Add Windows/clang-cl support for AMD HIP backend

python/triton/runtime/build.py

@@ -51,16 +51,26 @@ def is_msvc(cc):
     return cc == "cl" or cc == "cl.exe"
 
 
+def is_clang_cl(cc):
+    cc = os.path.basename(cc).lower()
+    return cc == "clang-cl" or cc == "clang-cl.exe"
+
+
 def is_clang(cc):
     cc = os.path.basename(cc).lower()
     return cc == "clang" or cc == "clang.exe"
 
 
 def _cc_cmd(cc: str, src: str, out: str, include_dirs: list[str], library_dirs: list[str], libraries: list[str],
             ccflags: list[str]) -> list[str]:
-    if is_msvc(cc):
+    if is_msvc(cc) or is_clang_cl(cc):
         out_base = os.path.splitext(out)[0]
-        cc_cmd = [cc, src, "/nologo", "/O2", "/LD", "/std:c11", "/wd4819"]
+        cc_cmd = [cc, src, "/nologo", "/O2", "/LD", "/wd4819"]
+        # clang-cl doesn't support /std:c11, use -std=c11 instead
+        if is_clang_cl(cc):
+            cc_cmd += ["-std=c11"]
+        else:
+            cc_cmd += ["/std:c11"]
         cc_cmd += [f"/I{dir}" for dir in include_dirs if dir is not None]
         cc_cmd += [f"/Fo{out_base + '.obj'}"]
         cc_cmd += ["/link"]
@@ -110,7 +120,7 @@ def _build(name: str, src: str, srcdir: str, library_dirs: list[str], include_di
         if sysconfig.get_config_var("Py_GIL_DISABLED"):
             version += "t"
         libraries = libraries + [f"python{version}"]
-    if is_msvc(cc):
+    if is_msvc(cc) or is_clang_cl(cc):
         _, msvc_winsdk_inc_dirs, msvc_winsdk_lib_dirs = find_msvc_winsdk()
         include_dirs = include_dirs + msvc_winsdk_inc_dirs
         library_dirs = library_dirs + msvc_winsdk_lib_dirs

third_party/amd/backend/compiler.py

@@ -5,12 +5,17 @@
 from typing import Any, Dict, Tuple
 from types import ModuleType
 import hashlib
+import os
+import platform
 import tempfile
 import re
 import functools
 import warnings
 from pathlib import Path
 
+def _is_windows():
+    return platform.system() == 'Windows'
+
 
 def get_min_dot_size(target: GPUTarget):
     # We fallback to use FMA and cast arguments if certain configurations is
@@ -437,13 +442,35 @@ def make_hsaco(src, metadata, options):
         if knobs.compilation.enable_asan:
             target_features = '+xnack'
         hsaco = amd.assemble_amdgcn(src, options.arch, target_features)
-        with tempfile.NamedTemporaryFile() as tmp_out:
-            with tempfile.NamedTemporaryFile() as tmp_in:
-                with open(tmp_in.name, "wb") as fd_in:
-                    fd_in.write(hsaco)
+        # On Windows, NamedTemporaryFile cannot be reopened while open, so we
+        # use delete=False and manually clean up.
+        if _is_windows():
+            tmp_in = tempfile.NamedTemporaryFile(delete=False, suffix='.o')
+            tmp_out = tempfile.NamedTemporaryFile(delete=False, suffix='.hsaco')
+            try:
+                tmp_in.write(hsaco)
+                tmp_in.close()
+                tmp_out.close()
                 amd.link_hsaco(tmp_in.name, tmp_out.name)
-            with open(tmp_out.name, "rb") as fd_out:
-                ret = fd_out.read()
+                with open(tmp_out.name, "rb") as fd_out:
+                    ret = fd_out.read()
+            finally:
+                try:
+                    os.unlink(tmp_in.name)
+                except OSError:
+                    pass
+                try:
+                    os.unlink(tmp_out.name)
+                except OSError:
+                    pass
+        else:
+            with tempfile.NamedTemporaryFile() as tmp_out:
+                with tempfile.NamedTemporaryFile() as tmp_in:
+                    with open(tmp_in.name, "wb") as fd_in:
+                        fd_in.write(hsaco)
+                    amd.link_hsaco(tmp_in.name, tmp_out.name)
+                with open(tmp_out.name, "rb") as fd_out:
+                    ret = fd_out.read()
         return ret
 
     def add_stages(self, stages, options, language):

third_party/amd/backend/driver.c

@@ -3,11 +3,42 @@
 #include <hip/hip_runtime_api.h>
 #define PY_SSIZE_T_CLEAN
 #include <Python.h>
-#include <dlfcn.h>
 #include <stdbool.h>
 #include <stdio.h>
 #include <stdlib.h>
 
+#ifdef _WIN32
+#include <windows.h>
+// Windows compatibility layer for dlopen/dlsym/dlclose/dlerror
+#define RTLD_NOW 0
+#define RTLD_LAZY 0
+#define RTLD_LOCAL 0
+static char dlerror_buf[512];
+static inline void *dlopen(const char *filename, int flags) {
+    (void)flags;
+    HMODULE h = LoadLibraryA(filename);
+    if (!h) {
+        snprintf(dlerror_buf, sizeof(dlerror_buf), "LoadLibrary failed with error %lu", GetLastError());
+    }
+    return (void *)h;
+}
+static inline void *dlsym(void *handle, const char *symbol) {
+    void *p = (void *)GetProcAddress((HMODULE)handle, symbol);
+    if (!p) {
+        snprintf(dlerror_buf, sizeof(dlerror_buf), "GetProcAddress failed for %s with error %lu", symbol, GetLastError());
+    }
+    return p;
+}
+static inline int dlclose(void *handle) {
+    return FreeLibrary((HMODULE)handle) ? 0 : -1;
+}
+static inline const char *dlerror(void) {
+    return dlerror_buf[0] ? dlerror_buf : NULL;
+}
+#else
+#include <dlfcn.h>
+#endif
+
 // The list of paths to search for the HIP runtime library. The caller Python
 // code should substitute the search path placeholder.
 static const char *hipLibSearchPaths[] = {"/*py_libhip_search_path*/"};