[ET][Portable][Build Size] Introduce uni-tensor and bi-tensor equivalents

Pull Request resolved: pytorch/executorch#6006

This introduces unitensor and bitensor equivalents to the tritensor template introduced by @swolchok.
They are applied to add.Scalar_out and add.out respectively, reducing op_add's build size.
The unitensor template is applied to clamp.Scalar_out, reducing op_clamp's build size further.

Build size reduction:
- add: 484K -> 21K
- clamp: 119K -> 28K
ghstack-source-id: 246919722
@exported-using-ghexport

Differential Revision: [D63838076](https://our.internmc.facebook.com/intern/diff/D63838076/)

Author: manuelcandales

kernels/portable/cpu/op_add.cpp

@@ -7,64 +7,14 @@
  */
 
 #include <executorch/kernels/portable/cpu/scalar_utils.h>
-#include <executorch/kernels/portable/cpu/util/broadcast_util.h>
-#include <executorch/kernels/portable/cpu/util/functional_util.h>
+#include <executorch/kernels/portable/cpu/util/elementwise_util.h>
 #include <executorch/kernels/portable/cpu/util/kernel_ops_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
 
 namespace torch {
 namespace executor {
 namespace native {
-namespace {
-
-template <
-    bool can_cast,
-    typename CTYPE_A,
-    typename CTYPE_B,
-    typename CTYPE_IN,
-    typename CTYPE_OUT>
-struct AddInner;
-
-template <
-    typename CTYPE_A,
-    typename CTYPE_B,
-    typename CTYPE_IN,
-    typename CTYPE_OUT>
-struct AddInner<true, CTYPE_A, CTYPE_B, CTYPE_IN, CTYPE_OUT> {
-  static void
-  run(const Tensor& a, const Tensor& b, CTYPE_IN alpha_val, Tensor& out) {
-    apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
-        // NOLINTNEXTLINE(facebook-hte-ConstantArgumentPassByValue)
-        [alpha_val](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);
-          CTYPE_IN value = a_casted + alpha_val * b_casted;
-
-          return static_cast<CTYPE_OUT>(value);
-        },
-        a,
-        b,
-        out);
-  }
-};
-
-template <typename CTYPE_IN>
-struct ReportCanCastBug {
-  static void run(const Tensor&, const Tensor&, CTYPE_IN, Tensor&) {
-    ET_DCHECK_MSG(false, "BUG: canCast should have been checked above");
-  }
-};
-
-template <
-    typename CTYPE_A,
-    typename CTYPE_B,
-    typename CTYPE_IN,
-    typename CTYPE_OUT>
-struct AddInner<false, CTYPE_A, CTYPE_B, CTYPE_IN, CTYPE_OUT>
-    : public ReportCanCastBug<CTYPE_IN> {};
-
-} // namespace
 
 Tensor& add_out(
     KernelRuntimeContext& ctx,
@@ -80,7 +30,9 @@ Tensor& add_out(
 
   ET_KERNEL_CHECK(
       ctx,
-      executorch::runtime::tensor_is_realhbbf16_type(out),
+      (executorch::runtime::tensor_is_realhbbf16_type(a) &&
+       executorch::runtime::tensor_is_realhbbf16_type(b) &&
+       executorch::runtime::tensor_is_realhbbf16_type(out)),
       InvalidArgument,
       out);
   ET_KERNEL_CHECK(
@@ -96,25 +48,20 @@ Tensor& add_out(
   ET_KERNEL_CHECK(
       ctx, check_alpha_type(alpha_type, common_type), InvalidArgument, out);
 
-  constexpr auto name = "add.out";
-
-  ET_SWITCH_REALHBBF16_TYPES(a_type, ctx, name, CTYPE_A, [&]() {
-    ET_SWITCH_REALHBBF16_TYPES(b_type, ctx, name, CTYPE_B, [&]() {
-      using CTYPE_IN = typename torch::executor::
-          promote_types<CTYPE_A, CTYPE_B, /*half_to_float*/ true>::type;
-      ET_DCHECK(CppTypeToScalarType<CTYPE_IN>::value == common_type);
-      CTYPE_IN alpha_val;
-      utils::extract_scalar(alpha, &alpha_val);
-
-      ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, name, CTYPE_OUT, [&]() {
-        AddInner<
-            can_cast<CTYPE_IN, CTYPE_OUT>::value,
-            CTYPE_A,
-            CTYPE_B,
-            CTYPE_IN,
-            CTYPE_OUT>::run(a, b, alpha_val, out);
-      });
-    });
+  static constexpr const char op_name[] = "add.out";
+
+  ET_SWITCH_REALB_TYPES(common_type, ctx, op_name, CTYPE_COMMON, [&]() {
+    utils::apply_bitensor_elementwise_fn<CTYPE_COMMON, op_name>(
+        [alpha](const CTYPE_COMMON val_a, const CTYPE_COMMON val_b) {
+          CTYPE_COMMON val_alpha = utils::scalar_to<CTYPE_COMMON>(alpha);
+          return val_a + val_alpha * val_b;
+        },
+        a,
+        utils::SupportedTensorDtypes::REALHBBF16,
+        b,
+        utils::SupportedTensorDtypes::REALHBBF16,
+        out,
+        utils::SupportedTensorDtypes::REALHBBF16);
   });
 
   return out;
@@ -138,14 +85,14 @@ Tensor& add_scalar_out(
 
   ET_KERNEL_CHECK(
       ctx,
-      executorch::runtime::tensor_is_realhbbf16_type(out),
+      (executorch::runtime::tensor_is_realhbbf16_type(a) &&
+       executorch::runtime::tensor_is_realhbbf16_type(out)),
       InvalidArgument,
       out);
   ET_KERNEL_CHECK(
       ctx, tensors_have_same_dim_order(a, out), InvalidArgument, out);
 
   ScalarType a_type = a.scalar_type();
-  ScalarType b_type = utils::get_scalar_dtype(b);
   ScalarType alpha_type = utils::get_scalar_dtype(alpha);
   ScalarType common_type =
       utils::promote_type_with_scalar(a_type, b, /*half_to_float*/ false);
@@ -155,42 +102,23 @@ Tensor& add_scalar_out(
   ET_KERNEL_CHECK(
       ctx, check_alpha_type(alpha_type, common_type), InvalidArgument, out);
 
-  if (common_type == ScalarType::Half) {
+  if (common_type == ScalarType::Half || common_type == ScalarType::BFloat16) {
     common_type = ScalarType::Float;
   }
 
-  constexpr auto name = "add.Scalar_out";
-
-  ET_SWITCH_REALHBBF16_TYPES(a_type, ctx, name, CTYPE_A, [&]() {
-    ET_SWITCH_SCALAR_OBJ_TYPES(b_type, ctx, name, CTYPE_B, [&]() {
-      using CTYPE_IN = typename utils::promote_type_with_scalar_type<
-          CTYPE_A,
-          CTYPE_B,
-          /*half_to_float*/ true>::type;
-      ET_DCHECK(CppTypeToScalarType<CTYPE_IN>::value == common_type);
-
-      CTYPE_B b_val;
-      utils::extract_scalar(b, &b_val);
-      CTYPE_IN b_casted = static_cast<CTYPE_IN>(b_val);
-
-      CTYPE_IN alpha_val;
-      utils::extract_scalar(alpha, &alpha_val);
-
-      using CTYPE_OUT = typename std::conditional<
-          std::is_same<CTYPE_A, internal::F2>::value,
-          internal::F2,
-          CTYPE_IN>::type;
-
-      apply_unary_map_fn(
-          [b_casted, alpha_val](const CTYPE_A val_a) {
-            CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a);
-            CTYPE_IN value = a_casted + alpha_val * b_casted;
-            return static_cast<CTYPE_OUT>(value);
-          },
-          a.const_data_ptr<CTYPE_A>(),
-          out.mutable_data_ptr<CTYPE_OUT>(),
-          out.numel());
-    });
+  static constexpr const char op_name[] = "add.Scalar_out";
+
+  ET_SWITCH_REALB_TYPES(common_type, ctx, op_name, CTYPE_COMMON, [&]() {
+    utils::apply_unitensor_elementwise_fn<CTYPE_COMMON, op_name>(
+        [b, alpha](const CTYPE_COMMON val_a) {
+          CTYPE_COMMON val_b = utils::scalar_to<CTYPE_COMMON>(b);
+          CTYPE_COMMON val_alpha = utils::scalar_to<CTYPE_COMMON>(alpha);
+          return val_a + val_alpha * val_b;
+        },
+        a,
+        utils::SupportedTensorDtypes::REALHBBF16,
+        out,
+        utils::SupportedTensorDtypes::REALHBBF16);
   });
 
   return out;

kernels/portable/cpu/op_clamp.cpp

@@ -13,7 +13,6 @@
 
 #include <executorch/kernels/portable/cpu/scalar_utils.h>
 #include <executorch/kernels/portable/cpu/util/elementwise_util.h>
-#include <executorch/kernels/portable/cpu/util/functional_util.h>
 #include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
@@ -122,43 +121,26 @@ Tensor& clamp_out(
 
   ET_KERNEL_CHECK(ctx, common_type == out_type, InvalidArgument, out);
 
-  ET_SWITCH_REALH_TYPES(out_type, ctx, "clamp", CTYPE_OUT, [&]() {
-    // Extract optional min value
-    CTYPE_OUT min = 0;
-    if (has_min) {
-      ET_SWITCH_SCALAR_OBJ_TYPES(min_type, ctx, "clamp", CTYPE_MIN, [&]() {
-        CTYPE_MIN min_val = 0;
-        utils::extract_scalar(min_opt.value(), &min_val);
-        min = static_cast<CTYPE_OUT>(min_val);
-      });
-    }
-
-    // Extract optional max value
-    CTYPE_OUT max = 0;
-    if (has_max) {
-      ET_SWITCH_SCALAR_OBJ_TYPES(max_type, ctx, "clamp", CTYPE_MAX, [&]() {
-        CTYPE_MAX max_val = 0;
-        utils::extract_scalar(max_opt.value(), &max_val);
-        max = static_cast<CTYPE_OUT>(max_val);
-      });
-    }
+  static constexpr const char op_name[] = "clamp.out";
 
-    ET_SWITCH_REALHB_TYPES(in_type, ctx, "clamp", CTYPE_IN, [&]() {
-      apply_unary_map_fn(
-          [has_min, min, has_max, max](const CTYPE_IN val_in) {
-            CTYPE_OUT val_out = static_cast<CTYPE_OUT>(val_in);
-            if (has_min) {
-              val_out = utils::max_override(val_out, min);
-            }
-            if (has_max) {
-              val_out = utils::min_override(val_out, max);
-            }
-            return val_out;
-          },
-          in.const_data_ptr<CTYPE_IN>(),
-          out.mutable_data_ptr<CTYPE_OUT>(),
-          in.numel());
-    });
+  ET_SWITCH_REALHB_TYPES(common_type, ctx, op_name, CTYPE_COMMON, [&]() {
+    utils::apply_unitensor_elementwise_fn<CTYPE_COMMON, op_name>(
+        [has_min, min_opt, has_max, max_opt](const CTYPE_COMMON val_in) {
+          CTYPE_COMMON val_out = val_in;
+          if (has_min) {
+            val_out = utils::max_override(
+                val_out, utils::scalar_to<CTYPE_COMMON>(min_opt.value()));
+          }
+          if (has_max) {
+            val_out = utils::min_override(
+                val_out, utils::scalar_to<CTYPE_COMMON>(max_opt.value()));
+          }
+          return val_out;
+        },
+        in,
+        utils::SupportedTensorDtypes::REALHBBF16,
+        out,
+        utils::SupportedTensorDtypes::REALHBBF16);
   });
 
   return out;

kernels/portable/cpu/util/elementwise_util.h

@@ -16,6 +16,33 @@ namespace executor {
 namespace native {
 namespace utils {
 
+/*
+ * Convert Scalar to C++ type
+ */
+
+template <typename T>
+T scalar_to(const Scalar& s) {
+  if (s.isBoolean()) {
+    return static_cast<T>(s.to<bool>());
+  } else if (s.isFloatingPoint()) {
+    return static_cast<T>(s.to<double>());
+  } else {
+    return static_cast<T>(s.to<int64_t>());
+  }
+}
+
+template <>
+inline double scalar_to<double>(const Scalar& s) {
+  return s.isFloatingPoint() ? s.to<double>()
+                             : static_cast<double>(s.to<int64_t>());
+}
+
+template <>
+inline int64_t scalar_to<int64_t>(const Scalar& s) {
+  return s.isFloatingPoint() ? static_cast<int64_t>(s.to<double>())
+                             : s.to<int64_t>();
+}
+
 namespace internal {
 
 template <typename To, typename From>
@@ -139,6 +166,86 @@ store_common_to_tensor_fn<CTYPE_COMMON> get_store_common_to_tensor_fn(
 
 } // namespace internal
 
+template <typename CTYPE_COMMON, const char* op_name, typename Op>
+inline void apply_unitensor_elementwise_fn(
+    const Op& compute_fun,
+    const Tensor& a,
+    SupportedTensorDtypes a_dtypes,
+    const Tensor& out,
+    SupportedTensorDtypes out_dtypes) {
+  const auto load_a_to_common =
+      internal::get_load_to_common_fn<CTYPE_COMMON, op_name>(a, a_dtypes);
+  const auto store_common_to_out =
+      internal::get_store_common_to_tensor_fn<CTYPE_COMMON, op_name>(
+          out, out_dtypes);
+  const char* const data_a = reinterpret_cast<const char*>(a.const_data_ptr());
+  const auto a_element_size = a.element_size();
+  const auto out_element_size = out.element_size();
+  char* const data_out = reinterpret_cast<char*>(out.mutable_data_ptr());
+
+  auto out_numel = out.numel();
+  for (size_t i = 0; i < out_numel; ++i) {
+    auto result = compute_fun(load_a_to_common(&data_a[i * a_element_size]));
+    store_common_to_out(result, &data_out[i * out_element_size]);
+  }
+}
+
+/**
+ * Useful for bi-tensor 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.
+ */
+template <typename CTYPE_COMMON, const char* op_name, typename Op>
+inline void apply_bitensor_elementwise_fn(
+    const Op& compute_fun,
+    const Tensor& a,
+    SupportedTensorDtypes a_dtypes,
+    const Tensor& b,
+    SupportedTensorDtypes b_dtypes,
+    const Tensor& out,
+    SupportedTensorDtypes out_dtypes) {
+  const bool a_is_broadcasted = !out.sizes().equals(a.sizes());
+  const bool b_is_broadcasted = !out.sizes().equals(b.sizes());
+  const bool any_is_broadcasted = (a_is_broadcasted || b_is_broadcasted);
+
+  const auto load_a_to_common =
+      internal::get_load_to_common_fn<CTYPE_COMMON, op_name>(a, a_dtypes);
+  const auto load_b_to_common =
+      internal::get_load_to_common_fn<CTYPE_COMMON, op_name>(b, b_dtypes);
+  const auto store_common_to_out =
+      internal::get_store_common_to_tensor_fn<CTYPE_COMMON, op_name>(
+          out, out_dtypes);
+  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 auto a_element_size = a.element_size();
+  const auto b_element_size = b.element_size();
+  const auto out_element_size = out.element_size();
+  char* const data_out = reinterpret_cast<char*>(out.mutable_data_ptr());
+
+  auto out_numel = out.numel();
+  for (size_t i = 0; i < out_numel; ++i) {
+    size_t a_linear_index = i;
+    size_t b_linear_index = i;
+
+    if (any_is_broadcasted) {
+      size_t out_indexes[kTensorDimensionLimit];
+      delinearize_index(i, out, out_indexes, kTensorDimensionLimit);
+
+      if (a_is_broadcasted) {
+        a_linear_index = linearize_access_indexes(out_indexes, out.dim(), a);
+      }
+      if (b_is_broadcasted) {
+        b_linear_index = linearize_access_indexes(out_indexes, out.dim(), b);
+      }
+    }
+
+    auto result = compute_fun(
+        load_a_to_common(&data_a[a_linear_index * a_element_size]),
+        load_b_to_common(&data_b[b_linear_index * b_element_size]));
+    store_common_to_out(result, &data_out[i * out_element_size]);
+  }
+}
+
 /**
  * Useful for tri-tensor elementwise operators. For each element of the inputs,
  * perform a computation and write to the corresponding element of the output.
@@ -194,7 +301,8 @@ inline void apply_tritensor_elementwise_fn(
   const auto out_element_size = out.element_size();
   char* const data_out = reinterpret_cast<char*>(out.mutable_data_ptr());
 
-  for (size_t i = 0; i < out.numel(); ++i) {
+  auto out_numel = out.numel();
+  for (size_t i = 0; i < out_numel; ++i) {
     size_t a_linear_index = i;
     size_t b_linear_index = i;
     size_t c_linear_index = i;