Always do dense_indexing on atomic_add (#950)

Author: voznesenskym

torchinductor/codegen/triton.py

@@ -717,7 +717,7 @@ def load(self, name: str, index: sympy.Expr, upcast: bool = False):
 
     def store(self, name, index, value, mode=None):
         var = self.args.output(name)
-        index, mask = self.indexing(index, value)
+        index, mask = self.indexing(index, value, dense_indexing=True)
         if mode is None:
             line = f"tl.store({var} + {index}, {value}, {mask})"
         elif mode == "atomic_add":