CI fixes

Author: bernhardmgruber

cub/cub/detail/launcher/cuda_runtime.cuh

@@ -34,13 +34,13 @@ struct TripleChevronFactory
 
   CUB_RUNTIME_FUNCTION ::cudaError_t PtxVersion(int& version)
   {
-    return ::cub::PtxVersion(version);
+    return cub::PtxVersion(version);
   }
 
   CUB_RUNTIME_FUNCTION ::cudaError_t ArchId(::cuda::arch_id& arch_id) const
   {
     int ptx_version;
-    if (const auto error = ::cub::PtxVersion(ptx_version))
+    if (const auto error = cub::PtxVersion(ptx_version))
     {
       return error;
     }

cub/cub/device/dispatch/tuning/tuning_reduce.cuh

@@ -21,7 +21,7 @@
 #include <cuda/std/optional>
 
 #if _CCCL_STD_VER >= 2020
-#  include <concepts>
+#  include <cuda/std/concepts>
 #endif // _CCCL_STD_VER >= 2020
 
 #if !_CCCL_COMPILER(NVRTC)

cub/test/catch2_test_env_launch_helper.h

@@ -108,6 +108,17 @@ struct stream_registry_factory_t
     return cub::PtxVersion(version);
   }
 
+  CUB_RUNTIME_FUNCTION cudaError_t ArchId(::cuda::arch_id& arch_id) const
+  {
+    int ptx_version;
+    if (const auto error = cub::PtxVersion(ptx_version))
+    {
+      return error;
+    }
+    arch_id = cuda::to_arch_id(cuda::compute_capability(ptx_version / 10));
+    return cudaSuccess;
+  }
+
   CUB_RUNTIME_FUNCTION cudaError_t MultiProcessorCount(int& sm_count) const
   {
     int device_ordinal;