Update cosma and a few related recipes (#35615)

* Add maintainers 

* Updated cosma archive checksum and costa version

- updated cosma version (in the cosma build system)
- updated costa version
- use the default generic url for downloading packages
- do not build tiled-mm when the cpu only version is needed


Signed-off-by: Dr. Mathieu Taillefumier <mathieu.taillefumier@free.fr>
Co-authored-by: Rocco Meli <r.meli@bluemail.ch>
Co-authored-by: Harmen Stoppels <harmenstoppels@gmail.com>

4 files changed, 128 insertions, 28 deletions

diff --git a/var/spack/repos/builtin/packages/cosma/package.py b/var/spack/repos/builtin/packages/cosma/package.py
index 04c126f439..2fccafe187 100644
--- a/var/spack/repos/builtin/packages/cosma/package.py
+++ b/var/spack/repos/builtin/packages/cosma/package.py
@@ -12,14 +12,20 @@ class Cosma(CMakePackage):
     Distributed Communication-Optimal Matrix-Matrix Multiplication Library
     """
 
-    maintainers("haampie", "kabicm", "teonnik")
+    maintainers("haampie", "kabicm", "teonnik", "simonpintarelli", "mtaillefumier")
     homepage = "https://github.com/eth-cscs/COSMA"
-    url = "https://github.com/eth-cscs/COSMA/releases/download/v2.5.1/COSMA-v2.5.1.tar.gz"
+    url = "https://github.com/eth-cscs/COSMA/archive/refs/tags/v2.6.6.tar.gz"
     git = "https://github.com/eth-cscs/COSMA.git"
 
     # note: The default archives produced with github do not have the archives
     #       of the submodules.
-    version("master", branch="master", submodules=True)
+    version("master", branch="master", submodules=False)
+    version("2.6.6", sha256="1604be101e77192fbcc5551236bc87888d336e402f5409bbdd9dea900401cc37")
+    version("2.6.5", sha256="10d9b7ecc1ce44ec5b9e0c0bf89278a63029912ec3ea99661be8576b553ececf")
+    version("2.6.4", sha256="6d7bd5e3005874af9542a329c93e7ccd29ca1a5573dae27618fac2704fa2b6ab")
+    version("2.6.3", sha256="8ca96ca41458f1e9d0da70d524c5a03c677dba7238d23a578f852163b6d45ac9")
+    version("2.6.2", sha256="2debb5123cc35aeebc5fd2f8a46cfd6356d1e27618c9bb57129ecd09aa400940")
+    version("2.6.1", sha256="69aa6634a030674f0d9be61e7b0bf0dc17acf0fc9e7a90b40e3179e2254c8d67")
     version("2.5.1", sha256="085b7787597374244bbb1eb89bc69bf58c35f6c85be805e881e1c0b25166c3ce")
     version("2.5.0", sha256="7f68bb0ee5c80f9b8df858afcbd017ad4ed87ac09439d13d7d890844dbdd3d54")
     version("2.4.0", sha256="5714315ce06d48037f86cfee2d7f19340643fee95e9d7f1e92dc1b623b67e395")
@@ -34,23 +40,36 @@ class Cosma(CMakePackage):
     variant("rocm", default=False, description="Build with rocBLAS support")
     variant("scalapack", default=False, description="Build with ScaLAPACK API")
     variant("shared", default=False, description="Build the shared library version")
+    variant("tests", default=False, description="Build tests")
+    variant("apps", default=False, description="Build miniapp")
+    variant("profiling", default=False, description="Enable profiling")
+    variant("gpu_direct", default=False, description="GPU aware MPI")
 
-    depends_on("cmake@3.12:", type="build")
+    with when("+cuda"):
+        variant("nccl", default=False, description="Use cuda nccl")
+
+    depends_on("cmake@3.22:", type="build")
     depends_on("mpi@3:")
     depends_on("blas", when="~cuda ~rocm")
     depends_on("scalapack", when="+scalapack")
     depends_on("cuda", when="+cuda")
     depends_on("rocblas", when="+rocm")
 
-    def url_for_version(self, version):
-        if version <= Version("2.3.0"):
-            return "https://github.com/eth-cscs/COSMA/releases/download/v{0}/cosma.tar.gz".format(
-                version
-            )
+    with when("@2.6.3:"):
+        depends_on("tiled-mm@2.2:+cuda", when="+cuda")
+        depends_on("tiled-mm@2.2:+rocm", when="+rocm")
 
-        return "https://github.com/eth-cscs/COSMA/releases/download/v{0}/COSMA-v{1}.tar.gz".format(
-            version, version
-        )
+    with when("@2.6.1:2.6.2"):
+        depends_on("tiled-mm@2.0+rocm", when="+rocm")
+        depends_on("tiled-mm@2.0+cuda", when="+cuda")
+
+    with when("@2.6.1:"):
+        depends_on("costa")
+        depends_on("costa+scalapack", when="+scalapack")
+        depends_on("cxxopts", when="+apps")
+        depends_on("cxxopts", when="+tests")
+        depends_on("semiprof", when="+profiling")
+        depends_on("costa+profiling", when="+profiling")
 
     def setup_build_environment(self, env):
         if "+cuda" in self.spec:
@@ -58,13 +77,13 @@ class Cosma(CMakePackage):
 
     def cosma_blas_cmake_arg(self):
         query_to_cmake_arg = [
+            ("+cuda", "CUDA"),
+            ("+rocm", "ROCM"),
             ("^intel-mkl", "MKL"),
             ("^intel-oneapi-mkl", "MKL"),
             ("^cray-libsci", "CRAY_LIBSCI"),
             ("^netlib-lapack", "CUSTOM"),
             ("^openblas", "OPENBLAS"),
-            ("+cuda", "CUDA"),
-            ("+rocm", "ROCM"),
         ]
 
         if self.version >= Version("2.4.0"):
@@ -92,9 +111,11 @@ class Cosma(CMakePackage):
 
     def cmake_args(self):
         return [
-            self.define("COSMA_WITH_TESTS", False),
-            self.define("COSMA_WITH_APPS", False),
-            self.define("COSMA_WITH_PROFILING", False),
+            self.define_from_variant("COSMA_WITH_TESTS", "tests"),
+            self.define_from_variant("COSMA_WITH_APPS", "apps"),
+            self.define_from_variant("COSMA_WITH_NCCL", "nccl"),
+            self.define_from_variant("COSMA_WITH_GPU_AWARE_MPI", "gpu_direct"),
+            self.define_from_variant("COSMA_WITH_PROFILING", "profiling"),
             self.define("COSMA_WITH_BENCHMARKS", False),
             self.define("COSMA_BLAS", self.cosma_blas_cmake_arg()),
             self.define("COSMA_SCALAPACK", self.cosma_scalapack_cmake_arg()),
diff --git a/var/spack/repos/builtin/packages/costa/package.py b/var/spack/repos/builtin/packages/costa/package.py
index 48a6fb8ec9..62e001fc39 100644
--- a/var/spack/repos/builtin/packages/costa/package.py
+++ b/var/spack/repos/builtin/packages/costa/package.py
@@ -13,32 +13,38 @@ class Costa(CMakePackage):
     Based on the paper: https://arxiv.org/abs/2106.06601
     """
 
-    maintainers("haampie", "kabicm", "RMeli")
+    maintainers("haampie", "kabicm", "RMeli", "mtaillefumier")
     homepage = "https://github.com/eth-cscs/COSTA"
     git = "https://github.com/eth-cscs/COSTA.git"
 
     # note: The default archives produced with github do not have the archives
     #       of the submodules.
     version("master", branch="master", submodules=True)
+    version("2.2.2", sha256="e87bc37aad14ac0c5922237be5d5390145c9ac6aef0350ed17d86cb2d994e67c")
+    version("2.2.1", sha256="aa8aa2a4a79de094f857c22293825de270ff72becd6bd736ff9f2dd8c192446d")
+    version("2.2", sha256="3e7333f012af76ec3508276ea90800313f6136504667021fe229e710bf6acdc7")
     version("2.1", sha256="c1e86452415083f7470b292d93ec60708b7c8dbafc2bac383636bb4b28135866")
     version("2.0", sha256="de250197f31f7d23226c6956a687c3ff46fb0ff6c621a932428236c3f7925fe4")
 
     variant("scalapack", default=False, description="Build with ScaLAPACK API")
     variant("shared", default=False, description="Build shared libraries")
+    variant("profiling", default=False, description="Enable profiling")
+    variant("tests", default=False, description="Enable tests")
+    variant("apps", default=False, description="Enable miniapp")
+    variant("benchmarks", default=False, description="Enable benchmarks")
 
-    depends_on("cmake@3.12:", type="build")
+    depends_on("cmake@3.22:", type="build")
     depends_on("mpi@3:")
     depends_on("scalapack", when="+scalapack")
+    depends_on("cxxopts", when="+apps")
+    depends_on("semiprof", when="+profiling")
 
     def url_for_version(self, version):
         if version == Version("2.0"):
             return "https://github.com/eth-cscs/COSTA/releases/download/v{0}/COSTA-v{1}.tar.gz".format(
                 version, version
             )
-        else:
-            return "https://github.com/eth-cscs/COSTA/archive/refs/tags/v{0}.tar.gz".format(
-                version
-            )
+        return "https://github.com/eth-cscs/COSTA/archive/refs/tags/v{0}.tar.gz".format(version)
 
     def setup_build_environment(self, env):
         return
@@ -57,10 +63,10 @@ class Costa(CMakePackage):
 
     def cmake_args(self):
         return [
-            self.define("COSTA_WITH_BENCHMARKS", "OFF"),
-            self.define("COSTA_WITH_APPS", "OFF"),
-            self.define("COSTA_WITH_TESTS", "OFF"),
-            self.define("COSTA_WITH_PROFILING", "OFF"),
+            self.define_from_variant("COSTA_WITH_BENCHMARKS", "benchmarks"),
+            self.define_from_variant("COSTA_WITH_APPS", "apps"),
+            self.define_from_variant("COSTA_WITH_TESTS", "tests"),
+            self.define_from_variant("COSTA_WITH_PROFILING", "profiling"),
+            self.define_from_variant("BUILD_SHARED_LIBS", "shared"),
             self.define("COSTA_SCALAPACK", self.costa_scalapack_cmake_arg()),
-            self.define("BUILD_SHARED_LIBS", "+shared" in self.spec),
         ]
diff --git a/var/spack/repos/builtin/packages/semiprof/package.py b/var/spack/repos/builtin/packages/semiprof/package.py
new file mode 100644
index 0000000000..0506ecf7a3
--- /dev/null
+++ b/var/spack/repos/builtin/packages/semiprof/package.py
@@ -0,0 +1,27 @@
+# Copyright 2013-2023 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)
+
+from spack.package import *
+
+
+class Semiprof(CMakePackage):
+    """Simple thread safe annotation based C++ profiler."""
+
+    homepage = "https://github.com/bcumming/semiprof"
+    url = "https://github.com/bcumming/semiprof/archive/refs/tags/v0.1.tar.gz"
+
+    maintainers = ["simonpintarelli"]
+
+    variant("examples", default=False)
+    variant("shared", default=True)
+
+    version("0.1", sha256="4fb3823c65a4f5dfbe05e8cbe1911dfd25cd7740597f82c7b3a84472de26f0dc")
+
+    def cmake_args(self):
+        return [
+            self.define("SEMIPROF_WITH_INSTALL", True),
+            self.define_from_variant("SEMIPROF_WITH_EXAMPLES", "examples"),
+            self.define_from_variant("BUILD_SHARED_LIBS", "shared"),
+        ]
diff --git a/var/spack/repos/builtin/packages/tiled-mm/package.py b/var/spack/repos/builtin/packages/tiled-mm/package.py
new file mode 100644
index 0000000000..c10b7a74fe
--- /dev/null
+++ b/var/spack/repos/builtin/packages/tiled-mm/package.py
@@ -0,0 +1,46 @@
+# Copyright 2013-2023 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)
+
+from spack.package import *
+
+
+class TiledMm(CMakePackage, CudaPackage, ROCmPackage):
+    """Matrix multiplication on GPUs for matrices stored on a CPU. Similar to cublasXt,
+    but ported to both NVIDIA and AMD GPUs."""
+
+    maintainers = ["mtaillefumier", "simonpintarelli"]
+    homepage = "https://github.com/eth-cscs/Tiled-MM/"
+    url = "https://github.com/eth-cscs/Tiled-MM/archive/refs/tags/v2.0.tar.gz"
+    git = "https://github.com/eth-cscs/Tiled-MM.git"
+
+    version("master", branch="master")
+    version("2.2", sha256="6d0b49c9588ece744166822fd44a7bc5bec3dc666b836de8bf4bf1a7bb675aac")
+    version("2.0", sha256="ea554aea8c53d7c8e40044e6d478c0e8137d7e8b09d7cb9650703430d92cf32e")
+
+    variant("shared", default=True)
+    variant("examples", default=False)
+    variant("tests", default=False)
+
+    depends_on("rocblas", when="+rocm")
+    depends_on("cxxopts", when="+tests")
+    depends_on("cxxopts", when="+examples")
+
+    conflicts("~cuda~rocm")
+    conflicts("+cuda", when="+rocm")
+
+    def cmake_args(self):
+        args = [
+            self.define_from_variant("BUILD_SHARED_LIBS", "shared"),
+            self.define_from_variant("TIELDMM_WITH_EXAMPLES", "examples"),
+            self.define_from_variant("TIELDMM_WITH_TESTS", "tests"),
+        ]
+
+        if "+rocm" in self.spec:
+            args.extend([self.define("TILEDMM_GPU_BACKEND", "ROCM")])
+
+        if "+cuda" in self.spec:
+            args.extend([self.define("TILEDMM_GPU_BACKEND", "CUDA")])
+
+        return args