diff --git a/tinygrad/llops/ops_gpu.py b/tinygrad/llops/ops_gpu.py
index 9da7730d1b..97b5f2d591 100644
--- a/tinygrad/llops/ops_gpu.py
+++ b/tinygrad/llops/ops_gpu.py
@@ -231,7 +231,7 @@ class CLASTKernel(ASTKernel):
   def codegen(self) -> Callable:
     self.process()
     self.upcast_in_mid_reduce = False
-    if not KOPT: self.hand_coded_optimizations()
+    if not KOPT or IMAGE == 2: self.hand_coded_optimizations()
 
     # add a local buffer for multistage reduce
     if len(self.group_for_reduce):