intel/llvm
1
0
Fork 0
mirror of https://github.com/intel/llvm.git synced 2026-01-14 11:57:39 +08:00
Code Issues Packages Projects Releases Wiki Activity

[libc][math] Refactor asinhf16 implementation to header-only in src/__support/math folder. (#150849)

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
2025-07-29 20:19:01 +03:00
committed by GitHub
parent a28e7f1aad
commit efbbc0b319
10 changed files with 199 additions and 107 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
libc/shared/math.h
View File

@@ -21,6 +21,7 @@
#include "math/asinf.h"
#include "math/asinf16.h"
#include "math/asinhf.h"
#include "math/asinhf16.h"
#include "math/erff.h"
#include "math/exp.h"
#include "math/exp10.h"

2
libc/shared/math/asinf16.h
View File

@@ -25,4 +25,4 @@ using math::asinf16;
#endif // LIBC_TYPES_HAS_FLOAT16
#endif // LLVM_LIBC_SHARED_MATH_ASINF_H
#endif // LLVM_LIBC_SHARED_MATH_ASINF16_H

28
libc/shared/math/asinhf16.h Normal file
View File

@@ -0,0 +1,28 @@
//===-- Shared asinhf16 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_ASINHF16_H
#define LLVM_LIBC_SHARED_MATH_ASINHF16_H
#include "shared/libc_common.h"
#ifdef LIBC_TYPES_HAS_FLOAT16
#include "src/__support/math/asinhf16.h"
namespace LIBC_NAMESPACE_DECL {
namespace shared {
using math::asinhf16;
} // namespace shared
} // namespace LIBC_NAMESPACE_DECL
#endif // LIBC_TYPES_HAS_FLOAT16
#endif // LLVM_LIBC_SHARED_MATH_ASINHF16_H

18
libc/src/__support/math/CMakeLists.txt
View File

@@ -154,6 +154,24 @@ add_header_library(
libc.src.__support.macros.optimization
)
add_header_library(
asinhf16
HDRS
asinhf16.h
DEPENDS
.acoshf_utils
libc.src.__support.FPUtil.fenv_impl
libc.src.__support.FPUtil.fp_bits
libc.src.__support.FPUtil.polyeval
libc.src.__support.FPUtil.cast
libc.src.__support.FPUtil.except_value_utils
libc.src.__support.FPUtil.multiply_add
libc.src.__support.FPUtil.rounding_mode
libc.src.__support.FPUtil.sqrt
libc.src.__support.macros.config
libc.src.__support.macros.optimization
)
add_header_library(
asinf
HDRS

121
libc/src/__support/math/asinhf16.h Normal file
View File

@@ -0,0 +1,121 @@
//===-- Implementation header for asinhf16 ----------------------*- 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_ASINHF16_H
#define LLVM_LIBC_SRC___SUPPORT_MATH_ASINHF16_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/rounding_mode.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 {
LIBC_INLINE static constexpr float16 asinhf16(float16 x) {
#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
constexpr size_t N_EXCEPTS = 8;
constexpr fputil::ExceptValues<float16, N_EXCEPTS> ASINHF16_EXCEPTS{{
// (input, RZ output, RU offset, RD offset, RN offset)
// x = 0x1.da4p-2, asinhf16(x) = 0x1.ca8p-2 (RZ)
{0x3769, 0x372a, 1, 0, 1},
// x = 0x1.d6cp-1, asinhf16(x) = 0x1.a58p-1 (RZ)
{0x3b5b, 0x3a96, 1, 0, 0},
// x = 0x1.c7cp+3, asinhf16(x) = 0x1.accp+1 (RZ)
{0x4b1f, 0x42b3, 1, 0, 0},
// x = 0x1.26cp+4, asinhf16(x) = 0x1.cd8p+1 (RZ)
{0x4c9b, 0x4336, 1, 0, 1},
// x = -0x1.da4p-2, asinhf16(x) = -0x1.ca8p-2 (RZ)
{0xb769, 0xb72a, 0, 1, 1},
// x = -0x1.d6cp-1, asinhf16(x) = -0x1.a58p-1 (RZ)
{0xbb5b, 0xba96, 0, 1, 0},
// x = -0x1.c7cp+3, asinhf16(x) = -0x1.accp+1 (RZ)
{0xcb1f, 0xc2b3, 0, 1, 0},
// x = -0x1.26cp+4, asinhf16(x) = -0x1.cd8p+1 (RZ)
{0xcc9b, 0xc336, 0, 1, 1},
}};
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
using namespace acoshf_internal;
using FPBits = fputil::FPBits<float16>;
FPBits xbits(x);
uint16_t x_u = xbits.uintval();
uint16_t x_abs = x_u & 0x7fff;
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();
}
return x;
}
#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
// Handle exceptional values
if (auto r = ASINHF16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value()))
return r.value();
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
float xf = x;
const float SIGN[2] = {1.0f, -1.0f};
float x_sign = SIGN[x_u >> 15];
// |x| <= 0.25
if (LIBC_UNLIKELY(x_abs <= 0x3400)) {
// when |x| < 0x1.718p-5, asinhf16(x) = x. Adjust by 1 ULP for certain
// rounding types.
if (LIBC_UNLIKELY(x_abs < 0x29c6)) {
int rounding = fputil::quick_get_round();
if ((rounding == FE_UPWARD || rounding == FE_TOWARDZERO) && xf < 0)
return fputil::cast<float16>(xf + 0x1p-24f);
if ((rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) && xf > 0)
return fputil::cast<float16>(xf - 0x1p-24f);
return fputil::cast<float16>(xf);
}
float x_sq = xf * xf;
// Generated by Sollya with:
// > P = fpminimax(asinh(x)/x, [|0, 2, 4, 6, 8|], [|SG...|], [0, 2^-2]);
// The last coefficient 0x1.bd114ep-6f has been changed to 0x1.bd114ep-5f
// for better accuracy.
float p = fputil::polyeval(x_sq, 1.0f, -0x1.555552p-3f, 0x1.332f6ap-4f,
-0x1.6c53dep-5f, 0x1.bd114ep-5f);
return fputil::cast<float16>(xf * p);
}
// General case: asinh(x) = ln(x + sqrt(x^2 + 1))
float sqrt_term = fputil::sqrt<float>(fputil::multiply_add(xf, xf, 1.0f));
return fputil::cast<float16>(
x_sign * log_eval(fputil::multiply_add(xf, x_sign, sqrt_term)));
}
} // namespace math
} // namespace LIBC_NAMESPACE_DECL
#endif // LIBC_TYPES_HAS_FLOAT16
#endif // LLVM_LIBC_SRC___SUPPORT_MATH_ASINHF16_H

13
libc/src/math/generic/CMakeLists.txt
View File

@@ -3912,18 +3912,7 @@ add_entrypoint_object(
HDRS
../asinhf16.h
DEPENDS
.explogxf
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.rounding_mode
libc.src.__support.FPUtil.sqrt
libc.src.__support.macros.optimization
libc.src.__support.macros.properties.types
libc.src.__support.math.asinhf16
)
add_entrypoint_object(

96
libc/src/math/generic/asinhf16.cpp
View File

@@ -7,102 +7,10 @@
//===----------------------------------------------------------------------===//
#include "src/math/asinhf16.h"
#include "explogxf.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/rounding_mode.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/asinhf16.h"
namespace LIBC_NAMESPACE_DECL {
#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
static constexpr size_t N_EXCEPTS = 8;
static constexpr fputil::ExceptValues<float16, N_EXCEPTS> ASINHF16_EXCEPTS{{
// (input, RZ output, RU offset, RD offset, RN offset)
// x = 0x1.da4p-2, asinhf16(x) = 0x1.ca8p-2 (RZ)
{0x3769, 0x372a, 1, 0, 1},
// x = 0x1.d6cp-1, asinhf16(x) = 0x1.a58p-1 (RZ)
{0x3b5b, 0x3a96, 1, 0, 0},
// x = 0x1.c7cp+3, asinhf16(x) = 0x1.accp+1 (RZ)
{0x4b1f, 0x42b3, 1, 0, 0},
// x = 0x1.26cp+4, asinhf16(x) = 0x1.cd8p+1 (RZ)
{0x4c9b, 0x4336, 1, 0, 1},
// x = -0x1.da4p-2, asinhf16(x) = -0x1.ca8p-2 (RZ)
{0xb769, 0xb72a, 0, 1, 1},
// x = -0x1.d6cp-1, asinhf16(x) = -0x1.a58p-1 (RZ)
{0xbb5b, 0xba96, 0, 1, 0},
// x = -0x1.c7cp+3, asinhf16(x) = -0x1.accp+1 (RZ)
{0xcb1f, 0xc2b3, 0, 1, 0},
// x = -0x1.26cp+4, asinhf16(x) = -0x1.cd8p+1 (RZ)
{0xcc9b, 0xc336, 0, 1, 1},
}};
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
LLVM_LIBC_FUNCTION(float16, asinhf16, (float16 x)) {
using namespace acoshf_internal;
using FPBits = fputil::FPBits<float16>;
FPBits xbits(x);
uint16_t x_u = xbits.uintval();
uint16_t x_abs = x_u & 0x7fff;
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();
}
return x;
}
#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
// Handle exceptional values
if (auto r = ASINHF16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value()))
return r.value();
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
float xf = x;
const float SIGN[2] = {1.0f, -1.0f};
float x_sign = SIGN[x_u >> 15];
// |x| <= 0.25
if (LIBC_UNLIKELY(x_abs <= 0x3400)) {
// when |x| < 0x1.718p-5, asinhf16(x) = x. Adjust by 1 ULP for certain
// rounding types.
if (LIBC_UNLIKELY(x_abs < 0x29c6)) {
int rounding = fputil::quick_get_round();
if ((rounding == FE_UPWARD || rounding == FE_TOWARDZERO) && xf < 0)
return fputil::cast<float16>(xf + 0x1p-24f);
if ((rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) && xf > 0)
return fputil::cast<float16>(xf - 0x1p-24f);
return fputil::cast<float16>(xf);
}
float x_sq = xf * xf;
// Generated by Sollya with:
// > P = fpminimax(asinh(x)/x, [|0, 2, 4, 6, 8|], [|SG...|], [0, 2^-2]);
// The last coefficient 0x1.bd114ep-6f has been changed to 0x1.bd114ep-5f
// for better accuracy.
float p = fputil::polyeval(x_sq, 1.0f, -0x1.555552p-3f, 0x1.332f6ap-4f,
-0x1.6c53dep-5f, 0x1.bd114ep-5f);
return fputil::cast<float16>(xf * p);
}
// General case: asinh(x) = ln(x + sqrt(x^2 + 1))
float sqrt_term = fputil::sqrt<float>(fputil::multiply_add(xf, xf, 1.0f));
return fputil::cast<float16>(
x_sign * log_eval(fputil::multiply_add(xf, x_sign, sqrt_term)));
}
LLVM_LIBC_FUNCTION(float16, asinhf16, (float16 x)) { return math::asinhf16(x); }
} // namespace LIBC_NAMESPACE_DECL

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

@@ -17,6 +17,7 @@ add_fp_unittest(
libc.src.__support.math.asinf
libc.src.__support.math.asinf16
libc.src.__support.math.asinhf
libc.src.__support.math.asinhf16
libc.src.__support.math.erff
libc.src.__support.math.exp
libc.src.__support.math.exp10

1
libc/test/shared/shared_math_test.cpp
View File

@@ -18,6 +18,7 @@ TEST(LlvmLibcSharedMathTest, AllFloat16) {
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::acoshf16(1.0f16));
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::acospif16(1.0f16));
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::asinf16(0.0f16));
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::asinhf16(0.0f16));
EXPECT_FP_EQ(0x1p+0f16, LIBC_NAMESPACE::shared::exp10f16(0.0f16));

25
utils/bazel/llvm-project-overlay/libc/BUILD.bazel
View File

@@ -2245,6 +2245,24 @@ libc_support_library(
],
)
libc_support_library(
name = "__support_math_asinhf16",
hdrs = ["src/__support/math/asinhf16.h"],
deps = [
":__support_math_acoshf_utils",
":__support_fputil_fenv_impl",
":__support_fputil_fp_bits",
":__support_fputil_polyeval",
":__support_fputil_cast",
":__support_fputil_except_value_utils",
":__support_fputil_multiply_add",
":__support_fputil_rounding_mode",
":__support_fputil_sqrt",
":__support_macros_config",
":__support_macros_optimization",
],
)
libc_support_library(
name = "__support_math_asinf",
hdrs = ["src/__support/math/asinf.h"],
@@ -2847,6 +2865,13 @@ libc_math_function(
],
)
libc_math_function(
name = "asinhf16",
additional_deps = [
":__support_math_asinhf16",
],
)
libc_math_function(
name = "atanf",
additional_deps = [