New package: NVIDIA cuTensor (#31397)

2 files changed, 58 insertions, 0 deletions

diff --git a/var/spack/repos/builtin/packages/cutensor/package.py b/var/spack/repos/builtin/packages/cutensor/package.py
new file mode 100644
index 0000000000..14734bd922
--- /dev/null
+++ b/var/spack/repos/builtin/packages/cutensor/package.py
@@ -0,0 +1,57 @@
+# Copyright 2013-2022 Lawrence Livermore National Security, LLC and other
+# Spack Project Developers. See the top-level COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+import platform
+
+from spack.package import *
+
+_versions = {
+    # cuTensor 1.5.0
+    '1.5.0.3': {
+        'Linux-x86_64': '4fdebe94f0ba3933a422cff3dd05a0ef7a18552ca274dd12564056993f55471d',
+        'Linux-ppc64le': 'ad736acc94e88673b04a3156d7d3a408937cac32d083acdfbd8435582cbe15db',
+        'Linux-aarch64': '5b9ac479b1dadaf40464ff3076e45f2ec92581c07df1258a155b5bcd142f6090'},
+
+}
+
+
+class Cutensor(Package):
+    """NVIDIA cuTENSOR Library is a GPU-accelerated tensor linear algebra
+    library providing tensor contraction, reduction and elementwise
+    operations."""
+
+    homepage = "https://developer.nvidia.com/cutensor"
+
+    maintainers = ['bvanessen']
+    url         = "cutensor"
+
+    for ver, packages in _versions.items():
+        key = "{0}-{1}".format(platform.system(), platform.machine())
+        pkg = packages.get(key)
+        cutensor_ver = ver
+
+        cuda_ver = '10.0'
+        if platform.machine() == 'aarch64':
+            cuda_ver = '11.0'
+
+        if pkg:
+            version(cutensor_ver, sha256=pkg)
+            # Add constraints matching CUDA version to cuTensor version
+            cuda_req = 'cuda@{0}:'.format(cuda_ver)
+            cutensor_ver_req = '@{0}'.format(cutensor_ver)
+            depends_on(cuda_req, when=cutensor_ver_req)
+
+    def url_for_version(self, version):
+        # Get the system and machine arch for building the file path
+        sys = "{0}-{1}".format(platform.system(), platform.machine())
+        # Munge it to match Nvidia's naming scheme
+        sys_key = sys.lower()
+        sys_key = sys_key.replace('aarch64', 'sbsa')
+
+        url = 'https://developer.download.nvidia.com/compute/cutensor/redist/libcutensor/{0}/libcutensor-{0}-{1}-archive.tar.xz'
+        return url.format(sys_key, version)
+
+    def install(self, spec, prefix):
+        install_tree('.', prefix)
diff --git a/var/spack/repos/builtin/packages/lbann/package.py b/var/spack/repos/builtin/packages/lbann/package.py
index 6c80069802..3500d3eb62 100644
--- a/var/spack/repos/builtin/packages/lbann/package.py
+++ b/var/spack/repos/builtin/packages/lbann/package.py
@@ -172,6 +172,7 @@ class Lbann(CMakePackage, CudaPackage, ROCmPackage):
     depends_on('cudnn', when='@0.90:0.100 +cuda')
     depends_on('cudnn@8.0.2:', when='@:0.90,0.101: +cuda')
     depends_on('cub', when='@0.94:0.98.2 +cuda ^cuda@:10')
+    depends_on('cutensor', when='@:0.90,0.102: +cuda')
     depends_on('hipcub', when='+rocm')
     depends_on('mpi')
     depends_on('hwloc@1.11:', when='@:0.90,0.102: +hwloc')