fix __global__ in gpu kernels

src/pystencils/backend/emission/base_printer.py

@@ -375,11 +375,16 @@ class BasePrinter(ABC):
                 )
 
     def print_signature(self, func: Kernel) -> str:
-        prefix = self._func_prefix
         params_str = ", ".join(
             f"{self._type_str(p.dtype)} {p.name}" for p in func.parameters
         )
-        sig_parts = ([prefix] if prefix is not None else []) + ["void", func.name, f"({params_str})"]
+
+        from ...codegen import GpuKernel
+        
+        sig_parts = [self._func_prefix] if self._func_prefix is not None else []
+        if isinstance(func, GpuKernel):
+            sig_parts.append("__global__")
+        sig_parts += ["void", func.name, f"({params_str})"]
         signature = " ".join(sig_parts)
         return signature