shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity

[libc][math] Refactor acoshf16 implementation to header-only in src/__support/math folder. (#148568)

Browse Source 
Part of #147386

in preparation for:
https://discourse.llvm.org/t/rfc-make-clang-builtin-math-functions-constexpr-with-llvm-libc-to-support-c-23-constexpr-math-functions/86450
This commit is contained in:
Muhammad Bassiouni 2025-07-22 00:18:15 +03:00 committed by GitHub
parent dcffa3d05c
commit 520398e752
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 200 additions and 109 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
libc/shared/math.h
View File

@ -15,6 +15,7 @@
#include "math/acosf.h"
#include "math/acosf16.h"
#include "math/acoshf.h"
#include "math/acoshf16.h"
#include "math/erff.h"
#include "math/exp.h"
#include "math/exp10.h"

29
libc/shared/math/acoshf16.h Normal file
View File

@ -0,0 +1,29 @@
//===-- Shared acoshf16 function --------------------------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_LIBC_SHARED_MATH_ACOSHF16_H
#define LLVM_LIBC_SHARED_MATH_ACOSHF16_H
#include "include/llvm-libc-macros/float16-macros.h"
#include "shared/libc_common.h"
#ifdef LIBC_TYPES_HAS_FLOAT16
#include "src/__support/math/acoshf16.h"
namespace LIBC_NAMESPACE_DECL {
namespace shared {
using math::acoshf16;
} // namespace shared
} // namespace LIBC_NAMESPACE_DECL
#endif // LIBC_TYPES_HAS_FLOAT16
#endif // LLVM_LIBC_SHARED_MATH_ACOSHF16_H

16
libc/src/__support/math/CMakeLists.txt
View File

@ -79,6 +79,22 @@ add_header_library(
libc.src.__support.macros.optimization
)
add_header_library(
acoshf16
HDRS
acoshf16.h
DEPENDS
.acoshf_utils
libc.src.__support.FPUtil.cast
libc.src.__support.FPUtil.except_value_utils
libc.src.__support.FPUtil.fenv_impl
libc.src.__support.FPUtil.fp_bits
libc.src.__support.FPUtil.multiply_add
libc.src.__support.FPUtil.polyeval
libc.src.__support.FPUtil.sqrt
libc.src.__support.macros.optimization
)
add_header_library(
asin_utils
HDRS

123
libc/src/__support/math/acoshf16.h Normal file
View File

@ -0,0 +1,123 @@
//===-- Implementation header for acoshf16 ----------------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_LIBC_SRC___SUPPORT_MATH_ACOSHF_H
#define LLVM_LIBC_SRC___SUPPORT_MATH_ACOSHF_H
#include "include/llvm-libc-macros/float16-macros.h"
#ifdef LIBC_TYPES_HAS_FLOAT16
#include "acoshf_utils.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/PolyEval.h"
#include "src/__support/FPUtil/cast.h"
#include "src/__support/FPUtil/except_value_utils.h"
#include "src/__support/FPUtil/multiply_add.h"
#include "src/__support/FPUtil/sqrt.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h"
namespace LIBC_NAMESPACE_DECL {
namespace math {
static constexpr float16 acoshf16(float16 x) {
using namespace acoshf_internal;
constexpr size_t N_EXCEPTS = 2;
constexpr fputil::ExceptValues<float16, N_EXCEPTS> ACOSHF16_EXCEPTS{{
// (input, RZ output, RU offset, RD offset, RN offset)
// x = 0x1.6dcp+1, acoshf16(x) = 0x1.b6p+0 (RZ)
{0x41B7, 0x3ED8, 1, 0, 0},
// x = 0x1.39p+0, acoshf16(x) = 0x1.4f8p-1 (RZ)
{0x3CE4, 0x393E, 1, 0, 1},
}};
using FPBits = fputil::FPBits<float16>;
FPBits xbits(x);
uint16_t x_u = xbits.uintval();
// Check for NaN input first.
if (LIBC_UNLIKELY(xbits.is_inf_or_nan())) {
if (xbits.is_signaling_nan()) {
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
if (xbits.is_neg()) {
fputil::set_errno_if_required(EDOM);
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
return x;
}
// Domain error for inputs less than 1.0.
if (LIBC_UNLIKELY(x <= 1.0f)) {
if (x == 1.0f)
return FPBits::zero().get_val();
fputil::set_errno_if_required(EDOM);
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
if (auto r = ACOSHF16_EXCEPTS.lookup(xbits.uintval());
LIBC_UNLIKELY(r.has_value()))
return r.value();
float xf = x;
// High-precision polynomial approximation for inputs close to 1.0
// ([1, 1.25)).
//
// Brief derivation:
// 1. Expand acosh(1 + delta) using Taylor series around delta=0:
// acosh(1 + delta) ≈ sqrt(2 * delta) * [1 - delta/12 + 3*delta^2/160
// - 5*delta^3/896 + 35*delta^4/18432 + ...]
// 2. Truncate the series to fit accurately for delta in [0, 0.25].
// 3. Polynomial coefficients (from sollya) used here are:
// P(delta) ≈ 1 - 0x1.555556p-4 * delta + 0x1.333334p-6 * delta^2
// - 0x1.6db6dcp-8 * delta^3 + 0x1.f1c71cp-10 * delta^4
// 4. The Sollya commands used to generate these coefficients were:
// > display = hexadecimal;
// > round(1/12, SG, RN);
// > round(3/160, SG, RN);
// > round(5/896, SG, RN);
// > round(35/18432, SG, RN);
// With hexadecimal display mode enabled, the outputs were:
// 0x1.555556p-4
// 0x1.333334p-6
// 0x1.6db6dcp-8
// 0x1.f1c71cp-10
// 5. The maximum absolute error, estimated using:
// dirtyinfnorm(acosh(1 + x) - sqrt(2*x) * P(x), [0, 0.25])
// is:
// 0x1.d84281p-22
if (LIBC_UNLIKELY(x_u < 0x3D00U)) {
float delta = xf - 1.0f;
float sqrt_2_delta = fputil::sqrt<float>(2.0 * delta);
float pe = fputil::polyeval(delta, 0x1p+0f, -0x1.555556p-4f, 0x1.333334p-6f,
-0x1.6db6dcp-8f, 0x1.f1c71cp-10f);
float approx = sqrt_2_delta * pe;
return fputil::cast<float16>(approx);
}
// acosh(x) = log(x + sqrt(x^2 - 1))
float sqrt_term = fputil::sqrt<float>(fputil::multiply_add(xf, xf, -1.0f));
float result = static_cast<float>(log_eval(xf + sqrt_term));
return fputil::cast<float16>(result);
}
} // namespace math
} // namespace LIBC_NAMESPACE_DECL
#endif // LIBC_TYPES_HAS_FLOAT16
#endif // LLVM_LIBC_SRC___SUPPORT_MATH_ACOSHF_H

14
libc/src/math/generic/CMakeLists.txt
View File

@ -3878,18 +3878,8 @@ add_entrypoint_object(
HDRS
../acoshf16.h
DEPENDS
.explogxf
libc.hdr.errno_macros
libc.hdr.fenv_macros
libc.src.__support.FPUtil.cast
libc.src.__support.FPUtil.except_value_utils
libc.src.__support.FPUtil.fenv_impl
libc.src.__support.FPUtil.fp_bits
libc.src.__support.FPUtil.multiply_add
libc.src.__support.FPUtil.polyeval
libc.src.__support.FPUtil.sqrt
libc.src.__support.macros.optimization
libc.src.__support.macros.properties.types
libc.src.__support.math.acoshf16
libc.src.errno.errno
)
add_entrypoint_object(

99
libc/src/math/generic/acoshf16.cpp
View File

@ -7,105 +7,10 @@
//===----------------------------------------------------------------------===//
#include "src/math/acoshf16.h"
#include "explogxf.h"
#include "hdr/errno_macros.h"
#include "hdr/fenv_macros.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/PolyEval.h"
#include "src/__support/FPUtil/cast.h"
#include "src/__support/FPUtil/except_value_utils.h"
#include "src/__support/FPUtil/multiply_add.h"
#include "src/__support/FPUtil/sqrt.h"
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h"
#include "src/__support/math/acoshf16.h"
namespace LIBC_NAMESPACE_DECL {
static constexpr size_t N_EXCEPTS = 2;
static constexpr fputil::ExceptValues<float16, N_EXCEPTS> ACOSHF16_EXCEPTS{{
// (input, RZ output, RU offset, RD offset, RN offset)
// x = 0x1.6dcp+1, acoshf16(x) = 0x1.b6p+0 (RZ)
{0x41B7, 0x3ED8, 1, 0, 0},
// x = 0x1.39p+0, acoshf16(x) = 0x1.4f8p-1 (RZ)
{0x3CE4, 0x393E, 1, 0, 1},
}};
LLVM_LIBC_FUNCTION(float16, acoshf16, (float16 x)) {
using namespace acoshf_internal;
using FPBits = fputil::FPBits<float16>;
FPBits xbits(x);
uint16_t x_u = xbits.uintval();
// Check for NaN input first.
if (LIBC_UNLIKELY(xbits.is_inf_or_nan())) {
if (xbits.is_signaling_nan()) {
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
if (xbits.is_neg()) {
fputil::set_errno_if_required(EDOM);
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
return x;
}
// Domain error for inputs less than 1.0.
if (LIBC_UNLIKELY(x <= 1.0f)) {
if (x == 1.0f)
return FPBits::zero().get_val();
fputil::set_errno_if_required(EDOM);
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
if (auto r = ACOSHF16_EXCEPTS.lookup(xbits.uintval());
LIBC_UNLIKELY(r.has_value()))
return r.value();
float xf = x;
// High-precision polynomial approximation for inputs close to 1.0
// ([1, 1.25)).
//
// Brief derivation:
// 1. Expand acosh(1 + delta) using Taylor series around delta=0:
// acosh(1 + delta) ≈ sqrt(2 * delta) * [1 - delta/12 + 3*delta^2/160
// - 5*delta^3/896 + 35*delta^4/18432 + ...]
// 2. Truncate the series to fit accurately for delta in [0, 0.25].
// 3. Polynomial coefficients (from sollya) used here are:
// P(delta) ≈ 1 - 0x1.555556p-4 * delta + 0x1.333334p-6 * delta^2
// - 0x1.6db6dcp-8 * delta^3 + 0x1.f1c71cp-10 * delta^4
// 4. The Sollya commands used to generate these coefficients were:
// > display = hexadecimal;
// > round(1/12, SG, RN);
// > round(3/160, SG, RN);
// > round(5/896, SG, RN);
// > round(35/18432, SG, RN);
// With hexadecimal display mode enabled, the outputs were:
// 0x1.555556p-4
// 0x1.333334p-6
// 0x1.6db6dcp-8
// 0x1.f1c71cp-10
// 5. The maximum absolute error, estimated using:
// dirtyinfnorm(acosh(1 + x) - sqrt(2*x) * P(x), [0, 0.25])
// is:
// 0x1.d84281p-22
if (LIBC_UNLIKELY(x_u < 0x3D00U)) {
float delta = xf - 1.0f;
float sqrt_2_delta = fputil::sqrt<float>(2.0 * delta);
float pe = fputil::polyeval(delta, 0x1p+0f, -0x1.555556p-4f, 0x1.333334p-6f,
-0x1.6db6dcp-8f, 0x1.f1c71cp-10f);
float approx = sqrt_2_delta * pe;
return fputil::cast<float16>(approx);
}
// acosh(x) = log(x + sqrt(x^2 - 1))
float sqrt_term = fputil::sqrt<float>(fputil::multiply_add(xf, xf, -1.0f));
float result = static_cast<float>(log_eval(xf + sqrt_term));
return fputil::cast<float16>(result);
}
LLVM_LIBC_FUNCTION(float16, acoshf16, (float16 x)) { return math::acoshf16(x); }
} // namespace LIBC_NAMESPACE_DECL

1
libc/test/shared/CMakeLists.txt
View File

@ -11,6 +11,7 @@ add_fp_unittest(
libc.src.__support.math.acosf
libc.src.__support.math.acosf16
libc.src.__support.math.acoshf
libc.src.__support.math.acoshf16
libc.src.__support.math.erff
libc.src.__support.math.exp
libc.src.__support.math.exp10

2
libc/test/shared/shared_math_test.cpp
View File

@ -14,6 +14,8 @@
TEST(LlvmLibcSharedMathTest, AllFloat16) {
int exponent;
EXPECT_FP_EQ(0x0p+0f, LIBC_NAMESPACE::shared::acoshf16(1.0f));
EXPECT_FP_EQ(0x1p+0f16, LIBC_NAMESPACE::shared::exp10f16(0.0f16));
EXPECT_FP_EQ(0x1p+0f16, LIBC_NAMESPACE::shared::expf16(0.0f16));

24
utils/bazel/llvm-project-overlay/libc/BUILD.bazel
View File

@ -2141,6 +2141,22 @@ libc_support_library(
],
)
libc_support_library(
name = "__support_math_acoshf16",
hdrs = ["src/__support/math/acoshf16.h"],
deps = [
":__support_math_acoshf_utils",
":__support_fputil_cast",
":__support_fputil_except_value_utils",
":__support_fputil_fenv_impl",
":__support_fputil_fp_bits",
":__support_fputil_multiply_add",
":__support_fputil_polyeval",
":__support_fputil_sqrt",
":__support_macros_optimization",
],
)
libc_support_library(
name = "__support_math_asin_utils",
hdrs = ["src/__support/math/asin_utils.h"],
@ -2682,6 +2698,14 @@ libc_math_function(
],
)
libc_math_function(
name = "acoshf16",
additional_deps = [
":__support_math_acoshf16",
":errno",
],
)
libc_math_function(
name = "asinf",
additional_deps = [