Added ability to build for local gpu by env Variable (#204)

* Added the ability to compile for Local gpu by environment variable

* adding gfx950 on default only on rocm 7.0 and above

* Updated docs

* removed xnack+ on specific gfx targets

---------

Co-authored-by: Yiltan Hassan Temucin <yiltan.temucin@amd.com>

CMakeLists.txt

@@ -92,6 +92,7 @@ include(ROCMCheckTargetIds)
 
 rocm_setup_version(VERSION ${VERSION_STRING})
 
+
 #############################################################################
 # SET GPU ARCHITECTURES
 #############################################################################
@@ -100,10 +101,15 @@ include(cmake/rocm_local_targets.cmake)
 set(DEFAULT_GPUS
     gfx90a:xnack-;
     gfx90a:xnack+;
-    gfx942:xnack-;
-    gfx942:xnack+;
-    gfx950:xnack-;
-    gfx950:xnack+)
+    gfx942)
+
+if(${ROCM_MAJOR_VERSION} GREATER 6)
+  list(APPEND DEFAULT_GPUS gfx950)
+endif()
+
+if($ENV{BUILD_LOCAL_GPU_TARGET_ONLY})
+  set(BUILD_LOCAL_GPU_TARGET_ONLY ON)
+endif()
 
 if (BUILD_LOCAL_GPU_TARGET_ONLY)
   message(STATUS "Building only for local GPU target")

README.md

@@ -28,6 +28,7 @@ rocSHMEM base requirements:
 * AMD GPUs
   * MI250X
   * MI300X
+  * MI350X (Requires ROCm 7.0 or later)
 * ROCm-aware Open MPI and UCX as described in
   [Building the Dependencies](#building-the-dependencies)
 

cmake/setup_project.cmake

@@ -31,6 +31,37 @@ set(CMAKE_BUILD_TYPE "Release" CACHE STRING
 ###############################################################################
 # GLOBAL COMPILE FLAGS
 ###############################################################################
+
+# Try to establish ROCM_PATH (for find_package)
+#==================================================================================================
+if(NOT DEFINED ROCM_PATH)
+  # Guess default location
+  set(ROCM_PATH "/opt/rocm")
+  message(WARNING "Unable to find ROCM_PATH: Falling back to ${ROCM_PATH}")
+else()
+  message(STATUS "ROCM_PATH found: ${ROCM_PATH}")
+endif()
+set(ENV{ROCM_PATH} ${ROCM_PATH})
+
+## Check for ROCm version
+
+if(ROCM_PATH)
+  message(STATUS "Reading ROCM version from ${ROCM_PATH}/.info/version")
+  file(READ "${ROCM_PATH}/.info/version" rocm_version_string)
+else()
+  message(FATAL_ERROR "Could not determine ROCM version (set EXPLICIT_ROCM_VERSION or set ROCM_PATH to a valid installation)")
+endif()
+string(REGEX MATCH "([0-9]+)\\.([0-9]+)\\.([0-9]+)" rocm_version_matches ${rocm_version_string})
+if (rocm_version_matches)
+  set(ROCM_MAJOR_VERSION ${CMAKE_MATCH_1})
+  set(ROCM_MINOR_VERSION ${CMAKE_MATCH_2})
+  set(ROCM_PATCH_VERSION ${CMAKE_MATCH_3})
+
+  message(STATUS "ROCm version: ${ROCM_MAJOR_VERSION}.${ROCM_MINOR_VERSION}.${ROCM_PATCH_VERSION}")
+else()
+  message(WARNING "Failed to extract ROCm version.")
+endif()
+
 foreach (root ${hip_ROOT} $ENV{hip_ROOT} ${ROCM_ROOT} $ENV{ROCM_ROOT} ${ROCM_PATH} $ENV{ROCM_PATH})
   if (IS_DIRECTORY ${root})
     list(PREPEND CMAKE_PREFIX_PATH ${root})

docs/install.rst

@@ -21,6 +21,8 @@ Requirements
 
   * MI300X
 
+  * MI350X (Requires ROCm 7.0 or later)
+
 * 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.
@@ -34,7 +36,7 @@ rocSHMEM supports two network backends:
 * 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. 
+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::