feat(windows/python_whl): make windows HAPPY for build megbrain python package

GitOrigin-RevId: 92b2c07bf939efcac6432bf4c93cd5c46910f129

CMakeLists.txt

@@ -439,6 +439,7 @@ endif()
 set(MGB_JIT ${MGE_WITH_JIT})
 set(MGB_JIT_HALIDE ${MGE_WITH_HALIDE})
 
+# Thread
 IF(APPLE)
     set(CMAKE_THREAD_LIBS_INIT "-lpthread")
     set(CMAKE_HAVE_THREADS_LIBRARY 1)
@@ -447,8 +448,14 @@ IF(APPLE)
     set(THREADS_PREFER_PTHREAD_FLAG ON)
 ENDIF()
 
-# Thread
-if(CMAKE_THREAD_LIBS_INIT)
+if(MSVC OR WIN32)
+    set(CMAKE_HAVE_THREADS_LIBRARY 1)
+    set(CMAKE_USE_WIN32_THREADS_INIT 1)
+    set(CMAKE_USE_PTHREADS_INIT 1)
+    set(THREADS_PREFER_PTHREAD_FLAG ON)
+endif()
+
+if(CMAKE_THREAD_LIBS_INIT OR CMAKE_USE_WIN32_THREADS_INIT)
     set(MGB_HAVE_THREAD 1)
 endif()
 
@@ -471,12 +478,6 @@ 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})
 

python_module/CMakeLists.txt

@@ -11,7 +11,13 @@ find_package(NumPy REQUIRED)
 
 find_package(SWIG REQUIRED)
 set(SWIG_SRC src/swig/mgb.i)
-set(CMAKE_SWIG_FLAGS -Wall -threads -py3 -modern -DSWIGWORDSIZE64)
+if(MSVC OR WIN32)
+    set(CMAKE_SWIG_FLAGS -Wall -threads -py3 -DSWIGWORDSIZE64)
+    message("WARN: swig have some define issue at windows(64) env")
+    message("Please refs scripts/whl/BUILD_PYTHON_WHL_README.md to init windows build env")
+else()
+    set(CMAKE_SWIG_FLAGS -Wall -threads -py3 -modern -DSWIGWORDSIZE64)
+endif()
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-parameter")
 
 file(GLOB_RECURSE OPR_DECL_SRCS "${PROJECT_SOURCE_DIR}/src/**/*.oprdecl")
@@ -69,6 +75,8 @@ set_target_properties(mgb PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BI
 if (APPLE)
     target_link_libraries(mgb megbrain megdnn)
     set_target_properties(mgb PROPERTIES LINK_FLAGS "-undefined dynamic_lookup")
+elseif (MSVC OR WIN32)
+    target_link_libraries(mgb megbrain megdnn)
 else()
     target_link_libraries(mgb megbrain megdnn -Wl,--version-script=${VERSION_SCRIPT})
 endif()

python_module/megengine/_internal/plugin.py

@@ -13,6 +13,7 @@ import atexit
 import collections
 import json
 import os
+import platform
 import signal
 import struct
 
@@ -163,7 +164,11 @@ class GlobalInfkernFinder:
     further investigation
     """
 
-    _signal = signal.SIGUSR1
+    _signal = None
+    if platform.system() != "Windows":
+        _signal = signal.SIGUSR1
+    else:
+        _signal = signal.CTRL_C_EVENT
     _registry = []
     _shell_maker = None
 
@@ -197,7 +202,7 @@ class GlobalInfkernFinder:
             from IPython.terminal.embed import InteractiveShellEmbed
 
             cls._shell_maker = InteractiveShellEmbed
-            fast_signal_hander(signal.SIGUSR1, cls._on_signal)
+            fast_signal_hander(cls._signal, cls._on_signal)
 
         cls._registry.append(finder)
 

python_module/megengine/utils/max_recursion_limit.py

@@ -7,10 +7,13 @@
 # software distributed under the License is distributed on an
 # "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 import platform
-import resource
 import sys
 import threading
 
+# Windows do not imp resource package
+if platform.system() != "Windows":
+    import resource
+
 
 class AlternativeRecursionLimit:
     r"""A reentrant context manager for setting global recursion limits.
@@ -29,6 +32,7 @@ class AlternativeRecursionLimit:
         with self.lock:
             if self.count == 0:
                 self.orig_py_limit = sys.getrecursionlimit()
+            if platform.system() != "Windows":
                 (
                     self.orig_rlim_stack_soft,
                     self.orig_rlim_stack_hard,
@@ -43,8 +47,9 @@ class AlternativeRecursionLimit:
                 except ValueError as exc:
                     if platform.system() != "Darwin":
                         raise exc
-                # increase recursion limit
-                sys.setrecursionlimit(self.new_py_limit)
+
+            # increase recursion limit
+            sys.setrecursionlimit(self.new_py_limit)
             self.count += 1
 
     def __exit__(self, type, value, traceback):
@@ -52,6 +57,8 @@ class AlternativeRecursionLimit:
             self.count -= 1
             if self.count == 0:
                 sys.setrecursionlimit(self.orig_py_limit)
+
+            if platform.system() != "Windows":
                 try:
                     resource.setrlimit(
                         resource.RLIMIT_STACK,

python_module/setup.py

@@ -5,6 +5,7 @@
 import os
 import re
 import pathlib
+import platform
 from distutils.file_util import copy_file
 from setuptools import setup, find_packages, Extension
 from setuptools.command.build_ext import build_ext as _build_ext
@@ -25,7 +26,10 @@ class build_ext(_build_ext):
             extdir.parent.mkdir(parents=True, exist_ok=True)
 
             modpath = self.get_ext_fullname(ext.name).split('.')
-            modpath[-1] += '.so'
+            if platform.system() == "Windows":
+                modpath[-1] += '.pyd'
+            else:
+                modpath[-1] += '.so'
             modpath = str(pathlib.Path(*modpath).resolve())
 
             copy_file(modpath, fullpath, verbose=self.verbose, dry_run=self.dry_run)

python_module/src/cpp/function_replace.cpp

@@ -21,7 +21,13 @@
 #include <cstdarg>
 
 #include <Python.h>
+
+#ifdef WIN32
+#include <windows.h>
+#include <stdio.h>
+#else
 #include <unistd.h>
+#endif
 
 namespace {
 
@@ -50,8 +56,35 @@ class Init {
 };
 Init Init::inst;
 
-int fork_exec_impl(const std::string &arg0, const std::string &arg1,
-        const std::string &arg2) {
+int fork_exec_impl(const std::string& arg0, const std::string& arg1,
+                   const std::string& arg2) {
+#ifdef WIN32
+    STARTUPINFO si;
+    PROCESS_INFORMATION pi;
+    ZeroMemory(&si, sizeof(si));
+    si.cb = sizeof(si);
+    ZeroMemory(&pi, sizeof(pi));
+    auto args_str = " " + arg1 + " " + arg2;
+
+    // Start the child process.
+    if (!CreateProcess(arg0.c_str(),                         // exe name
+                       const_cast<char*>(args_str.c_str()),  // Command line
+                       NULL,   // Process handle not inheritable
+                       NULL,   // Thread handle not inheritable
+                       FALSE,  // Set handle inheritance to FALSE
+                       0,      // No creation flags
+                       NULL,   // Use parent's environment block
+                       NULL,   // Use parent's starting directory
+                       &si,    // Pointer to STARTUPINFO structure
+                       &pi)    // Pointer to PROCESS_INFORMATION structure
+    ) {
+        mgb_log_warn("CreateProcess failed (%lu).\n", GetLastError());
+        fprintf(stderr, "[megbrain] failed to execl %s [%s, %s]\n",
+                arg0.c_str(), arg1.c_str(), arg2.c_str());
+        __builtin_trap();
+    }
+    return pi.dwProcessId;
+#else
     auto pid = fork();
     if (!pid) {
         execl(arg0.c_str(), arg0.c_str(), arg1.c_str(), arg2.c_str(), nullptr);
@@ -62,6 +95,7 @@ int fork_exec_impl(const std::string &arg0, const std::string &arg1,
     }
     mgb_assert(pid > 0, "failed to fork: %s", std::strerror(errno));
     return pid;
+#endif
 }
 
 } // anonymous namespace

python_module/src/cpp/megbrain_config.cpp

@@ -17,7 +17,32 @@
 
 #include <set>
 
+#if defined(WIN32)
+#include <io.h>
+#include <windows.h>
+#define F_OK 0
+#define RTLD_LAZY 0
+#define RTLD_GLOBAL 0
+#define RTLD_NOLOAD 0
+#define access(a, b) false
+
+static void* dlopen(const char* file, int) {
+    return static_cast<void*>(LoadLibrary(file));
+}
+
+static void* dlerror() {
+    const char* errmsg = "dlerror not aviable in windows";
+    return const_cast<char*>(errmsg);
+}
+
+static void* dlsym(void* handle, const char* name) {
+    FARPROC symbol = GetProcAddress((HMODULE)handle, name);
+    return reinterpret_cast<void*>(symbol);
+}
+
+#else
 #include <dlfcn.h>
+#endif
 
 #if MGB_ENABLE_OPR_MM
 #include "megbrain/opr/mm_handler.h"
@@ -274,8 +299,7 @@ void _config::load_opr_library(const char* self_path, const char* lib_path) {
     }
 }
 
-std::vector<std::pair<unsigned long int, std::string>>
-_config::dump_registered_oprs() {
+std::vector<std::pair<size_t, std::string>> _config::dump_registered_oprs() {
 #if MGB_ENABLE_DEBUG_UTIL
     return serialization::OprRegistry::dump_registries();
 #else

python_module/src/cpp/megbrain_config.h

@@ -62,8 +62,8 @@ class _config {
         static void load_opr_library(
                 const char* self_path, const char* lib_path);
 
-        static std::vector<std::pair<unsigned long int, std::string>>
-            dump_registered_oprs();
+        static std::vector<std::pair<size_t, std::string>>
+        dump_registered_oprs();
 
         static int create_mm_server(const std::string& server_addr, int port);
 

python_module/src/cpp/megbrain_wrap.cpp

@@ -404,11 +404,18 @@ void CompGraphCallbackValueProxy::do_copy() {
     m_copy_event->record();
 }
 
+#if defined(WIN32)
+#include <windows.h>
+#include <stdio.h>
+#undef CONST
+#define usleep Sleep
+#endif
 void CompGraphCallbackValueProxy::sync() const {
     mgb_assert(!m_use_raw_hv);
     RealTimer t0;
     double next_warn_time = 2, warn_time_delta = 1;
     while (!m_copy_event->finished()) {
+        //! sleep 1ms or sleep 1us no difference for performance on win32
         usleep(1);
         if (t0.get_secs() >= next_warn_time) {
             mgb_log_warn("wait d2h copy for more than %.3f secs",

python_module/src/cpp/plugin.cpp

@@ -16,10 +16,15 @@
 
 #include <thread>
 #include <cstring>
+#include <sstream>
 
+#ifdef WIN32
+#include <windows.h>
+#else
 #include <pthread.h>
-#include <signal.h>
 #include <unistd.h>
+#endif
+#include <signal.h>
 
 /* ================= _InfkernFinderImpl ================= */
 size_t _InfkernFinderImpl::sm_id = 0;
@@ -72,23 +77,28 @@ class _FastSignal::Impl {
     bool m_worker_started = false;
     std::mutex m_mtx;
     std::thread m_worker_hdl;
+#ifdef WIN32
+    SECURITY_ATTRIBUTES win_sa = {sizeof(SECURITY_ATTRIBUTES), NULL, TRUE};
+    HANDLE pipe_r, pipe_w;
+    DWORD bytes_r_w;
+#else
     int m_pfd[2]; //! pipe fds; write signal handlers, -1 for exit
+#endif
     std::unordered_map<int, HandlerCallback> m_handler_callbacks;
 
     void worker() {
-#ifdef __APPLE__
-        uint64_t tid;
-        pthread_threadid_np(NULL, &tid);
-        mgb_log("fast signal worker started in thread 0x%zx",
-                static_cast<size_t>(tid));
-#else
-        mgb_log("fast signal worker started in thread 0x%zx",
-                static_cast<size_t>(pthread_self()));
-#endif
+        std::ostringstream oss;
+        oss << std::this_thread::get_id() << std::endl;
+        mgb_log("fast signal worker started in thread %s", oss.str().c_str());
         mgb::sys::set_thread_name("fastsgl");
         int signum;
-        for (; ; ) {
+        for (;;) {
+#ifdef WIN32
+            if (ReadFile(pipe_r, &signum, sizeof(int), &bytes_r_w, NULL) ==
+                NULL) {
+#else
             if (read(m_pfd[0], &signum, sizeof(int)) != sizeof(int)) {
+#endif
                 if (errno == EINTR)
                     continue;
                 mgb_log_error("fast signal worker: "
@@ -114,10 +124,17 @@ class _FastSignal::Impl {
         if (m_worker_started)
             return;
 
+#ifdef WIN32
+        if (!CreatePipe(&pipe_r, &pipe_w, &win_sa, 0)) {
+            throw mgb::MegBrainError(mgb::ssprintf("failed to create pipe: %s",
+                                                   strerror(errno)));
+        }
+#else
         if (pipe(m_pfd)) {
-            throw mgb::MegBrainError(mgb::ssprintf(
-                        "failed to create pipe: %s", strerror(errno)));
+            throw mgb::MegBrainError(mgb::ssprintf("failed to create pipe: %s",
+                                                   strerror(errno)));
         }
+#endif
         std::thread t(std::bind(&Impl::worker, this));
         m_worker_hdl.swap(t);
         m_worker_started = true;
@@ -125,7 +142,11 @@ class _FastSignal::Impl {
 
     void write_pipe(int v) {
         mgb_assert(m_worker_started);
+#ifdef WIN32
+        if (WriteFile(pipe_w, &v, sizeof(int), &bytes_r_w, NULL) == NULL) {
+#else
         if (write(m_pfd[1], &v, sizeof(int)) != sizeof(int)) {
+#endif
             mgb_log_error("fast signal: failed to write to self pipe: %s",
                     strerror(errno));
         }
@@ -169,8 +190,13 @@ class _FastSignal::Impl {
                 return;
             write_pipe(-1);
             m_worker_hdl.join();
+#ifdef WIN32
+            CloseHandle(pipe_r);
+            CloseHandle(pipe_w);
+#else
             close(m_pfd[0]);
             close(m_pfd[1]);
+#endif
             m_handler_callbacks.clear();
             m_worker_started = false;
         }
@@ -192,11 +218,19 @@ void _FastSignal::signal_hander(int signum) {
 }
 
 void _FastSignal::register_handler(int signum, PyObject *func) {
+#ifdef WIN32
+    //! up to now we can only use CTRL_C_EVENT to unix signal.SIGUSR1/2
+    //FIXME: how to coherence signal number at python side
+    // https://docs.microsoft.com/en-gb/cpp/c-runtime-library/reference/signal?view=vs-2017
+    mgb_assert(signum == CTRL_C_EVENT, "only allow register CTRL_C_EVENT as unix signal.SIGUSR1/2 now");
+    signal(signum, signal_hander);
+#else
     struct sigaction action;
     memset(&action, 0, sizeof(action));
     action.sa_handler = &signal_hander;
     int ret = sigaction(signum, &action, nullptr);
     mgb_assert(!ret, "sigaction failed: %s", strerror(errno));
+#endif
 
     sm_impl.register_handler(signum, func);
 }

python_module/src/cpp/python_helper.cpp

@@ -45,7 +45,7 @@ std::string demangle_typeid(const char* name) {
 
 namespace {
 // does nothing if not g++
-std::string mgb::demangle_typeid(const char* name) {
+std::string demangle_typeid(const char* name) {
     return name;
 }
 }

python_module/src/swig/mgb.i

@@ -31,13 +31,7 @@ void _init_bfloat16_types(PyObject *m); // implemented in bfloat16.cpp
 %template(_VectorString) std::vector<std::string>;
 %template(_PairStringSizeT) std::pair<std::string, size_t>;
 %template(_PairSizeTSizeT) std::pair<size_t, size_t>;
-/*
- * swig use uint64_t have compat build issue with
- * clang at osx env, so we use unsigned long to
- * replace uint64_t,more detail refs stdint.i
- *
- */
-%template(_VectorPairUint64String) std::vector<std::pair<unsigned long int, std::string>>;
+%template(_VectorPairSizeTString) std::vector<std::pair<size_t, std::string>>;
 
 %pythoncode %{
 import numpy as np

python_module/test/integration/test_distributed.py

@@ -89,6 +89,9 @@ def start_workers(worker, world_size, trace=False):
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 def test_distributed():
     start_workers(worker, 2, trace=True)
     start_workers(worker, 2, trace=False)

python_module/test/run.sh

@@ -10,10 +10,10 @@ ignore_list="--ignore test/unit/module/test_pytorch.py \
 test_dirs="megengine test"
 
 pushd $(dirname "${BASH_SOURCE[0]}")/.. >/dev/null
-    pytest -xv -m 'isolated_distributed' \
+    python3 -m pytest -xv -m 'isolated_distributed' \
         --json-report --json-report-file=time_python_test.json \
         $ignore_list $test_dirs
-    pytest -xv -m 'not internet and not isolated_distributed' \
+    python3 -m pytest -xv -m 'not internet and not isolated_distributed' \
         --json-report --json-report-file=time_python_test.json \
         $ignore_list $test_dirs
 popd >/dev/null

python_module/test/unit/distributed/test_functional.py

@@ -29,6 +29,9 @@ def _init_process_group_wrapper(world_size, rank, dev, backend, q):
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_reduce_sum():
     world_size = 2
@@ -68,6 +71,9 @@ def test_reduce_sum():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_gather():
     world_size = 2
@@ -107,6 +113,9 @@ def test_gather():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_broadcast():
     world_size = 2
@@ -142,6 +151,9 @@ def test_broadcast():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_scatter():
     world_size = 2
@@ -181,6 +193,9 @@ def test_scatter():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_all_to_all():
     world_size = 2
@@ -218,6 +233,9 @@ def test_all_to_all():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_all_gather():
     world_size = 2
@@ -254,6 +272,9 @@ def test_all_gather():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_reduce_scatter_sum():
     world_size = 2
@@ -294,6 +315,9 @@ def test_reduce_scatter_sum():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_all_reduce_sum():
     world_size = 2
@@ -330,6 +354,9 @@ def test_all_reduce_sum():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_all_reduce_max():
     world_size = 2
@@ -366,6 +393,9 @@ def test_all_reduce_max():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_all_reduce_min():
     world_size = 2
@@ -402,6 +432,9 @@ def test_all_reduce_min():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_bcast_param():
     world_size = 2

python_module/test/unit/distributed/test_util.py

@@ -45,6 +45,9 @@ def _init_process_group_wrapper(world_size, rank, dev, backend, q):
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_create_mm_server():
     def worker():
@@ -67,6 +70,9 @@ def test_create_mm_server():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_init_process_group():
     world_size = 2
@@ -102,6 +108,9 @@ def test_init_process_group():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_group_barrier():
     world_size = 2
@@ -137,6 +146,9 @@ def test_group_barrier():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_synchronized():
     world_size = 2

python_module/test/unit/module/test_batchnorm.py

@@ -22,6 +22,9 @@ from megengine.test import assertTensorClose
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 @pytest.mark.isolated_distributed
 def test_syncbn():
     nr_chan = 8
@@ -143,6 +146,9 @@ def test_batchnorm():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 def test_syncbn1d():
     nr_chan = 8
     data_shape = (3, nr_chan, 4)
@@ -241,6 +247,9 @@ def test_batchnorm2d():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 def test_syncbn2d():
     nr_chan = 8
     data_shape = (3, nr_chan, 16, 16)
@@ -315,6 +324,9 @@ def test_batchnorm_no_stats():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 def test_syncbn_no_stats():
     nr_chan = 8
     data_shape = (3, nr_chan, 4)
@@ -367,6 +379,9 @@ def test_batchnorm2d_no_stats():
 @pytest.mark.skipif(
     platform.system() == "Darwin", reason="do not imp GPU mode at macos now"
 )
+@pytest.mark.skipif(
+    platform.system() == "Windows", reason="do not imp GPU mode at Windows now"
+)
 def test_syncbn2d_no_stats():
     nr_chan = 8
     data_shape = (3, nr_chan, 16, 16)

python_module/test/unit/module/test_module.py

@@ -6,6 +6,7 @@
 # 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.
+import os
 import tempfile
 from collections import OrderedDict
 from io import BytesIO
@@ -367,8 +368,13 @@ def test_dump_model():
     data.set_value(np.random.random(data_shape))
     mlp = MLP()
     pred = mlp(data)
-    with tempfile.NamedTemporaryFile() as f:
-        mge.dump(pred, f.name)
+    f = tempfile.NamedTemporaryFile(delete=False)
+    f_name = f.name
+    try:
+        mge.dump(pred, f_name)
+    finally:
+        f.close()
+        os.unlink(f_name)
 
 
 def test_load_quantized():

scripts/cmake-build/host_build.sh

@@ -166,6 +166,18 @@ function prepare_env_for_windows_build() {
     echo "put vcvarsall.bat path to PATH env.."
 }
 
+WINDOWS_BUILD_TARGET="Ninja all > build.log"
+if [[ -z ${MAKE_DEVELOP} ]]
+then
+    MAKE_DEVELOP="false"
+fi
+function config_windows_build_target() {
+    if [ ${MAKE_DEVELOP} = "true" ]; then
+        echo "build all and develop for pytest test"
+        WINDOWS_BUILD_TARGET="Ninja all > build.log && Ninja develop"
+    fi
+}
+
 function cmake_build_windows() {
     # windows do not support long path, so we cache the BUILD_DIR ASAP
     prepare_env_for_windows_build
@@ -201,11 +213,12 @@ function cmake_build_windows() {
         ${EXTRA_CMAKE_ARGS} \
         ../../.. && \
         echo \"start Ninja build log to build.log, may take serval min...\" && \
-        Ninja load_and_run > build.log"
+        ${WINDOWS_BUILD_TARGET}"
 }
 
 
 if [[ $OS =~ "NT" ]]; then
+    config_windows_build_target
     cmake_build_windows $MGE_WITH_CUDA $MGE_INFERENCE_ONLY $BUILD_TYPE
 else
     cmake_build $MGE_WITH_CUDA $MGE_INFERENCE_ONLY $BUILD_TYPE

scripts/whl/BUILD_PYTHON_WHL_README.md

 # python whl package build support status
-*  windows build (not ok)
-*  linux build   (ok, cpu or gpu)
-*  macos build   (ok,cpu only)
+*  windows build (ok,cpu only)
+*  linux   build (ok, cpu or gpu)
+*  macos   build (ok,cpu only)
 # build env prepare
 ## linux
 
@@ -17,6 +17,48 @@
     ```
     ./scripts/whl/macos/macos_whl_env_prepare.sh
     ```
+
+## windows
+    ```
+    1: refs scripts/cmake-build/BUILD_README.md windows section build for base windows build
+    2: install several python or install u care python version, default install dir: /c/Users/${USER}/mge_whl_python_env
+       a: mkdir /c/Users/${USER}/mge_whl_python_env
+       b: download python 64-bit install exe
+          https://www.python.org/ftp/python/3.5.4/python-3.5.4-amd64.exe
+          https://www.python.org/ftp/python/3.6.8/python-3.6.8-amd64.exe
+          https://www.python.org/ftp/python/3.7.7/python-3.7.7-amd64.exe
+          https://www.python.org/ftp/python/3.8.3/python-3.8.3-amd64.exe
+       c: install python-3.5.4-amd64.exe to /c/Users/${USER}/mge_whl_python_env/3.5.4 from install gui
+       d: install python-3.6.8-amd64.exe to /c/Users/${USER}/mge_whl_python_env/3.6.8 from install gui
+       e: install python-3.7.7-amd64.exe to /c/Users/${USER}/mge_whl_python_env/3.7.7 from install gui
+       f: install python-3.8.3-amd64.exe to /c/Users/${USER}/mge_whl_python_env/3.8.3 from install gui
+    3: rename python.exe to python3.exe
+       a: mv /c/Users/${USER}/mge_whl_python_env/3.5.4/python.exe /c/Users/${USER}/mge_whl_python_env/3.5.4/python3.exe
+       b: mv /c/Users/${USER}/mge_whl_python_env/3.6.8/python.exe /c/Users/${USER}/mge_whl_python_env/3.6.8/python3.exe
+       c: mv /c/Users/${USER}/mge_whl_python_env/3.7.7/python.exe /c/Users/${USER}/mge_whl_python_env/3.7.7/python3.exe
+       d: mv /c/Users/${USER}/mge_whl_python_env/3.8.3/python.exe /c/Users/${USER}/mge_whl_python_env/3.8.3/python3.exe
+    4: install needed package for build python whl package
+       a0: /c/Users/${USER}/mge_whl_python_env/3.5.4/python3.exe -m pip install --upgrade pip
+       a1: /c/Users/${USER}/mge_whl_python_env/3.5.4/python3.exe -m pip install -r python_module/requires-test.txt
+       a2: /c/Users/${USER}/mge_whl_python_env/3.5.4/python3.exe -m pip install numpy wheel requests tqdm tabulate
+
+       b0: /c/Users/${USER}/mge_whl_python_env/3.6.8/python3.exe -m pip install --upgrade pip
+       b1: /c/Users/${USER}/mge_whl_python_env/3.6.8/python3.exe -m pip install -r python_module/requires-test.txt
+       b2: /c/Users/${USER}/mge_whl_python_env/3.6.8/python3.exe -m pip install numpy wheel requests tqdm tabulate
+
+       c0: /c/Users/${USER}/mge_whl_python_env/3.7.7/python3.exe -m pip install --upgrade pip
+       c1: /c/Users/${USER}/mge_whl_python_env/3.7.7/python3.exe -m pip install -r python_module/requires-test.txt
+       c2: /c/Users/${USER}/mge_whl_python_env/3.7.7/python3.exe -m pip install numpy wheel requests tqdm tabulate
+
+       d0: /c/Users/${USER}/mge_whl_python_env/3.8.3/python3.exe -m pip install --upgrade pip
+       d1: /c/Users/${USER}/mge_whl_python_env/3.8.3/python3.exe -m pip install -r python_module/requires-test.txt
+       d2: /c/Users/${USER}/mge_whl_python_env/3.8.3/python3.exe -m pip install numpy wheel requests tqdm tabulate
+    5: install swig from install gui
+       a: download swig: https://nchc.dl.sourceforge.net/project/swig/swigwin/swigwin-4.0.2/swigwin-4.0.2.zip
+       b: install swig to /c/Users/${USER}/swigwin-4.0.2
+       c: apply scripts/whl/windows/fix-ptr-define-issue.patch to c/Users/${USER}/swigwin-4.0.2
+    ```
+
 # how to build
 ## build for linux
     MegBrain delivers `wheel` package with `manylinux2010` tag defined in [PEP-571](https://www.python.org/dev/peps/pep-0571/).
@@ -36,9 +78,22 @@
 
     ```
     ALL_PYTHON=35m ./build_wheel.sh cpu
-    
     ```
 ## build for macos
     ```
     ./scripts/whl/macos/macos_build_whl.sh
     ```
+    If you just want to build for a specific Python verison, you can use `ALL_PYTHON` environment variable. eg:
+
+    ```
+    ALL_PYTHON=3.5.9 ./scripts/whl/macos/macos_build_whl.sh
+    ```
+## build for windows
+    ```
+    ./scripts/whl/windows/windows_build_whl.sh
+    ```
+    If you just want to build for a specific Python verison, you can use `ALL_PYTHON` environment variable. eg:
+
+    ```
+    ALL_PYTHON=3.5.4 ./scripts/whl/windows/windows_build_whl.sh
+    ```

scripts/whl/macos/macos_build_whl.sh

@@ -149,13 +149,16 @@ function do_build() {
 }
 
 function third_party_prepare() {
-    if [[ -z ${ALREADY_INSTALL_THIRD_PARTY} ]]
+    echo "init third_party..."
+    ${SRC_DIR}/third_party/prepare.sh
+
+
+    if [[ -z ${ALREADY_INSTALL_MKL} ]]
     then
         echo "init third_party..."
-        ${SRC_DIR}/third_party/prepare.sh
         ${SRC_DIR}/third_party/install-mkl.sh
     else
-        echo "skip init third_party..."
+        echo "skip init mkl internal"
     fi
 }
 

scripts/whl/windows/fix-ptr-define-issue.patch

+diff --git a/Lib/stdint.i b/Lib/stdint.i
+index 14fe619..45337c0 100644
+--- a/Lib/stdint.i
++++ b/Lib/stdint.i
+@@ -86,8 +86,8 @@ typedef unsigned long long int	uint_fast64_t;
+ 
+ /* Types for `void *' pointers.  */
+ #if defined(SWIGWORDSIZE64)
+-typedef long int		intptr_t;
+-typedef unsigned long int	uintptr_t;
++typedef long long		intptr_t;
++typedef unsigned long long	uintptr_t;
+ #else
+ typedef int			intptr_t;
+ typedef unsigned int		uintptr_t;
+diff --git a/Lib/swigarch.i b/Lib/swigarch.i
+index bf4ee8e..bf6b5c3 100644
+--- a/Lib/swigarch.i
++++ b/Lib/swigarch.i
+@@ -53,7 +53,7 @@
+ #ifndef LONG_MAX
+ #include <limits.h>
+ #endif
+-#if (__WORDSIZE == 32) || (LONG_MAX == INT_MAX)
++#if (__WORDSIZE == 32)
+ # error "SWIG wrapped code invalid in 32 bit architecture, regenerate code using -DSWIGWORDSIZE32"
+ #endif
+ %}

scripts/whl/windows/windows_build_whl.sh

+#!/bin/bash -e
+
+NT=$(echo `uname` | grep "NT")
+echo $NT
+if [ -z "$NT" ];then
+    echo "only run at windows bash env"
+    echo "pls consider install bash-like tools, eg MSYS or git-cmd, etc"
+    exit -1
+fi
+
+function err_env() {
+    echo "check_env failed: pls refs ${SRC_DIR}/scripts/whl/BUILD_PYTHON_WHL_README.md to init env"
+    exit -1
+}
+
+function append_path_env_and_check() {
+    echo "export swig pwd to PATH"
+    export PATH=/c/Users/${USER}/swigwin-4.0.2::$PATH
+    echo  "export vs2019 install path"
+    export VS_PATH=/c/Program\ Files\ \(x86\)/Microsoft\ Visual\ Studio/2019/Enterprise
+    # for llvm-strip
+    export PATH=$VS_PATH/VC/Tools/Llvm/bin/:$PATH
+}
+
+append_path_env_and_check
+
+SRC_DIR=$(READLINK -f "`dirname $0`/../../../")
+ALL_PYTHON=${ALL_PYTHON}
+FULL_PYTHON_VER="3.5.4 3.6.8 3.7.7 3.8.3"
+if [[ -z ${ALL_PYTHON} ]]
+then
+    ALL_PYTHON=${FULL_PYTHON_VER}
+fi
+
+PYTHON_DIR=
+PYTHON_LIBRARY=
+PYTHON_INCLUDE_DIR=
+WINDOWS_WHL_HOME=${SRC_DIR}/scripts/whl/windows/windows_whl_home
+if [ -e "${WINDOWS_WHL_HOME}" ]; then
+    echo "remove old windows whl file"
+    rm -rf ${WINDOWS_WHL_HOME}
+fi
+mkdir -p ${WINDOWS_WHL_HOME}
+
+function config_python_env() {
+    PYTHON_DIR=/c/Users/${USER}/mge_whl_python_env/$1
+    PYTHON_BIN=${PYTHON_DIR}
+    if [ ! -f "$PYTHON_BIN/python3.exe" ]; then
+        echo "ERR: can not find $PYTHON_BIN , Invalid python package"
+        echo "now support list: ${FULL_PYTHON_VER}"
+        err_env
+    else
+        echo "put python3 to env..."
+        export PATH=${PYTHON_BIN}:$PATH
+        which python3
+    fi
+    echo ${ver}
+
+    PYTHON_LIBRARY=${PYTHON_DIR}/libs/python3.lib
+    PYTHON_INCLUDE_DIR=${PYTHON_DIR}/include
+}
+
+function do_build() {
+    for ver in ${ALL_PYTHON}
+    do
+        #config
+        config_python_env ${ver}
+
+        #check env
+        if [ ! -f "$PYTHON_LIBRARY" ]; then
+            echo "ERR: can not find $PYTHON_LIBRARY , Invalid python package"
+            err_env
+        fi
+        if [ ! -d "$PYTHON_INCLUDE_DIR" ]; then
+            echo "ERR: can not find $PYTHON_INCLUDE_DIR , Invalid python package"
+            err_env
+        fi
+        echo "PYTHON_LIBRARY: ${PYTHON_LIBRARY}"
+        echo "PYTHON_INCLUDE_DIR: ${PYTHON_INCLUDE_DIR}"
+        #append cmake args for config python
+        #FIXME: ninja handle err with cmake 3.17 when assgin PYTHON_LIBRARY
+        #But after put python3.exe to HEAD of PATH by config_python_env, cmake can also handle the
+        #right PYTHON_LIBRARY and PYTHON_INCLUDE_DIR, at the same time, clang-cl need swig target
+        #force LINK a real PYTHON_LIBRARY file, after test we do not find the symbols conflict with python
+        #export EXTRA_CMAKE_ARGS="-DPYTHON_LIBRARY=${PYTHON_LIBRARY} -DPYTHON_INCLUDE_DIR=${PYTHON_INCLUDE_DIR} "
+        #config build type to RelWithDebInfo to enable MGB_ENABLE_DEBUG_UTIL etc
+        export EXTRA_CMAKE_ARGS=${EXTRA_CMAKE_ARGS}" -DCMAKE_BUILD_TYPE=RelWithDebInfo "
+
+        #call build and install
+        #FIXME: cmake do not triger update python config, after
+        #change PYTHON_LIBRARY and PYTHON_INCLUDE_DIR, so add
+        #-r to remove build cache after a new ver build, which
+        #will be more slow build than without -r
+        ${SRC_DIR}/scripts/cmake-build/host_build.sh -t -r
+
+        #call setup.py
+        BUILD_DIR=${SRC_DIR}/build_dir/host/build/
+        cd ${BUILD_DIR}
+
+        if [ -d "staging" ]; then
+            echo "remove old build cache file"
+            rm -rf staging
+        fi
+        mkdir -p staging
+
+
+        cp -a python_module/{megengine,setup.py,requires.txt,requires-style.txt,requires-test.txt} staging/
+        cd ${BUILD_DIR}/staging/megengine/_internal
+        llvm-strip -s _mgb.pyd
+        cd ${BUILD_DIR}/staging
+        ${PYTHON_DIR}/python3 setup.py bdist_wheel
+        cp ${BUILD_DIR}/staging/dist/Meg*.whl ${WINDOWS_WHL_HOME}/
+
+        echo ""
+        echo "##############################################################################################"
+        echo "windows whl package location: ${WINDOWS_WHL_HOME}"
+        ls ${WINDOWS_WHL_HOME}
+        echo "##############################################################################################"
+    done
+}
+
+function third_party_prepare() {
+    echo "init third_party..."
+    ${SRC_DIR}/third_party/prepare.sh
+
+
+    if [[ -z ${ALREADY_INSTALL_MKL} ]]
+    then
+        echo "init third_party..."
+        ${SRC_DIR}/third_party/install-mkl.sh
+    else
+        echo "skip init mkl internal"
+    fi
+}
+
+######################
+third_party_prepare
+do_build

sdk/load-and-run/src/mgblar.cpp

@@ -598,11 +598,9 @@ void run_test_st(Args &env) {
 
     auto format =
             serialization::GraphLoader::identify_graph_dump_format(*inp_file);
-    if (!format.valid()) {
-        printf("invalid model: unknown model format, please make sure input "
-               "file is generated by GraphDumper\n");
-        return;
-    }
+    mgb_assert(format.valid(),
+               "invalid model: unknown model format, please make sure input "
+               "file is generated by GraphDumper");
     auto loader =
             serialization::GraphLoader::make(std::move(inp_file), format.val());
     RealTimer timer;

src/core/impl/utils/debug.cpp

@@ -35,16 +35,21 @@ using namespace debug;
 #include <cuda_runtime.h>
 #endif
 
+#ifndef WIN32
 #include <pthread.h>
+#include <unistd.h>
+#endif
+
 #include <signal.h>
 #include <sys/types.h>
-#include <unistd.h>
 
 #ifdef __ANDROID__
 #include <unwind.h>
 #else
+#ifndef WIN32
 #include <execinfo.h>
 #endif
+#endif
 
 #ifdef __ANDROID__
 namespace {
@@ -121,6 +126,8 @@ void get_mem_map(
     fclose(fin);
 }
 
+#ifndef WIN32
+//FIXME: imp SigHandlerInit backtrace for windows
 class SigHandlerInit {
     static void death_handler(int signum) {
         char msg0[] =
@@ -157,6 +164,7 @@ public:
         std::set_terminate([]() { death_handler(-1); });
     }
 };
+#endif
 
 #if MGB_CUDA
 class CudaCheckOnFork {
@@ -201,7 +209,9 @@ class InitCaller {
     static InitCaller inst;
 
     InitCaller() {
+#ifndef WIN32
         SigHandlerInit::init_for_segv();
+#endif
 #if MGB_CUDA
         CudaCheckOnFork::init();
 #endif
@@ -216,6 +226,7 @@ void (*ForkAfterCudaError::throw_)() = throw_fork_cuda_exc;
 std::atomic_size_t ScopedForkWarningSupress::sm_depth{0};
 
 BacktraceResult mgb::debug::backtrace(int nr_exclude) {
+#ifndef WIN32
     static bool thread_local recursive_call = false;
     if (recursive_call) {
         fprintf(stderr, "recursive call to backtrace()!\n");
@@ -265,6 +276,11 @@ BacktraceResult mgb::debug::backtrace(int nr_exclude) {
 
     recursive_call = false;
     return result;
+#else
+    //FIXME: imp Backtrace for windows
+    BacktraceResult result;
+    return result;
+#endif
 }
 
 void BacktraceResult::fmt_to_str(std::string& dst) {

src/core/include/megbrain/common.h

@@ -197,6 +197,11 @@ void __log__(LogLevel level, const char *file, const char *func, int line,
 #define MGB_GETENV(_name)  static_cast<char*>(nullptr)
 #endif
 
+#ifdef WIN32
+#define unsetenv(_name) _putenv_s(_name, "");
+#define setenv(name,value,overwrite) _putenv_s(name,value)
+#endif
+
 // use some macro tricks to get lock guard with unique variable name
 #define MGB_TOKENPASTE(x, y) x ## y
 #define MGB_TOKENPASTE2(x, y) MGB_TOKENPASTE(x, y)

src/serialization/impl/opr_registry.cpp

@@ -159,10 +159,9 @@ void OprRegistry::add_using_dynamic_loader(
 }
 
 #if MGB_ENABLE_DEBUG_UTIL
-std::vector<std::pair<unsigned long int, std::string>>
-OprRegistry::dump_registries() {
+std::vector<std::pair<size_t, std::string>> OprRegistry::dump_registries() {
     auto&& id2reg = static_data().id2reg;
-    std::vector<std::pair<unsigned long int, std::string>> result;
+    std::vector<std::pair<size_t, std::string>> result;
     for (auto iter = id2reg.begin(); iter != id2reg.end(); ++iter) {
         if (iter->second.name.size() == 0)
             result.push_back({iter->first, "<special>"});

src/serialization/include/megbrain/serialization/opr_registry.h

@@ -76,8 +76,7 @@ namespace serialization {
 
 #if MGB_ENABLE_DEBUG_UTIL
         //! dump registered oprs
-        static std::vector<std::pair<unsigned long int, std::string>>
-        dump_registries();
+        static std::vector<std::pair<size_t, std::string>> dump_registries();
 #endif
     };
 

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

@@ -17,13 +17,6 @@
 
 #include <iostream>
 
-#if defined(WIN32)
-static inline void unsetenv(std::string name) {
-    name += "=";
-    _putenv(name.c_str());
-}
-#define setenv(name,value,overwrite) _putenv_s(name,value)
-#endif
 #if !MGB_ENABLE_EXCEPTION
 #pragma GCC diagnostic ignored  "-Wunused-variable"
 #endif