Enable RDC detection with clang-cuda

Author: davebayer

ci/build_cub.sh

@@ -18,8 +18,8 @@ if [[ "$CUDA_COMPILER" == *nvcc* ]]; then
         ENABLE_CCCL_BENCHMARKS="true"
         echo "nvcc version is $NVCC_VERSION. Building CUB benchmarks."
     fi
-else
-    echo "Not building with NVCC, disabling RDC and benchmarks."
+elif [[ "$CUDA_COMPILER" == *clang* ]]; then
+    ENABLE_CUB_RDC="true"
 fi
 
 if [[ "$HOST_COMPILER" == *icpc* || "$HOST_COMPILER" == *nvhpc* ]]; then

libcudacxx/include/cuda/std/__cccl/visibility.h

@@ -146,7 +146,7 @@
 // Some functions can be called from host or device code and launch kernels inside. Thus, they use CUDA Dynamic
 // Parallelism (CDP) and require compiling with Relocatable Device Code (RDC).
 // TODO(bgruber): remove CUB_DISABLE_CDP in CCCL 4.0
-#if defined(__CUDACC_RDC__) && !defined(CCCL_DISABLE_CDP) && !defined(CUB_DISABLE_CDP)
+#if (defined(__CUDACC_RDC__) || defined(__CLANG_RDC__)) && !defined(CCCL_DISABLE_CDP) && !defined(CUB_DISABLE_CDP)
 #  define _CCCL_HAS_RDC() 1
 // We have RDC, so host and device APIs can call kernels
 #  define _CCCL_CDP_API _CCCL_API