cuda: add v12.9 and new Blackwell targets (#50284)

* cuda: Add 12.9 and new Blackwell targets

Including family-specific architecture features like sm_100f

* tau, caliper: Add cuda v12.9 conflict

Legacy nvtx was dropped.

* petsc: "libnvToolsExt.a" -> "libnvtx3interop.a" for cuda@12.9: and petsc@:3.23.1

* legion: Add upper bound to cuda version range

Avoid build failure due to CUDA driver API incompatibility.
See https://github.com/StanfordLegion/legion/issues/1864

---------

Co-authored-by: Satish Balay <balay@mcs.anl.gov>

lib/spack/spack/build_systems/cuda.py

@@ -49,10 +49,19 @@ class CudaPackage(PackageBase):
         "90a",
         "100",
         "100a",
+        "100f",
         "101",
         "101a",
+        "101f",
+        "103",
+        "103a",
+        "103f",
         "120",
         "120a",
+        "120f",
+        "121",
+        "121a",
+        "121f",
     )
 
     # FIXME: keep cuda and cuda_arch separate to make usage easier until
@@ -155,6 +164,15 @@ def compute_capabilities(arch_list: Iterable[str]) -> List[str]:
     depends_on("cuda@12.8:", when="cuda_arch=101a")
     depends_on("cuda@12.8:", when="cuda_arch=120")
     depends_on("cuda@12.8:", when="cuda_arch=120a")
+    depends_on("cuda@12.9:", when="cuda_arch=100f")
+    depends_on("cuda@12.9:", when="cuda_arch=101f")
+    depends_on("cuda@12.9:", when="cuda_arch=120f")
+    depends_on("cuda@12.9:", when="cuda_arch=103")
+    depends_on("cuda@12.9:", when="cuda_arch=103a")
+    depends_on("cuda@12.9:", when="cuda_arch=103f")
+    depends_on("cuda@12.9:", when="cuda_arch=121")
+    depends_on("cuda@12.9:", when="cuda_arch=121a")
+    depends_on("cuda@12.9:", when="cuda_arch=121f")
     # From the NVIDIA install guide we know of conflicts for particular
     # platforms (linux, darwin), architectures (x86, powerpc) and compilers
     # (gcc, clang). We don't restrict %gcc and %clang conflicts to
@@ -186,7 +204,7 @@ def compute_capabilities(arch_list: Iterable[str]) -> List[str]:
         conflicts("%gcc@12:", when="+cuda ^cuda@:11.8")
         conflicts("%gcc@13:", when="+cuda ^cuda@:12.3")
         conflicts("%gcc@14:", when="+cuda ^cuda@:12.6")
-        conflicts("%gcc@15:", when="+cuda ^cuda@:12.8")
+        conflicts("%gcc@15:", when="+cuda ^cuda@:12.9")
         conflicts("%clang@12:", when="+cuda ^cuda@:11.4.0")
         conflicts("%clang@13:", when="+cuda ^cuda@:11.5")
         conflicts("%clang@14:", when="+cuda ^cuda@:11.7")
@@ -195,7 +213,7 @@ def compute_capabilities(arch_list: Iterable[str]) -> List[str]:
         conflicts("%clang@17:", when="+cuda ^cuda@:12.3")
         conflicts("%clang@18:", when="+cuda ^cuda@:12.5")
         conflicts("%clang@19:", when="+cuda ^cuda@:12.6")
-        conflicts("%clang@20:", when="+cuda ^cuda@:12.8")
+        conflicts("%clang@20:", when="+cuda ^cuda@:12.9")
 
         # https://gist.github.com/ax3l/9489132#gistcomment-3860114
         conflicts("%gcc@10", when="+cuda ^cuda@:11.4.0")

var/spack/repos/spack_repo/builtin/packages/caliper/package.py

@@ -137,6 +137,8 @@ class Caliper(CachedCMakePackage, CudaPackage, ROCmPackage):
     conflicts("+libdw", "@:2.4")
     conflicts("+rocm", "@:2.7")
     conflicts("+rocm+cuda")
+    # Legacy nvtx is only supported until cuda@12.8, newer cuda only provides nvtx3.
+    conflicts("^cuda@12.9:", "@:2.12.1")
 
     patch("for_aarch64.patch", when="@:2.11 target=aarch64:")
     patch(

var/spack/repos/spack_repo/builtin/packages/cuda/package.py

@@ -21,6 +21,16 @@
 #    format returned by platform.system() and 'arch' by platform.machine()
 
 _versions = {
+    "12.9.0": {
+        "Linux-aarch64": (
+            "f3b7ae71f95d11de0a03ccfa1c0aff7be336d2199b50b1a15b03695fd15a6409",
+            "https://developer.download.nvidia.com/compute/cuda/12.9.0/local_installers/cuda_12.9.0_575.51.03_linux_sbsa.run",
+        ),
+        "Linux-x86_64": (
+            "bbce2b760fe2096ca1c86f729e03bf377c1519add7b2755ecc4e9b0a9e07ee43",
+            "https://developer.download.nvidia.com/compute/cuda/12.9.0/local_installers/cuda_12.9.0_575.51.03_linux.run",
+        ),
+    },
     "12.8.1": {
         "Linux-aarch64": (
             "353cbab1b57282a1001071796efd95c1e40ec27a3375e854d12637eaa1c6107c",

var/spack/repos/spack_repo/builtin/packages/legion/package.py

@@ -121,10 +121,10 @@ class Legion(CMakePackage, ROCmPackage):
     depends_on("mpi", when="conduit=mpi")
     depends_on("cuda@10.0:11.9", when="+cuda_unsupported_compiler @21.03.0:23.03.0")
     depends_on("cuda@10.0:11.9", when="+cuda @21.03.0:23.03.0")
-    depends_on("cuda@11.7:", when="+cuda_unsupported_compiler @23.06.0:")
+    depends_on("cuda@11.7:12.8", when="+cuda_unsupported_compiler @23.06.0:")
-    depends_on("cuda@11.7:", when="+cuda @23.06.0:")
+    depends_on("cuda@11.7:12.8", when="+cuda @23.06.0:")
-    depends_on("cuda@11.7:", when="+cuda_unsupported_compiler @stable")
+    depends_on("cuda@11.7:12.8", when="+cuda_unsupported_compiler @stable")
-    depends_on("cuda@11.7:", when="+cuda @stable")
+    depends_on("cuda@11.7:12.8", when="+cuda @stable")
     depends_on("hip@5.1:5.7", when="+rocm @23.03.0:23.12.0")
     depends_on("hip@5.1:", when="+rocm")
     depends_on("hdf5", when="+hdf5")

var/spack/repos/spack_repo/builtin/packages/petsc/package.py

@@ -750,6 +750,13 @@ def revert_kokkos_nvcc_wrapper(self):
             env["MPICXX_CXX"] = env["CXX"]
 
     def configure(self, spec, prefix):
+        if spec.satisfies("@:3.23.1 +cuda ^cuda@12.9:"):
+            filter_file(
+                "libnvToolsExt.a",
+                "libnvtx3interop.a",
+                "config/BuildSystem/config/packages/cuda.py",
+                string=True,
+            )
         self.revert_kokkos_nvcc_wrapper()
         python("configure", "--prefix=%s" % prefix, *self.configure_options())
 

var/spack/repos/spack_repo/builtin/packages/tau/package.py

@@ -159,7 +159,8 @@ class Tau(Package):
         depends_on("python@:3.10", when="@:2.32.1")
     depends_on("libunwind", when="+libunwind")
     depends_on("mpi", when="+mpi", type=("build", "run", "link"))
-    depends_on("cuda", when="+cuda")
+    # Legacy nvtx is only supported until cuda@12.8, newer cuda only provides nvtx3.
+    depends_on("cuda@:12.8", when="+cuda")
     depends_on("gasnet", when="+gasnet")
     depends_on("adios2", when="+adios2")
     depends_on("sqlite", when="+sqlite")