Use this skill when the user asks to configure or build hipDNN through the rocm-libraries repository superbuild. It builds only; use hipdnn-superbuild-test for tests after a successful build.
-
Determine the repository root:
git rev-parse --show-toplevel
-
Choose the build and log locations:
- First honor any active workspace or repository instructions for artifact directories and build output safety.
- If no such instructions exist, use
BUILD_DIR=<repo-root>/build.
- Keep full configure/build output in a log file and show only a short tail on failure.
-
Locate this skill's helper directory. Skills are host-level, not tied to a repo checkout — default to the scripts bundled with the skill you were invoked from (<skill-directory>/scripts), even when you are working inside a repo or worktree. Do NOT run the <repo-root>/projects/hipdnn/tools/ai/skills/hipdnn-superbuild/scripts copy just because a checkout is present: that copy can be a stale stub (on develop) or an unmerged in-progress version (on a feature branch). Use the source-checkout copy only when you are actively developing this skill itself and intend to exercise your in-progress edits, or when the invoked skill has no bundled scripts/ directory.
-
Resolve ROCm and Clang paths (Windows also provisions the ROCm SDK wheels when missing):
python3 <scripts>/windows_rocm_setup.py --repo-root <repo-root> [--venv-path <path>] [--rocm-path <path>] [--clang-path <path>] [--gpu-targets <arch>] [--sha <commit>] [--provision auto|always|never]
On Linux this echoes only provided overrides. On Windows it validates the wheel-based ROCm install and, when the SDK is absent (or --provision always), creates the venv and pip-installs the ROCm SDK wheels before printing KEY=VALUE lines on stdout. Progress goes to stderr, so stdout carries only the ROCM_PATH=/CLANG_PATH=/GPU_TARGETS= lines. Clang is a prerequisite and is not provisioned; a missing clang is reported as an error.
-
If a clean rebuild was requested, remove the selected build directory using the active host's normal approval/safety flow.
-
Configure from the repository root. Always bind the preset configure to the selected build directory so configure and build operate on the same tree:
cmake --preset <preset> -B <build-dir> [extra -D options]
Add -DROCM_PATH=<path> when a ROCm path is resolved or provided. On Windows also add -DCMAKE_PROGRAM_PATH=<clang-path> and -DGPU_TARGETS=<arch>.
-
Build with output redirected to a log:
cmake --build <build-dir> > <log> 2>&1
If explicit jobs are allowed and requested, pass them through to CMake/Ninja. On failure, report the log path and tail the last relevant lines.
-
If the build fails with a stale CMake cache error such as does not match the source, clean the selected build directory once, reconfigure with the same -B <build-dir> command, and retry once. Do not loop.
-
On Windows, always stage the wheel's amd_comgr.dll app-local into <build-dir>/bin after a successful build:
python3 <scripts>/comgr_stage.py --rocm-bin <rocm-bin> --build-dir <build-dir> --verbose
The AMD driver leaves an old amd_comgr.dll in C:\Windows\System32 that outranks the wheel's copy on PATH, so MIOpen otherwise loads stale comgr and can fail to JIT-build kernels at runtime (GCN-assembly Winograd solvers are the common example, but the mismatch is not limited to them). Do this on every Windows build rather than only when a specific kernel path is expected. The Win32 loader checks the executable's own directory before System32, so an app-local copy in <build-dir>/bin wins; PATH manipulation alone cannot. The helper compares the wheel comgr's PE version against any already-staged copy and skips the copy when the versions match (content-hash fallback when version metadata is absent), so it is cheap to re-run. This step is a no-op on Linux. The test runner (cmake_run.py) stages comgr on its own as well, so this build step is belt-and-suspenders that makes the app-local copy present immediately after build.
