Add GPU-aware MPI easyconfigs for ROCm

easybuild/easyconfigs/c/cassini-headers/cassini-headers-1.1.1.eb

@@ -0,0 +1,32 @@
+# Author:   Aayush Joglekar <aayush.joglekar@surf.nl>
+
+easyblock = 'Tarball'
+
+name = 'cassini-headers'
+version = '1.1.1'
+
+homepage = 'https://github.com/HewlettPackard/shs-cassini-headers'
+description = """Cassini ASIC register and user-API headers for the HPE
+Slingshot-11 interconnect. Consumed by libcxi and the libfabric cxi provider."""
+
+toolchain = {'name': 'system', 'version': 'system'}
+
+# git tag release/shs-12.0.1 corresponds to cray-cassini-headers-user 1.1.1
+sources = [
+    {
+        'filename': SOURCE_TAR_GZ,
+        'git_config': {
+            'url': 'https://github.com/HewlettPackard',
+            'repo_name': 'shs-cassini-headers',
+            'tag': 'release/shs-12.0.1',
+            'keep_git_dir': False,
+        },
+    },
+]
+
+sanity_check_paths = {
+    'files': ['include/cassini_user_defs.h', 'include/cxi_prov_hw.h'],
+    'dirs': ['include', 'share/cassini-headers'],
+}
+
+moduleclass = 'lib'

easybuild/easyconfigs/c/cxi-driver-headers/cxi-driver-headers-1.0.0.eb

@@ -0,0 +1,34 @@
+# Author:   Aayush Joglekar <aayush.joglekar@surf.nl>
+
+easyblock = 'Tarball'
+
+name = 'cxi-driver-headers'
+version = '1.0.0'
+
+homepage = 'https://github.com/HewlettPackard/shs-cxi-driver'
+description = """Userspace (UAPI) headers for the HPE Slingshot-11 CXI driver
+(uapi/misc/cxi.h and linux/cxi.h), defining the ioctl interface that libcxi and
+libfabric's cxi provider use to talk to the NIC. This module installs the headers
+only; the CXI kernel driver itself is supplied by the host operating system."""
+
+toolchain = {'name': 'system', 'version': 'system'}
+
+# git tag release/shs-12.0.2 corresponds to cray-cxi-driver 1.0.0
+sources = [
+    {
+        'filename': SOURCE_TAR_GZ,
+        'git_config': {
+            'url': 'https://github.com/HewlettPackard',
+            'repo_name': 'shs-cxi-driver',
+            'tag': 'release/shs-12.0.2',
+            'keep_git_dir': False,
+        },
+    },
+]
+
+sanity_check_paths = {
+    'files': ['include/uapi/misc/cxi.h'],
+    'dirs': ['include'],
+}
+
+moduleclass = 'lib'

easybuild/easyconfigs/h/hwloc-ROCm/hwloc-ROCm-2.11.2-rocm-compilers-19.0.0-ROCm-6.4.1.eb

@@ -0,0 +1,69 @@
+# Author:   Aayush Joglekar <aayush.joglekar@surf.nl>
+
+easyblock = 'ConfigureMake'
+
+name = 'hwloc-ROCm'
+version = '2.11.2'
+_rocm_version = '6.4.1'
+
+homepage = 'https://www.open-mpi.org/projects/hwloc/'
+
+description = """This provides the RSMI plugin for hwloc, which discovers AMD GPUs
+through the ROCm SMI library. It is built as a plugin on top of an existing hwloc
+installation and is loaded at runtime via HWLOC_PLUGINS_PATH; hwloc 2.11 has no
+separate ROCm plugin, RSMI is the only AMD GPU backend."""
+
+toolchain = {'name': 'rocm-compilers', 'version': '19.0.0-ROCm-6.4.1'}
+
+source_urls = ['https://www.open-mpi.org/software/hwloc/v2.11/downloads/']
+sources = ['hwloc-2.11.2.tar.gz']
+checksums = ['866ac8ef07b350a6a2ba0c6826c37d78e8994dcbcd443bdd2b436350de19d540']
+
+builddependencies = [
+    ('pkgconf', '2.3.0'),
+]
+
+dependencies = [
+    ('hwloc', version),
+    ('numactl', '2.0.19'),
+    ('rocm-smi', '7.6.0', f'-ROCm-{_rocm_version}'),  # GCCcore
+    ('HIP', '6.4.1'),
+]
+
+configopts = ' '.join([
+    '--enable-plugins',
+    '--disable-cuda',
+    '--disable-nvml',
+    '--with-rocm=$EBROOTHIP',
+    '--enable-rsmi',
+    '--disable-cairo',
+    '--disable-opencl',
+    '--disable-levelzero',
+    '--disable-gl',
+    '--disable-libxml2',
+    '--disable-pci',
+    '--disable-libudev',
+])
+
+# Build ONLY the AMD RSMI plugin
+build_cmd = 'cd hwloc && make hwloc_rsmi.la'
+install_cmd = f'cp hwloc/.libs/hwloc_rsmi.{SHLIB_EXT} %(installdir)s'
+
+modextrapaths = {
+    'HWLOC_PLUGINS_PATH': '.',
+}
+
+sanity_check_paths = {
+    'files': [f'hwloc_rsmi.{SHLIB_EXT}'],
+    'dirs': [],
+}
+
+# Verify that hwloc can successfully load the new AMD plugin
+sanity_check_commands = [
+    # rsmi backend registers at runtime
+    "HWLOC_COMPONENTS_VERBOSE=1 lstopo 2>&1 | grep -q 'rsmi'",
+    # lstopo binary runs cleanly
+    "lstopo --version",
+]
+
+moduleclass = 'system'

easybuild/easyconfigs/l/libconfig/libconfig-1.8.1-GCCcore-14.2.0.eb

@@ -0,0 +1,29 @@
+# Author:   Aayush Joglekar <aayush.joglekar@surf.nl>
+
+easyblock = 'ConfigureMake'
+
+name = 'libconfig'
+version = '1.8.1'
+
+homepage = 'https://hyperrealm.github.io/libconfig'
+description = "Libconfig is a simple library for processing structured configuration files"
+
+toolchain = {'name': 'GCCcore', 'version': '14.2.0'}
+
+source_urls = ['https://hyperrealm.github.io/libconfig/dist/']
+sources = [SOURCE_TAR_GZ]
+checksums = [{'libconfig-1.8.1.tar.gz': 'c73ee3d914ec68c99b61e864832931e9a7112eeabfb449dad217fd83e385cbdf'}]
+
+builddependencies = [
+    ('binutils', '2.42'),
+    ('Bison', '3.8.2'),
+    ('flex', '2.6.4'),
+]
+
+sanity_check_paths = {
+    'files': ['include/libconfig.h', 'include/libconfig.h++', 'lib/libconfig.a', 'lib/libconfig++.a',
+              'lib/libconfig.so', 'lib/libconfig++.so'],
+    'dirs': ['lib/pkgconfig'],
+}
+
+moduleclass = 'lib'

easybuild/easyconfigs/l/libcxi/libcxi-1.0.2-GCCcore-14.2.0.eb

@@ -0,0 +1,59 @@
+# Author:   Aayush Joglekar <aayush.joglekar@surf.nl>
+
+easyblock = 'ConfigureMake'
+
+name = 'libcxi'
+version = '1.0.2'
+
+homepage = 'https://github.com/HewlettPackard/shs-libcxi'
+description = """libcxi is the userspace library for the HPE Slingshot-11
+Cassini (CXI) NIC. It is the link-time dependency of libfabric's cxi provider."""
+
+toolchain = {'name': 'GCCcore', 'version': '14.2.0'}
+toolchainopts = {'pic': True}
+
+# git tag release/shs-12.0.2 corresponds to cray-libcxi 1.0.2
+sources = [
+    {
+        'filename': SOURCE_TAR_GZ,
+        'git_config': {
+            'url': 'https://github.com/HewlettPackard',
+            'repo_name': 'shs-libcxi',
+            'tag': 'release/shs-12.0.2',
+            'keep_git_dir': False,
+        },
+    },
+]
+patches = ['libcxi-1.0.2_build-library-only.patch']
+checksums = [
+    None,  # .tgz not reproducible
+    {'libcxi-1.0.2_build-library-only.patch': '2c83195383bc95751a63362674e201716b1bb21801b5c80a131f22380e491d4b'},
+]
+
+builddependencies = [
+    ('binutils', '2.42'),
+    ('pkgconf', '2.3.0'),
+    ('Autotools', '20240712'),
+]
+dependencies = [
+    ('cassini-headers', '1.1.1', '', SYSTEM),
+    ('cxi-driver-headers', '1.0.0', '', SYSTEM),
+    ('libconfig', '1.8.1'),
+    ('libuv', '1.51.0'),
+    ('libnl', '3.11.0'),
+    ('numactl', '2.0.19'),
+]
+
+preconfigopts = "./autogen.sh && "
+
+# --without-rocm does NOT disable GPU-direct over CXI. In libcxi all GPU code
+# lives in the (now-stripped) utils/ and tests/; The actual ROCm HMEM (GPU-direct)
+# backend is implemented in libfabric's cxi provider and enabled there via --with-rocr.
+configopts = "--disable-static --without-systemd --without-rocm --without-cuda --without-ze "
+
+sanity_check_paths = {
+    'files': ['lib/libcxi.so'],
+    'dirs': ['include', 'lib/pkgconfig'],
+}
+
+moduleclass = 'lib'