[ExecuTorch] Simplify function pointers for apply_ternary_elementwise_fn

Cleaning up some of the required boilerplate. I updated op_clamp and op_where, but continued to not optimize op_where for size/build time. Ideal usage optimizing for size/build time would look like op_clamp.

Differential Revision: [D63790004](https://our.internmc.facebook.com/intern/diff/D63790004/)

[ghstack-poisoned]

Author: swolchok

kernels/portable/cpu/op_clamp.cpp

@@ -66,15 +66,6 @@ ET_NODISCARD bool check_bounds(
   return is_valid;
 }
 
-template <typename To, typename From>
-To load_and_convert(const void* fromPtr) {
-  return static_cast<To>(*reinterpret_cast<const From*>(fromPtr));
-}
-
-template <typename To, typename From>
-void convert_and_store(From f, void* dst) {
-  *reinterpret_cast<To*>(dst) = static_cast<To>(f);
-}
 } // namespace
 
 Tensor& clamp_out(
@@ -221,26 +212,9 @@ Tensor& clamp_tensor_out(
 
   ET_KERNEL_CHECK(ctx, canCast(common_type, out_type), InvalidArgument, out);
 
-  constexpr auto name = "clamp.Tensor_out";
+  static constexpr const char op_name[] = "clamp.Tensor_out";
 
-  ET_SWITCH_REALHB_TYPES(common_type, ctx, name, CTYPE_COMMON, [&]() {
-    using ToCtypeCommonFn = CTYPE_COMMON (*)(const void*);
-    ToCtypeCommonFn in_to_common;
-    ET_SWITCH_REALHB_TYPES(in_type, ctx, name, CTYPE_IN, [&]() {
-      in_to_common = load_and_convert<CTYPE_COMMON, CTYPE_IN>;
-    });
-    ToCtypeCommonFn min_to_common;
-    ET_SWITCH_REALHB_TYPES(min_type, ctx, name, CTYPE_MIN, [&]() {
-      min_to_common = load_and_convert<CTYPE_COMMON, CTYPE_MIN>;
-    });
-    ToCtypeCommonFn max_to_common;
-    ET_SWITCH_REALHB_TYPES(max_type, ctx, name, CTYPE_MAX, [&]() {
-      max_to_common = load_and_convert<CTYPE_COMMON, CTYPE_MAX>;
-    });
-    void (*common_to_out)(CTYPE_COMMON, void*);
-    ET_SWITCH_REALHB_TYPES(out_type, ctx, name, CTYPE_OUT, [&]() {
-      common_to_out = convert_and_store<CTYPE_OUT, CTYPE_COMMON>;
-    });
+  ET_SWITCH_REALHB_TYPES(common_type, ctx, op_name, CTYPE_COMMON, [&]() {
     apply_ternary_elementwise_fn<CTYPE_COMMON>(
         [has_min, has_max](
             const CTYPE_COMMON val_in,
@@ -259,10 +233,10 @@ Tensor& clamp_tensor_out(
         min,
         max,
         out,
-        in_to_common,
-        min_to_common,
-        max_to_common,
-        common_to_out);
+        get_load_to_common_fn_realhbbf16<CTYPE_COMMON, op_name>(in),
+        get_load_to_common_fn_realhbbf16<CTYPE_COMMON, op_name>(min),
+        get_load_to_common_fn_realhbbf16<CTYPE_COMMON, op_name>(max),
+        get_store_common_to_tensor_fn_realhbbf16<CTYPE_COMMON, op_name>(out));
   });
 
   return out;

kernels/portable/cpu/op_where.cpp

@@ -56,21 +56,10 @@ Tensor& where_out(
           b,
           cond,
           out,
-          [](const void* a_ptr) {
-            return static_cast<CTYPE_OUT>(
-                *reinterpret_cast<const CTYPE_A*>(a_ptr));
-          },
-          [](const void* b_ptr) {
-            return static_cast<CTYPE_OUT>(
-                *reinterpret_cast<const CTYPE_B*>(b_ptr));
-          },
-          [](const void* c_ptr) {
-            return static_cast<CTYPE_OUT>(
-                *reinterpret_cast<const uint8_t*>(c_ptr));
-          },
-          [](CTYPE_OUT result, void* out) {
-            *reinterpret_cast<CTYPE_OUT*>(out) = result;
-          });
+          internal::load_and_convert<CTYPE_OUT, CTYPE_A>,
+          internal::load_and_convert<CTYPE_OUT, CTYPE_B>,
+          internal::load_and_convert<CTYPE_OUT, uint8_t>,
+          internal::convert_and_store<CTYPE_OUT, CTYPE_OUT>);
     });
   });
 

kernels/portable/cpu/util/broadcast_util.h

@@ -270,6 +270,46 @@ size_t linearize_access_indexes(
 // Mapping with broadcasting
 //
 
+namespace internal {
+template <typename To, typename From>
+To load_and_convert(const void* fromPtr) {
+  return static_cast<To>(*reinterpret_cast<const From*>(fromPtr));
+}
+
+template <typename To, typename From>
+void convert_and_store(From f, void* dst) {
+  *reinterpret_cast<To*>(dst) = static_cast<To>(f);
+}
+} // namespace internal
+
+template <typename CTYPE_COMMON>
+using load_to_common_fn = CTYPE_COMMON (*)(const void*);
+
+template <typename CTYPE_COMMON, const char* op_name>
+load_to_common_fn<CTYPE_COMMON> get_load_to_common_fn_realhbbf16(
+    const Tensor& t) {
+  CTYPE_COMMON (*result)(const void*) = nullptr;
+  ET_SWITCH_REALHBBF16_TYPES(
+      t.scalar_type(), unused, op_name, TENSOR_CTYPE, [&]() {
+        result = internal::load_and_convert<CTYPE_COMMON, TENSOR_CTYPE>;
+      });
+  return result;
+}
+
+template <typename CTYPE_COMMON>
+using store_common_to_tensor_fn = void (*)(CTYPE_COMMON, void*);
+
+template <typename CTYPE_COMMON, const char* op_name>
+store_common_to_tensor_fn<CTYPE_COMMON>
+get_store_common_to_tensor_fn_realhbbf16(const Tensor& t) {
+  void (*result)(CTYPE_COMMON, void*) = nullptr;
+  ET_SWITCH_REALHBBF16_TYPES(
+      t.scalar_type(), unused, op_name, TENSOR_CTYPE, [&]() {
+        result = internal::convert_and_store<TENSOR_CTYPE, CTYPE_COMMON>;
+      });
+  return result;
+}
+
 /**
  * Useful for binary elementwise operators. For each element of the inputs,
  * perform a computation and write to the corresponding element of the output.