[ExecuTorch] Rework apply_ternary_elementwise_fn to allow fixing op_clamp build time (1/2)

This diff is a simple refactor that shouldn't materially
change the generated code -- we still create a "kernel" for each
combination of input1/input2/input3/output dtypes. The following diff
will use this to improve op_clamp (but not op_where, because one of
the input types is fixed, so it's not as bad a build time outlier)
build time.

Differential Revision: [D63681033](https://our.internmc.facebook.com/intern/diff/D63681033/)

ghstack-source-id: 245613706
Pull Request resolved: #5783

Author: swolchok

kernels/portable/cpu/op_clamp.cpp

@@ -218,30 +218,36 @@ Tensor& clamp_tensor_out(
     ET_SWITCH_REALHB_TYPES(min_type, ctx, name, CTYPE_MIN, [&]() {
       ET_SWITCH_REALHB_TYPES(max_type, ctx, name, CTYPE_MAX, [&]() {
         ET_SWITCH_REALHB_TYPES(out_type, ctx, name, CTYPE_OUT, [&]() {
-          apply_ternary_elementwise_fn<
-              CTYPE_IN,
-              CTYPE_MIN,
-              CTYPE_MAX,
-              CTYPE_OUT>(
+          apply_ternary_elementwise_fn<CTYPE_OUT>(
               [has_min, has_max](
-                  const CTYPE_IN val_in,
-                  const CTYPE_MIN val_min,
-                  const CTYPE_MAX val_max) {
-                CTYPE_OUT val_out = static_cast<CTYPE_OUT>(val_in);
+                  const CTYPE_OUT val_in,
+                  const CTYPE_OUT val_min,
+                  const CTYPE_OUT val_max) {
+                CTYPE_OUT val_out = val_in;
                 if (has_min) {
-                  val_out = utils::max_override(
-                      val_out, static_cast<CTYPE_OUT>(val_min));
+                  val_out = utils::max_override(val_out, val_min);
                 }
                 if (has_max) {
-                  val_out = utils::min_override(
-                      val_out, static_cast<CTYPE_OUT>(val_max));
+                  val_out = utils::min_override(val_out, val_max);
                 }
                 return val_out;
               },
               in,
               min,
               max,
-              out);
+              out,
+              [](const void* inPtr) {
+                return static_cast<CTYPE_OUT>(
+                    *reinterpret_cast<const CTYPE_IN*>(inPtr));
+              },
+              [](const void* minPtr) {
+                return static_cast<CTYPE_OUT>(
+                    *reinterpret_cast<const CTYPE_MIN*>(minPtr));
+              },
+              [](const void* maxPtr) {
+                return static_cast<CTYPE_OUT>(
+                    *reinterpret_cast<const CTYPE_MAX*>(maxPtr));
+              });
         });
       });
     });

kernels/portable/cpu/op_where.cpp

@@ -48,16 +48,26 @@ Tensor& where_out(
     ET_SWITCH_REALHBBF16_TYPES(b_type, ctx, name, CTYPE_B, [&]() {
       using CTYPE_OUT =
           typename torch::executor::promote_types<CTYPE_A, CTYPE_B>::type;
-      apply_ternary_elementwise_fn<CTYPE_A, CTYPE_B, uint8_t, CTYPE_OUT>(
-          [](const CTYPE_A val_a, const CTYPE_B val_b, const uint8_t val_c) {
-            CTYPE_OUT a_casted = static_cast<CTYPE_OUT>(val_a);
-            CTYPE_OUT b_casted = static_cast<CTYPE_OUT>(val_b);
-            return val_c ? a_casted : b_casted;
-          },
+      apply_ternary_elementwise_fn<CTYPE_OUT>(
+          [](const CTYPE_OUT val_a,
+             const CTYPE_OUT val_b,
+             const CTYPE_OUT val_c) { return val_c ? val_a : val_b; },
           a,
           b,
           cond,
-          out);
+          out,
+          [](const void* aPtr) {
+            return static_cast<CTYPE_OUT>(
+                *reinterpret_cast<const CTYPE_A*>(aPtr));
+          },
+          [](const void* bPtr) {
+            return static_cast<CTYPE_OUT>(
+                *reinterpret_cast<const CTYPE_B*>(bPtr));
+          },
+          [](const void* cPtr) {
+            return static_cast<CTYPE_OUT>(
+                *reinterpret_cast<const uint8_t*>(cPtr));
+          });
     });
   });
 

kernels/portable/cpu/util/broadcast_util.h

@@ -313,28 +313,42 @@ inline void apply_binary_elementwise_fn(
  * Useful for ternary elementwise operators. For each element of the inputs,
  * perform a computation and write to the corresponding element of the output.
  * Tensor broadcasting is applied wherever it is required.
+ *
+ * In order to mitigate build time cost (straightforwardly |CTYPE_A| *
+ * |CTYPE_B| * |CTYPE_C| * |CTYPE_OUT|), all arguments to compute_fun
+ * are passed as CTYPE_OUT and we require conversion functions from
+ * each input type to the output type be provided. Each conversion
+ * function must take a void* pointing to an element of the
+ * corresponding tensor, load that element, and convert it to
+ * CTYPE_OUT.
  */
 template <
-    typename CTYPE_A,
-    typename CTYPE_B,
-    typename CTYPE_C,
     typename CTYPE_OUT,
-    typename Op>
+    typename Op,
+    typename AToOutFunc,
+    typename BToOutFunc,
+    typename CToOutFunc>
 inline void apply_ternary_elementwise_fn(
     const Op& compute_fun,
     const Tensor& a,
     const Tensor& b,
     const Tensor& c,
-    const Tensor& out) {
+    const Tensor& out,
+    AToOutFunc aToOut,
+    BToOutFunc bToOut,
+    CToOutFunc cToOut) {
   const bool a_is_broadcasted = !out.sizes().equals(a.sizes());
   const bool b_is_broadcasted = !out.sizes().equals(b.sizes());
   const bool c_is_broadcasted = !out.sizes().equals(c.sizes());
   const bool any_is_broadcasted =
       (a_is_broadcasted || b_is_broadcasted || c_is_broadcasted);
 
-  const CTYPE_A* const data_a = a.const_data_ptr<CTYPE_A>();
-  const CTYPE_B* const data_b = b.const_data_ptr<CTYPE_B>();
-  const CTYPE_C* const data_c = c.const_data_ptr<CTYPE_C>();
+  const char* const data_a = reinterpret_cast<const char*>(a.const_data_ptr());
+  const char* const data_b = reinterpret_cast<const char*>(b.const_data_ptr());
+  const char* const data_c = reinterpret_cast<const char*>(c.const_data_ptr());
+  const auto a_element_size = a.element_size();
+  const auto b_element_size = b.element_size();
+  const auto c_element_size = c.element_size();
   CTYPE_OUT* const data_out = out.mutable_data_ptr<CTYPE_OUT>();
 
   for (size_t i = 0; i < out.numel(); ++i) {
@@ -358,7 +372,9 @@ inline void apply_ternary_elementwise_fn(
     }
 
     data_out[i] = compute_fun(
-        data_a[a_linear_index], data_b[b_linear_index], data_c[c_linear_index]);
+        aToOut(&data_a[a_linear_index * a_element_size]),
+        bToOut(&data_b[b_linear_index * b_element_size]),
+        cToOut(&data_c[c_linear_index * c_element_size]));
   }
 }