diff --git a/lib/gpu/lal_neighbor_shared.cpp b/lib/gpu/lal_neighbor_shared.cpp
index 1547eac4f2..aeac76062a 100644
--- a/lib/gpu/lal_neighbor_shared.cpp
+++ b/lib/gpu/lal_neighbor_shared.cpp
@@ -50,7 +50,8 @@ void NeighborShared::compile_kernels(UCL_Device &dev, const int gpu_nbor) {
   	return;
   	
   _gpu_nbor=gpu_nbor;
-  std::string flags="-cl-fast-relaxed-math -cl-mad-enable -D"+
+  std::string flags="-cl-fast-relaxed-math -cl-mad-enable "+
+                    std::string(OCL_PRECISION_COMPILE)+" -D"+
                     std::string(OCL_VENDOR);
 
   if (_gpu_nbor==0) {
diff --git a/lib/gpu/lal_preprocessor.h b/lib/gpu/lal_preprocessor.h
index 722860f512..7c1019ce90 100644
--- a/lib/gpu/lal_preprocessor.h
+++ b/lib/gpu/lal_preprocessor.h
@@ -255,6 +255,10 @@ typedef struct _double4 double4;
 #define ucl_ceil ceil
 #define ucl_abs fabs
 
+#ifdef _DOUBLE_DOUBLE
+#define NO_HARDWARE_TRANSCENDENTALS
+#endif
+
 #ifdef NO_HARDWARE_TRANSCENDENTALS
 
 #define ucl_exp exp