no unroll in Laswp

Signed-off-by: Yuuichi Asahi <[email protected]>

batched/dense/impl/KokkosBatched_Laswp_Serial_Internal.hpp

@@ -94,17 +94,16 @@ struct SerialLaswpVectorBackwardInternal {
   template <typename IntType, typename ValueType>
   KOKKOS_INLINE_FUNCTION static int invoke(const int plen, const IntType *KOKKOS_RESTRICT p, const int ps0,
                                            /* */ 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
+// Disabling loop unrolling fixes the issue
 #if defined(KOKKOS_ENABLE_CUDA) && defined(KOKKOS_ARCH_HOPPER90)
-#if CUDA_VERSION == 12000
-      if (piv < 0) return 0;
+#if CUDA_VERSION >= 12000 && CUDA_VERSION < 12100
+#pragma unroll 1
 #endif
 #endif
+    for (int i = (plen - 1); i >= 0; --i) {
+      const int piv = p[i * ps0];
       if (piv != i) {
         const int idx_i = i * as0, idx_p = piv * as0;
         const ValueType tmp = A[idx_i];