Use compile-time promotion to reduce comparison op size & build time

Summary: Yet another 6 improved ops (see previous diffs).

Differential Revision: D56744787

Author: swolchok

kernels/portable/cpu/pattern/comparison_op.h

@@ -35,25 +35,23 @@ Tensor& comparison_op_out(
 
   ET_SWITCH_REAL_TYPES_AND(Bool, a_type, ctx, op_name, CTYPE_A, [&]() {
     ET_SWITCH_REAL_TYPES_AND(Bool, b_type, ctx, op_name, CTYPE_B, [&]() {
-      ET_SWITCH_REAL_TYPES_AND(
-          Bool, common_type, ctx, op_name, CTYPE_IN, [&]() {
-            ET_SWITCH_REAL_TYPES_AND(
-                Bool, out_type, ctx, op_name, CTYPE_OUT, [&]() {
-                  apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
-                      [](const CTYPE_A val_a, const CTYPE_B val_b) {
-                        CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a);
-                        CTYPE_IN b_casted = static_cast<CTYPE_IN>(val_b);
-                        bool value = OpFunc<CTYPE_IN>()(a_casted, b_casted);
-                        return static_cast<CTYPE_OUT>(value);
-                      },
-                      a,
-                      b,
-                      out);
-                });
-          });
+      using CTYPE_IN =
+          typename torch::executor::promote_types<CTYPE_A, CTYPE_B>::type;
+      ET_DCHECK(CppTypeToScalarType<CTYPE_IN>::value == common_type);
+      ET_SWITCH_REAL_TYPES_AND(Bool, out_type, ctx, op_name, CTYPE_OUT, [&]() {
+        apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
+            [](const CTYPE_A val_a, const CTYPE_B val_b) {
+              CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a);
+              CTYPE_IN b_casted = static_cast<CTYPE_IN>(val_b);
+              bool value = OpFunc<CTYPE_IN>()(a_casted, b_casted);
+              return static_cast<CTYPE_OUT>(value);
+            },
+            a,
+            b,
+            out);
+      });
     });
   });
-
   return out;
 }
 } // namespace native