Respect selective build for dtype_specialized_elementwise_fn_impl in elementwise_util (#11975)

This fancy fast path I added didn't respect selective build. Now it
should.

Author: swolchok

kernels/portable/cpu/util/elementwise_util.h

@@ -9,6 +9,7 @@
 #pragma once
 
 #include <c10/util/irange.h>
+#include <executorch/kernels/portable/cpu/selective_build.h>
 #include <executorch/kernels/portable/cpu/util/broadcast_indexes_range.h>
 #include <executorch/kernels/portable/cpu/util/broadcast_util.h>
 #include <executorch/kernels/portable/cpu/util/dtype_util.h>
@@ -345,20 +346,22 @@ inline void apply_elementwise_fn(
   }
 
   constexpr auto compute_type = CppTypeToScalarType<CTYPE_COMPUTE>::value;
-  const bool all_inputs_compute_dtype =
-      ((inputs.first->scalar_type() == compute_type) && ...);
-
-  constexpr ScalarType out_specialized_scalar_type =
-      specialized_output_scalar_type<CTYPE_COMPUTE>(out_dtypes);
-  if (all_inputs_compute_dtype &&
-      out.scalar_type() == out_specialized_scalar_type) {
-    using CTYPE_OUT =
-        typename ScalarTypeToCppType<out_specialized_scalar_type>::type;
-    dtype_specialized_elementwise_fn_impl<
-        CTYPE_COMPUTE,
-        CTYPE_OUT,
-        support_noncontiguous_tensors>(compute_fun, ctx, out, inputs...);
-    return;
+  if constexpr (should_include_kernel_dtype(op_name, compute_type)) {
+    const bool all_inputs_compute_dtype =
+        ((inputs.first->scalar_type() == compute_type) && ...);
+
+    constexpr ScalarType out_specialized_scalar_type =
+        specialized_output_scalar_type<CTYPE_COMPUTE>(out_dtypes);
+    if (all_inputs_compute_dtype &&
+        out.scalar_type() == out_specialized_scalar_type) {
+      using CTYPE_OUT =
+          typename ScalarTypeToCppType<out_specialized_scalar_type>::type;
+      dtype_specialized_elementwise_fn_impl<
+          CTYPE_COMPUTE,
+          CTYPE_OUT,
+          support_noncontiguous_tensors>(compute_fun, ctx, out, inputs...);
+      return;
+    }
   }
 
   apply_elementwise_fn_generic_impl<

kernels/portable/cpu/util/targets.bzl

@@ -115,10 +115,10 @@ def define_common_targets():
             ":vectorized_math",
             "//executorch/runtime/core/portable_type/c10/c10:aten_headers_for_executorch",
             "//executorch/runtime/kernel:kernel_runtime_context",
+            "//executorch/kernels/portable/cpu:scalar_utils",
             "//executorch/extension/threadpool:threadpool",
         ],
         deps = [
-            "//executorch/kernels/portable/cpu:scalar_utils",
             "//executorch/runtime/kernel:kernel_includes",
         ],
         visibility = ["//executorch/kernels/portable/cpu/...", "//executorch/kernels/optimized/cpu/...", "@EXECUTORCH_CLIENTS"],