Adding more Threadblock Tiles for Mixed-input TensorOp (BF16 * S8) in cutlass_library (#1132)

* Adding more tiles in the cutlass_library for mixed-input support.

* fix rebase issue

* more tiles to upcast a

tools/library/include/cutlass/library/operation_table.h

@@ -215,6 +215,18 @@ struct GemmPreferenceKey {
     return compute_capability == rhs.compute_capability;
   }
 };
+/////////////////////////////////////////////////////////////////////////////////////////////////
+inline
+std::ostream& operator<< (std::ostream& out, const cutlass::library::GemmPreferenceKey& key) {
+    out << "{\n"
+      << "compute_capability : " << key.compute_capability << std::endl
+      << "alignment          : " << key.alignment << std::endl
+      << "}";
+  
+  return out;
+}
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
 
 /////////////////////////////////////////////////////////////////////////////////////////////////
 

tools/library/include/cutlass/library/types.h

@@ -172,6 +172,7 @@ enum class MathOperationID {
   kAdd,
   kMultiplyAdd,
   kMultiplyAddSaturate,
+  kMultiplyAddMixedInputUpcast,
   kMultiplyAddFastBF16,
   kMultiplyAddFastF16,
   kMultiplyAddFastF32,