Work around NVHPC 23.x not dealing with __isGlobal

nliber · May 9, 2023 · 417a6ee · 417a6ee
1 parent 60b982a
commit 417a6ee
Showing 1 changed file with 7 additions and 2 deletions.
diff --git a/tpls/desul/include/desul/atomics/cuda/CUDA_asm.hpp b/tpls/desul/include/desul/atomics/cuda/CUDA_asm.hpp
@@ -2,10 +2,15 @@
 namespace desul {
 namespace Impl {
 // Choose the variant of atomics we are using later
+// The __isGlobal intrinsic was only introduced in CUDA 11.2
+// It also stopped working in NVC++ 23.1 - it works in 22.11
+// this is a bug in NVHPC, not treating CUDA intrinsics correctly
+// FIXME_NVHPC
 #if !defined(DESUL_IMPL_ATOMIC_CUDA_PTX_PREDICATE) && \
     !defined(DESUL_IMPL_ATOMIC_CUDA_PTX_ISGLOBAL)
-#if (__CUDACC_VER_MAJOR__ > 11) || \
-    ((__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ > 1))
+#if ((__CUDACC_VER_MAJOR__ > 11) ||                                   \
+     ((__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ > 1))) && \
+    (!defined(__NVCOMPILER) || __NVCOMPILER_MAJOR__ < 23)
 #define DESUL_IMPL_ATOMIC_CUDA_PTX_ISGLOBAL
 #else
 #define DESUL_IMPL_ATOMIC_CUDA_PTX_PREDICATE