Minor fix for a regressing tuning in reduce.by_key (#3723)

NVIDIA · Feb 6, 2025 · 3795966 · 3795966
1 parent 0d6ab79
commit 3795966
Showing 1 changed file with 5 additions and 0 deletions.
diff --git a/cub/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh b/cub/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh
@@ -717,6 +717,11 @@ struct sm100_tuning<KeyT, AccumT, primitive_op::yes, primitive_key::yes, primiti
   static constexpr CacheLoadModifier load_modifier   = LOAD_DEFAULT;
 };
 
+// I16, F32, I32 regresses, default it back.
+template <class KeyT>
+struct sm100_tuning<KeyT, float, primitive_op::yes, primitive_key::yes, primitive_accum::yes, key_size::_2, accum_size::_4>
+{};
+
 // todo(gonidelis): Add tunings for I128.
 // template <class KeyT, class AccumT>
 // struct sm100_tuning<KeyT, AccumT, primitive_op::yes, primitive_key::yes, primitive_accum::no, key_size::_2,