Roll-up of minimal changes to TensileLite needed to build the library on Windows.

CMakeLists.txt

@@ -200,7 +200,11 @@ else()
 
     set( Tensile_LOGIC "asm_full" CACHE STRING "Tensile to use which logic?")
     set( Tensile_CODE_OBJECT_VERSION "4" CACHE STRING "Tensile code_object_version")
-    set( Tensile_COMPILER "amdclang++" CACHE STRING "Tensile compiler")
+    if(WIN32)
+      set( Tensile_COMPILER "clang++.exe" CACHE STRING "Tensile compiler")
+    else()
+      set( Tensile_COMPILER "amdclang++" CACHE STRING "Tensile compiler")
+    endif()
     set( Tensile_LIBRARY_FORMAT "msgpack" CACHE STRING "Tensile library format")
     set( Tensile_CPU_THREADS "" CACHE STRING "Number of threads for Tensile parallel build")
 

tensilelite/Tensile/Common/GlobalParameters.py

@@ -551,7 +551,8 @@ def assignGlobalParameters(config, isaInfoMap: Dict[IsaVersion, IsaInfo]):
     globalParameters["ROCmBinPath"] = os.path.join(globalParameters["ROCmPath"], "bin")
     globalParameters["ROCmSMIPath"] = locateExe(globalParameters["ROCmBinPath"], "rocm-smi")
     globalParameters["ROCmLdPath"] = locateExe(
-        os.path.join(globalParameters["ROCmPath"], "llvm/bin"), "ld.lld"
+        os.path.join(globalParameters["ROCmPath"], "lib/llvm/bin"),
+        "ld.lld" if os.name != "nt" else "ld.lld.exe"
     )
 
     if "AsanBuild" in config:
@@ -582,7 +583,10 @@ def assignGlobalParameters(config, isaInfoMap: Dict[IsaVersion, IsaInfo]):
     try:
         compiler = "hipcc"
         output = subprocess.run(
-            [compiler, "--version"], check=True, stdout=subprocess.PIPE
+            [compiler, "--version"], check=True,
+            stdout=subprocess.PIPE,
+            # Avoids some warning spam on Windows.
+            stderr=subprocess.DEVNULL,
         ).stdout.decode()
 
         for line in output.split("\n"):

tensilelite/Tensile/Common/Parallel.py

@@ -48,7 +48,10 @@ def CPUThreadCount(enable=True):
         return 1
     else:
         if os.name == "nt":
-            cpu_count = os.cpu_count()
+            # Windows supports at most 61 workers because the scheduler uses
+            # WaitForMultipleObjects directly, which has the limit (the limit
+            # is actually 64, but some handles are needed for accounting).
+            cpu_count = min(os.cpu_count(), 61)
         else:
             cpu_count = len(os.sched_getaffinity(0))
         cpuThreads = globalParameters["CpuThreads"]

tensilelite/Tensile/Common/Utilities.py

@@ -92,9 +92,10 @@ def isExe(filePath):
 
 def locateExe(defaultPath, exeName):  # /opt/rocm/bin, hip-clang
     # look in defaultPath first
-    exePath = os.path.join(defaultPath, exeName)
-    if isExe(exePath):
-        return exePath
+    if defaultPath:
+        exePath = os.path.join(defaultPath, exeName)
+        if isExe(exePath):
+            return exePath
     # look in PATH second
     for path in os.environ["PATH"].split(os.pathsep):
         exePath = os.path.join(path, exeName)

tensilelite/Tensile/Toolchain/Component.py

@@ -286,13 +286,14 @@ def compress(self, srcPath: str, destPath: str, target: str):
         Raises:
             RuntimeError: If compressing the code object file fails.
         """
+        devnull = "/dev/null" if os_name != "nt" else "NUL"
         args = [
             self._component_path,
             "--compress",
             "--type=o",
             "--bundle-align=4096",
             f"--targets=host-x86_64-unknown-linux-gnu,hipv4-amdgcn-amd-amdhsa-unknown-{target}",
-            "--input=/dev/null",
+            f"--input={devnull}",
             f"--input={srcPath}",
             f"--output={destPath}",
         ]
@@ -359,8 +360,7 @@ def __call__(self, srcPaths: List[str], destPath: str):
         if os_name == "nt":
             # Use args file on Windows b/c the command may exceed the limit of 8191 characters
             with open(Path.cwd() / "clang_args.txt", "wt") as file:
-                file.write(" ".join(srcPaths))
-                file.flush()
+                file.write(" ".join(srcPaths).replace('\\', '\\\\'))
             args = [*(self.default_args), "-o", destPath, "@clang_args.txt"]
         else:
             args = [*(self.default_args), *srcPaths, "-o", destPath]

tensilelite/Tensile/Toolchain/Validators.py

@@ -113,7 +113,7 @@ class ToolchainDefaults(NamedTuple):
     OFFLOAD_BUNDLER = osSelect(linux="clang-offload-bundler", windows="clang-offload-bundler.exe")
     DEVICE_ENUMERATOR = osSelect(linux="rocm_agent_enumerator" if isRhel8() else "amdgpu-arch", windows="hipinfo")
     ASSEMBLER = osSelect(linux="amdclang++", windows="clang++.exe")
-    HIP_CONFIG = osSelect(linux="hipconfig", windows="hipconfig")
+    HIP_CONFIG = osSelect(linux="hipconfig", windows="hipconfig.exe")
 
 
 def _supportedComponent(component: str, targets: List[str]) -> bool:

tensilelite/Tensile/cmake/TensileConfig.cmake

@@ -232,6 +232,7 @@ function(TensileCreateLibraryFiles
 
       add_custom_command(
         COMMENT "Generating Tensile Libraries"
+        USES_TERMINAL
         OUTPUT ${Tensile_OUTPUT_PATH}/library
         COMMAND ${CommandLine}
       )

tensilelite/rocisa/CMakeLists.txt

@@ -21,7 +21,7 @@
 #
 # ########################################################################
 cmake_minimum_required(VERSION 3.15)
-project(rocisa LANGUAGES HIP CXX)
+project(rocisa LANGUAGES CXX)
 set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)