diff --git a/AMSS_NCKU_source/bssn_class.C b/AMSS_NCKU_source/bssn_class.C
index 6c0e749..15f4d74 100644
--- a/AMSS_NCKU_source/bssn_class.C
+++ b/AMSS_NCKU_source/bssn_class.C
@@ -98,7 +98,9 @@ static int cuda_compute_rhs_bssn_ss(
                      Symmetry, Lev, eps, sst, co);
 }
 }
-// All call sites below that use f_compute_rhs_bssn_ss get redirected to GPU
+// Use GPU for shell RHS in Step and SHStep.
+// Compute_Constraint / Interp_Constraint / Constraint_Out keep
+// CPU Fortran (GPU alloc-per-call overhead dominates there).
 #define f_compute_rhs_bssn_ss cuda_compute_rhs_bssn_ss
 #endif
 
@@ -9066,6 +9068,12 @@ void bssn_class::AnalysisStuff(int lev, double dT_lev)
 
 //================================================================================================
 
+#if USE_CUDA_BSSN && defined(WithShell)
+#undef f_compute_rhs_bssn_ss
+// Restore the original Fortran name mapping from bssn_rhs.h (fortran3 convention)
+#define f_compute_rhs_bssn_ss compute_rhs_bssn_ss_
+#endif
+
 // This member function computes and outputs constraint violations
 
 //================================================================================================