fix(build/windows): fix windows build:

* compat clang-cl 11 build at windows env
* fix cuda/cudnn/trt copy env build failed on windows

GitOrigin-RevId: 7fe2d2c0dcaabf56df1725994a41a15dd09692b6

CMakeLists.txt

@@ -182,9 +182,8 @@ if(MSVC OR WIN32)
     endif()
 
     add_compile_definitions(NOMINMAX=1 _USE_MATH_DEFINES=1 WIN32=1)
-    message(STATUS "into windows build...")
-    message(VERBOSE "CMAKE_C_COMPILER_ID: ${CMAKE_C_COMPILER_ID}")
-    if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang-cl")
+    message(STATUS "into windows build CMAKE_C_COMPILER_ID: ${CMAKE_C_COMPILER_ID}")
+    if (NOT ${CMAKE_C_COMPILER_ID} STREQUAL "Clang" AND NOT ${CMAKE_C_COMPILER_ID} STREQUAL "Clang-cl")
         message(FATAL_ERROR "only support clang-cl for windows build, pls check detail: scripts/cmake-build/BUILD_README.md")
     endif()
     # add flags for enable sse instruction optimize for X86, enable avx header to compile avx code
@@ -397,7 +396,7 @@ if(MGE_WITH_CUDA)
     set(CMAKE_CUDA_FLAGS_MINSIZEREL "-Os")
     if(MSVC OR WIN32)
         set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xfatbin -compress-all")
-        set(CCBIN_FLAG "${CCBIN_FLAG} /wd4819 /wd4334 /wd4267 /wd4002 /wd4244 /wd4068")
+        set(CCBIN_FLAG "${CCBIN_FLAG} /wd4819 /wd4334 /wd4267 /wd4002 /wd4244 /wd4068 /std:c++14")
         if(${CMAKE_BUILD_TYPE} STREQUAL "Debug")
             set(CCBIN_FLAG "${CCBIN_FLAG} -D_ITERATOR_DEBUG_LEVEL=2 -MTd")
         endif()

cmake/flatbuffers.cmake

@@ -16,6 +16,9 @@ if(MSVC OR WIN32)
     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}")

scripts/cmake-build/BUILD_README.md

@@ -18,7 +18,7 @@
 ### Windows host build
 * commands:
 ```
-1: installl Visual Studio (need support LLVM/clang-cl), eg 2019. Please install LLVM-10, VS LLVM linker have issue, please replace lld-link.exe, which can be download from https://releases.llvm.org/download.html#10.0.0
+1: installl Visual Studio (need support LLVM/clang-cl), eg 2019. Please install LLVM-10/11, VS LLVM linker have issue, please replace lld-link.exe, which can be download from https://releases.llvm.org/download.html#10.0.0, what`s more,  Visual Studio cl.exe version >=14.28.29910 do not compat with cuda 10.1, please do not use this issue version!
 2: install extension of VS: Python/Cmake/LLVM/Ninja
 3: now we support cuda10.1+cudnn7.6+TensorRT6.0 on Windows, as Windows can only use DLL in fact with cudnn/TensorRT, so please install the same version;
     3a: install cuda10.1 to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1

scripts/cmake-build/host_build.sh

@@ -203,6 +203,8 @@ function prepare_env_for_windows_build() {
     export CPATH=$CPATH:$NIVIDA_INSTALL_PRE/${TRT_V}/include:$NIVIDA_INSTALL_PRE/CUDA/${CUDA_V}/include:$NIVIDA_INSTALL_PRE/CUDA/${CUDA_V}/include/nvtx3:$PC_CUDNN_INCLUDE_DIRS
     export LIBRARY_PATH=$LIBRARY_PATH:$LD_LIBRARY_PATH
     export INCLUDE=$INCLUDE:$CPATH
+    export CUDNN_ROOT_DIR=${NIVIDA_INSTALL_PRE}/${CUDNN_V}/cuda
+    export TRT_ROOT_DIR=${NIVIDA_INSTALL_PRE}/${TRT_V}
 
     # python version will be config by whl build script or ci script, we need
     # a DFT version for build success when we just call host_build.sh

src/CMakeLists.txt

@@ -59,6 +59,16 @@ add_library(megbrain OBJECT ${SOURCES})
 target_link_libraries(megbrain PUBLIC mgb_opr_param_defs)
 if(MGE_WITH_CUDA)
     target_include_directories(megbrain PUBLIC ${TRT_INCLUDE_DIR})
+    target_include_directories(megbrain PRIVATE ${CUDNN_INCLUDE_DIR})
+    find_path(NVTX3_INCLUDE
+        NAMES nvToolsExtCudaRt.h
+        HINTS $ENV{CUDA_ROOT_DIR} $ENV{CUDA_PATH} $ENV{CUDA_BIN_PATH}
+        PATH_SUFFIXES include/nvtx3
+        DOC "NVTX3_INCLUDE" )
+    if(NVTX3_INCLUDE STREQUAL "NVTX3_INCLUDE-NOTFOUND")
+        message(FATAL_ERROR "Can not find NVTX3 INCLUDE, please export cuda sdk path to CUDA_ROOT_DIR or CUDA_PATH or CUDA_BIN_PATH")
+    endif()
+    target_include_directories(megbrain PRIVATE ${NVTX3_INCLUDE})
 endif()
 target_include_directories(megbrain
     PUBLIC $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>