megdnn
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_2x2.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_2x2.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_2x2.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(1, 2);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_3x3.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_3x3.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_3x3.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(1, 3);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_5x5.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_5x5.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_5x5.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(1, 5);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_7x7.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_7x7.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1_7x7.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s1.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(1, 7);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.cpp
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.cpp
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.h
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -266,7 +266,8 @@ void conv_direct_sdot_int8_nchw44(
 #define INSTANTIATION(dst_type, stride, filter_size, bias_mode, Op)                \
    template void                                                                  \
    conv_direct_sdot_int8_nchw44<dst_type, stride, bias_mode, Op, filter_size>(    \
    megdnn::arm_common::direct_dotprod_nchw44::conv_direct_sdot_int8_nchw44<       \
            dst_type, stride, bias_mode, Op, filter_size>(                         \
            dst_type * dst, const int oh, const int ow, const int8_t* src,         \
            const int ih, const int iw, const int8_t* weight, const int32_t* bias, \
            const int oh_size, const int oc, const int ic, const Op& op);
@@ -285,22 +286,6 @@ void conv_direct_sdot_int8_nchw44(
    FOR_OP(stride, i, BiasMode::NO_BIAS) \
    FOR_OP(stride, i, BiasMode::BROADCAST_CHANNEL_BIAS)
 #define FOR_FILTER(stride) \
    FOR_BIAS(stride, 2)    \
    FOR_BIAS(stride, 3)    \
    FOR_BIAS(stride, 5)    \
    FOR_BIAS(stride, 7)
 FOR_FILTER(2)
 #undef FOR_STRIDE
 #undef FOR_FILTER
 #undef FOR_IC
 #undef FOR_BIAS
 #undef FOR_NONLINEAR
 #undef FOR_REMAIN
 #undef INSTANTIATION
 }  // namespace direct_dotprod_nchw44
 }  // namespace arm_common
 }  // namespace megdnn
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_2x2.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_2x2.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_2x2.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(2, 2);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_3x3.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_3x3.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_3x3.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(2, 3);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_5x5.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_5x5.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_5x5.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(2, 5);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_7x7.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_7x7.cpp
@@ -0,0 +1,21 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2_7x7.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/dot_direct_nchw44_s2.h"
 #if MGB_ENABLE_DOT
 using namespace megdnn;
 using namespace arm_common;
 FOR_BIAS(2, 7);
 #endif
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_common.h
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_common.h
@@ -1,6 +1,6 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1.cpp
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_common.h
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
@@ -45,4 +45,4 @@ public:
 }  // namespace arm_common
 }  // namespace megdnn
 // vim: syntax=cpp.doxygen
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1.cpp
@@ -13,336 +13,9 @@
 #include "src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_common.h"
 #include "src/arm_common/conv_bias/int8/direct_nchw_nchw44_kern.h"
 namespace megdnn {
 namespace arm_common {
 namespace {
 /**
 * @brief core code for calculation patten
 *
 * @tparam src_idx is offset of src reg
 * @tparam weight_idx is offset of weight reg
 * @tparam c_dim is output channel
 * @tparam Func mla operation funcion
 * @tparam stride
 * @tparam T outpur regs type
 * @tparam T2 src regs type
 * @tparam T3 weight regs type
 * @tparam T4 temp regs type
 */
 template <
        int src_idx, int weight_idx, int c_dim, int stride, typename T, typename T2,
        typename T3, typename T4>
 struct ShiftCalHelper {
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight, T4& temp);
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight);
 };
 template <
        int src_idx, int weight_idx, int c_dim, int stride, typename T, typename T2,
        typename T3, typename T4>
 MEGDNN_ALWAYS_INLINE void cal_helper(T& c, T2& src, T3& weight, T4& temp) {
    ShiftCalHelper<src_idx, weight_idx, c_dim, stride, T, T2, T3, T4>::impl(
            c, src, weight, temp);
 }
 template <
        int src_idx, int weight_idx, int c_dim, int stride, typename T, typename T2,
        typename T3>
 MEGDNN_ALWAYS_INLINE void cal_helper(T& c, T2& src, T3& weight) {
    ShiftCalHelper<src_idx, weight_idx, c_dim, stride, T, T2, T3, int>::impl(
            c, src, weight);
 };
 template <
        int src_idx, int weight_idx, typename T, typename T2, typename T3, typename T4>
 struct ShiftCalHelper<src_idx, weight_idx, 2, 1, T, T2, T3, T4> {
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight, T4& temp) {
        c[0][0] = vdotq_s32_h(
                src[(0 + src_idx) % 8], weight[0][weight_idx], c[0][0], temp[0]);
        c[1][0] = vdotq_s32_h(
                src[(0 + src_idx) % 8], weight[1][weight_idx], c[1][0], temp[1]);
        c[0][1] = vdotq_s32_h(
                src[(1 + src_idx) % 8], weight[0][weight_idx], c[0][1], temp[2]);
        c[1][1] = vdotq_s32_h(
                src[(1 + src_idx) % 8], weight[1][weight_idx], c[1][1], temp[3]);
        c[0][2] = vdotq_s32_h(
                src[(2 + src_idx) % 8], weight[0][weight_idx], c[0][2], temp[0]);
        c[1][2] = vdotq_s32_h(
                src[(2 + src_idx) % 8], weight[1][weight_idx], c[1][2], temp[1]);
        c[0][3] = vdotq_s32_h(
                src[(3 + src_idx) % 8], weight[0][weight_idx], c[0][3], temp[2]);
        c[1][3] = vdotq_s32_h(
                src[(3 + src_idx) % 8], weight[1][weight_idx], c[1][3], temp[3]);
        c[0][4] = vdotq_s32_h(
                src[(4 + src_idx) % 8], weight[0][weight_idx], c[0][4], temp[0]);
        c[1][4] = vdotq_s32_h(
                src[(4 + src_idx) % 8], weight[1][weight_idx], c[1][4], temp[1]);
        c[0][5] = vdotq_s32_h(
                src[(5 + src_idx) % 8], weight[0][weight_idx], c[0][5], temp[2]);
        c[1][5] = vdotq_s32_h(
                src[(5 + src_idx) % 8], weight[1][weight_idx], c[1][5], temp[3]);
        c[0][6] = vdotq_s32_h(
                src[(6 + src_idx) % 8], weight[0][weight_idx], c[0][6], temp[0]);
        c[1][6] = vdotq_s32_h(
                src[(6 + src_idx) % 8], weight[1][weight_idx], c[1][6], temp[1]);
        c[0][7] = vdotq_s32_h(
                src[(7 + src_idx) % 8], weight[0][weight_idx], c[0][7], temp[2]);
        c[1][7] = vdotq_s32_h(
                src[(7 + src_idx) % 8], weight[1][weight_idx], c[1][7], temp[3]);
    }
    static MEGDNN_ALWAYS_INLINE void impl(T&, T2&, T3&);
 };
 template <
        int src_idx, int weight_idx, typename T, typename T2, typename T3, typename T4>
 struct ShiftCalHelper<src_idx, weight_idx, 1, 1, T, T2, T3, T4> {
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight, T4& temp) {
        c[0][0] = vdotq_s32_h(
                src[(0 + src_idx) % 8], weight[0][weight_idx], c[0][0], temp[0]);
        c[0][1] = vdotq_s32_h(
                src[(1 + src_idx) % 8], weight[0][weight_idx], c[0][1], temp[1]);
        c[0][2] = vdotq_s32_h(
                src[(2 + src_idx) % 8], weight[0][weight_idx], c[0][2], temp[2]);
        c[0][3] = vdotq_s32_h(
                src[(3 + src_idx) % 8], weight[0][weight_idx], c[0][3], temp[3]);
        c[0][4] = vdotq_s32_h(
                src[(4 + src_idx) % 8], weight[0][weight_idx], c[0][4], temp[0]);
        c[0][5] = vdotq_s32_h(
                src[(5 + src_idx) % 8], weight[0][weight_idx], c[0][5], temp[1]);
        c[0][6] = vdotq_s32_h(
                src[(6 + src_idx) % 8], weight[0][weight_idx], c[0][6], temp[2]);
        c[0][7] = vdotq_s32_h(
                src[(7 + src_idx) % 8], weight[0][weight_idx], c[0][7], temp[3]);
    }
    static MEGDNN_ALWAYS_INLINE void impl(T&, T2&, T3&);
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 1, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 1;
        constexpr int filter_width = 4;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 1;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 0 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 2, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 2;
        constexpr int filter_width = 4;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 1;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 0 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr + 1 * filter_width * oc_step, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 1 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 3, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 3;
        constexpr int filter_width = 4;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 1;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 0 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr + 1 * filter_width * oc_step, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 1 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr + 2 * filter_width * oc_step, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 2 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 5, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 5;
        constexpr int filter_width = 8;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 2;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
 #define cb(step)                                                                     \
    load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(                           \
            dot4_weight, weight_ptr + step * filter_width * oc_step, ld_weight_oc);  \
    load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(                                  \
            src, nchw_src_ptr + step * iw * pack_iw_len, 0);                         \
    cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);                    \
    load_helper<4, 0, simd_len, 0, Vld1q_s8>(                                        \
            src, nchw_src_ptr + step * iw * pack_iw_len + src_reg * pack_iw_len, 0); \
    cal_helper<4, 1, c_dim, stride>(c, src, dot4_weight, temp_c);
            UNROLL_CALL_RAW(5, cb);
 #undef cb
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 7, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 7;
        constexpr int filter_width = 8;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 2;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
 #define cb(step)                                                                     \
    load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(                           \
            dot4_weight, weight_ptr + step * filter_width * oc_step, ld_weight_oc);  \
    load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(                                  \
            src, nchw_src_ptr + step * iw * pack_iw_len, 0);                         \
    cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);                    \
    load_helper<4, 0, simd_len, 0, Vld1q_s8>(                                        \
            src, nchw_src_ptr + step * iw * pack_iw_len + src_reg * pack_iw_len, 0); \
    cal_helper<4, 1, c_dim, stride>(c, src, dot4_weight, temp_c);
            UNROLL_CALL_RAW(7, cb);
 #undef cb
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 }  // namespace
 namespace int8_direct_nchw_nchw44 {
 /**
 * pack {oc / 4, fh, fw, ic, 4(oc)} to {oc / 4, ic, fh ,fw/4, 4(oc)*4(fw)}
@@ -444,115 +117,9 @@ void pack_nchw_src_for_nchw44_conv<1>(
    }
 }
 template <BiasMode bias_mode, typename Op, size_t filter_size>
 struct ConvDiectStrideInt8NchwNchw44<bias_mode, Op, filter_size, 1> {
    static void impl(
            const int8_t* src, const int8_t* filter, const int32_t* bias, int32_t* temp,
            int8_t* dst, const size_t oc, const size_t ic, const size_t ih,
            const size_t iw, const size_t oh, const size_t ow, const Op& op) {
        MEGDNN_MARK_USED_VAR(temp);
        constexpr int stride = 1;
        constexpr size_t fh = filter_size;
        constexpr size_t fw = (filter_size + 3) / 4 * 4;
        constexpr size_t ic_step = 1;
        constexpr size_t big_oc_step = 8;
        constexpr size_t oc_step = 4;
        constexpr size_t ih_step = 1;
        constexpr size_t oh_step = 1;
        constexpr size_t ow_step = 8;
        constexpr size_t stride_h = stride;
        constexpr size_t stride_w = stride;
        constexpr int pack_iw_len = 16;
        const size_t img_stride = oh * ow;
        const size_t ow_end = ow / ow_step * ow_step;
        const size_t ow_remain = ow - ow_end;
        const size_t oc_end = oc / big_oc_step * big_oc_step;
        const size_t oc_remain = oc - oc_end;
        const int ld_dst_oc = oc_step * img_stride;
        using remain_fun = std::function<void(
                const int8_t* src_ptr, const int8_t* weight_ptr,
                const int32_t* bias_ptr, int8_t* dst_ptr, int ic, int ih, int iw,
                int ld_dst_oc, const Op& op)>;
        remain_fun kern_big_oc_remain = nullptr;
        remain_fun kern_small_oc_remain = nullptr;
        switch (ow_remain) {
 #define cb(step)                                                              \
    case step:                                                                \
        kern_big_oc_remain = KerNeonXXs2NchwNchw44<                           \
                bias_mode, Op, step, filter_size, big_oc_step, stride>::impl; \
        kern_small_oc_remain = KerNeonXXs2NchwNchw44<                         \
                bias_mode, Op, step, filter_size, oc_step, stride>::impl;     \
        break;
            UNROLL_CALL_RAW(8, cb);
            default:
                megdnn_assert(0, "no remain %zu for kern", ow_remain);
        }
        for (size_t oc_idx = 0; oc_idx < oc_end; oc_idx += big_oc_step) {
            const size_t weight_offset = oc_idx * ic * fh * fw;
            for (size_t oh_idx = 0; oh_idx < oh; oh_idx += oh_step) {
                for (size_t ow_idx = 0; ow_idx < ow_end; ow_idx += ow_step) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_idx * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_idx) * oc_step;
                    KerNeonXXs2NchwNchw44<
                            bias_mode, Op, ow_step, filter_size, big_oc_step, stride>::
                            impl(src + src_offset, filter + weight_offset,
                                 bias + oc_idx, dst + dst_offset, ic, ih, iw, ld_dst_oc,
                                 op);
                }
                if (ow_remain > 0) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_end * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_end) * oc_step;
                    kern_big_oc_remain(
                            src + src_offset, filter + weight_offset, bias + oc_idx,
                            dst + dst_offset, ic, ih, iw, ld_dst_oc, op);
                }
            }
        }
        if (oc_remain > 0) {
            size_t oc_idx = oc_end;
            const size_t weight_offset = oc_idx * ic * fh * fw;
            for (size_t oh_idx = 0; oh_idx < oh; oh_idx += oh_step) {
                for (size_t ow_idx = 0; ow_idx < ow_end; ow_idx += ow_step) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_idx * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_idx) * oc_step;
                    KerNeonXXs2NchwNchw44<
                            bias_mode, Op, ow_step, filter_size, oc_step, stride>::
                            impl(src + src_offset, filter + weight_offset,
                                 bias + oc_idx, dst + dst_offset, ic, ih, iw, ld_dst_oc,
                                 op);
                }
                if (ow_remain > 0) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_end * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_end) * oc_step;
                    kern_small_oc_remain(
                            src + src_offset, filter + weight_offset, bias + oc_idx,
                            dst + dst_offset, ic, ih, iw, ld_dst_oc, op);
                }
            }
        }
    }
 };
 #define INSTANCE_CONV_KERN_FUN(stride, filter_size, bias_mode, Op) \
    template struct ConvDiectStrideInt8NchwNchw44<bias_mode, Op, filter_size, stride>;
    template struct megdnn::arm_common::int8_direct_nchw_nchw44::  \
            ConvDiectStrideInt8NchwNchw44<bias_mode, Op, filter_size, stride>;
 #define INSTANCE_OP_PARAM(stride, filter, bias_mode)                               \
    INSTANCE_CONV_KERN_FUN(                                                        \
@@ -566,17 +133,10 @@ struct ConvDiectStrideInt8NchwNchw44<bias_mode, Op, filter_size, 1> {
    INSTANCE_OP_PARAM(stride, filter, BiasMode::NO_BIAS) \
    INSTANCE_OP_PARAM(stride, filter, BiasMode::BROADCAST_CHANNEL_BIAS)
 #define INSTANCE_CONV_KERN(stride)      \
    INSTANCE_BIAS_MODE_PARAM(stride, 1) \
    INSTANCE_BIAS_MODE_PARAM(stride, 2) \
    INSTANCE_BIAS_MODE_PARAM(stride, 3) \
    INSTANCE_BIAS_MODE_PARAM(stride, 5) \
    INSTANCE_BIAS_MODE_PARAM(stride, 7)
 INSTANCE_CONV_KERN(1);
 #define INSTANCE_CONV_KERN(stride, filter) INSTANCE_BIAS_MODE_PARAM(stride, filter)
 }  // namespace int8_direct_nchw_nchw44
 }  // namespace arm_common
 }  // namespace megdnn
 // vim: syntax=cpp.doxygen
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1.h
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1.h
@@ -0,0 +1,481 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1.h
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_common.h"
 #include "src/arm_common/conv_bias/int8/direct_nchw_nchw44_kern.h"
 namespace megdnn {
 namespace arm_common {
 namespace {
 /**
 * @brief core code for calculation patten
 *
 * @tparam src_idx is offset of src reg
 * @tparam weight_idx is offset of weight reg
 * @tparam c_dim is output channel
 * @tparam Func mla operation funcion
 * @tparam stride
 * @tparam T outpur regs type
 * @tparam T2 src regs type
 * @tparam T3 weight regs type
 * @tparam T4 temp regs type
 */
 template <
        int src_idx, int weight_idx, int c_dim, int stride, typename T, typename T2,
        typename T3, typename T4>
 struct ShiftCalHelper {
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight, T4& temp);
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight);
 };
 template <
        int src_idx, int weight_idx, int c_dim, int stride, typename T, typename T2,
        typename T3, typename T4>
 MEGDNN_ALWAYS_INLINE void cal_helper(T& c, T2& src, T3& weight, T4& temp) {
    ShiftCalHelper<src_idx, weight_idx, c_dim, stride, T, T2, T3, T4>::impl(
            c, src, weight, temp);
 }
 template <
        int src_idx, int weight_idx, int c_dim, int stride, typename T, typename T2,
        typename T3>
 MEGDNN_ALWAYS_INLINE void cal_helper(T& c, T2& src, T3& weight) {
    ShiftCalHelper<src_idx, weight_idx, c_dim, stride, T, T2, T3, int>::impl(
            c, src, weight);
 };
 template <
        int src_idx, int weight_idx, typename T, typename T2, typename T3, typename T4>
 struct ShiftCalHelper<src_idx, weight_idx, 2, 1, T, T2, T3, T4> {
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight, T4& temp) {
        c[0][0] = vdotq_s32_h(
                src[(0 + src_idx) % 8], weight[0][weight_idx], c[0][0], temp[0]);
        c[1][0] = vdotq_s32_h(
                src[(0 + src_idx) % 8], weight[1][weight_idx], c[1][0], temp[1]);
        c[0][1] = vdotq_s32_h(
                src[(1 + src_idx) % 8], weight[0][weight_idx], c[0][1], temp[2]);
        c[1][1] = vdotq_s32_h(
                src[(1 + src_idx) % 8], weight[1][weight_idx], c[1][1], temp[3]);
        c[0][2] = vdotq_s32_h(
                src[(2 + src_idx) % 8], weight[0][weight_idx], c[0][2], temp[0]);
        c[1][2] = vdotq_s32_h(
                src[(2 + src_idx) % 8], weight[1][weight_idx], c[1][2], temp[1]);
        c[0][3] = vdotq_s32_h(
                src[(3 + src_idx) % 8], weight[0][weight_idx], c[0][3], temp[2]);
        c[1][3] = vdotq_s32_h(
                src[(3 + src_idx) % 8], weight[1][weight_idx], c[1][3], temp[3]);
        c[0][4] = vdotq_s32_h(
                src[(4 + src_idx) % 8], weight[0][weight_idx], c[0][4], temp[0]);
        c[1][4] = vdotq_s32_h(
                src[(4 + src_idx) % 8], weight[1][weight_idx], c[1][4], temp[1]);
        c[0][5] = vdotq_s32_h(
                src[(5 + src_idx) % 8], weight[0][weight_idx], c[0][5], temp[2]);
        c[1][5] = vdotq_s32_h(
                src[(5 + src_idx) % 8], weight[1][weight_idx], c[1][5], temp[3]);
        c[0][6] = vdotq_s32_h(
                src[(6 + src_idx) % 8], weight[0][weight_idx], c[0][6], temp[0]);
        c[1][6] = vdotq_s32_h(
                src[(6 + src_idx) % 8], weight[1][weight_idx], c[1][6], temp[1]);
        c[0][7] = vdotq_s32_h(
                src[(7 + src_idx) % 8], weight[0][weight_idx], c[0][7], temp[2]);
        c[1][7] = vdotq_s32_h(
                src[(7 + src_idx) % 8], weight[1][weight_idx], c[1][7], temp[3]);
    }
    static MEGDNN_ALWAYS_INLINE void impl(T&, T2&, T3&);
 };
 template <
        int src_idx, int weight_idx, typename T, typename T2, typename T3, typename T4>
 struct ShiftCalHelper<src_idx, weight_idx, 1, 1, T, T2, T3, T4> {
    static MEGDNN_ALWAYS_INLINE void impl(T& c, T2& src, T3& weight, T4& temp) {
        c[0][0] = vdotq_s32_h(
                src[(0 + src_idx) % 8], weight[0][weight_idx], c[0][0], temp[0]);
        c[0][1] = vdotq_s32_h(
                src[(1 + src_idx) % 8], weight[0][weight_idx], c[0][1], temp[1]);
        c[0][2] = vdotq_s32_h(
                src[(2 + src_idx) % 8], weight[0][weight_idx], c[0][2], temp[2]);
        c[0][3] = vdotq_s32_h(
                src[(3 + src_idx) % 8], weight[0][weight_idx], c[0][3], temp[3]);
        c[0][4] = vdotq_s32_h(
                src[(4 + src_idx) % 8], weight[0][weight_idx], c[0][4], temp[0]);
        c[0][5] = vdotq_s32_h(
                src[(5 + src_idx) % 8], weight[0][weight_idx], c[0][5], temp[1]);
        c[0][6] = vdotq_s32_h(
                src[(6 + src_idx) % 8], weight[0][weight_idx], c[0][6], temp[2]);
        c[0][7] = vdotq_s32_h(
                src[(7 + src_idx) % 8], weight[0][weight_idx], c[0][7], temp[3]);
    }
    static MEGDNN_ALWAYS_INLINE void impl(T&, T2&, T3&);
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 1, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 1;
        constexpr int filter_width = 4;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 1;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 0 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 2, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 2;
        constexpr int filter_width = 4;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 1;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 0 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr + 1 * filter_width * oc_step, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 1 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 3, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 3;
        constexpr int filter_width = 4;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 1;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 0 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr + 1 * filter_width * oc_step, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 1 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(
                    dot4_weight, weight_ptr + 2 * filter_width * oc_step, ld_weight_oc);
            load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(
                    src, nchw_src_ptr + 2 * iw * pack_iw_len, 0);
            cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 5, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 5;
        constexpr int filter_width = 8;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 2;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
 #define cb(step)                                                                     \
    load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(                           \
            dot4_weight, weight_ptr + step * filter_width * oc_step, ld_weight_oc);  \
    load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(                                  \
            src, nchw_src_ptr + step * iw * pack_iw_len, 0);                         \
    cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);                    \
    load_helper<4, 0, simd_len, 0, Vld1q_s8>(                                        \
            src, nchw_src_ptr + step * iw * pack_iw_len + src_reg * pack_iw_len, 0); \
    cal_helper<4, 1, c_dim, stride>(c, src, dot4_weight, temp_c);
            UNROLL_CALL_RAW(5, cb);
 #undef cb
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 template <BiasMode bias_mode, typename Op, int remain_w, int oc_block>
 struct KerNeonXXs2NchwNchw44<bias_mode, Op, remain_w, 7, oc_block, 1> {
    static void impl(
            const int8_t* src_ptr, const int8_t* weight_ptr, const int32_t* bias_ptr,
            int8_t* dst_ptr, int ic, int ih, int iw, int ld_dst_oc, const Op& op) {
        constexpr int stride = 1;
        constexpr int filter_height = 7;
        constexpr int filter_width = 8;
        constexpr int oc_step = 4;
        constexpr int loop_ic_step = 1;
        constexpr int simd_len = 16;
        constexpr int pack_iw_len = 16;
        constexpr int src_reg = 8;
        constexpr int weight_reg = 2;
        const int ic_stride = ih * iw * pack_iw_len;
        const int ld_weight_oc = oc_step * filter_height * filter_width * ic;
        constexpr int c_dim = OCHelper<oc_block>::val;
        int32x4_t c[c_dim][8];
        init_ocx_ow8<c_dim, bias_mode, remain_w>(c, bias_ptr, oc_step);
        for (int ic_idx = 0; ic_idx < ic; ic_idx += loop_ic_step) {
            const int8_t* nchw_src_ptr = src_ptr + ic_idx * ic_stride;
            int8x16_t src[src_reg];
            int8x16_t dot4_weight[c_dim][weight_reg];
            int16x8_t temp_c[4];
 #define cb(step)                                                                     \
    load_helper<weight_reg, 0, simd_len, c_dim, Vld1q_s8>(                           \
            dot4_weight, weight_ptr + step * filter_width * oc_step, ld_weight_oc);  \
    load_helper<src_reg, 0, simd_len, 0, Vld1q_s8>(                                  \
            src, nchw_src_ptr + step * iw * pack_iw_len, 0);                         \
    cal_helper<0, 0, c_dim, stride>(c, src, dot4_weight, temp_c);                    \
    load_helper<4, 0, simd_len, 0, Vld1q_s8>(                                        \
            src, nchw_src_ptr + step * iw * pack_iw_len + src_reg * pack_iw_len, 0); \
    cal_helper<4, 1, c_dim, stride>(c, src, dot4_weight, temp_c);
            UNROLL_CALL_RAW(7, cb);
 #undef cb
            weight_ptr += oc_step * filter_height * filter_width;
        }
        store_ocx_ow8_remain_static_dt<c_dim, remain_w, Op, dt_qint8*>(
                c, op, dst_ptr, ld_dst_oc);
    }
 };
 }  // namespace
 namespace int8_direct_nchw_nchw44 {
 /**
 * pack {oc / 4, fh, fw, ic, 4(oc)} to {oc / 4, ic, fh ,fw/4, 4(oc)*4(fw)}
 * pack interleave two adjacent row in filter to one row
 * */
 template <BiasMode bias_mode, typename Op, size_t filter_size>
 struct ConvDiectStrideInt8NchwNchw44<bias_mode, Op, filter_size, 1> {
    static void impl(
            const int8_t* src, const int8_t* filter, const int32_t* bias, int32_t* temp,
            int8_t* dst, const size_t oc, const size_t ic, const size_t ih,
            const size_t iw, const size_t oh, const size_t ow, const Op& op) {
        MEGDNN_MARK_USED_VAR(temp);
        constexpr int stride = 1;
        constexpr size_t fh = filter_size;
        constexpr size_t fw = (filter_size + 3) / 4 * 4;
        constexpr size_t ic_step = 1;
        constexpr size_t big_oc_step = 8;
        constexpr size_t oc_step = 4;
        constexpr size_t ih_step = 1;
        constexpr size_t oh_step = 1;
        constexpr size_t ow_step = 8;
        constexpr size_t stride_h = stride;
        constexpr size_t stride_w = stride;
        constexpr int pack_iw_len = 16;
        const size_t img_stride = oh * ow;
        const size_t ow_end = ow / ow_step * ow_step;
        const size_t ow_remain = ow - ow_end;
        const size_t oc_end = oc / big_oc_step * big_oc_step;
        const size_t oc_remain = oc - oc_end;
        const int ld_dst_oc = oc_step * img_stride;
        using remain_fun = std::function<void(
                const int8_t* src_ptr, const int8_t* weight_ptr,
                const int32_t* bias_ptr, int8_t* dst_ptr, int ic, int ih, int iw,
                int ld_dst_oc, const Op& op)>;
        remain_fun kern_big_oc_remain = nullptr;
        remain_fun kern_small_oc_remain = nullptr;
        switch (ow_remain) {
 #define cb(step)                                                              \
    case step:                                                                \
        kern_big_oc_remain = KerNeonXXs2NchwNchw44<                           \
                bias_mode, Op, step, filter_size, big_oc_step, stride>::impl; \
        kern_small_oc_remain = KerNeonXXs2NchwNchw44<                         \
                bias_mode, Op, step, filter_size, oc_step, stride>::impl;     \
        break;
            UNROLL_CALL_RAW(8, cb);
            default:
                megdnn_assert(0, "no remain %zu for kern", ow_remain);
        }
        for (size_t oc_idx = 0; oc_idx < oc_end; oc_idx += big_oc_step) {
            const size_t weight_offset = oc_idx * ic * fh * fw;
            for (size_t oh_idx = 0; oh_idx < oh; oh_idx += oh_step) {
                for (size_t ow_idx = 0; ow_idx < ow_end; ow_idx += ow_step) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_idx * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_idx) * oc_step;
                    KerNeonXXs2NchwNchw44<
                            bias_mode, Op, ow_step, filter_size, big_oc_step, stride>::
                            impl(src + src_offset, filter + weight_offset,
                                 bias + oc_idx, dst + dst_offset, ic, ih, iw, ld_dst_oc,
                                 op);
                }
                if (ow_remain > 0) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_end * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_end) * oc_step;
                    kern_big_oc_remain(
                            src + src_offset, filter + weight_offset, bias + oc_idx,
                            dst + dst_offset, ic, ih, iw, ld_dst_oc, op);
                }
            }
        }
        if (oc_remain > 0) {
            size_t oc_idx = oc_end;
            const size_t weight_offset = oc_idx * ic * fh * fw;
            for (size_t oh_idx = 0; oh_idx < oh; oh_idx += oh_step) {
                for (size_t ow_idx = 0; ow_idx < ow_end; ow_idx += ow_step) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_idx * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_idx) * oc_step;
                    KerNeonXXs2NchwNchw44<
                            bias_mode, Op, ow_step, filter_size, oc_step, stride>::
                            impl(src + src_offset, filter + weight_offset,
                                 bias + oc_idx, dst + dst_offset, ic, ih, iw, ld_dst_oc,
                                 op);
                }
                if (ow_remain > 0) {
                    const size_t src_offset =
                            (oh_idx * stride_h * iw + ow_end * stride_w * ih_step) *
                            ic_step * pack_iw_len;
                    const size_t dst_offset =
                            oc_idx * img_stride + (oh_idx * ow + ow_end) * oc_step;
                    kern_small_oc_remain(
                            src + src_offset, filter + weight_offset, bias + oc_idx,
                            dst + dst_offset, ic, ih, iw, ld_dst_oc, op);
                }
            }
        }
    }
 };
 #define INSTANCE_CONV_KERN_FUN(stride, filter_size, bias_mode, Op) \
    template struct megdnn::arm_common::int8_direct_nchw_nchw44::  \
            ConvDiectStrideInt8NchwNchw44<bias_mode, Op, filter_size, stride>;
 #define INSTANCE_OP_PARAM(stride, filter, bias_mode)                               \
    INSTANCE_CONV_KERN_FUN(                                                        \
            stride, filter, bias_mode, TypeCvtOp<dt_qint32 MEGDNN_COMMA dt_qint8>) \
    INSTANCE_CONV_KERN_FUN(                                                        \
            stride, filter, bias_mode, ReluOp<dt_qint32 MEGDNN_COMMA dt_qint8>)    \
    INSTANCE_CONV_KERN_FUN(                                                        \
            stride, filter, bias_mode, HSwishOp<dt_qint32 MEGDNN_COMMA dt_qint8>)
 #define INSTANCE_BIAS_MODE_PARAM(stride, filter)         \
    INSTANCE_OP_PARAM(stride, filter, BiasMode::NO_BIAS) \
    INSTANCE_OP_PARAM(stride, filter, BiasMode::BROADCAST_CHANNEL_BIAS)
 #define INSTANCE_CONV_KERN(stride, filter) INSTANCE_BIAS_MODE_PARAM(stride, filter)
 }  // namespace int8_direct_nchw_nchw44
 }  // namespace arm_common
 }  // namespace megdnn
 // vim: syntax=cpp.doxygen
--- a/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1_1x1.cpp
+++ b/dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1_1x1.cpp
@@ -0,0 +1,19 @@
 /**
 * \file
 * dnn/src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1_1x1.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
 #include "src/arm_common/conv_bias/int8/direct_kernels/int8_direct_nchw_nchw44_s1.h"
 using namespace megdnn;
 using namespace arm_common;
 INSTANCE_CONV_KERN(1, 1);
 // vim: syntax=cpp.doxygen