Fixes for building on Windows (#382)

* make C++ code compatible with Windows + MSVC

* added dlfcn-win32 for cross-platform dlopen

* fixed building and pip install on Windows

* fixed shared library file name under Windows

.gitignore

@@ -1,6 +1,9 @@
+build/
+
 __pycache__
 .pytest_cache
 
 python/build/
 python/triton.egg-info/
+python/triton/_C/libtriton.pyd
 python/triton/_C/libtriton.so

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "deps/dlfcn-win32"]
+	path = deps/dlfcn-win32
+	url = https://github.com/dlfcn-win32/dlfcn-win32.git

CMakeLists.txt

@@ -1,6 +1,8 @@
 cmake_minimum_required(VERSION 3.6)
 include(ExternalProject)
 
+set(CMAKE_CXX_STANDARD 17)
+
 if(NOT TRITON_LLVM_BUILD_DIR)
     set(TRITON_LLVM_BUILD_DIR ${CMAKE_BINARY_DIR})
 endif()
@@ -8,7 +10,9 @@ endif()
 
 project(triton)
 include(CTest)
+if(NOT WIN32)
   list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
+endif()
 
 # Options
 option(BUILD_TUTORIALS "Build C++ Triton tutorials" ON)
@@ -20,10 +24,19 @@ if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE "Release")
 endif()
 
+if(NOT WIN32)
     find_library(TERMINFO_LIBRARY tinfo)
+endif()
 
 # Compiler flags
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include)
+
+if(WIN32)
+    SET(BUILD_SHARED_LIBS OFF)
+    include_directories(${CMAKE_CURRENT_SOURCE_DIR}/deps/dlfcn-win32/src)
+    add_subdirectory(deps/dlfcn-win32/src ${CMAKE_BINARY_DIR}/dlfcn-win32)
+endif()
+
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__STDC_FORMAT_MACROS  -std=gnu++17")
 
 
@@ -31,7 +44,20 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__STDC_FORMAT_MACROS  -std=gnu++17")
 # LLVM
 ##########
 if("${LLVM_LIBRARY_DIR}" STREQUAL "")
+    if(WIN32)
+      find_package(LLVM 13 REQUIRED COMPONENTS nvptx amdgpu)
+
+      include_directories(${LLVM_INCLUDE_DIRS})
+      separate_arguments(LLVM_DEFINITIONS_LIST NATIVE_COMMAND ${LLVM_DEFINITIONS})
+      add_definitions(${LLVM_DEFINITIONS_LIST})
+
+      llvm_map_components_to_libnames(LLVM_LIBRARIES support core
+        NVPTXInfo nvptxcodegen
+        AMDGPUInfo AMDGPUcodegen
+      )
+    else()
       find_package(LLVM 11 REQUIRED COMPONENTS "nvptx;amdgpu")
+    endif()
     message(STATUS "Found LLVM ${LLVM_PACKAGE_VERSION}")
     if(APPLE)
       set(CMAKE_OSX_DEPLOYMENT_TARGET "10.14")
@@ -108,12 +134,25 @@ endif()
 
 # Triton
 file(GLOB_RECURSE LIBTRITON_SRC lib/*.cc)
+if (WIN32 AND BUILD_PYTHON_MODULE)
+    find_package(Python3 REQUIRED COMPONENTS Development)
+    Python3_add_library(triton SHARED ${LIBTRITON_SRC} ${PYTHON_SRC})
+    set_target_properties(triton PROPERTIES SUFFIX ".pyd")
+    set_target_properties(triton PROPERTIES PREFIX "lib")
+else()
     add_library(triton SHARED ${LIBTRITON_SRC} ${PYTHON_SRC})
+endif()
+
 target_link_options(triton PRIVATE ${LLVM_LDFLAGS})
+
+if(WIN32)
+    target_link_libraries(triton PRIVATE ${LLVM_LIBRARIES} dl) # dl is from dlfcn-win32
+else()
     target_link_libraries(triton ${LLVM_LIBRARIES} z ${TERMINFO_LIBRARY})
+endif()
 
 
-if(BUILD_PYTHON_MODULE)
+if(BUILD_PYTHON_MODULE AND NOT WIN32)
     set(CMAKE_SHARED_LIBRARY_SUFFIX ".so")
     # Check if the platform is MacOS
     if(APPLE)

include/triton/tools/sys/exec.hpp

@@ -13,6 +13,14 @@ namespace tools
 {
 
 
+#ifdef _WIN32
+#define popen _popen
+#define pclose _pclose
+#endif
+
+#ifndef WEXITSTATUS
+#define WEXITSTATUS(stat_val) ((unsigned)(stat_val) & 255)
+#endif
 
 int exec(const std::string& cmd, std::string& result) {
   char buffer[128];

include/triton/tools/sys/getenv.hpp

@@ -33,19 +33,10 @@ namespace tools
 
     inline std::string getenv(const char * name)
     {
-        #ifdef _MSC_VER
-            char* cache_path = 0;
-            std::size_t sz = 0;
-            _dupenv_s(&cache_path, &sz, name);
-        #else
         const char * cstr = std::getenv(name);
-        #endif
         if(!cstr)
             return "";
         std::string result(cstr);
-        #ifdef _MSC_VER
-            free(cache_path);
-        #endif
         return result;
     }
 

lib/codegen/selection/generator.cc

@@ -441,18 +441,18 @@ std::tuple<Value*, Value*, Value*, Value*> generator::fp8x4_to_fp16x4(Value *in0
   "lop3.b32 $1, b1, 0x80008000, a1, 0xf8; \n\t"
   "}", "=r,=r,r", false);
   Value *packed_in = UndefValue::get(vec_ty(i8_ty, 4));
-  packed_in = insert_elt(packed_in, in0, (int)0);
-  packed_in = insert_elt(packed_in, in1, (int)1);
-  packed_in = insert_elt(packed_in, in2, (int)2);
-  packed_in = insert_elt(packed_in, in3, (int)3);
+  packed_in = insert_elt(packed_in, in0, (uint64_t)0);
+  packed_in = insert_elt(packed_in, in1, (uint64_t)1);
+  packed_in = insert_elt(packed_in, in2, (uint64_t)2);
+  packed_in = insert_elt(packed_in, in3, (uint64_t)3);
   Value *in = bit_cast(packed_in, i32_ty);
   Value *ret = call(ptx, {in});
   Value *packed_ret0 = extract_val(ret, {0});
   Value *packed_ret1 = extract_val(ret, {1});
-  Value *ret0 = extract_elt(packed_ret0, (int)0);
-  Value *ret1 = extract_elt(packed_ret0, (int)1);
-  Value *ret2 = extract_elt(packed_ret1, (int)0);
-  Value *ret3 = extract_elt(packed_ret1, (int)1);
+  Value *ret0 = extract_elt(packed_ret0, (uint64_t)0);
+  Value *ret1 = extract_elt(packed_ret0, (uint64_t)1);
+  Value *ret2 = extract_elt(packed_ret1, (uint64_t)0);
+  Value *ret3 = extract_elt(packed_ret1, (uint64_t)1);
   return std::make_tuple(ret0, ret1, ret2, ret3);
 }
 
@@ -694,11 +694,11 @@ void generator::visit_load_inst(ir::load_inst* x){
     // ---
     // finally call inline ASM
     // ---
-    InlineAsm *_asm = InlineAsm::get(asm_ty, asm_oss.str(), asm_cstrt, true);
+    InlineAsm *inlineAsm = InlineAsm::get(asm_ty, asm_oss.str(), asm_cstrt, true);
     std::vector<Value*> args = {pred, ptr};
     for(Value *v: others)
         args.push_back(v);
-    Value *_ret = call(_asm, args);
+    Value *_ret = call(inlineAsm, args);
     // ---
     // extract and store return values
     // ---

lib/driver/llvm.cc

@@ -20,7 +20,9 @@
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 */
 #include <fstream>
+#if __has_include(<unistd.h>)
     #include <unistd.h>
+#endif
 #include <memory>
 #include <regex>
 #include "triton/driver/llvm.h"
@@ -165,10 +167,10 @@ std::string ptx_to_cubin(const std::string& ptx, int cc) {
     return "";
 
   // compile ptx with ptxas
-  char _fsrc[] = "/tmp/triton_k_XXXXXX";
-  char _flog[] = "/tmp/triton_l_XXXXXX";
-  mkstemp(_fsrc);
-  mkstemp(_flog);
+  char _fsrc[L_tmpnam];
+  char _flog[L_tmpnam];
+  std::tmpnam(_fsrc);
+  std::tmpnam(_flog);
   std::string fsrc = _fsrc;
   std::string flog = _flog;
   std::string fbin = fsrc + ".o";
@@ -202,10 +204,10 @@ CUmodule ptx_to_cumodule(const std::string& ptx, int cc) {
     // Use PTXAS via system call
     if(use_system_ptxas){
       // compile ptx with ptxas
-      char _fsrc[] = "/tmp/triton_k_XXXXXX";
-      char _flog[] = "/tmp/triton_l_XXXXXX";
-      mkstemp(_fsrc);
-      mkstemp(_flog);
+      char _fsrc[L_tmpnam];
+      char _flog[L_tmpnam];
+      std::tmpnam(_fsrc);
+      std::tmpnam(_flog);
       std::string fsrc = _fsrc;
       std::string flog = _flog;
       std::string fbin = fsrc + ".o";
@@ -232,8 +234,8 @@ CUmodule ptx_to_cumodule(const std::string& ptx, int cc) {
     CUjit_option opt[] = {CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES, CU_JIT_ERROR_LOG_BUFFER,
                           CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES, CU_JIT_INFO_LOG_BUFFER,
                           CU_JIT_LOG_VERBOSE};
-    unsigned int errbufsize = 8192;
-    unsigned int logbufsize = 8192;
+    const unsigned int errbufsize = 8192;
+    const unsigned int logbufsize = 8192;
     char _err[errbufsize];
     char _log[logbufsize];
     void* optval[] = {(void*)(uintptr_t)errbufsize, (void*)_err, (void*)(uintptr_t)logbufsize, (void*)_log, (void*)1};
@@ -344,8 +346,8 @@ hipModule_t amdgpu_to_hipmodule(const std::string& path) {
   hipJitOption opt[] = {hipJitOptionErrorLogBufferSizeBytes, hipJitOptionErrorLogBuffer,
                             hipJitOptionInfoLogBufferSizeBytes, hipJitOptionInfoLogBuffer,
                             hipJitOptionLogVerbose};
-  unsigned int errbufsize = 8192;
-  unsigned int logbufsize = 8192;
+  const unsigned int errbufsize = 8192;
+  const unsigned int logbufsize = 8192;
   char _err[errbufsize];
   char _log[logbufsize];
   void* optval[] = {(void*)(uintptr_t)errbufsize,

python/setup.py

@@ -24,6 +24,8 @@ def get_llvm():
     paths = [p for p in paths if p is not None]
     if paths:
       return '', ''
+    if platform.system() == "Windows":
+      return '', ''
     # download if nothing is installed
     name = 'clang+llvm-11.0.1-x86_64-linux-gnu-ubuntu-16.04'
     dir = '/tmp'
@@ -105,7 +107,7 @@ class CMakeBuild(build_ext):
         build_args = ["--config", cfg]
 
         if platform.system() == "Windows":
-            cmake_args += ["-DCMAKE_LIBRARY_OUTPUT_DIRECTORY_{}={}".format(cfg.upper(), extdir)]
+            cmake_args += ["-DCMAKE_RUNTIME_OUTPUT_DIRECTORY_{}={}".format(cfg.upper(), extdir)]
             if sys.maxsize > 2**32:
                 cmake_args += ["-A", "x64"]
             build_args += ["--", "/m"]

python/src/triton.cc

@@ -14,6 +14,7 @@
 #include <pybind11/pybind11.h>
 #include <pybind11/stl.h>
 #include <regex>
+#include <sstream>
 #include <string>
 #include "llvm/IR/Module.h"
 #include "llvm/IR/LegacyPassManager.h"