feat(cmake/cross_build/host_build/windows): imp windows

host build and cross build

now cmake status:
a: host build
    1: windows build -- ok
    2: linux build -- ok
    3: macos build -- ok
b: cross build
    1: windows cross build arm-android -- ok
    2: windows cross build arm-linux   -- ok
    3: linux cross build arm-android -- ok
    4: linux cross build arm-linux   -- ok
    5: macos cross build arm-android -- ok
    6: macos cross build arm-linux   -- ok
    7: macos cross build ios         -- ok
GitOrigin-RevId: f7f376fe8c522410568c304a62310df823c70b65

CMakeLists.txt

@@ -7,7 +7,7 @@ set(CMAKE_CXX_EXTENSIONS OFF)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake/Modules)
 
-if(NOT MSVC AND NOT APPLE)
+if(NOT MSVC AND NOT APPLE AND NOT WIN32)
     set(CMAKE_CXX_ARCHIVE_CREATE "<CMAKE_AR> Dqc <TARGET> <LINK_FLAGS> <OBJECTS>")
     set(CMAKE_CXX_ARCHIVE_APPEND "<CMAKE_AR> Dq  <TARGET> <LINK_FLAGS> <OBJECTS>")
     set(CMAKE_CXX_ARCHIVE_FINISH "<CMAKE_RANLIB> -D <TARGET>")
@@ -72,7 +72,7 @@ if(CMAKE_TOOLCHAIN_FILE)
 endif()
 
 if(${MGE_ARCH} STREQUAL "AUTO")
-    if(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "x86_64")
+    if(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "x86_64" OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL "AMD64")
         set(MGE_ARCH "x86_64")
     elseif(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "i386" OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL "i686")
         set(MGE_ARCH "i386")
@@ -85,8 +85,42 @@ if(${MGE_ARCH} STREQUAL "AUTO")
     endif()
 endif()
 
+if(MSVC OR WIN32)
+    add_compile_definitions(NOMINMAX=1 _USE_MATH_DEFINES=1 WIN32=1)
+    message("-- into windows build...")
+    message(" -- CMAKE_C_COMPILER_ID: ${CMAKE_C_COMPILER_ID}")
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang-cl")
+        message(FATAL_ERROR "only support clang-cl for windows build, pls check detail: scripts/cmake-build/BUILD_README.md")
+    endif()
+    # add flags for sse/avx for X86
+    set(WIN_FLAGS "-msse4.2 -mavx -mavx2")
+    # if u CPU is cascadelake series, u can enable for performance
+    # set(WIN_FLAGS "{WIN_FLAGS} -march=cascadelake -mtune=cascadelake")
+    # set(WIN_FLAGS "{WIN_FLAGS} -mavx512cd -mavx512vl -mavx512dq -mavx512bw -mavx512vbmi -mavx512vnni")
+
+    # for windows build
+    set(WIN_FLAGS "${WIN_FLAGS} -Wno-error=implicit-int-conversion -Wno-error=double-promotion")
+    set(WIN_FLAGS "${WIN_FLAGS} -Wno-error=zero-as-null-pointer-constant -Wno-error=implicit-int-conversion")
+    set(WIN_FLAGS "${WIN_FLAGS} -Wno-error=float-conversion -Wno-error=shadow-field -Wno-error=covered-switch-default")
+    set(WIN_FLAGS "${WIN_FLAGS} -Wno-error=deprecated  -Wno-error=documentation  -Wno-error=unreachable-code-break")
+    set(WIN_FLAGS "${WIN_FLAGS} /DWIN32 -Wno-macro-redefined /D_WIN32_WINNT=0x0601")
+    set(WIN_FLAGS "${WIN_FLAGS} /D_CRT_SECURE_NO_DEPRECATE /D_CRT_SECURE_NO_WARNINGS /DNOGDI /D_USE_MATH_DEFINES /bigobj")
+    set(WIN_FLAGS "${WIN_FLAGS} /Zm500 /EHs /wd4351 /wd4291 /wd4250 /wd4996 /wd4819 -Wno-inconsistent-dllimport")
+
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${WIN_FLAGS}")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${WIN_FLAGS}")
+else()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra")
+    set(CMAKE_CXX_FLAGS_DEBUG "-O0 -g")
+    if(ANDROID)
+        set(CMAKE_CXX_FLAGS_RELEASE "-Ofast -DNDEBUG")
+    else()
+        set(CMAKE_CXX_FLAGS_RELEASE "-O3 -DNDEBUG")
+    endif()
+endif()
+
 CHECK_CXX_COMPILER_FLAG(-fuse-ld=gold CXX_SUPPORT_GOLD)
-if(CXX_SUPPORT_GOLD AND NOT ANDROID AND NOT APPLE)
+if(CXX_SUPPORT_GOLD AND NOT ANDROID AND NOT APPLE AND NOT MSVC AND NOT WIN32)
     message("-- Using GNU gold linker.")
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fuse-ld=gold")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fuse-ld=gold")
@@ -188,18 +222,6 @@ if(MGB_WITH_FLATBUFFERS)
     include(cmake/flatbuffers.cmake)
 endif()
 
-if(MSVC)
-    add_compile_definitions(NOMINMAX=1 _USE_MATH_DEFINES=1 WIN32=1)
-else()
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra")
-    set(CMAKE_CXX_FLAGS_DEBUG "-O0 -g")
-    if(ANDROID)
-        set(CMAKE_CXX_FLAGS_RELEASE "-Ofast -DNDEBUG")
-    else()
-        set(CMAKE_CXX_FLAGS_RELEASE "-O3 -DNDEBUG")
-    endif()
-endif()
-
 if(MGE_WITH_CUDA)
     include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
     foreach(path ${CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES})
@@ -393,6 +415,12 @@ else()
     set(MGB_ENABLE_DEBUG_UTIL 0)
 endif()
 
+# FIXME: remove this after imp DEBUG UTIL for windows
+if(MSVC OR WIN32)
+    set(MGB_ENABLE_DEBUG_UTIL 0)
+    message(" -- disable MGB_ENABLE_DEBUG_UTIL in windows build")
+endif()
+
 # TensorRT
 set(MGB_ENABLE_TENSOR_RT ${MGE_WITH_TRT})
 

cmake/MKL_DNN.cmake

-include(ExternalProject)
-include(GNUInstallDirs)
-
-set(MKLDNN_DIR "${PROJECT_SOURCE_DIR}/third_party/intel-mkl-dnn" CACHE STRING "mkldnn directory")
-set(MKLDNN_BUILD_DIR ${PROJECT_BINARY_DIR}/third_party/intel-mkl-dnn)
-set(MKLDNN_LIB ${MKLDNN_BUILD_DIR}/${CMAKE_INSTALL_LIBDIR}/libdnnl.a)
-
+option(DNNL_BUILD_TESTS "" OFF)
+option(DNNL_BUILD_EXAMPLES "" OFF)
+set(DNNL_LIBRARY_TYPE STATIC CACHE STRING "config dnnl to STATIC")
+# 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 DNNL_RUNTIME_SEQ CACHE STRING "config dnnl to DNNL_RUNTIME_SEQ")
 if(MGE_BLAS STREQUAL "MKL")
-    list(APPEND MKLDNN_BUILD_ARGS -D_DNNL_USE_MKL=ON -DMKLROOT=${MKL_ROOT_DIR})
+    option(_DNNL_USE_MKL "" ON)
+    set(MKLROOT ${MKL_ROOT_DIR} CACHE STRING "MKL ROOT FOR DNNL")
+    if (WIN32)
+        set(MKLLIB "mkl_core" CACHE STRING "MKLLIB NAME FOR DNNL")
+    else()
+        set(MKLLIB "libmkl_core.a" CACHE STRING "MKLLIB NAME FOR DNNL")
+    endif()
+    # workround for DNNL link failed, we do not want to modify
+    # dnnl MKL.cmake of intel-mkl-dnn when include intel-mkl-dnn
+    # via add_subdirectory api
+    link_directories(${MKL_ROOT_DIR}/lib)
 else()
-    list(APPEND MKLDNN_BUILD_ARGS -D_DNNL_USE_MKL=OFF)
+    option(_DNNL_USE_MKL "" OFF)
 endif()
 
-ExternalProject_add(
-    mkl_dnn
-    SOURCE_DIR ${MKLDNN_DIR}
-    PREFIX ${MKLDNN_BUILD_DIR}
-    CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER} -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER} -DDNNL_BUILD_TESTS=OFF -DDNNL_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} -DCMAKE_INSTALL_PREFIX=${MKLDNN_BUILD_DIR} -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} -DDNNL_LIBRARY_TYPE=STATIC -DDNNL_CPU_RUNTIME=DNNL_RUNTIME_SEQ ${MKLDNN_BUILD_ARGS}
-    BUILD_BYPRODUCTS ${MKLDNN_LIB}
-)
-
-set(MKLDNN_INC ${MKLDNN_BUILD_DIR}/include)
-file(MAKE_DIRECTORY ${MKLDNN_INC})
-
-add_library(libmkl_dnn STATIC IMPORTED GLOBAL)
-add_dependencies(libmkl_dnn mkl_dnn)
-set_target_properties(
-    libmkl_dnn PROPERTIES
-    IMPORTED_LOCATION ${MKLDNN_LIB}
-    INTERFACE_INCLUDE_DIRECTORIES ${MKLDNN_INC}
-)
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-unused-parameter -Wno-extra")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-parameter -Wno-extra")
+add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/intel-mkl-dnn)

cmake/flatbuffers.cmake

@@ -2,8 +2,24 @@ if (MGE_USE_SYSTEM_LIB)
     find_package(FlatBuffers REQUIRED)
     return()
 endif()
+if(MSVC OR WIN32)
+    message("-- 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 "${FLATC_FLAGS} -Wno-error=switch-enum -Wno-error=missing-noreturn -Wno-error=float-equal")
+
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${FLATC_FLAGS}")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${FLATC_FLAGS}")
+endif()
 
 option(FLATBUFFERS_BUILD_TESTS "" OFF)
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/flatbuffers
                  ${CMAKE_CURRENT_BINARY_DIR}/flatbuffers
-                 EXCLUDE_FROM_ALL)
+                 EXCLUDE_FROM_ALL)
\ No newline at end of file

cmake/mkl.cmake

@@ -2,6 +2,8 @@ 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
     $ENV{MKLDIR}
     /opt/intel/mkl/*/
     /opt/intel/cmkl/*/
@@ -23,20 +25,20 @@ find_path(MKL_INCLUDE_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 libmkl_core.lib
+        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 libmkl_sequential.lib
+        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 libmkl_intel_ilp64.lib
+            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 "x86_32")
+    elseif(${MGE_ARCH} STREQUAL "i386")
         find_library(MKL_IPL_LIBRARY
-            NAMES libmkl_intel_32.a libmkl_intel_32.lib
+            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()
 

dnn/CMakeLists.txt

@@ -11,7 +11,7 @@ add_custom_command(
     COMMAND ${PYTHON_EXECUTABLE} ${OPR_PARAM_DEFS_SCRIPT} ${OPR_PARAM_DEFS_SRCS}
         ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_defs.h
     COMMAND ${PYTHON_EXECUTABLE} ${OPR_PARAM_DEFS_SCRIPT} ${OPR_PARAM_DEFS_SRCS}
-        /dev/null --write-cppjson ${OPR_PARAM_DEFS_OUT_DIR}/megdnn/opr_param_json.h
+       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
 )

dnn/include/megdnn/dtype/half_common_prologue.h

@@ -189,6 +189,8 @@
 #include <climits>
 #include <cmath>
 #include <cstring>
+#include <ostream>
+#include <istream>
 #if HALF_ENABLE_CPP11_TYPE_TRAITS
 	#include <type_traits>
 #endif

dnn/src/CMakeLists.txt

@@ -67,7 +67,7 @@ endif()
 target_compile_definitions(megdnn INTERFACE ${LIBMEGDNN_DEF})
 
 if(MGE_WITH_MKLDNN AND ${MGE_ARCH} STREQUAL "x86_64")
-    target_link_libraries(megdnn libmkl_dnn)
+    target_link_libraries(megdnn dnnl)
 endif()
 target_link_libraries(megdnn ${MGE_CUDA_LIBS})
 target_link_libraries(megdnn ${MGE_BLAS_LIBS})

dnn/src/common/cv/aligned_allocator.h

@@ -18,14 +18,15 @@
 
 #include "megdnn/arch.h"
 
-#ifdef _MSC_VER
+#if defined(_MSC_VER) || defined(__MINGW32__)
 #include "malloc.h"
 #endif
 
+
 #if defined(__ANDROID__) || defined(ANDROID)
 #include "malloc.h"
 #define HAS_MEMALIGN
-#elif !defined(_MSC_VER)
+#elif !defined(_MSC_VER) && !defined(__MINGW32__)
 #define HAS_POSIX_MEMALIGN
 #endif
 
@@ -77,7 +78,7 @@ public:
         return result;
 #elif defined(HAS_MEMALIGN)
         return (_Tp*)memalign(_align, __n * sizeof(_Tp));
-#elif defined(_MSC_VER)
+#elif defined(_MSC_VER) || defined(__MINGW32__)
         return (_Tp*)_aligned_malloc(__n * sizeof(_Tp), _align);
 #else
 #warning \

dnn/src/x86/add_update/opr_impl.cpp

@@ -17,7 +17,7 @@
 #include "src/x86/utils.h"
 
 #include <immintrin.h>
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avxintrin.h>
 #include <fmaintrin.h>
 #endif

dnn/src/x86/conv_bias/f32/strategy_2x3_8x8.cpp

@@ -18,7 +18,7 @@
 #include "src/x86/avx_helper.h"
 
 #include <x86intrin.h>
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avxintrin.h>
 #include <smmintrin.h>
 #include <avx2intrin.h>

dnn/src/x86/conv_bias/f32/strategy_6x3_8x8.cpp

@@ -19,7 +19,7 @@
 #include "src/x86/avx_helper.h"
 
 #include <x86intrin.h>
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avx2intrin.h>
 #include <avxintrin.h>
 #include <fmaintrin.h>

dnn/src/x86/conv_bias/int8/avx2_chanwise_kern.cpp

@@ -15,7 +15,7 @@
 #include "src/common/unroll_macro.h"
 #include "src/x86/conv_bias/int8/common_helper.h"
 #include "src/x86/elemwise_op.h"
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <smmintrin.h>
 #endif
 

dnn/src/x86/conv_bias/int8/common_helper.h

@@ -13,7 +13,7 @@
 #include <immintrin.h>
 #include "src/common/unroll_macro.h"
 #include "megdnn/arch.h"
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <smmintrin.h>
 #endif
 

dnn/src/x86/elemwise/avx_util/avx_mathfun.h

@@ -12,7 +12,7 @@
 #include "megdnn/arch.h"
 #include "megdnn/basic_types.h"
 #include <immintrin.h>
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avxintrin.h>
 #include <smmintrin.h>
 #include <avx2intrin.h>

dnn/src/x86/elemwise/fma_util/fma_util.cpp

@@ -9,7 +9,7 @@
  * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  */
 #include <immintrin.h>
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avxintrin.h>
 #include <fmaintrin.h>
 #endif

dnn/src/x86/elemwise_helper/kimpl/typecvt.h

@@ -11,7 +11,7 @@
 #pragma once
 
 #include <immintrin.h>
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avx2intrin.h>
 #include <avxintrin.h>
 #include <fmaintrin.h>

dnn/src/x86/matrix_mul/common/common.h

@@ -11,7 +11,7 @@
 #pragma once
 #include <x86intrin.h>
 
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avxintrin.h>
 #include <smmintrin.h>
 #include <avx2intrin.h>

dnn/src/x86/quantized_converter.h

@@ -11,7 +11,7 @@
 #pragma once
 
 #include <immintrin.h>
-#ifdef WIN32CMAKE
+#ifdef WIN32
 #include <avx2intrin.h>
 #include <avxintrin.h>
 #include <fmaintrin.h>

scripts/cmake-build/BUILD_README.md

+# build support status
+##  host build
+*  windows build (ok)
+*  linux build   (ok)
+*  macos build   (ok)
+##  cross build
+*  windows cross build arm-android (ok)
+*  windows cross build arm-linux   (ok)
+*  linux cross build arm-android   (ok)
+*  linux cross build arm-linux     (ok)
+*  macos cross build arm-android   (ok)
+*  macos cross build arm-linux     (ok but experimental)
+*  macos cross build ios           (ok)
+
+# build env prepare
+## package install
+### windows host build
+    ```
+    1: installl Visual Studio (need support LLVM/clang-cl), eg 2019
+    clang-cl 9 linker have crash issue, pls install 7/8/10
+    2: install extension of VS: python/cmake/LLVM
+    3: CUDA env(if enable CUDA), version detail: project_root_dir/README.md
+    ```
+### linux host build
+    ```
+    1: cmake, which version > 3.14.4
+    2: gcc/g++, which version > 6
+    3: install build-essential git git-lfs gfortran libgfortran-6-dev autoconf gnupg flex bison gperf curl 
+    4: zlib1g-dev gcc-multilib g++-multilib lib32ncurses5-dev libxml2-utils xsltproc unzip libtool:
+    5: librdmacm-dev rdmacm-utils python3-dev swig python3-numpy texinfo
+    6: CUDA env(if enable CUDA), version detail: project_root_dir/README.md
+    ```
+### macos host build
+    ```
+    1: cmake, which version > 3.14.4
+    2: install brew: /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install.sh)"
+    3: brew install python python3 swig coreutils
+    4: install at least xcode command line tool: https://developer.apple.com/xcode/
+    5: about cuda: we do not support CUDA on macos
+    ```
+### cross build for arm-android
+    now we support windows/linux/macos cross build to arm-android
+    ```
+    1: install unix-like tools, eg MSYS if you are using windows(recommend)
+    we also support CMD.exe or powershell on windows
+    1: download NDK from https://developer.android.google.cn/ndk/downloads/
+    for diff OS platform package, suggested NDK20 or NDK21
+    2: export NDK_ROOT=NDK_DIR at bash-like env
+    3: config NDK_ROOT to PATH env at windows control board if use CMD/powershell
+    ```
+### cross build for arm-linux
+    now we support arm-linux on linux and windows fully, also experimental on MACOS
+    ```
+    1: download toolchains on https://releases.linaro.org/components/toolchain/gcc-linaro/
+    or https://developer.arm.com/tools-and-software/open-source-software/developer-tools/gnu-toolchain/gnu-a/downloads
+    if use windows or linux
+    2: download https://github.com/thinkski/osx-arm-linux-toolchains if use MACOS
+    ```
+### cross build for ios
+    now we only support cross build to ios on MACOS
+    ```
+    1: install full xcode: https://developer.apple.com/xcode/
+    ```
+## third-party code prepare
+### with bash env(linux/macos/unix-like tools on windows, eg: msys etc)
+    ```
+    ./third_party/prepare.sh
+    ./third_party/install-mkl.sh
+    ```
+### windows shell env(eg: cmd, powershell etc)
+    infact if you can use git command on windows, which means you always install
+    bash.exe at the same dir of git.exe, find it, then you can prepare third-party
+    code by command:
+    ```
+    bash.exe ./third_party/prepare.sh
+    bash.exe ./third_party/install-mkl.sh
+    ```
+# how to build
+## with bash env(linux/macos/unix-like tools on windows, eg: msys etc)
+    ```
+    1: host build just use scripts:scripts/cmake-build/host_build.sh
+    2: cross build to arm-android: scripts/cmake-build/cross_build_android_arm_inference.sh
+    3: cross build to arm-linux:   scripts/cmake-build/cross_build_linux_arm_inference.sh
+    4: cross build to ios:         scripts/cmake-build/cross_build_ios_arm_inference.sh
+    ```
+## windows shell env(eg: cmd, powershell etc)
+    ```
+    1: we do not provide BAT for cmd/powershlel scripts, BUT u can refs for scripts/cmake-build/*.sh
+    ```
+## Visual Studio GUI(only for windows host)
+    ```
+    1: import megengine src to Visual Studio as a project
+    2: right click CMakeLists.txt, choose config 'cmake config'
+       choose clang_cl_x86 or clang_cl_x64
+    3: config other CMAKE config, eg, CUDA ON OR OFF
+    ```
+
+
+# other arm-linux-like board support
+it`s easy to support other customized arm-linux-like board, example:
+1: HISI 3516/3519, infact u can just use toolchains from arm developer or linaro
+then call scripts/cmake-build/cross_build_linux_arm_inference.sh to build a ELF
+binary, or if you get HISI official toolschain, you just need modify CMAKE_CXX_COMPILER
+and CMAKE_C_COMPILER in toolchains/arm-linux-gnueabi* to a real name
+
+2: about Raspberry, just use scripts/cmake-build/cross_build_linux_arm_inference.sh

scripts/cmake-build/cross_build_android_arm_inference.sh

@@ -76,13 +76,18 @@ echo "ARCH: $ARCH"
 echo "----------------------------------------------------"
 
 READLINK=readlink
+MAKEFILE_TYPE="Unix"
 OS=$(uname -s)
 
 if [ $OS = "Darwin" ];then
     READLINK=greadlink
+elif [[ $OS =~ "NT" ]]; then
+    echo "BUILD in NT ..."
+    MAKEFILE_TYPE="Unix"
 fi
 
 SRC_DIR=$($READLINK -f "`dirname $0`/../../")
+source $SRC_DIR/scripts/cmake-build/utils/host_build_flatc.sh
 
 if [ -z $NDK_ROOT ];then
     echo "can not find NDK_ROOT env, pls export you NDK root dir to NDK_ROOT"
@@ -99,6 +104,7 @@ function cmake_build() {
     echo "build type: $BUILD_TYPE"
     echo "build ABI: $BUILD_ABI"
     echo "build native level: $BUILD_NATIVE_LEVEL"
+    echo "BUILD MAKEFILE_TYPE: $MAKEFILE_TYPE"
     if [ -e $BUILD_DIR ];then
         echo "clean old dir: $BUILD_DIR"
         rm -rf $BUILD_DIR
@@ -112,7 +118,8 @@ function cmake_build() {
     mkdir -p $BUILD_DIR
     mkdir -p $INSTALL_DIR
     cd $BUILD_DIR
-    cmake -DCMAKE_TOOLCHAIN_FILE="$NDK_ROOT/build/cmake/android.toolchain.cmake" \
+    cmake -G "$MAKEFILE_TYPE Makefiles" \
+        -DCMAKE_TOOLCHAIN_FILE="$NDK_ROOT/build/cmake/android.toolchain.cmake" \
         -DANDROID_NDK="$NDK_ROOT" \
         -DCMAKE_BUILD_TYPE=$BUILD_TYPE \
         -DANDROID_ABI=$BUILD_ABI \
@@ -129,34 +136,7 @@ function cmake_build() {
     make install/strip
 }
 
-function build_flatc() {
-    BUILD_DIR=$SRC_DIR/build_dir/host_flatc/build
-    INSTALL_DIR=$BUILD_DIR/../install
-    if [ -e $BUILD_DIR ];then
-        echo "clean old dir: $BUILD_DIR"
-        rm -rf $BUILD_DIR
-    fi
-    if [ -e $INSTALL_DIR ];then
-        echo "clean old dir: $INSTALL_DIR"
-        rm -rf $INSTALL_DIR
-    fi
-
-    echo "create build dir"
-    mkdir -p $BUILD_DIR
-    mkdir -p $INSTALL_DIR
-    cd $BUILD_DIR
-    cmake -DCMAKE_BUILD_TYPE=Release \
-        -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR \
-        -DFLATBUFFERS_BUILD_TESTS=OFF \
-        -DFLATBUFFERS_BUILD_FLATHASH=OFF \
-        -DFLATBUFFERS_BUILD_FLATLIB=OFF \
-        -DFLATBUFFERS_LIBCXX_WITH_CLANG=OFF \
-        $SRC_DIR/third_party/flatbuffers
-
-    make -j$(nproc)
-    make install/strip
-}
-build_flatc
+build_flatc $SRC_DIR
 
 api_level=16
 abi="armeabi-v7a with NEON"

scripts/cmake-build/cross_build_ios_arm_inference.sh

@@ -86,6 +86,7 @@ else
 fi
 
 SRC_DIR=$($READLINK -f "`dirname $0`/../../")
+source $SRC_DIR/scripts/cmake-build/utils/host_build_flatc.sh
 
 function cmake_build() {
     BUILD_DIR=$SRC_DIR/build_dir/apple/$3/$4/$1/$BUILD_TYPE/build
@@ -133,34 +134,7 @@ function cmake_build() {
     make install
 }
 
-function build_flatc() {
-    BUILD_DIR=$SRC_DIR/build_dir/host_flatc/build
-    INSTALL_DIR=$BUILD_DIR/../install
-    if [ -e $BUILD_DIR ];then
-        echo "clean old dir: $BUILD_DIR"
-        rm -rf $BUILD_DIR
-    fi
-    if [ -e $INSTALL_DIR ];then
-        echo "clean old dir: $INSTALL_DIR"
-        rm -rf $INSTALL_DIR
-    fi
-
-    echo "create build dir"
-    mkdir -p $BUILD_DIR
-    mkdir -p $INSTALL_DIR
-    cd $BUILD_DIR
-    cmake -DCMAKE_BUILD_TYPE=Release \
-        -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR \
-        -DFLATBUFFERS_BUILD_TESTS=OFF \
-        -DFLATBUFFERS_BUILD_FLATHASH=OFF \
-        -DFLATBUFFERS_BUILD_FLATLIB=OFF \
-        -DFLATBUFFERS_LIBCXX_WITH_CLANG=OFF \
-        $SRC_DIR/third_party/flatbuffers
-
-    make -j$(nproc)
-    make install/strip
-}
-build_flatc
+build_flatc $SRC_DIR
 
 # refs for ../../toolchains/ios.toolchain.cmake
 # to config this, if u want to build other,

scripts/cmake-build/cross_build_linux_arm_inference.sh

@@ -76,13 +76,19 @@ echo "ARCH: $ARCH"
 echo "----------------------------------------------------"
 
 READLINK=readlink
+MAKEFILE_TYPE="Unix"
 OS=$(uname -s)
 
 if [ $OS = "Darwin" ];then
     READLINK=greadlink
+elif [[ $OS =~ "NT" ]]; then
+    echo "BUILD in NT ..."
+    MAKEFILE_TYPE="Unix"
 fi
 
 SRC_DIR=$($READLINK -f "`dirname $0`/../../")
+source $SRC_DIR/scripts/cmake-build/utils/host_build_flatc.sh
+
 function cmake_build() {
     BUILD_DIR=$SRC_DIR/build_dir/gnu-linux/$1/$BUILD_TYPE/build
     INSTALL_DIR=$BUILD_DIR/../install
@@ -91,6 +97,7 @@ function cmake_build() {
     echo "install dir: $INSTALL_DIR"
     echo "build type: $BUILD_TYPE"
     echo "build toolchain: $TOOLCHAIN"
+    echo "BUILD MAKEFILE_TYPE: $MAKEFILE_TYPE"
     if [ -e $BUILD_DIR ];then
         echo "clean old dir: $BUILD_DIR"
         rm -rf $BUILD_DIR
@@ -104,7 +111,8 @@ function cmake_build() {
     mkdir -p $BUILD_DIR
     mkdir -p $INSTALL_DIR
     cd $BUILD_DIR
-    cmake -DCMAKE_TOOLCHAIN_FILE=$TOOLCHAIN \
+    cmake -G "$MAKEFILE_TYPE Makefiles" \
+        -DCMAKE_TOOLCHAIN_FILE=$TOOLCHAIN \
         -DCMAKE_BUILD_TYPE=$BUILD_TYPE \
         -DMGE_INFERENCE_ONLY=ON \
         -DMGE_WITH_CUDA=OFF \
@@ -118,34 +126,7 @@ function cmake_build() {
     make install/strip
 }
 
-function build_flatc() {
-    BUILD_DIR=$SRC_DIR/build_dir/host_flatc/build
-    INSTALL_DIR=$BUILD_DIR/../install
-    if [ -e $BUILD_DIR ];then
-        echo "clean old dir: $BUILD_DIR"
-        rm -rf $BUILD_DIR
-    fi
-    if [ -e $INSTALL_DIR ];then
-        echo "clean old dir: $INSTALL_DIR"
-        rm -rf $INSTALL_DIR
-    fi
-
-    echo "create build dir"
-    mkdir -p $BUILD_DIR
-    mkdir -p $INSTALL_DIR
-    cd $BUILD_DIR
-    cmake -DCMAKE_BUILD_TYPE=Release \
-        -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR \
-        -DFLATBUFFERS_BUILD_TESTS=OFF \
-        -DFLATBUFFERS_BUILD_FLATHASH=OFF \
-        -DFLATBUFFERS_BUILD_FLATLIB=OFF \
-        -DFLATBUFFERS_LIBCXX_WITH_CLANG=OFF \
-        $SRC_DIR/third_party/flatbuffers
-
-    make -j$(nproc)
-    make install/strip
-}
-build_flatc
+build_flatc $SRC_DIR
 
 toolchain=null
 if [ "$ARCH" = "arm64-v8a" ]; then

scripts/cmake-build/host_build.sh

@@ -7,6 +7,8 @@ function usage() {
     echo "-d : Build with Debug mode, defaule Release mode"
     echo "-c : Build with CUDA, default without CUDA"
     echo "-t : Build with training mode, default inference only"
+    echo "-m : Build with m32 mode(only for windows build), default m64"
+    echo "-h : show usage"
     echo "example: $0 -d"
     exit -1
 }
@@ -14,8 +16,11 @@ function usage() {
 BUILD_TYPE=Release
 MGE_WITH_CUDA=OFF
 MGE_INFERENCE_ONLY=ON
+MGE_WINDOWS_BUILD_ARCH=x64
+MGE_WINDOWS_BUILD_MARCH=m64
+MGE_ARCH=x86_64
 
-while getopts "dct" arg
+while getopts "hdctm" arg
 do
     case $arg in
         d)
@@ -30,6 +35,16 @@ do
             echo "Build with training mode"
             MGE_INFERENCE_ONLY=OFF
             ;;
+        h)
+            echo "show usage"
+            usage
+            ;;
+        m)
+            echo "build for m32(only use for windows)"
+            MGE_WINDOWS_BUILD_ARCH=x86
+            MGE_WINDOWS_BUILD_MARCH=m32
+            MGE_ARCH=i386
+            ;;
         ?)
             echo "unkonw argument"
             usage
@@ -51,6 +66,8 @@ if [ $OS = "Darwin" ];then
         echo "MACOS DO NOT SUPPORT TensorRT, ABORT NOW!!"
         exit -1
     fi
+elif [[ $OS =~ "NT" ]]; then
+    echo "BUILD in NT ..."
 fi
 
 SRC_DIR=$($READLINK -f "`dirname $0`/../../")
@@ -88,7 +105,111 @@ function cmake_build() {
 
     make -j$(nproc)
     make install/strip
-    }
+}
+
+function windows_env_err() {
+    echo "check windows env failed!!"
+    echo "please install LLVM/clang-cl/cmake/python at Visual Studio Extensions"
+    exit -1
+}
+
+function prepare_env_for_windows_build() {
+    echo "check Visual Studio install path env..."
+    if [[ -z $VS_PATH ]];then
+        echo "can not find visual_studio_path env, pls export you Visual Studio install dir to VS_PATH"
+        echo "examle for export Visual Studio 2019 Enterprise default install dir"
+        echo "export VS_PATH=/c/Program\ Files\ \(x86\)/Microsoft\ Visual\ Studio/2019/Enterprise"
+        exit -1
+    fi
+    echo $VS_PATH
 
-cmake_build $MGE_WITH_CUDA $MGE_INFERENCE_ONLY $BUILD_TYPE
+    # only use cmake/clang-cl/Ninja install from Visual Studio, if not, may build failed
+    # some user env may install cmake/clang-cl/Ninja at MSYS env, so we put Visual Studio
+    # path at the head of PATH, and check the valid
+    echo "check cmake install..."
+    export PATH=$VS_PATH/Common7/IDE/CommonExtensions/Microsoft/CMake/CMake/bin/:$PATH
+    which cmake
+    cmake_loc=`which cmake`
+    if [[ $cmake_loc =~ "Visual" ]]; then
+        echo "cmake valid ..."
+    else
+        echo "cmake Invalid: ..."
+        windows_env_err
+    fi
 
+    echo "check clang-cl install..."
+    export PATH=$VS_PATH/VC/Tools/Llvm/bin/:$PATH
+    which clang-cl
+    clang_loc=`which clang-cl`
+    if [[ $clang_loc =~ "Visual" ]]; then
+        echo "clang-cl valid ..."
+    else
+        echo "clang-cl Invalid: ..."
+        windows_env_err
+    fi
+
+    echo "check Ninja install..."
+    export PATH=$VS_PATH/Common7/IDE/CommonExtensions/Microsoft/CMake/Ninja/:$PATH
+    which Ninja
+    ninja_loc=`which Ninja`
+    if [[ $ninja_loc =~ "Visual" ]]; then
+        echo "Ninja valid ..."
+    else
+        echo "Ninja Invalid: ..."
+        windows_env_er
+    fi
+
+    export PATH=$VS_PATH/VC/Auxiliary/Build:$PATH
+    echo "put vcvarsall.bat path to PATH env.."
+}
+
+function cmake_build_windows() {
+    # windows do not support long path, so we cache the BUILD_DIR ASAP
+    prepare_env_for_windows_build
+    BUILD_DIR=$SRC_DIR/build_dir/host/build
+    INSTALL_DIR=$BUILD_DIR/../install
+    MGE_WITH_CUDA=$1
+    MGE_INFERENCE_ONLY=$2
+    BUILD_TYPE=$3
+    echo "build dir: $BUILD_DIR"
+    echo "install dir: $INSTALL_DIR"
+    echo "build type: $BUILD_TYPE"
+    echo "MGE_WITH_CUDA: $MGE_WITH_CUDA"
+    echo "MGE_INFERENCE_ONLY: $MGE_INFERENCE_ONLY"
+    if [ -e $BUILD_DIR ];then
+        echo "clean old dir: $BUILD_DIR"
+        rm -rf $BUILD_DIR
+    fi
+    if [ -e $INSTALL_DIR ];then
+        echo "clean old dir: $INSTALL_DIR"
+        rm -rf $INSTALL_DIR
+    fi
+
+    echo "create build dir"
+    mkdir -p $BUILD_DIR
+    mkdir -p $INSTALL_DIR
+    cd $BUILD_DIR
+    echo "now try build windows native with cmake/clang-ci/Ninja/Visual Studio ....."
+    export CFLAGS=-$MGE_WINDOWS_BUILD_MARCH
+    export CXXFLAGS=-$MGE_WINDOWS_BUILD_MARCH
+    cmd.exe /c " \
+        vcvarsall.bat $MGE_WINDOWS_BUILD_ARCH && cmake  -G "Ninja" \
+        -DMGE_ARCH=$MGE_ARCH \
+        -DMGE_INFERENCE_ONLY=$MGE_INFERENCE_ONLY \
+        -DMGE_WITH_CUDA=$MGE_WITH_CUDA \
+        -DCMAKE_BUILD_TYPE=$CMAKE_BUILD_TYPE \
+        -DCMAKE_INSTALL_PREFIX:PATH=$INSTALL_DIR  \
+        -DCMAKE_C_COMPILER=clang-cl.exe \
+        -DCMAKE_CXX_COMPILER=clang-cl.exe \
+        -DCMAKE_MAKE_PROGRAM=ninja.exe \
+        ../../.. && \
+        echo \"start Ninja build log to build.log, may take serval min...\" && \
+        Ninja load_and_run > build.log"
+}
+
+
+if [[ $OS =~ "NT" ]]; then
+    cmake_build_windows $MGE_WITH_CUDA $MGE_INFERENCE_ONLY $BUILD_TYPE
+else
+    cmake_build $MGE_WITH_CUDA $MGE_INFERENCE_ONLY $BUILD_TYPE
+fi

scripts/cmake-build/utils/host_build_flatc.sh

+#!/usr/bin/env bash
+set -e
+
+MAKEFILE_TYPE="Unix"
+OS=$(uname -s)
+
+if [[ $OS =~ "NT" ]]; then
+    echo "BUILD in NT ..."
+    MAKEFILE_TYPE="Unix"
+fi
+
+function build_flatc() {
+    BUILD_DIR=$1/build_dir/host_flatc/build
+    INSTALL_DIR=$BUILD_DIR/../install
+    if [ -e $BUILD_DIR ];then
+        echo "clean old dir: $BUILD_DIR"
+        rm -rf $BUILD_DIR
+    fi
+    if [ -e $INSTALL_DIR ];then
+        echo "clean old dir: $INSTALL_DIR"
+        rm -rf $INSTALL_DIR
+    fi
+
+    echo "create build dir"
+    mkdir -p $BUILD_DIR
+    mkdir -p $INSTALL_DIR
+    cd $BUILD_DIR
+    cmake -G "$MAKEFILE_TYPE Makefiles" \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR \
+        -DFLATBUFFERS_BUILD_TESTS=OFF \
+        -DFLATBUFFERS_BUILD_FLATHASH=OFF \
+        -DFLATBUFFERS_BUILD_FLATLIB=OFF \
+        -DFLATBUFFERS_LIBCXX_WITH_CLANG=OFF \
+        $SRC_DIR/third_party/flatbuffers
+
+    make -j$(nproc)
+    make install/strip
+}

src/CMakeLists.txt

@@ -106,7 +106,7 @@ if (MGB_WITH_FLATBUFFERS)
     target_compile_definitions(megbrain PUBLIC MGB_ENABLE_FBS_SERIALIZATION=1)
     target_link_libraries(megbrain flatbuffers)
     
-    set (GENERATED_FLATBUFFERS_CONVERTER_PATH ${CMAKE_CURRENT_BINARY_DIR}/genfiles/this/should/be/added/to/sereg_caller/target/only/but/hey/yolo/there/are/no/fine/grained/targets/let/us/just/make/this/path/unreasonably/long/to/avoid/collision)
+    set (GENERATED_FLATBUFFERS_CONVERTER_PATH ${CMAKE_CURRENT_BINARY_DIR}/genfiles)
     set (GEN_FLATBUFFERS_CONVERTER_PY ${PROJECT_SOURCE_DIR}/dnn/scripts/gen_flatbuffers_converter.py)
     file (MAKE_DIRECTORY ${GENERATED_FLATBUFFERS_CONVERTER_PATH})
     add_custom_command(

test/src/include/megbrain/test/helper.h

@@ -17,7 +17,7 @@
 
 #include <iostream>
 
-#ifdef _WIN32
+#if defined(WIN32)
 static inline void unsetenv(std::string name) {
     name += "=";
     _putenv(name.c_str());

third_party/install-mkl.sh

@@ -2,19 +2,50 @@
 
 cd $(dirname $0)
 
+#FIXME: anaconda just upload serval lastest version, so this version may lose efficacy
+echo "this script only for linux/macos/windows-unix-like-env(MSYS etc) prepare MKL env"
+echo "if you build windows for native at cmd.exe, powershell env or Visual Studio GUI,"
+echo  "u need download MKL package and untar manually"
+echo "refs: https://software.intel.com/content/www/us/en/develop/tools/math-kernel-library/choose-download/windows.html"
 MKL_VERSION=2019.5
 MKL_PATCH=281
 CONDA_BASE_URL=https://anaconda.org/intel
+OS=$(uname -s)
+FILE_PREFIX=null
+TAR=tar
+
+if [ $OS = "Darwin" ];then
+    FILE_PREFIX=osx
+elif [ $OS = "Linux" ];then
+    FILE_PREFIX=linux
+elif [[ $OS =~ "NT" ]]; then
+    FILE_PREFIX=win
+    # NT use /c/Windows/system32/tar will stuck for big file
+    # so we back to GNU tar
+    TAR=/usr/bin/tar
+else
+    echo "DO NOT SUPPORT OS NOW"
+    exit -1
+fi
+
+echo "config FILE_PREFIX to: $FILE_PREFIX"
 
 rm -rf mkl
 
 for platform in 32 64
 do
+    if [ $OS = "Darwin" ]&&[ $platform = 32 ];then
+        echo "strip 32 bit file for Darwin"
+        continue
+    fi
     mkdir -p mkl/x86_${platform}
     for package in "mkl-include" "mkl-static"
     do
-        echo "Installing ${package} for x86_${platform}..."
-        URL=${CONDA_BASE_URL}/${package}/${MKL_VERSION}/download/linux-${platform}/${package}-${MKL_VERSION}-intel_${MKL_PATCH}.tar.bz2
-        wget -q --show-progress "${URL}" -O - | tar xj -C mkl/x86_${platform}
+        DOWNLOAD_FILE=${package}-${MKL_VERSION}-intel_${MKL_PATCH}.tar.bz2
+        echo "Installing ${DOWNLOAD_FILE} for x86_${platform}..."
+        URL=${CONDA_BASE_URL}/${package}/${MKL_VERSION}/download/$FILE_PREFIX-${platform}/${DOWNLOAD_FILE}
+        echo "try download mkl package from: ${URL}"
+        wget -q --show-progress "${URL}" -O mkl/x86_${platform}/${DOWNLOAD_FILE}
+        $TAR xvj -C mkl/x86_${platform} -f mkl/x86_${platform}/${DOWNLOAD_FILE}
     done
 done