ci: align relocatable package flow with RVS reference

.github/workflows/README_BUILD_PACKAGES.md

@@ -93,13 +93,26 @@ sudo rpm -i --replacefiles --nodeps build/amdrocm7-transferbench-*.rpm
 
 ### Any Linux (TGZ — relocatable install tree, requires ROCm runtime on target)
 
+End-user instructions (pre-install ROCm, runtime dependencies, extract,
+`PATH` / `LD_LIBRARY_PATH`, troubleshooting) live in the project docs at
+[docs/install/INSTALL_TGZ.rst](../../docs/install/INSTALL_TGZ.rst).
+
+Quick smoke test from the repo root after a successful build:
+
 ```bash
 sudo mkdir -p /opt/rocm/extras-7
 sudo tar -xzf build/amdrocm7-transferbench-*.tar.gz -C /opt/rocm/extras-7 --strip-components=1
 export PATH=/opt/rocm/extras-7/bin:$PATH
+# With no args, TransferBench prints version, usage, available presets,
+# and detected topology — a fast end-to-end check that the binary loads
+# its ROCm libs correctly.
 TransferBench
 ```
 
+> Once the `help` and `smoketest` presets land on `develop` (currently on
+> `candidate`), prefer `TransferBench help` for usage and `TransferBench
+> smoketest` for a real correctness check.
+
 ## S3 upload (OIDC)
 
 S3 upload runs only when:

CMakeLists.txt

@@ -9,7 +9,41 @@ if (NOT CMAKE_TOOLCHAIN_FILE)
   message(STATUS "CMAKE_TOOLCHAIN_FILE: ${CMAKE_TOOLCHAIN_FILE}")
 endif()
 
-set(VERSION_STRING "1.66.02")
+set(TRANSFERBENCH_VERSION_MAJOR 1)
+set(TRANSFERBENCH_VERSION_MINOR 66)
+set(TRANSFERBENCH_VERSION_PATCH_FALLBACK "02")
+
+# Auto-compute patch from git: count commits since the last v<MAJOR>.<MINOR>.* tag.
+# Falls back to TRANSFERBENCH_VERSION_PATCH_FALLBACK when git is unavailable,
+# this is not a git checkout, or no matching tag exists. Mirrors the RVS flow.
+set(TRANSFERBENCH_VERSION_PATCH "${TRANSFERBENCH_VERSION_PATCH_FALLBACK}")
+find_package(Git QUIET)
+if(GIT_FOUND AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
+  execute_process(
+    COMMAND "${GIT_EXECUTABLE}" describe --tags --abbrev=0 --match
+            "v${TRANSFERBENCH_VERSION_MAJOR}.${TRANSFERBENCH_VERSION_MINOR}.*"
+    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
+    OUTPUT_VARIABLE _tb_last_tag
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+    RESULT_VARIABLE _tb_describe_rc
+    ERROR_QUIET)
+  if(_tb_describe_rc EQUAL 0 AND _tb_last_tag)
+    execute_process(
+      COMMAND "${GIT_EXECUTABLE}" rev-list --count "${_tb_last_tag}..HEAD"
+      WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
+      OUTPUT_VARIABLE _tb_commit_count
+      OUTPUT_STRIP_TRAILING_WHITESPACE
+      RESULT_VARIABLE _tb_count_rc
+      ERROR_QUIET)
+    if(_tb_count_rc EQUAL 0 AND _tb_commit_count MATCHES "^[0-9]+$")
+      set(TRANSFERBENCH_VERSION_PATCH "${_tb_commit_count}")
+    endif()
+  endif()
+endif()
+
+set(VERSION_STRING
+    "${TRANSFERBENCH_VERSION_MAJOR}.${TRANSFERBENCH_VERSION_MINOR}.${TRANSFERBENCH_VERSION_PATCH}")
+message(STATUS "TransferBench version: ${VERSION_STRING}")
 project(TransferBench VERSION ${VERSION_STRING} LANGUAGES CXX)
 
 ## Load CMake modules
@@ -272,6 +306,22 @@ if(BUILD_RELOCATABLE_PACKAGE)
     set(ROCM_MAJOR_VERSION "7")
   endif()
 
+  # Relocatable RPATH (matches the RVS reference flow). Mirrors what
+  # build_packages_local.sh used to inject via -DCMAKE_INSTALL_RPATH=, so
+  # plain `cmake -DBUILD_RELOCATABLE_PACKAGE=ON ..` now produces the same
+  # RPATH as a CI/packaged build.
+  set(CMAKE_SKIP_RPATH FALSE)
+  set(CMAKE_INSTALL_RPATH_USE_LINK_PATH FALSE)
+  set(CMAKE_INSTALL_RPATH
+      "\$ORIGIN:\$ORIGIN/../lib:/opt/rocm/extras-${ROCM_MAJOR_VERSION}/lib:/opt/rocm/lib:/opt/rocm/lib/llvm/lib:/opt/rocm/core-${ROCM_MAJOR_VERSION}/lib:/opt/rocm/core-${ROCM_MAJOR_VERSION}/lib/llvm/lib")
+  set(CMAKE_BUILD_RPATH "${CMAKE_INSTALL_RPATH}")
+  # Strip implicit SDK-from-build-host paths on install so the ephemeral
+  # $HOME/rocm-sdk/install path the CI script uses does not leak into the
+  # packaged binary's RPATH.
+  if(NOT CMAKE_VERSION VERSION_LESS "3.16")
+    set(CMAKE_INSTALL_REMOVE_ENVIRONMENT_RPATH TRUE)
+  endif()
+
   install(TARGETS TransferBench RUNTIME DESTINATION bin COMPONENT devel)
 
   set(CPACK_PACKAGE_NAME "amdrocm${ROCM_MAJOR_VERSION}-transferbench")

build_packages_local.sh

@@ -167,29 +167,38 @@ printf -v ROCM_LIBPATCH_VERSION '%02d%02d' "${ROCM_MAJOR}" "${ROCM_MINOR}"
 export ROCM_MAJOR ROCM_MINOR ROCM_LIBPATCH_VERSION
 log "ROCm major=${ROCM_MAJOR} minor=${ROCM_MINOR} libpatch=${ROCM_LIBPATCH_VERSION}"
 
-# Package release string: branch.commit for dev, run_number for release branches
+# Package release string. Format mirrors the RVS reference flow:
+#   default (push/schedule/dispatch/local): r<libpatch>.<yyyymmdd>
+#   pull request:                           r<libpatch>.<yyyymmdd>.<src-branch>.<commit>
+#   release/* branch (non-PR):              ${GITHUB_RUN_NUMBER} (fallback 1)
 GIT_BRANCH="${GITHUB_REF_NAME:-$(git -C "${REPO_ROOT}" rev-parse --abbrev-ref HEAD 2>/dev/null || echo unknown)}"
 GIT_COMMIT="$(git -C "${REPO_ROOT}" rev-parse --short HEAD 2>/dev/null || echo unknown)"
-if [[ "${GIT_BRANCH}" == rel* ]] || [[ "${GIT_BRANCH}" == release/* ]]; then
-  PKG_RELEASE="${GITHUB_RUN_NUMBER}"
+BUILD_DATE_UTC="$(date -u +%Y%m%d)"
+
+# Collapse non-alphanumerics into single dots and trim — DEB/RPM release
+# fields reject most punctuation.
+sanitize_release() {
+  local s
+  s="$(printf '%s' "$1" | sed -E 's/[^[:alnum:]]+/./g; s/^\.+//; s/\.+$//')"
+  printf '%s' "${s:-unknown}"
+}
+
+if [[ "${GITHUB_EVENT_NAME:-}" == "pull_request" ]]; then
+  PR_BRANCH="$(sanitize_release "${GITHUB_HEAD_REF:-${GIT_BRANCH}}")"
+  PKG_RELEASE="r${ROCM_LIBPATCH_VERSION}.${BUILD_DATE_UTC}.${PR_BRANCH}.${GIT_COMMIT}"
+elif [[ "${GIT_BRANCH}" == release/* ]]; then
+  PKG_RELEASE="${GITHUB_RUN_NUMBER:-1}"
 else
-  # Sanitize: DEB/RPM release fields disallow many punctuation chars.
-  # Collapse anything that's not [A-Za-z0-9] into a single dot, then trim.
-  SAFE_BRANCH="$(printf '%s' "${GIT_BRANCH}" | sed -E 's/[^[:alnum:]]+/./g; s/^\.+//; s/\.+$//')"
-  SAFE_BRANCH="${SAFE_BRANCH:-unknown}"
-  PKG_RELEASE="${SAFE_BRANCH}.${GIT_COMMIT}"
+  PKG_RELEASE="r${ROCM_LIBPATCH_VERSION}.${BUILD_DATE_UTC}"
 fi
 export CPACK_DEBIAN_PACKAGE_RELEASE="${CPACK_DEBIAN_PACKAGE_RELEASE:-$PKG_RELEASE}"
 export CPACK_RPM_PACKAGE_RELEASE="${CPACK_RPM_PACKAGE_RELEASE:-$PKG_RELEASE}"
 log "Package release tag: ${PKG_RELEASE}"
 
 # -------- configure --------
 INSTALL_PREFIX="/opt/rocm/extras-${ROCM_MAJOR}"
-# Relocatable RPATH: $ORIGIN-relative + install prefix + the conventional
-# install-time ROCm locations. Do NOT embed ${ROCM_PATH} (the ephemeral
-# build-time SDK download path) — that would leak CI paths into the
-# packaged binary and break relocatability.
-RPATH_LIST="\$ORIGIN:\$ORIGIN/../lib:${INSTALL_PREFIX}/lib:/opt/rocm/lib:/opt/rocm/lib64"
+# Relocatable RPATH defaults live in CMakeLists.txt under
+# if(BUILD_RELOCATABLE_PACKAGE); enabling that option below activates them.
 
 log "Configuring CMake..."
 rm -rf "${BUILD_DIR}"
@@ -204,9 +213,6 @@ CMAKE_ARGS=(
   -DHIP_PLATFORM=amd
   -DCMAKE_INSTALL_PREFIX="${INSTALL_PREFIX}"
   -DCPACK_PACKAGING_INSTALL_PREFIX="${INSTALL_PREFIX}"
-  -DCMAKE_SKIP_RPATH=FALSE
-  -DCMAKE_INSTALL_RPATH_USE_LINK_PATH=FALSE
-  -DCMAKE_INSTALL_RPATH="${RPATH_LIST}"
   -DCMAKE_VERBOSE_MAKEFILE=ON
   -DBUILD_RELOCATABLE_PACKAGE=ON
   -DBUILD_LOCAL_GPU_TARGET_ONLY=OFF

docs/install/INSTALL_TGZ.rst

@@ -0,0 +1,136 @@
+:orphan:
+
+.. meta::
+  :description: Install the relocatable TransferBench TGZ archive on any Linux distribution
+  :keywords: TransferBench, TGZ, tarball, install, relocatable
+
+.. _install-transferbench-tgz:
+
+------------------------------------------------
+Installing TransferBench from the TGZ archive
+------------------------------------------------
+
+The TransferBench TGZ archive (``amdrocm<MAJOR>-transferbench-*.tar.gz``) is a
+relocatable install tree that works on any Linux distribution where a
+compatible ROCm runtime is already present. Use it when you cannot or do not
+want to install the DEB or RPM package — for example on a distribution
+without a native ROCm package, or inside a non-root container.
+
+The TGZ ships only the ``TransferBench`` binary and its supporting files. It
+does **not** bundle ROCm; the host system must already provide the ROCm
+runtime libraries (``hsa-rocr`` and the HIP runtime).
+
+Pre-install: ROCm
+-----------------
+
+Install ROCm on the target system before extracting the TGZ. Follow the
+official AMD documentation:
+
+* `ROCm documentation <https://rocm.docs.amd.com/>`_
+* `Linux install guide <https://rocm.docs.amd.com/projects/install-on-linux/en/latest/>`_
+
+After installing, ``ROCM_PATH`` (typically ``/opt/rocm``) must be set
+correctly and the ROCm libraries must be loadable by the dynamic linker.
+
+Install runtime dependencies
+----------------------------
+
+The DEB and RPM packages declare these runtime dependencies; TGZ users must
+install them manually on the target host.
+
+.. list-table::
+   :header-rows: 1
+   :widths: 30 70
+
+   * - Family
+     - Required packages
+   * - Debian / Ubuntu
+     - ``numactl``, ``libnuma1``, plus the ROCm runtime (``hsa-rocr``)
+   * - RHEL / Rocky / AlmaLinux
+     - ``numactl``, plus the ROCm runtime (``hsa-rocr``)
+
+Install commands:
+
+.. code-block:: bash
+
+  # Ubuntu / Debian
+  sudo apt update && sudo apt install -y numactl libnuma1
+
+  # RHEL / Rocky / AlmaLinux
+  sudo dnf install -y numactl
+
+The ROCm packages (``hsa-rocr`` and friends) come from the ROCm repo
+configured in the pre-install step above.
+
+Extract the TGZ
+---------------
+
+Extract the archive into ``/opt/rocm/extras-<MAJOR>``, where ``<MAJOR>`` is
+the ROCm major version the package was built against (encoded in the package
+name, for example ``amdrocm7-transferbench-*.tar.gz`` → major ``7``).
+
+.. code-block:: bash
+
+  # Example for ROCm major 7 — match your package
+  sudo mkdir -p /opt/rocm/extras-7
+  sudo tar -xzf amdrocm7-transferbench-*.tar.gz -C /opt/rocm/extras-7 --strip-components=1
+
+The ``--strip-components=1`` option discards the top-level directory inside
+the tarball so files land directly under ``/opt/rocm/extras-7/{bin,lib,...}``.
+
+Configure ``PATH`` and ``LD_LIBRARY_PATH``
+------------------------------------------
+
+Point the shell at the extracted prefix and your ROCm install. Copy and paste
+the block as one unit (replace paths with your real ``ROCM_PATH`` and major
+version):
+
+.. code-block:: bash
+
+  export ROCM_PATH=/opt/rocm                          # or your real ROCm root
+  export PATH=/opt/rocm/extras-7/bin:$ROCM_PATH/bin:$PATH
+  export LD_LIBRARY_PATH=/opt/rocm/extras-7/lib:$ROCM_PATH/lib:$ROCM_PATH/lib/llvm/lib:$LD_LIBRARY_PATH
+
+The ``TransferBench`` binary embeds an ``RPATH`` covering ``$ORIGIN``,
+``$ORIGIN/../lib``, ``/opt/rocm/extras-<MAJOR>/lib``, ``/opt/rocm/lib``,
+``/opt/rocm/lib/llvm/lib``, ``/opt/rocm/core-<MAJOR>/lib``, and
+``/opt/rocm/core-<MAJOR>/lib/llvm/lib``. The ``LD_LIBRARY_PATH`` export above
+is mainly defensive — useful if your ROCm tree lives somewhere non-standard
+or if you want to override which copy of a library is loaded for
+troubleshooting.
+
+Verify the install
+------------------
+
+.. code-block:: bash
+
+  TransferBench
+
+Run with no arguments, ``TransferBench`` prints its version, usage, the
+list of available preset benchmarks, and the detected GPU/CPU topology,
+then exits. Seeing that output confirms the binary loaded its ROCm
+libraries correctly.
+
+If the binary fails to load a shared library, inspect:
+
+.. code-block:: bash
+
+  ldd /opt/rocm/extras-7/bin/TransferBench
+  readelf -d /opt/rocm/extras-7/bin/TransferBench | grep -E 'RPATH|RUNPATH'
+
+Make a persistent shell setup
+-----------------------------
+
+To avoid re-exporting every shell, drop the variables into a profile script:
+
+.. code-block:: bash
+
+  sudo tee /etc/profile.d/transferbench.sh >/dev/null <<'EOF'
+  export ROCM_PATH=/opt/rocm
+  export PATH=/opt/rocm/extras-7/bin:$ROCM_PATH/bin:$PATH
+  export LD_LIBRARY_PATH=/opt/rocm/extras-7/lib:$ROCM_PATH/lib:$ROCM_PATH/lib/llvm/lib:$LD_LIBRARY_PATH
+  EOF
+  sudo chmod 0644 /etc/profile.d/transferbench.sh
+
+Log out and back in (or ``source /etc/profile.d/transferbench.sh``) for the
+changes to apply.

docs/install/install.rst

@@ -83,3 +83,9 @@ TransferBench looks for NVCC in ``/usr/local/cuda`` by default. To modify the lo
 .. code-block:: bash
 
   CUDA_PATH=/usr/local/cuda make
+
+Installing from the relocatable TGZ archive
+-------------------------------------------
+
+If you want to install a pre-built TransferBench binary on a system where
+you cannot install the DEB or RPM package, see :ref:`install-transferbench-tgz`.