Fixing issue with pragma unroll in laswp

Signed-off-by: Luc Berger-Vergiat <lberge@sandia.gov>
kokkos · Mar 6, 2025 · 6713c36 · 6713c36
1 parent 23efb6c
commit 6713c36
Showing 1 changed file with 2 additions and 0 deletions.
diff --git a/batched/dense/impl/KokkosBatched_Laswp_Serial_Internal.hpp b/batched/dense/impl/KokkosBatched_Laswp_Serial_Internal.hpp
@@ -97,10 +97,12 @@ struct SerialLaswpVectorBackwardInternal {
 // On H100 with Cuda 12.0.0, the compiler seems to apply
 // an aggressive optimization which crashes this function
 // Disabling loop unrolling fixes the issue
+#if defined(KOKKOS_ENABLE_PRAGMA_UNROLL)
 #if defined(KOKKOS_ENABLE_CUDA) && defined(KOKKOS_ARCH_HOPPER90)
 #if CUDA_VERSION >= 12000 && CUDA_VERSION < 12100
 #pragma unroll 1
 #endif
+#endif
 #endif
     for (int i = (plen - 1); i >= 0; --i) {
       const int piv = p[i * ps0];