Windows native port (#2478)

Fixes #2407. Current state is that code builds on windows and is able to
pass many unit tests. More fixes will be done when this patch is
integrated into main branch.

---------

Signed-off-by: Gregory Shimansky <[email protected]>

Author: gshimansky

CMakeLists.txt

@@ -8,23 +8,24 @@ endif()
 
 include(ExternalProject)
 
-set(CMAKE_CXX_STANDARD 17)
-
 set(CMAKE_INCLUDE_CURRENT_DIR ON)
 
 project(triton CXX)
 include(CTest)
 
-if(NOT WIN32)
-  list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
-endif()
-
-
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 
 # Options
+if(WIN32)
+  set(DEFAULT_BUILD_PROTON OFF)
+else()
+  set(DEFAULT_BUILD_PROTON ON)
+endif()
+
+# Define the option with the determined default value
+option(TRITON_BUILD_PROTON "Build the Triton Proton profiler" ${DEFAULT_BUILD_PROTON})
 option(TRITON_BUILD_TUTORIALS "Build C++ Triton tutorials" ON)
 option(TRITON_BUILD_PYTHON_MODULE "Build Python Triton bindings" OFF)
-option(TRITON_BUILD_PROTON "Build the Triton Proton profiler" ON)
 option(TRITON_BUILD_UT "Build C++ Triton Unit Tests" ON)
 option(TRITON_BUILD_WITH_CCACHE "Build with ccache (if available)" ON)
 set(TRITON_CODEGEN_BACKENDS "" CACHE STRING "Enable different codegen backends")
@@ -49,10 +50,21 @@ endif()
 # used conditionally in this file and by lit tests
 
 # Customized release build type with assertions: TritonRelBuildWithAsserts
-set(CMAKE_C_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g")
-set(CMAKE_CXX_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g")
-set(CMAKE_C_FLAGS_TRITONBUILDWITHO1 "-O1")
-set(CMAKE_CXX_FLAGS_TRITONBUILDWITHO1 "-O1")
+if(NOT MSVC)
+  set(CMAKE_CXX_STANDARD 17)
+  set(CMAKE_C_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g")
+  set(CMAKE_CXX_FLAGS_TRITONRELBUILDWITHASSERTS "-O2 -g")
+  set(CMAKE_C_FLAGS_TRITONBUILDWITHO1 "-O1")
+  set(CMAKE_CXX_FLAGS_TRITONBUILDWITHO1 "-O1")
+else()
+  set(CMAKE_CXX_STANDARD 20)
+  set(CMAKE_C_FLAGS_TRITONRELBUILDWITHASSERTS "/Zi /Ob0 /Od /RTC1 /bigobj /Zc:preprocessor")
+  set(CMAKE_CXX_FLAGS_TRITONRELBUILDWITHASSERTS "/Zi /Ob0 /Od /RTC1 /bigobj /Zc:preprocessor")
+  set(CMAKE_EXE_LINKER_FLAGS_TRITONRELBUILDWITHASSERTS "/debug:fastlink /INCREMENTAL")
+  set(CMAKE_MODULE_LINKER_FLAGS_TRITONRELBUILDWITHASSERTS "/debug:fastlink /INCREMENTAL")
+  set(CMAKE_SHARED_LINKER_FLAGS_TRITONRELBUILDWITHASSERTS "/debug:fastlink /INCREMENTAL")
+  set(CMAKE_STATIC_LINKER_FLAGS_TRITONRELBUILDWITHASSERTS "/debug:fastlink /INCREMENTAL")
+endif()
 
 # Default build type
 if(NOT CMAKE_BUILD_TYPE)
@@ -70,7 +82,15 @@ endif()
 
 # Compiler flags
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include)
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__STDC_FORMAT_MACROS  -fPIC -std=gnu++17")
+if(NOT MSVC)
+  if(NOT WIN32)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__STDC_FORMAT_MACROS -fPIC")
+  else()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__STDC_FORMAT_MACROS -Wno-deprecated")
+  endif()
+else()
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__STDC_FORMAT_MACROS /wd4244 /wd4624 /wd4715 /wd4530")
+endif()
 
 
 # #########
@@ -124,7 +144,11 @@ endfunction()
 
 
 # Disable warnings that show up in external code (gtest;pybind11)
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror -Wno-covered-switch-default -fvisibility=hidden")
+if(NOT MSVC)
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror -Wno-covered-switch-default -fvisibility=hidden")
+else()
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /WX-")
+endif()
 
 include_directories(".")
 include_directories(${MLIR_INCLUDE_DIRS})
@@ -134,7 +158,8 @@ include_directories(${PROJECT_BINARY_DIR}/include) # Tablegen'd files
 include_directories(${PROJECT_SOURCE_DIR}/third_party)
 include_directories(${PROJECT_BINARY_DIR}/third_party) # Tablegen'd files
 
-# link_directories(${LLVM_LIBRARY_DIR})
+link_directories(${LLVM_LIBRARY_DIR})
+
 add_subdirectory(include)
 add_subdirectory(lib)
 
@@ -163,6 +188,8 @@ if(TRITON_BUILD_PYTHON_MODULE)
     # using pip install.
     include_directories(${PYTHON_INCLUDE_DIRS})
     include_directories(${PYBIND11_INCLUDE_DIR})
+    message(STATUS "PYTHON_LIB_DIRS ${PYTHON_LIB_DIRS}")
+    link_directories(${PYTHON_LIB_DIRS})
   else()
     # Otherwise, we might be building from top CMakeLists.txt directly.
     # Try to find Python and pybind11 packages.
@@ -245,7 +272,7 @@ if(TRITON_BUILD_PYTHON_MODULE)
           LLVMAArch64CodeGen
           LLVMAArch64AsmParser
       )
-  elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
+  elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64" OR CMAKE_SYSTEM_PROCESSOR MATCHES "AMD64")
       list(APPEND TRITON_LIBRARIES
           LLVMX86CodeGen
           LLVMX86AsmParser
@@ -280,6 +307,8 @@ if(TRITON_BUILD_PYTHON_MODULE)
   target_link_libraries(triton PUBLIC ${TRITON_LIBRARIES})
   if(WIN32)
     target_link_libraries(triton PRIVATE ${CMAKE_DL_LIBS})
+    set_target_properties(triton PROPERTIES SUFFIX ".pyd")
+    set_target_properties(triton PROPERTIES PREFIX "lib")
   else()
     target_link_libraries(triton PRIVATE z)
   endif()
@@ -306,6 +335,10 @@ if(NOT TRITON_BUILD_PYTHON_MODULE)
     add_subdirectory(third_party/${CODEGEN_BACKEND})
   endforeach()
 endif()
+if(WIN32)
+  option(CMAKE_USE_WIN32_THREADS_INIT "using WIN32 threads" ON)
+  option(gtest_disable_pthreads "Disable uses of pthreads in gtest." ON)
+endif()
 
 add_subdirectory(third_party/f2reduce)
 add_subdirectory(bin)

python/setup.py

@@ -103,6 +103,49 @@ def copy_externals():
         ]
 
 
+def find_vswhere():
+    program_files = os.environ.get("ProgramFiles(x86)", "C:\\Program Files (x86)")
+    vswhere_path = Path(program_files) / "Microsoft Visual Studio" / "Installer" / "vswhere.exe"
+    if vswhere_path.exists():
+        return vswhere_path
+    return None
+
+
+def find_visual_studio(version_ranges):
+    vswhere = find_vswhere()
+    if not vswhere:
+        raise FileNotFoundError("vswhere.exe not found.")
+
+    for version_range in version_ranges:
+        command = [
+            str(vswhere), "-version", version_range, "-requires", "Microsoft.VisualStudio.Component.VC.Tools.x86.x64",
+            "-property", "installationPath", "-prerelease"
+        ]
+
+        try:
+            output = subprocess.check_output(command, text=True).strip()
+            if output:
+                return output
+        except subprocess.CalledProcessError:
+            continue
+
+    return None
+
+
+def set_env_vars(vs_path, arch="x64"):
+    vcvarsall_path = Path(vs_path) / "VC" / "Auxiliary" / "Build" / "vcvarsall.bat"
+    if not vcvarsall_path.exists():
+        raise FileNotFoundError(f"vcvarsall.bat not found in expected path: {vcvarsall_path}")
+
+    command = ["call", vcvarsall_path, arch, "&&", "set"]
+    output = subprocess.check_output(command, shell=True, text=True)
+
+    for line in output.splitlines():
+        if '=' in line:
+            var, value = line.split('=', 1)
+            os.environ[var] = value
+
+
 # Taken from https://github.com/pytorch/pytorch/blob/master/tools/setup_helpers/env.py
 def check_env_flag(name: str, default: str = "") -> bool:
     return os.getenv(name, default).upper() in ["ON", "1", "YES", "TRUE", "Y"]
@@ -196,6 +239,8 @@ def get_llvm_package_info():
                 f"LLVM pre-compiled image is not available for {system}-{arch}. Proceeding with user-configured LLVM from source build."
             )
             return Package("llvm", "LLVM-C.lib", "", "LLVM_INCLUDE_DIRS", "LLVM_LIBRARY_DIR", "LLVM_SYSPATH")
+    elif system == 'Windows':
+        system_suffix = "windows-x64"
     else:
         print(
             f"LLVM pre-compiled image is not available for {system}-{arch}. Proceeding with user-configured LLVM from source build."
@@ -281,10 +326,10 @@ def download_and_copy(name, src_path, dst_path, variable, version, url_func):
     base_dir = os.path.dirname(__file__)
     system = platform.system()
     try:
-        arch = {"x86_64": "64", "arm64": "aarch64", "aarch64": "aarch64"}[platform.machine()]
+        arch = {"x86_64": "64", "AMD64": "64", "arm64": "aarch64", "aarch64": "aarch64"}[platform.machine()]
     except KeyError:
         arch = platform.machine()
-    supported = {"Linux": "linux", "Darwin": "linux"}
+    supported = {"Linux": "linux", "Darwin": "linux", "Windows": "win"}
     url = url_func(supported[system], arch, version)
     tmp_path = os.path.join(triton_cache_path, "nvidia", name)  # path to cache the download
     dst_path = os.path.join(base_dir, os.pardir, "third_party", "nvidia", "backend", dst_path)  # final binary path
@@ -401,6 +446,11 @@ def get_proton_cmake_args(self):
     def build_extension(self, ext):
         lit_dir = shutil.which('lit')
         ninja_dir = shutil.which('ninja')
+        if platform.system() == "Windows":
+            vs_path = find_visual_studio(["[17.0,18.0)", "[16.0,17.0)"])
+            env = set_env_vars(vs_path)
+            if not vs_path:
+                raise EnvironmentError("Visual Studio 2019 or 2022 not found.")
         # lit is used by the test suite
         thirdparty_cmake_args = get_thirdparty_packages([get_llvm_package_info()])
         thirdparty_cmake_args += self.get_pybind11_cmake_args()
@@ -421,6 +471,10 @@ def build_extension(self, ext):
             "-DTRITON_CODEGEN_BACKENDS=" + ';'.join([b.name for b in backends if not b.is_external]),
             "-DTRITON_PLUGIN_DIRS=" + ';'.join([b.src_dir for b in backends if b.is_external])
         ]
+        if platform.system() == "Windows":
+            installed_base = sysconfig.get_config_var('installed_base')
+            py_lib_dirs = os.getenv("PYTHON_LIB_DIRS", os.path.join(installed_base, "libs"))
+            cmake_args.append("-DPYTHON_LIB_DIRS=" + py_lib_dirs)
         if lit_dir is not None:
             cmake_args.append("-DLLVM_EXTERNAL_LIT=" + lit_dir)
         cmake_args.extend(thirdparty_cmake_args)

python/triton/runtime/CLFinder.py

@@ -0,0 +1,55 @@
+import os
+import subprocess
+from pathlib import Path
+
+
+def find_vswhere():
+    program_files = os.environ.get("ProgramFiles(x86)", "C:\\Program Files (x86)")
+    vswhere_path = Path(program_files) / "Microsoft Visual Studio" / "Installer" / "vswhere.exe"
+    if vswhere_path.exists():
+        return vswhere_path
+    return None
+
+
+def find_visual_studio(version_ranges):
+    vswhere = find_vswhere()
+    if not vswhere:
+        raise FileNotFoundError("vswhere.exe not found.")
+
+    for version_range in version_ranges:
+        command = [
+            str(vswhere), "-version", version_range, "-requires", "Microsoft.VisualStudio.Component.VC.Tools.x86.x64",
+            "-property", "installationPath", "-prerelease"
+        ]
+
+        try:
+            output = subprocess.check_output(command, text=True).strip()
+            if output:
+                return output
+        except subprocess.CalledProcessError:
+            continue
+
+    return None
+
+
+def set_env_vars(vs_path, arch="x64"):
+    vcvarsall_path = Path(vs_path) / "VC" / "Auxiliary" / "Build" / "vcvarsall.bat"
+    if not vcvarsall_path.exists():
+        raise FileNotFoundError(f"vcvarsall.bat not found in expected path: {vcvarsall_path}")
+
+    command = f'call "{vcvarsall_path}" {arch} && set'
+    output = subprocess.check_output(command, shell=True, text=True)
+
+    for line in output.splitlines():
+        if '=' in line:
+            var, value = line.split('=', 1)
+            os.environ[var] = value
+
+
+def initialize_visual_studio_env(version_ranges, arch="x64"):
+    # Check if the environment variable that vcvarsall.bat sets is present
+    if os.environ.get('VSCMD_ARG_TGT_ARCH') != arch:
+        vs_path = find_visual_studio(version_ranges)
+        if not vs_path:
+            raise EnvironmentError("Visual Studio not found in specified version ranges.")
+        set_env_vars(vs_path, arch)

python/triton/runtime/build.py

@@ -6,6 +6,8 @@
 import shutil
 import subprocess
 import setuptools
+import platform
+from .CLFinder import initialize_visual_studio_env
 
 
 def is_xpu():
@@ -23,6 +25,29 @@ def quiet():
         sys.stdout, sys.stderr = old_stdout, old_stderr
 
 
+def _cc_cmd(cc, src, out, include_dirs, library_dirs, libraries):
+    if cc in ["cl", "clang-cl"]:
+        cc_cmd = [cc, src, "/nologo", "/O2", "/LD"]
+        cc_cmd += [f"/I{dir}" for dir in include_dirs]
+        cc_cmd += [f"/Fo{os.path.join(os.path.dirname(out), 'main.obj')}"]
+        cc_cmd += ["/link"]
+        cc_cmd += [f"/OUT:{out}"]
+        cc_cmd += [f"/IMPLIB:{os.path.join(os.path.dirname(out), 'main.lib')}"]
+        cc_cmd += [f"/PDB:{os.path.join(os.path.dirname(out), 'main.pdb')}"]
+        cc_cmd += [f"/LIBPATH:{dir}" for dir in library_dirs]
+        cc_cmd += [f'{lib}.lib' for lib in libraries]
+    else:
+        cc_cmd = [cc, src, "-O3", "-shared", "-Wno-psabi"]
+        if os.name != "nt":
+            cc_cmd += ["-fPIC"]
+        cc_cmd += [f'-l{lib}' for lib in libraries]
+        cc_cmd += [f"-L{dir}" for dir in library_dirs]
+        cc_cmd += [f"-I{dir}" for dir in include_dirs]
+        cc_cmd += ["-o", out]
+
+    return cc_cmd
+
+
 def _build(name, src, srcdir, library_dirs, include_dirs, libraries, extra_compile_args=[]):
     suffix = sysconfig.get_config_var('EXT_SUFFIX')
     so = os.path.join(srcdir, '{name}{suffix}'.format(name=name, suffix=suffix))
@@ -33,6 +58,9 @@ def _build(name, src, srcdir, library_dirs, include_dirs, libraries, extra_compi
         clang = shutil.which("clang")
         gcc = shutil.which("gcc")
         cc = gcc if gcc is not None else clang
+        if platform.system() == "Windows":
+            cc = "cl"
+            initialize_visual_studio_env(["[17.0,18.0)", "[16.0,17.0)"])
         if cc is None:
             raise RuntimeError("Failed to find C compiler. Please specify via CC environment variable.")
     # This function was renamed and made public in Python 3.10
@@ -55,25 +83,24 @@ def _build(name, src, srcdir, library_dirs, include_dirs, libraries, extra_compi
             clangpp = shutil.which("clang++")
             gxx = shutil.which("g++")
             icpx = shutil.which("icpx")
-            cxx = icpx or clangpp or gxx
+            cxx = icpx if os.name == "nt" else icpx or clangpp or gxx
             if cxx is None:
                 raise RuntimeError("Failed to find C++ compiler. Please specify via CXX environment variable.")
+        cc = cxx
         import numpy as np
         numpy_include_dir = np.get_include()
         include_dirs = include_dirs + [numpy_include_dir]
-        cc_cmd = [cxx]
         if icpx is not None:
-            cc_cmd += ["-fsycl"]
+            extra_compile_args += ["-fsycl"]
         else:
-            cc_cmd += ["--std=c++17"]
+            extra_compile_args += ["--std=c++17"]
+        if os.name == "nt":
+            library_dirs += [os.path.join(sysconfig.get_paths(scheme=scheme)["stdlib"], "..", "libs")]
     else:
         cc_cmd = [cc]
 
     # for -Wno-psabi, see https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111047
-    cc_cmd += [src, "-O3", "-shared", "-fPIC", "-Wno-psabi", "-o", so]
-    cc_cmd += [f'-l{lib}' for lib in libraries]
-    cc_cmd += [f"-L{dir}" for dir in library_dirs]
-    cc_cmd += [f"-I{dir}" for dir in include_dirs if dir is not None]
+    cc_cmd = _cc_cmd(cc, src, so, include_dirs, library_dirs, libraries)
     cc_cmd += extra_compile_args
 
     if os.getenv("VERBOSE"):
@@ -90,7 +117,7 @@ def _build(name, src, srcdir, library_dirs, include_dirs, libraries, extra_compi
         language='c',
         sources=[src],
         include_dirs=include_dirs,
-        extra_compile_args=extra_compile_args + ['-O3'],
+        extra_compile_args=extra_compile_args + ['-O3' if "-O3" in cc_cmd else "/O2"],
         extra_link_args=extra_link_args,
         library_dirs=library_dirs,
         libraries=libraries,