disallow optimization to fix serial getrs

Signed-off-by: Yuuichi Asahi <y.asahi@nr.titech.ac.jp>
kokkos · Jan 28, 2025 · 3e27b0d · 3e27b0d
1 parent 9c8eedf
commit 3e27b0d
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 1 deletion.
diff --git a/batched/dense/impl/KokkosBatched_Laswp_Serial_Impl.hpp b/batched/dense/impl/KokkosBatched_Laswp_Serial_Impl.hpp
@@ -83,7 +83,7 @@ struct SerialLaswp<Direct::Forward> {
 template <>
 struct SerialLaswp<Direct::Backward> {
   template <typename PivViewType, typename AViewType>
-  KOKKOS_INLINE_FUNCTION static int invoke(const PivViewType piv, const AViewType &A) {
+  KOKKOS_INLINE_FUNCTION static int invoke(const PivViewType &piv, const AViewType &A) {
     auto info = KokkosBatched::Impl::checkLaswpInput(piv, A);
     if (info) return info;
 

diff --git a/batched/dense/impl/KokkosBatched_Laswp_Serial_Internal.hpp b/batched/dense/impl/KokkosBatched_Laswp_Serial_Internal.hpp
@@ -96,6 +96,15 @@ struct SerialLaswpVectorBackwardInternal {
                                            /* */ ValueType *KOKKOS_RESTRICT A, const int as0) {
     for (int i = (plen - 1); i >= 0; --i) {
       const int piv = p[i * ps0];
+
+// On H100 with Cuda 12.0.0, the compiler seems to apply
+// an aggressive optimization which crashes this function
+// Insert unnecessary operation to disallow optimization
+#if defined(KOKKOS_ENABLE_CUDA) && defined(KOKKOS_ARCH_HOPPER90)
+#if CUDA_VERSION == 12000
+      if (piv < 0) return 0;
+#endif
+#endif
       if (piv != i) {
         const int idx_i = i * as0, idx_p = piv * as0;
         const ValueType tmp = A[idx_i];