Find non-default HIP SDK on Windows

Tensile/BuildCommands/AssemblyCommands.py

@@ -74,7 +74,6 @@ def buildAssemblyCodeObjectFiles(
     maxLineLength = (
         8191 if os.name == "nt" else int(subprocess.check_output(["getconf", "ARG_MAX"]).strip())
     )
-    tPrint(0, f"Maximum command line length: {maxLineLength}")
 
     assemblyKernels = [k for k in kernels if k["KernelLanguage"] == "Assembly"]
     if len(assemblyKernels) == 0:

Tensile/Common.py

@@ -2424,7 +2424,7 @@ def assignGlobalParameters( config, capabilitiesCache: Optional[dict] = None ):
   if "TENSILE_ROCM_PATH" in os.environ:
     globalParameters["ROCmPath"] = os.environ.get("TENSILE_ROCM_PATH")
   if os.name == "nt" and "HIP_DIR" in os.environ:
-    globalParameters["ROCmPath"] = os.environ.get("HIP_DIR") # windows has no ROCM
+    globalParameters["ROCmPath"] = os.environ.get("HIP_PATH") # windows has no ROCM
   globalParameters["CmakeCxxCompiler"] = None
   if "CMAKE_CXX_COMPILER" in os.environ:
     globalParameters["CmakeCxxCompiler"] = os.environ.get("CMAKE_CXX_COMPILER")

Tensile/TensileCreateLib/ParseArguments.py

@@ -79,7 +79,8 @@ def parseArguments(input: Optional[List[str]] = None) -> Dict[str, Any]:
         default=ToolchainDefaults.CXX_COMPILER,
         type=str,
         help="C++ compiler used when generating binaries."
-        " On linux, amdclang++ (default) or hipcc. On Windows clang++ (default) or hipcc.",
+        " On linux, amdclang++ (default) or hipcc. On Windows clang++.exe (default) or hipcc. "
+        "If clang++.exe is specified, it must include the extension `.exe`.",
     )
     parser.add_argument(
         "--c-compiler",

Tensile/Utilities/Toolchain.py

@@ -1,50 +1,91 @@
 import os
 import re
 from pathlib import Path
+from subprocess import PIPE, run
 from typing import List, NamedTuple, Union
 from warnings import warn
-from subprocess import run, PIPE
-
-ROCM_BIN_PATH = Path("/opt/rocm/bin")
-ROCM_LLVM_BIN_PATH = Path("/opt/rocm/lib/llvm/bin")
-
-if os.name == "nt":
-    def _windowsLatestRocmBin(path: Union[Path, str]) -> Path:
-        """Get the path to the latest ROCm bin directory, on Windows.
-
-        This function assumes that ROCm versions are differentiated with the form ``X.Y``.
-
-        Args:
-            path: The path to the ROCm root directory, typically ``C:/Program Files/AMD/ROCm``.
-
-        Returns:
-            The path to the ROCm bin directory for the latest ROCm version.
-            Typically of the form ``C:/Program Files/AMD/ROCm/X.Y/bin``.
-        """
-        path = Path(path)
-        pattern = re.compile(r'^\d+\.\d+$')
-        versions = filter(lambda d: d.is_dir() and pattern.match(d.name), path.iterdir())
-        latest = max(versions, key=lambda d: tuple(map(int, d.name.split('.'))))
-        return latest / "bin"
-    # LLVM binaries are in the same directory as ROCm binaries on Windows
-    ROCM_BIN_PATH = _windowsLatestRocmBin("C:/Program Files/AMD/ROCm")
-    ROCM_LLVM_BIN_PATH = _windowsLatestRocmBin("C:/Program Files/AMD/ROCm")
+
+DEFAULT_ROCM_BIN_PATH_POSIX = Path("/opt/rocm/bin")
+DEFAULT_ROCM_LLVM_BIN_PATH_POSIX = Path("/opt/rocm/lib/llvm/bin")
+DEFAULT_ROCM_BIN_PATH_WINDOWS = Path("C:/Program Files/AMD/ROCm")
 
 
 osSelect = lambda linux, windows: linux if os.name != "nt" else windows
 
 
+def _windowsLatestRocmBin(path: Union[Path, str]) -> Path:
+    """Get the path to the latest ROCm bin directory, on Windows.
+
+    This function assumes that ROCm versions are differentiated with the form ``X.Y``.
+
+    Args:
+        path: The path to the ROCm root directory, typically ``C:/Program Files/AMD/ROCm``.
+
+    Returns:
+        The path to the ROCm bin directory for the latest ROCm version.
+        Typically of the form ``C:/Program Files/AMD/ROCm/X.Y/bin``.
+    """
+    path = Path(path)
+    pattern = re.compile(r"^\d+\.\d+$")
+    versions = filter(lambda d: d.is_dir() and pattern.match(d.name), path.iterdir())
+    latest = max(versions, key=lambda d: tuple(map(int, d.name.split("."))))
+    return latest / "bin"
+
+
+def _windowsSearchPaths() -> List[Path]:
+    defaultPath = DEFAULT_ROCM_BIN_PATH_WINDOWS
+    searchPaths = []
+
+    if os.environ.get("HIP_PATH"):
+        hipPaths = [Path(p) / "bin" for p in os.environ["HIP_PATH"].split(os.pathsep)]
+        searchPaths.extend(hipPaths)
+
+    if Path(defaultPath).exists():
+        searchPaths.append(_windowsLatestRocmBin(defaultPath))
+
+    if os.environ.get("PATH"):
+        envPath = [Path(p) for p in os.environ["PATH"].split(os.pathsep)]
+        searchPaths.extend(envPath)
+
+    return searchPaths
+
+
+def _posixSearchPaths() -> List[Path]:
+
+    searchPaths = []
+
+    if os.environ.get("ROCM_PATH"):
+        for p in os.environ["ROCM_PATH"].split(os.pathsep):
+            searchPaths.append(Path(p) / "bin")
+            searchPaths.append(Path(p) / "lib" / "llvm" / "bin")
+
+    searchPaths.extend(
+        [
+            DEFAULT_ROCM_BIN_PATH_POSIX,
+            DEFAULT_ROCM_LLVM_BIN_PATH_POSIX,
+        ]
+    )
+
+    if os.environ.get("PATH"):
+        envPath = [Path(p) for p in os.environ["PATH"].split(os.pathsep)]
+        searchPaths.extend(envPath)
+
+    return searchPaths
+
+
 class ToolchainDefaults(NamedTuple):
-    CXX_COMPILER= osSelect(linux="amdclang++", windows="clang++.exe")
-    C_COMPILER= osSelect(linux="amdclang", windows="clang.exe")
-    OFFLOAD_BUNDLER= osSelect(linux="clang-offload-bundler", windows="clang-offload-bundler.exe")
+    CXX_COMPILER = osSelect(linux="amdclang++", windows="clang++.exe")
+    C_COMPILER = osSelect(linux="amdclang", windows="clang.exe")
+    OFFLOAD_BUNDLER = osSelect(linux="clang-offload-bundler", windows="clang-offload-bundler.exe")
     ASSEMBLER = osSelect(linux="amdclang++", windows="clang++.exe")
     HIP_CONFIG = osSelect(linux="hipconfig", windows="hipconfig")
-    DEVICE_ENUMERATOR= osSelect(linux="rocm_agent_enumerator", windows="hipinfo.exe")
+    DEVICE_ENUMERATOR = osSelect(linux="rocm_agent_enumerator", windows="hipinfo.exe")
 
 
 def _supportedComponent(component: str, targets: List[str]) -> bool:
-    isSupported = any([component == t for t in targets]) or any([Path(component).name == t for t in targets])
+    isSupported = any([component == t for t in targets]) or any(
+        [Path(component).name == t for t in targets]
+    )
     return isSupported
 
 
@@ -119,7 +160,7 @@ def _exeExists(file: Path) -> bool:
     """
     if os.access(file, os.X_OK):
         if "rocm" not in map(str.lower, file.parts):
-            warn(f"Found non-ROCm install of `{file.name}`: {file}")
+            warn(f"Found `{file.name}` but in a non-default ROCm location: {file}")
         return True
     return False
 
@@ -134,48 +175,56 @@ def _validateExecutable(file: str, searchPaths: List[Path]) -> str:
     Returns:
         The validated executable with an absolute path.
     """
-    if not any((
-        supportedCxxCompiler(file),
-        supportedCCompiler(file),
-        supportedOffloadBundler(file),
-        supportedHip(file),
-        supportedDeviceEnumerator(file)
-    )):
+    if not any(
+        (
+            supportedCxxCompiler(file),
+            supportedCCompiler(file),
+            supportedOffloadBundler(file),
+            supportedHip(file),
+            supportedDeviceEnumerator(file),
+        )
+    ):
         raise ValueError(f"{file} is not a supported toolchain component for OS: {os.name}")
 
-    if _exeExists(Path(file)): return file
+    if _exeExists(Path(file)):
+        return file
     for path in searchPaths:
-        path /= file 
-        if _exeExists(path): return str(path)
-    raise FileNotFoundError(f"`{file}` either not found or not executable in any search path: {':'.join(map(str, searchPaths))}")
+        path /= file
+        if _exeExists(path):
+            return str(path)
+    raise FileNotFoundError(
+        f"`{file}` either not found or not executable in any search path: "
+        f"{':'.join(map(str, searchPaths))}\n"
+        "Please refer to the troubleshooting section of the Tensile documentation at "
+        "https://rocm.docs.amd.com/projects/Tensile/en/latest/src/#"
+    )
 
 
 def validateToolchain(*args: str):
     """Validate that the given toolchain components are in the PATH and executable.
 
     Args:
         args: List of executable toolchain components to validate.
-     
+
     Returns:
         List of validated executables with absolute paths.
-    
+
     Raises:
         ValueError: If no toolchain components are provided.
         FileNotFoundError: If a toolchain component is not found in the PATH.
     """
     if not args:
         raise ValueError("No toolchain components to validate, at least one argument is required")
 
-    searchPaths = [
-        ROCM_BIN_PATH,
-        ROCM_LLVM_BIN_PATH,
-    ] + [Path(p) for p in os.environ["PATH"].split(os.pathsep)]
+    searchPaths = _windowsSearchPaths() if os.name == "nt" else _posixSearchPaths()
 
     out = (_validateExecutable(x, searchPaths) for x in args)
-    return next(out) if len(args) == 1 else tuple(out) 
+    return next(out) if len(args) == 1 else tuple(out)
 
 
-def getVersion(executable: str, versionFlag: str="--version", regex: str=r'version\s+([\d.]+)') -> str:
+def getVersion(
+    executable: str, versionFlag: str = "--version", regex: str = r"version\s+([\d.]+)"
+) -> str:
     """Print the version of a toolchain component.
 
     Args:

docs/sphinx/_toc.yml.in

@@ -33,6 +33,10 @@ subtrees:
   - file: src/how-to/programmers-guide
   - file: src/how-to/contribution-guidelines
 
+- caption: Support
+  entries:
+  - file: src/support/troubleshooting
+
 - caption: About
   entries:
   - file: license.rst

docs/src/install/installation.rst

@@ -8,6 +8,8 @@
 Installation
 ********************************************************************
 
+.. _install-rocm:
+
 Install ROCm
 ============
 

docs/src/support/troubleshooting.rst

@@ -0,0 +1,38 @@
+.. meta::
+  :description: Tensile is a tool for creating a benchmark-driven backend library for GEMM
+  :keywords: Tensile developers guide, Tensile contributors guide, Tensile programmers guide, GEMM, Tensor
+.. highlight:: none
+
+.. _troubleshooting:
+
+********************************************************************
+Troubleshooting
+********************************************************************
+
+This topic provides information for programmers and users on how to resolve common issues in Tensile.
+
+============================
+Missing toolchain components
+============================
+
+.. code-block::
+
+    FileNotFoundError: `amdclang++` either not found or not executable in any search path
+
+In this situation, Tensile cannot locate one or more binaries required for proper program execution. This includes compilers, assemblers, linkers, and bundlers.
+
+.. note::
+
+   On Linux, the default installation location is */opt/rocm*. 
+
+.. note::
+
+   On Windows, the default installation location is *C:\\Program Files\\AMD\\ROCm\\X.Y*, where *X.Y* identifies the major and minor version of the current ROCm installation.
+   When the HIP SDK is installed on Windows, the variable ``HIP_PATH`` is set to the installation location of the HIP SDK.
+
+There are two possible reasons for this:
+
+1. ROCm is not installed on the system. To resolve this issue, install ROCm by following the instructions at :ref:`install-rocm`.
+2. ROCm is installed, but in a non-default location and the binaries cannot be found in the system PATH.
+   In this case, add the installation location to the ROCM_PATH on Linux, HIP_PATH on Windows, or the system PATH on either.
+   For example, on Linux use ``export ROCM_PATH=<path_to_rocm>/bin`` or on Windows PowerShell use ``$env:HIP_PATH = "<path_to_rocm>\bin"``.

tox.ini

@@ -71,6 +71,7 @@ commands =
       {toxinidir}/Tensile/TensileCreateLibrary.py \
       {toxinidir}/Tensile/TensileCreateLib/ \
       {toxinidir}/Tensile/BuildCommands/ \
+      {toxinidir}/Tensile/Utilities/Toolchain.py \
       {toxinidir}/Tensile/Tests/unit/test_TensileCreateLibrary.py \
       {toxinidir}/Tensile/Tests/unit/test_KernelFileContext.py \
       {toxinidir}/Tensile/Tests/unit/test_AsmRegisterPool.py \
@@ -88,6 +89,7 @@ commands =
       {toxinidir}/Tensile/TensileCreateLibrary.py \
       {toxinidir}/Tensile/TensileCreateLib/ \
       {toxinidir}/Tensile/BuildCommands/ \
+      {toxinidir}/Tensile/Utilities/Toolchain.py \
       {toxinidir}/Tensile/Tests/unit/test_TensileCreateLibrary.py \
       {toxinidir}/Tensile/Tests/unit/test_KernelFileContext.py \
       {toxinidir}/Tensile/Tests/unit/test_AsmRegisterPool.py \