Add float16 exp/log math functions via promotion to float32

BUILD

@@ -433,6 +433,7 @@ cc_library(
     copts = COPTS,
     textual_hdrs = [
         "hwy/contrib/math/math-inl.h",
+        "hwy/contrib/math/f16_math-inl.h",
         "hwy/contrib/math/fast_math-inl.h",
         "hwy/contrib/math/fp_arith-inl.h",
     ],

CMakeLists.txt

@@ -240,6 +240,7 @@ list(APPEND HWY_CONTRIB_SOURCES
     hwy/contrib/hash/cuckoo2x2-inl.h
     hwy/contrib/image/image.cc
     hwy/contrib/image/image.h
+    hwy/contrib/math/f16_math-inl.h
     hwy/contrib/math/fast_math-inl.h
     hwy/contrib/intdiv/intdiv-inl.h
     hwy/contrib/math/fp_arith-inl.h
@@ -1029,6 +1030,7 @@ list(APPEND HWY_TEST_FILES
   hwy/contrib/hash/shardmul_test.cc
   hwy/contrib/image/image_test.cc
   hwy/contrib/intdiv/intdiv_test.cc
+  hwy/contrib/math/f16_math_test.cc
   hwy/contrib/math/fast_math_test.cc
   hwy/contrib/math/math_hyper_test.cc
   hwy/contrib/math/math_tan_test.cc

hwy.gni

@@ -52,6 +52,7 @@ hwy_contrib_public = [
   "$_hwy/contrib/algo/transform-inl.h",
   "$_hwy/contrib/dot/dot-inl.h",
   "$_hwy/contrib/image/image.h",
+  "$_hwy/contrib/math/f16_math-inl.h",
   "$_hwy/contrib/math/fast_math-inl.h",
   "$_hwy/contrib/math/fp_arith-inl.h",
   "$_hwy/contrib/intdiv/intdiv-inl.h",

hwy/contrib/math/f16_math-inl.h

@@ -0,0 +1,236 @@
+// Copyright 2026 Google LLC
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Include guard (still compiled once per target)
+#if defined(HIGHWAY_HWY_CONTRIB_MATH_F16_MATH_INL_H_) == \
+    defined(HWY_TARGET_TOGGLE)  // NOLINT
+#ifdef HIGHWAY_HWY_CONTRIB_MATH_F16_MATH_INL_H_
+#undef HIGHWAY_HWY_CONTRIB_MATH_F16_MATH_INL_H_
+#else
+#define HIGHWAY_HWY_CONTRIB_MATH_F16_MATH_INL_H_
+#endif
+
+#include "hwy/contrib/math/math-inl.h"
+#include "hwy/highway.h"
+
+HWY_BEFORE_NAMESPACE();
+namespace hwy {
+namespace HWY_NAMESPACE {
+
+// float16_t overloads of the math functions in math-inl.h. There are no
+// float16 kernels yet, so these promote to float32, evaluate the float32
+// kernel, and demote the result. The ops used here support float16_t lanes
+// on all targets, even when HWY_HAVE_FLOAT16 is 0.
+
+// Not named `impl`: unqualified calls such as Log() from inside that
+// namespace would also find math-inl.h's impl::Log and be ambiguous.
+namespace f16_impl {
+
+// Promotes lower/upper halves of the float16 vector `v` to float32,
+// evaluates `kernel` (a functor wrapping one float32 math function) twice,
+// then demotes and recombines. Never instantiated on HWY_SCALAR, so
+// PromoteUpperTo and Combine are safe to name here. There is no
+// OrderedDemote2To for f32->f16, hence DemoteTo + Combine.
+template <class D, class Kernel, class V = VFromD<D>>
+HWY_INLINE V F16ViaF32PerHalf(D d, V v, Kernel kernel) {
+  const Half<D> dh;
+  const RepartitionToWide<D> df32;
+  const VFromD<decltype(df32)> lo = kernel(df32, PromoteLowerTo(df32, v));
+  const VFromD<decltype(df32)> hi = kernel(df32, PromoteUpperTo(df32, v));
+  return Combine(d, DemoteTo(dh, hi), DemoteTo(dh, lo));
+}
+
+// Evaluates `kernel` on all lanes of `v` with a single promotion and one
+// kernel evaluation. This overload handles vectors that fit in half of the
+// widest vector - including all partial vectors, N=1, and the entire
+// HWY_SCALAR target. The kPow2 bound ensures Rebind<float, D> (one more
+// kPow2 because float lanes are twice as large) is a valid tag.
+template <class D, class Kernel, class V = VFromD<D>, HWY_IF_F16_D(D),
+          HWY_IF_V_SIZE_LE_D(D, HWY_MAX_BYTES / 2),
+          HWY_IF_POW2_LE_D(D, HWY_MAX_POW2 - 1)>
+HWY_INLINE V F16ViaF32(D d, V v, Kernel kernel) {
+  const Rebind<float, D> df32;
+  return DemoteTo(d, kernel(df32, PromoteTo(df32, v)));
+}
+
+// Wider vectors take the per-half path.
+template <class D, class Kernel, class V = VFromD<D>, HWY_IF_F16_D(D),
+          HWY_IF_V_SIZE_GT_D(D, HWY_MAX_BYTES / 2)>
+HWY_INLINE V F16ViaF32(D d, V v, Kernel kernel) {
+  return F16ViaF32PerHalf(d, v, kernel);
+}
+
+// Small vectors already at the maximum kPow2 (possible on RVV, e.g.
+// CappedTag<float16_t, 8, 3>): Rebind<float, D> would exceed HWY_MAX_POW2,
+// so use the per-half path, whose RepartitionToWide keeps kPow2 unchanged.
+template <class D, class Kernel, class V = VFromD<D>, HWY_IF_F16_D(D),
+          HWY_IF_V_SIZE_LE_D(D, HWY_MAX_BYTES / 2),
+          HWY_IF_POW2_GT_D(D, HWY_MAX_POW2 - 1)>
+HWY_INLINE V F16ViaF32(D d, V v, Kernel kernel) {
+  return F16ViaF32PerHalf(d, v, kernel);
+}
+
+struct ExpKernel {
+  template <class DF, class VF>
+  HWY_INLINE VF operator()(DF df, VF x) const {
+    return Exp(df, x);
+  }
+};
+
+struct Exp2Kernel {
+  template <class DF, class VF>
+  HWY_INLINE VF operator()(DF df, VF x) const {
+    return Exp2(df, x);
+  }
+};
+
+struct Expm1Kernel {
+  template <class DF, class VF>
+  HWY_INLINE VF operator()(DF df, VF x) const {
+    return Expm1(df, x);
+  }
+};
+
+struct LogKernel {
+  template <class DF, class VF>
+  HWY_INLINE VF operator()(DF df, VF x) const {
+    return Log(df, x);
+  }
+};
+
+struct Log10Kernel {
+  template <class DF, class VF>
+  HWY_INLINE VF operator()(DF df, VF x) const {
+    return Log10(df, x);
+  }
+};
+
+struct Log1pKernel {
+  template <class DF, class VF>
+  HWY_INLINE VF operator()(DF df, VF x) const {
+    return Log1p(df, x);
+  }
+};
+
+struct Log2Kernel {
+  template <class DF, class VF>
+  HWY_INLINE VF operator()(DF df, VF x) const {
+    return Log2(df, x);
+  }
+};
+
+}  // namespace f16_impl
+
+// The generic templates in math-inl.h are constrained with
+// HWY_IF_NOT_SPECIAL_FLOAT_D, so these HWY_IF_F16_D overloads partition the
+// overload set rather than being ambiguous.
+
+/**
+ * Highway SIMD version of std::exp(x) for float16 lanes.
+ *
+ * Valid Lane Types: float16
+ *        Max Error: ULP = 2
+ *      Valid Range: float16[-65504, +104]
+ * @return e^x
+ */
+template <class D, class V, HWY_IF_F16_D(D)>
+HWY_INLINE V Exp(D d, V x) {
+  return f16_impl::F16ViaF32(d, x, f16_impl::ExpKernel());
+}
+
+/**
+ * Highway SIMD version of std::exp2(x) for float16 lanes.
+ *
+ * Valid Lane Types: float16
+ *        Max Error: ULP = 2
+ *      Valid Range: float16[-65504, +128]
+ * @return 2^x
+ */
+template <class D, class V, HWY_IF_F16_D(D)>
+HWY_INLINE V Exp2(D d, V x) {
+  return f16_impl::F16ViaF32(d, x, f16_impl::Exp2Kernel());
+}
+
+/**
+ * Highway SIMD version of std::expm1(x) for float16 lanes.
+ *
+ * Valid Lane Types: float16
+ *        Max Error: ULP = 2
+ *      Valid Range: float16[-65504, +104]
+ * @return e^x - 1
+ */
+template <class D, class V, HWY_IF_F16_D(D)>
+HWY_INLINE V Expm1(D d, V x) {
+  return f16_impl::F16ViaF32(d, x, f16_impl::Expm1Kernel());
+}
+
+/**
+ * Highway SIMD version of std::log(x) for float16 lanes.
+ *
+ * Valid Lane Types: float16
+ *        Max Error: ULP = 1
+ *      Valid Range: float16(0, +65504]
+ * @return natural logarithm of 'x'
+ */
+template <class D, class V, HWY_IF_F16_D(D)>
+HWY_INLINE V Log(D d, V x) {
+  return f16_impl::F16ViaF32(d, x, f16_impl::LogKernel());
+}
+
+/**
+ * Highway SIMD version of std::log10(x) for float16 lanes.
+ *
+ * Valid Lane Types: float16
+ *        Max Error: ULP = 1
+ *      Valid Range: float16(0, +65504]
+ * @return base 10 logarithm of 'x'
+ */
+template <class D, class V, HWY_IF_F16_D(D)>
+HWY_INLINE V Log10(D d, V x) {
+  return f16_impl::F16ViaF32(d, x, f16_impl::Log10Kernel());
+}
+
+/**
+ * Highway SIMD version of std::log1p(x) for float16 lanes.
+ *
+ * Valid Lane Types: float16
+ *        Max Error: ULP = 2
+ *      Valid Range: float16[0, +65504]
+ * @return log(1 + x)
+ */
+template <class D, class V, HWY_IF_F16_D(D)>
+HWY_INLINE V Log1p(D d, V x) {
+  return f16_impl::F16ViaF32(d, x, f16_impl::Log1pKernel());
+}
+
+/**
+ * Highway SIMD version of std::log2(x) for float16 lanes.
+ *
+ * Valid Lane Types: float16
+ *        Max Error: ULP = 1
+ *      Valid Range: float16(0, +65504]
+ * @return base 2 logarithm of 'x'
+ */
+template <class D, class V, HWY_IF_F16_D(D)>
+HWY_INLINE V Log2(D d, V x) {
+  return f16_impl::F16ViaF32(d, x, f16_impl::Log2Kernel());
+}
+
+// NOLINTNEXTLINE(google-readability-namespace-comments)
+}  // namespace HWY_NAMESPACE
+}  // namespace hwy
+HWY_AFTER_NAMESPACE();
+
+#endif  // HIGHWAY_HWY_CONTRIB_MATH_F16_MATH_INL_H_

hwy/contrib/math/f16_math_test.cc

@@ -0,0 +1,91 @@
+// Copyright 2026 Google LLC
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <stdint.h>
+#include <stdio.h>
+
+#include <cmath>  // std::exp
+
+// For faster tests. Not using AES, hence NEON_WITHOUT_AES is sufficient.
+// SVE is mostly superseded by SVE2.
+#ifndef HWY_DISABLED_TARGETS
+#define HWY_DISABLED_TARGETS (HWY_NEON | HWY_SVE)
+#endif  // HWY_DISABLED_TARGETS
+
+#include "hwy/base.h"
+
+// clang-format off
+#undef HWY_TARGET_INCLUDE
+#define HWY_TARGET_INCLUDE "hwy/contrib/math/f16_math_test.cc"
+#include "hwy/foreach_target.h"  // IWYU pragma: keep
+#include "hwy/highway.h"
+#include "hwy/contrib/math/f16_math-inl.h"
+#include "hwy/contrib/math/math-inl.h"  // CallExp
+#include "hwy/contrib/math/math_test-inl.h"
+// clang-format on
+
+HWY_BEFORE_NAMESPACE();
+namespace hwy {
+namespace HWY_NAMESPACE {
+namespace {
+
+// The float16 min/max bounds mirror the float32 bounds in math_test.cc where
+// the kernel is validated (e.g. +104 for Exp), clamped to the float16 finite
+// range [-65504, +65504]. The smallest positive float16 subnormal is 2^-24.
+//
+// Exp, Exp2, Expm1 and Log1p allow 2 ULP because they are the functions whose
+// results reach the float16 subnormals, where the spacing is a fixed 2^-24 and
+// the smallest results hold only about four significant bits, so a float32
+// result that differs slightly between targets can land two float16 values
+// away. The logarithms never return a subnormal and stay at 1 ULP.
+// clang-format off
+DEFINE_F16_MATH_TEST(Exp,
+  std::exp,   CallExp,   -65504.0f,        +104.0f,   2)
+DEFINE_F16_MATH_TEST(Exp2,
+  std::exp2,  CallExp2,  -65504.0f,        +128.0f,   2)
+DEFINE_F16_MATH_TEST(Expm1,
+  std::expm1, CallExpm1, -65504.0f,        +104.0f,   2)
+DEFINE_F16_MATH_TEST(Log,
+  std::log,   CallLog,   +5.960464478E-8f, +65504.0f, 1)
+DEFINE_F16_MATH_TEST(Log10,
+  std::log10, CallLog10, +5.960464478E-8f, +65504.0f, 1)
+DEFINE_F16_MATH_TEST(Log1p,
+  std::log1p, CallLog1p, +0.0f,            +65504.0f, 2)
+DEFINE_F16_MATH_TEST(Log2,
+  std::log2,  CallLog2,  +5.960464478E-8f, +65504.0f, 1)
+// clang-format on
+
+}  // namespace
+// NOLINTNEXTLINE(google-readability-namespace-comments)
+}  // namespace HWY_NAMESPACE
+}  // namespace hwy
+HWY_AFTER_NAMESPACE();
+
+#if HWY_ONCE
+namespace hwy {
+namespace {
+HWY_BEFORE_TEST(HwyF16MathTest);
+HWY_EXPORT_AND_TEST_P(HwyF16MathTest, TestAllF16Exp);
+HWY_EXPORT_AND_TEST_P(HwyF16MathTest, TestAllF16Exp2);
+HWY_EXPORT_AND_TEST_P(HwyF16MathTest, TestAllF16Expm1);
+HWY_EXPORT_AND_TEST_P(HwyF16MathTest, TestAllF16Log);
+HWY_EXPORT_AND_TEST_P(HwyF16MathTest, TestAllF16Log10);
+HWY_EXPORT_AND_TEST_P(HwyF16MathTest, TestAllF16Log1p);
+HWY_EXPORT_AND_TEST_P(HwyF16MathTest, TestAllF16Log2);
+HWY_AFTER_TEST();
+}  // namespace
+}  // namespace hwy
+HWY_TEST_MAIN();
+#endif  // HWY_ONCE