[EE/BE] add float variant to unary_ufunc_realhbbf16_to_floathbf16

manuelcandales · manuelcandales · commit c1a941768354 · 2025-07-08T11:33:48.000-07:00
Differential Revision: [D77759060](https://our.internmc.facebook.com/intern/diff/D77759060/) ghstack-source-id: 294877446 Pull Request resolved: #12277
diff --git a/backends/cadence/fusion_g3/operators/op_exp.cpp b/backends/cadence/fusion_g3/operators/op_exp.cpp
@@ -60,7 +60,7 @@ Tensor& exp_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(std::exp, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(std::exp, std::exp, ctx, in, out);
   }
 }
 
diff --git a/backends/cadence/fusion_g3/operators/op_rsqrt.cpp b/backends/cadence/fusion_g3/operators/op_rsqrt.cpp
@@ -27,7 +27,8 @@ namespace native {
 
 namespace {
 
-double rsqrt(double x) {
+template <typename T>
+T rsqrt(T x) {
   return 1.0 / std::sqrt(x);
 }
 
@@ -61,11 +62,11 @@ Tensor& rsqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(rsqrt, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(rsqrt, rsqrt, ctx, in, out);
   }
 }
 
 } // namespace native
 } // namespace G3
 } // namespace impl
-} // namespace cadence
+} // namespace cadence
diff --git a/backends/cadence/fusion_g3/operators/op_sqrt.cpp b/backends/cadence/fusion_g3/operators/op_sqrt.cpp
@@ -55,7 +55,8 @@ Tensor& sqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(std::sqrt, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(
+            std::sqrt, std::sqrt, ctx, in, out);
   }
 }
 
diff --git a/backends/cadence/fusion_g3/operators/op_tanh.cpp b/backends/cadence/fusion_g3/operators/op_tanh.cpp
@@ -55,7 +55,8 @@ Tensor& tanh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(std::tanh, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(
+            std::tanh, std::tanh, ctx, in, out);
   }
 }
 
diff --git a/backends/cadence/hifi/operators/op_rsqrt.cpp b/backends/cadence/hifi/operators/op_rsqrt.cpp
@@ -21,7 +21,8 @@ namespace HiFi {
 namespace native {
 namespace {
 
-double rsqrt(double x) {
+template <typename T>
+T rsqrt(T x) {
   return 1.0 / std::sqrt(x);
 }
 
@@ -46,7 +47,7 @@ Tensor& rsqrt_out(RuntimeContext& ctx, const Tensor& in, Tensor& out) {
   }
 
   return torch::executor::native::internal::
-      unary_ufunc_realhbbf16_to_floathbf16(rsqrt, ctx, in, out);
+      unary_ufunc_realhbbf16_to_floathbf16(rsqrt, rsqrt, ctx, in, out);
 }
 
 } // namespace native
diff --git a/backends/cadence/hifi/operators/op_tanh.cpp b/backends/cadence/hifi/operators/op_tanh.cpp
@@ -35,10 +35,10 @@ Tensor& tanh_out(RuntimeContext& ctx, const Tensor& in, Tensor& out) {
   }
 
   return torch::executor::native::internal::
-      unary_ufunc_realhbbf16_to_floathbf16(std::tanh, ctx, in, out);
+      unary_ufunc_realhbbf16_to_floathbf16(std::tanh, std::tanh, ctx, in, out);
 }
 
 } // namespace native
 } // namespace HiFi
 } // namespace impl
-} // namespace cadence
+} // namespace cadence
diff --git a/kernels/portable/cpu/op_acos.cpp b/kernels/portable/cpu/op_acos.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(acos_out, std::acos)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_acosh.cpp b/kernels/portable/cpu/op_acosh.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(acosh_out, std::acosh)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_asin.cpp b/kernels/portable/cpu/op_asin.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(asin_out, std::asin)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_asinh.cpp b/kernels/portable/cpu/op_asinh.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(asinh_out, std::asinh)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_atan.cpp b/kernels/portable/cpu/op_atan.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(atan_out, std::atan)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_atanh.cpp b/kernels/portable/cpu/op_atanh.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(atanh_out, std::atanh)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_cos.cpp b/kernels/portable/cpu/op_cos.cpp
@@ -14,9 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(cos_out, std::cos)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_cosh.cpp b/kernels/portable/cpu/op_cosh.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(cosh_out, std::cosh)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_erf.cpp b/kernels/portable/cpu/op_erf.cpp
@@ -14,9 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(erf_out, std::erf)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_exp.cpp b/kernels/portable/cpu/op_exp.cpp
@@ -14,9 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(exp_out, std::exp)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_expm1.cpp b/kernels/portable/cpu/op_expm1.cpp
@@ -14,10 +14,7 @@ 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(expm1_out, std::expm1)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_log.cpp b/kernels/portable/cpu/op_log.cpp
@@ -14,9 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(log_out, std::log)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_log10.cpp b/kernels/portable/cpu/op_log10.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(log10_out, std::log10)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_log1p.cpp b/kernels/portable/cpu/op_log1p.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(log1p_out, std::log1p)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_log2.cpp b/kernels/portable/cpu/op_log2.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(log2_out, std::log2)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_reciprocal.cpp b/kernels/portable/cpu/op_reciprocal.cpp
@@ -14,17 +14,14 @@ namespace executor {
 namespace native {
 namespace {
 
-double reciprocal(double x) {
+template <typename T>
+T reciprocal(T 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(reciprocal_out, reciprocal)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_rsqrt.cpp b/kernels/portable/cpu/op_rsqrt.cpp
@@ -14,15 +14,14 @@ namespace executor {
 namespace native {
 namespace {
 
-double rsqrt(double x) {
+template <typename T>
+T rsqrt(T 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(rsqrt_out, rsqrt)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_sin.cpp b/kernels/portable/cpu/op_sin.cpp
@@ -14,9 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(sin_out, std::sin)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_sinh.cpp b/kernels/portable/cpu/op_sinh.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(sinh_out, std::sinh)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_sqrt.cpp b/kernels/portable/cpu/op_sqrt.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(sqrt_out, std::sqrt)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_tan.cpp b/kernels/portable/cpu/op_tan.cpp
@@ -14,9 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(tan_out, std::tan)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/op_tanh.cpp b/kernels/portable/cpu/op_tanh.cpp
@@ -14,10 +14,7 @@ namespace torch {
 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);
-}
+DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(tanh_out, std::tanh)
 
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/pattern/pattern.h b/kernels/portable/cpu/pattern/pattern.h
@@ -84,11 +84,18 @@ Tensor& unary_ufunc_realhb_to_bool(
  * the math operation which is applied to the input tensor element-wise.
  */
 Tensor& unary_ufunc_realhbbf16_to_floathbf16(
-    double (*fn)(double),
+    float (*fn_float)(float),
+    double (*fn_double)(double),
     KernelRuntimeContext& ctx,
     const Tensor& in,
     Tensor& out);
 
+#define DEFINE_UNARY_UFUNC_REALHBBF16_TO_FLOATHBF16(op_name, fn)              \
+  Tensor& op_name(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) { \
+    return internal::unary_ufunc_realhbbf16_to_floathbf16(                    \
+        fn, fn, ctx, in, out);                                                \
+  }
+
 } // namespace internal
 } // namespace native
 } // namespace executor
diff --git a/kernels/portable/cpu/pattern/unary_ufunc_realhbbf16_to_floathbf16.cpp b/kernels/portable/cpu/pattern/unary_ufunc_realhbbf16_to_floathbf16.cpp

Original file line number	Diff line number	Diff line change
`@@ -60,7 +60,7 @@ Tensor& exp_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`60`	`60`	`return out;`
`61`	`61`	`} else {`
`62`	`62`	`return torch::executor::native::internal::`
`63`		`- unary_ufunc_realhbbf16_to_floathbf16(std::exp, ctx, in, out);`
	`63`	`+ unary_ufunc_realhbbf16_to_floathbf16(std::exp, std::exp, ctx, in, out);`
`64`	`64`	`}`
`65`	`65`	`}`
`66`	`66`
Original file line number	Diff line number	Diff line change
`@@ -27,7 +27,8 @@ namespace native {`
`27`	`27`
`28`	`28`	`namespace {`
`29`	`29`
`30`		`-double rsqrt(double x) {`
	`30`	`+template <typename T>`
	`31`	`+T rsqrt(T x) {`
`31`	`32`	`return 1.0 / std::sqrt(x);`
`32`	`33`	`}`
`33`	`34`
`@@ -61,11 +62,11 @@ Tensor& rsqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`61`	`62`	`return out;`
`62`	`63`	`} else {`
`63`	`64`	`return torch::executor::native::internal::`
`64`		`- unary_ufunc_realhbbf16_to_floathbf16(rsqrt, ctx, in, out);`
	`65`	`+ unary_ufunc_realhbbf16_to_floathbf16(rsqrt, rsqrt, ctx, in, out);`
`65`	`66`	`}`
`66`	`67`	`}`
`67`	`68`
`68`	`69`	`} // namespace native`
`69`	`70`	`} // namespace G3`
`70`	`71`	`} // namespace impl`
`71`		`-} // namespace cadence`
	`72`	`+} // namespace cadence`
Original file line number	Diff line number	Diff line change
`@@ -55,7 +55,8 @@ Tensor& sqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`55`	`55`	`return out;`
`56`	`56`	`} else {`
`57`	`57`	`return torch::executor::native::internal::`
`58`		`- unary_ufunc_realhbbf16_to_floathbf16(std::sqrt, ctx, in, out);`
	`58`	`+ unary_ufunc_realhbbf16_to_floathbf16(`
	`59`	`+ std::sqrt, std::sqrt, ctx, in, out);`
`59`	`60`	`}`
`60`	`61`	`}`
`61`	`62`
Original file line number	Diff line number	Diff line change
`@@ -55,7 +55,8 @@ Tensor& tanh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`55`	`55`	`return out;`
`56`	`56`	`} else {`
`57`	`57`	`return torch::executor::native::internal::`
`58`		`- unary_ufunc_realhbbf16_to_floathbf16(std::tanh, ctx, in, out);`
	`58`	`+ unary_ufunc_realhbbf16_to_floathbf16(`
	`59`	`+ std::tanh, std::tanh, ctx, in, out);`
`59`	`60`	`}`
`60`	`61`	`}`
`61`	`62`
Original file line number	Diff line number	Diff line change
`@@ -21,7 +21,8 @@ namespace HiFi {`
`21`	`21`	`namespace native {`
`22`	`22`	`namespace {`
`23`	`23`
`24`		`-double rsqrt(double x) {`
	`24`	`+template <typename T>`
	`25`	`+T rsqrt(T x) {`
`25`	`26`	`return 1.0 / std::sqrt(x);`
`26`	`27`	`}`
`27`	`28`
`@@ -46,7 +47,7 @@ Tensor& rsqrt_out(RuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`46`	`47`	`}`
`47`	`48`
`48`	`49`	`return torch::executor::native::internal::`
`49`		`- unary_ufunc_realhbbf16_to_floathbf16(rsqrt, ctx, in, out);`
	`50`	`+ unary_ufunc_realhbbf16_to_floathbf16(rsqrt, rsqrt, ctx, in, out);`
`50`	`51`	`}`
`51`	`52`
`52`	`53`	`} // namespace native`
Original file line number	Diff line number	Diff line change
`@@ -35,10 +35,10 @@ Tensor& tanh_out(RuntimeContext& ctx, const Tensor& in, Tensor& out) {`
`35`	`35`	`}`
`36`	`36`
`37`	`37`	`return torch::executor::native::internal::`
`38`		`- unary_ufunc_realhbbf16_to_floathbf16(std::tanh, ctx, in, out);`
	`38`	`+ unary_ufunc_realhbbf16_to_floathbf16(std::tanh, std::tanh, ctx, in, out);`
`39`	`39`	`}`
`40`	`40`
`41`	`41`	`} // namespace native`
`42`	`42`	`} // namespace HiFi`
`43`	`43`	`} // namespace impl`
`44`		`-} // namespace cadence`
	`44`	`+} // namespace cadence`