* Update documentation to include RO and how to use it

* Clarify supported configuration

Co-authored-by: yugang-amd <yugang.wang@amd.com>
Этот коммит содержится в:
Aurelien Bouteiller
2025-07-10 18:49:10 -04:00
коммит произвёл GitHub
родитель a66f782540
Коммит 42e28835ad
2 изменённых файлов: 61 добавлений и 4 удалений
+3
Просмотреть файл
@@ -84,3 +84,6 @@ You can control the behavior of rocSHMEM by using the following environment vari
* - ROCSHMEM_UNIQUEID_WITH_MPI
- 0
- Defines whether rocSHMEM is expected to use MPI when using the uniqueId based initialization.
* - ROCSHMEM_RO_DISABLE_IPC
- 0
- Defines whether to force using the RO conduit even when IPC is available.
+58 -4
Просмотреть файл
@@ -23,6 +23,23 @@ Requirements
* ROCm-aware Open MPI and UCX. For more information, see :ref:`install-dependencies`.
* Inter-node communication requires MPI, and is tested with Open MPI and CX7 Infiniband NICs.
Available network backends
--------------------------
rocSHMEM supports two network backends:
* The **IPC (Inter-Process Communication)** backend enables fast communication between GPUs on the same host using ROCm inter-process mechanisms. It does not support inter-node communication.
* The **RO (Reverse Offload)** backend enables communication between GPUs on different nodes through a NIC, using a host-based proxy to forward communication orders to and from the GPU. In this release, RO is the only inter-node communication backend and is built on an MPI-RMA compatibility layer.
You can activate IPC and RO backends in the same rocSHMEM build. In this case, IPC handles intra-node communication, while RO handles inter-node communication.
.. note::
When RO is active, all atomic operations use the RO backend, even for intra-node communication.
Installing from a package manager
---------------------------------
@@ -82,6 +99,37 @@ For more information about OpenMPI-UCX support, see
Installing from source
--------------------------------
You can select between two communication backends at build time for rocSHMEM: RO and IPC.
The default configuration enables both backends, using IPC for intra-node communication
RO for inter-node communication at runtime. In this configuration, rocSHMEM atomic operations always use the RO backend.
rocSHMEM also supports the IPC-only configuration, which allows atomic operations to use the IPC backend only.
RO and IPC backend build
^^^^^^^^^^^^^^^^^^^^
To build and install rocSHMEM with the hybrid RO (off-node) and IPC (on-node) backends, run:
.. code-block:: bash
git clone git@github.com:ROCm/rocSHMEM.git
cd rocSHMEM
mkdir build
cd build
../scripts/build_configs/ro_ipc
The build script passes configuration options to CMake to set up a canonical build.
.. note::
The only officially supported configuration for the RO backend uses Open MPI and UCX with a CX7 InfiniBand adapter. For more information, see :ref:`install-dependencies`. Other configurations, such as MPI implementations that are thread-safe and support GPU buffers, might work but are considered experimental.
IPC only backend build
^^^^^^^^^^^^^^^^^^^^^^
To build and install rocSHMEM with the IPC on-node, GPU-to-GPU backend, run:
.. code-block:: bash
@@ -92,16 +140,22 @@ To build and install rocSHMEM with the IPC on-node, GPU-to-GPU backend, run:
cd build
../scripts/build_configs/ipc_single
The build script passes configuration options to CMake to setup a canonical build.
The build script passes configuration options to CMake to setup a single-node build.
This is similar to the default build in ROCm 6.4.
.. note::
Other experimental configuration scripts are available in ``./scripts/build_configs``, but only ``ipc_single`` is currently supported.
The default configuration changed from IPC only in ROCm 6.4 (built with the ``ipc_single`` script) to RO and IPC in ROCm 7.0 (built with the ``ro_ipc`` script).
Other experimental configuration scripts are available in ``./scripts/build_configs``, but only ``ipc_single`` and ``ro_ipc``
are officially supported.
Installation prefix
^^^^^^^^^^^^^^^^^^^
By default, the library is installed in ``~/rocshmem``. You can customize the installation path by running:
By default, the build scripts install the library to ``~/rocshmem``. You can customize the installation path by adding
the desired path as the script parameter. For example, to relocate the default configuration:
.. code-block:: bash
../scripts/build_configs/ipc_single /path/to/install
../scripts/build_configs/ro_ipc /path/to/install