Implement unary_ufunc functions using elementwise_util (#9386)

swolchok · web-flow · commit dffbd426a4c5 · 2025-06-09T13:28:44.000-07:00
One less set of independent implementations to worry about going forward
(e.g., we don't have to vectorize these separately from elementwise_util
and they get all benefits of elementwise_util).
diff --git a/kernels/portable/cpu/op_acos.cpp b/kernels/portable/cpu/op_acos.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& acos_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::acos, ctx, in, out);
+  static constexpr const char op_name[] = "acos.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::acos(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_acosh.cpp b/kernels/portable/cpu/op_acosh.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& acosh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::acosh, ctx, in, out);
+  static constexpr const char op_name[] = "acosh.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::acosh(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_asin.cpp b/kernels/portable/cpu/op_asin.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& asin_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::asin, ctx, in, out);
+  static constexpr const char op_name[] = "asin.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::asin(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_asinh.cpp b/kernels/portable/cpu/op_asinh.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& asinh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::asinh, ctx, in, out);
+  static constexpr const char op_name[] = "asinh.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::asinh(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_atan.cpp b/kernels/portable/cpu/op_atan.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& atan_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::atan, ctx, in, out);
+  static constexpr const char op_name[] = "atan.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::atan(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_atanh.cpp b/kernels/portable/cpu/op_atanh.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& atanh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::atanh, ctx, in, out);
+  static constexpr const char op_name[] = "atanh.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::atanh(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_ceil.cpp b/kernels/portable/cpu/op_ceil.cpp
@@ -17,7 +17,9 @@ namespace native {
 using executorch::aten::Tensor;
 
 Tensor& ceil_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbf16(std::ceil, ctx, in, out);
+  static constexpr const char op_name[] = "ceil.out";
+  return internal::unary_ufunc_realhbf16<op_name>(
+      [](auto x) { return executorch::math::ceil(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_cos.cpp b/kernels/portable/cpu/op_cos.cpp
@@ -15,7 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& cos_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(std::cos, ctx, in, out);
+  static constexpr const char op_name[] = "cos.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::cos(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_cosh.cpp b/kernels/portable/cpu/op_cosh.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& cosh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::cosh, ctx, in, out);
+  static constexpr const char op_name[] = "cosh.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::cosh(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_erf.cpp b/kernels/portable/cpu/op_erf.cpp
@@ -15,7 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& erf_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(std::erf, ctx, in, out);
+  static constexpr const char op_name[] = "erf.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::erf(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_exp.cpp b/kernels/portable/cpu/op_exp.cpp
@@ -15,7 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& exp_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(std::exp, ctx, in, out);
+  static constexpr const char op_name[] = "exp.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::exp(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_expm1.cpp b/kernels/portable/cpu/op_expm1.cpp
@@ -7,16 +7,19 @@
  */
 
 #include <executorch/kernels/portable/cpu/pattern/pattern.h>
+#include <executorch/kernels/portable/cpu/util/elementwise_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <cmath>
+#include <type_traits>
 
 namespace torch {
 namespace executor {
 namespace native {
 
 Tensor& expm1_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::expm1, ctx, in, out);
+  static constexpr const char op_name[] = "expm1.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::expm1(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_floor.cpp b/kernels/portable/cpu/op_floor.cpp
@@ -17,7 +17,9 @@ namespace native {
 using executorch::aten::Tensor;
 
 Tensor& floor_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbf16(std::floor, ctx, in, out);
+  static constexpr const char op_name[] = "floor.out";
+  return internal::unary_ufunc_realhbf16<op_name>(
+      [](auto x) { return executorch::math::floor(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_isinf.cpp b/kernels/portable/cpu/op_isinf.cpp
@@ -17,8 +17,9 @@ namespace native {
 Tensor& isinf_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
   // Lambda is syntactic sugar needed to workaround compilation on some older
   // non-compatible distros where isnan is returning int rather than bool
-  return internal::unary_ufunc_realhb_to_bool(
-      [](double x) -> bool { return std::isinf(x); }, ctx, in, out);
+  static constexpr const char op_name[] = "isinf.out";
+  return internal::unary_ufunc_realhb_to_bool<op_name>(
+      [](auto x) -> bool { return std::isinf(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_isnan.cpp b/kernels/portable/cpu/op_isnan.cpp
@@ -17,8 +17,9 @@ namespace native {
 Tensor& isnan_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
   // Lambda is syntactic sugar needed to workaround compilation on some older
   // non-compatible distros where isnan is returning int rather than bool
-  return internal::unary_ufunc_realhb_to_bool(
-      [](double x) -> bool { return std::isnan(x); }, ctx, in, out);
+  static constexpr const char op_name[] = "isnan.out";
+  return internal::unary_ufunc_realhb_to_bool<op_name>(
+      [](auto x) -> bool { return std::isnan(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_log.cpp b/kernels/portable/cpu/op_log.cpp
@@ -15,7 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& log_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(std::log, ctx, in, out);
+  static constexpr const char op_name[] = "log.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::log(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_log10.cpp b/kernels/portable/cpu/op_log10.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& log10_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::log10, ctx, in, out);
+  static constexpr const char op_name[] = "log10.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::log10(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_log1p.cpp b/kernels/portable/cpu/op_log1p.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& log1p_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::log1p, ctx, in, out);
+  static constexpr const char op_name[] = "log1p.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::log1p(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_log2.cpp b/kernels/portable/cpu/op_log2.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& log2_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::log2, ctx, in, out);
+  static constexpr const char op_name[] = "log2.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::log2(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_reciprocal.cpp b/kernels/portable/cpu/op_reciprocal.cpp
@@ -12,18 +12,11 @@
 namespace torch {
 namespace executor {
 namespace native {
-namespace {
-
-double reciprocal(double x) {
-  return 1.0 / x;
-}
-
-} // namespace
-
 Tensor&
 reciprocal_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      reciprocal, ctx, in, out);
+  static constexpr const char op_name[] = "reciprocal.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::reciprocal(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_rsqrt.cpp b/kernels/portable/cpu/op_rsqrt.cpp
@@ -12,16 +12,11 @@
 namespace torch {
 namespace executor {
 namespace native {
-namespace {
-
-double rsqrt(double x) {
-  return 1.0 / std::sqrt(x);
-}
-
-} // namespace
 
 Tensor& rsqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(rsqrt, ctx, in, out);
+  static constexpr const char op_name[] = "rsqrt.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::rsqrt(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_sin.cpp b/kernels/portable/cpu/op_sin.cpp
@@ -15,7 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& sin_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(std::sin, ctx, in, out);
+  static constexpr const char op_name[] = "sin.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::sin(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_sinh.cpp b/kernels/portable/cpu/op_sinh.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& sinh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::sinh, ctx, in, out);
+  static constexpr const char op_name[] = "sinh.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::sinh(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_sqrt.cpp b/kernels/portable/cpu/op_sqrt.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& sqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::sqrt, ctx, in, out);
+  static constexpr const char op_name[] = "sqrt.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::sqrt(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_tan.cpp b/kernels/portable/cpu/op_tan.cpp
@@ -15,7 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& tan_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(std::tan, ctx, in, out);
+  static constexpr const char op_name[] = "tan.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::tan(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_tanh.cpp b/kernels/portable/cpu/op_tanh.cpp
@@ -15,8 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& tanh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbbf16_to_floathbf16(
-      std::tanh, ctx, in, out);
+  static constexpr const char op_name[] = "tanh.out";
+  return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(
+      [](auto x) { return executorch::math::tanh(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/op_trunc.cpp b/kernels/portable/cpu/op_trunc.cpp
@@ -15,7 +15,9 @@ namespace executor {
 namespace native {
 
 Tensor& trunc_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
-  return internal::unary_ufunc_realhbf16(std::trunc, ctx, in, out);
+  static constexpr const char op_name[] = "trunc.out";
+  return internal::unary_ufunc_realhbf16<op_name>(
+      [](auto x) { return executorch::math::trunc(x); }, ctx, in, out);
 }
 
 } // namespace native
diff --git a/kernels/portable/cpu/pattern/pattern.cpp b/kernels/portable/cpu/pattern/pattern.cpp
@@ -0,0 +1,28 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/kernels/portable/cpu/pattern/pattern.h>
+
+namespace torch::executor::native::internal {
+
+bool check_and_resize_inputs(
+    KernelRuntimeContext& ctx,
+    const Tensor& in,
+    Tensor& out) {
+  ET_KERNEL_CHECK(
+      ctx, tensors_have_same_dim_order(in, out), InvalidArgument, false);
+  ET_KERNEL_CHECK_MSG(
+      ctx,
+      resize_tensor(out, in.sizes()) == Error::Ok,
+      InvalidArgument,
+      false,
+      "Failed to resize output tensor.");
+  return true;
+}
+
+} // namespace torch::executor::native::internal
diff --git a/kernels/portable/cpu/pattern/pattern.h b/kernels/portable/cpu/pattern/pattern.h
diff --git a/kernels/portable/cpu/pattern/targets.bzl b/kernels/portable/cpu/pattern/targets.bzl
diff --git a/kernels/portable/cpu/pattern/unary_ufunc_realhb_to_bool.cpp b/kernels/portable/cpu/pattern/unary_ufunc_realhb_to_bool.cpp
diff --git a/kernels/portable/cpu/pattern/unary_ufunc_realhbbf16_to_floathbf16.cpp b/kernels/portable/cpu/pattern/unary_ufunc_realhbbf16_to_floathbf16.cpp
diff --git a/kernels/portable/cpu/pattern/unary_ufunc_realhbf16.cpp b/kernels/portable/cpu/pattern/unary_ufunc_realhbf16.cpp
diff --git a/kernels/portable/cpu/util/vectorized_math.h b/kernels/portable/cpu/util/vectorized_math.h
diff --git a/shim_et/xplat/executorch/kernels/portable/op_registration_util.bzl b/shim_et/xplat/executorch/kernels/portable/op_registration_util.bzl

Original file line number	Diff line number	Diff line change
`@@ -17,7 +17,9 @@ namespace native {`
`17`	`17`	`using executorch::aten::Tensor;`
`18`	`18`
`19`	`19`	`Tensor& ceil_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`20`		`- return internal::unary_ufunc_realhbf16(std::ceil, ctx, in, out);`
	`20`	`+ static constexpr const char op_name[] = "ceil.out";`
	`21`	`+ return internal::unary_ufunc_realhbf16<op_name>(`
	`22`	`+ [](auto x) { return executorch::math::ceil(x); }, ctx, in, out);`
`21`	`23`	`}`
`22`	`24`
`23`	`25`	`} // namespace native`
Original file line number	Diff line number	Diff line change
`@@ -15,7 +15,9 @@ namespace executor {`
`15`	`15`	`namespace native {`
`16`	`16`
`17`	`17`	`Tensor& cos_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`18`		`- return internal::unary_ufunc_realhbbf16_to_floathbf16(std::cos, ctx, in, out);`
	`18`	`+ static constexpr const char op_name[] = "cos.out";`
	`19`	`+ return internal::unary_ufunc_realhbbf16_to_floathbf16<op_name>(`
	`20`	`+ [](auto x) { return executorch::math::cos(x); }, ctx, in, out);`
`19`	`21`	`}`
`20`	`22`
`21`	`23`	`} // namespace native`