intel/llvm
1
0
Fork 0
mirror of https://github.com/intel/llvm.git synced 2026-01-30 14:07:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

[libc] Add sinpif16 function (#110994)

Browse Source 
Half-precision floating point (16-bit) implementation of the
trigonometric function Sin for inputs scaled by pi
This commit is contained in:
wldfngrs
2024-10-15 23:40:08 +01:00
committed by GitHub
parent 7cbb365903
commit ddc3f2dd26
13 changed files with 307 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
libc/config/linux/aarch64/entrypoints.txt
View File

@@ -679,6 +679,7 @@ if(LIBC_TYPES_HAS_FLOAT16)
libc.src.math.scalbnf16
libc.src.math.setpayloadf16
libc.src.math.setpayloadsigf16
libc.src.math.sinpif16
libc.src.math.totalorderf16
libc.src.math.totalordermagf16
libc.src.math.truncf16

1
libc/config/linux/x86_64/entrypoints.txt
View File

@@ -677,6 +677,7 @@ if(LIBC_TYPES_HAS_FLOAT16)
libc.src.math.scalbnf16
libc.src.math.setpayloadf16
libc.src.math.setpayloadsigf16
libc.src.math.sinpif16
libc.src.math.totalorderf16
libc.src.math.totalordermagf16
libc.src.math.truncf16

2
libc/docs/math/index.rst
View File

@@ -342,7 +342,7 @@ Higher Math Functions
+-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+
| sinh | |check| | | | | | 7.12.5.5 | F.10.2.5 |
+-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+
| sinpi | |check| | | | | | 7.12.4.13 | F.10.1.13 |
| sinpi | |check| | | | |check| | | 7.12.4.13 | F.10.1.13 |
+-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+
| sqrt | |check| | |check| | |check| | | |check| | 7.12.7.10 | F.10.4.10 |
+-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+

7
libc/newhdrgen/yaml/math.yaml
View File

@@ -2297,6 +2297,13 @@ functions:
return_type: float
arguments:
- type: float
- name: sinpif16
standards:
- stdc
return_type: _Float16
arguments:
- type: _Float16
guard: LIBC_TYPES_HAS_FLOAT16
- name: sqrt
standards:
- stdc

1
libc/src/math/CMakeLists.txt
View File

@@ -475,6 +475,7 @@ add_math_entrypoint_object(sincosf)
add_math_entrypoint_object(sin)
add_math_entrypoint_object(sinf)
add_math_entrypoint_object(sinpif)
add_math_entrypoint_object(sinpif16)
add_math_entrypoint_object(sinh)
add_math_entrypoint_object(sinhf)

19
libc/src/math/generic/CMakeLists.txt
View File

@@ -528,6 +528,25 @@ add_entrypoint_object(
-O3
)
add_entrypoint_object(
sinpif16
SRCS
sinpif16.cpp
HDRS
../sinpif16.h
DEPENDS
libc.src.__support.common
libc.src.__support.FPUtil.cast
libc.src.__support.FPUtil.fenv_impl
libc.src.__support.FPUtil.fp_bits
libc.src.__support.FPUtil.multiply_add
libc.src.__support.FPUtil.nearest_integer
libc.src.__support.FPUtil.polyeval
libc.src.__support.macros.properties.types
COMPILE_OPTIONS
-O3
)
add_entrypoint_object(
tan
SRCS

136
libc/src/math/generic/sinpif16.cpp Normal file
View File

@@ -0,0 +1,136 @@
//===-- Half-precision sinpif function ------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "src/math/sinpif16.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/multiply_add.h"
#include "src/__support/FPUtil/nearest_integer.h"
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
namespace LIBC_NAMESPACE_DECL {
// Lookup table for sin(k * pi / 32) with k = 0, ..., 63.
// Table is generated with Sollya as follows:
// > display = hexadecimmal;
// > for k from 0 to 63 do { round(sin(k * pi/32), SG, RN); };
static constexpr float SIN_K_PI_OVER_32[64] = {
0x0.0p0, 0x1.917a6cp-4, 0x1.8f8b84p-3, 0x1.294062p-2,
0x1.87de2ap-2, 0x1.e2b5d4p-2, 0x1.1c73b4p-1, 0x1.44cf32p-1,
0x1.6a09e6p-1, 0x1.8bc806p-1, 0x1.a9b662p-1, 0x1.c38b3p-1,
0x1.d906bcp-1, 0x1.e9f416p-1, 0x1.f6297cp-1, 0x1.fd88dap-1,
0x1p0, 0x1.fd88dap-1, 0x1.f6297cp-1, 0x1.e9f416p-1,
0x1.d906bcp-1, 0x1.c38b3p-1, 0x1.a9b662p-1, 0x1.8bc806p-1,
0x1.6a09e6p-1, 0x1.44cf32p-1, 0x1.1c73b4p-1, 0x1.e2b5d4p-2,
0x1.87de2ap-2, 0x1.294062p-2, 0x1.8f8b84p-3, 0x1.917a6cp-4,
0x0.0p0, -0x1.917a6cp-4, -0x1.8f8b84p-3, -0x1.294062p-2,
-0x1.87de2ap-2, -0x1.e2b5d4p-2, -0x1.1c73b4p-1, -0x1.44cf32p-1,
-0x1.6a09e6p-1, -0x1.8bc806p-1, -0x1.a9b662p-1, -0x1.c38b3p-1,
-0x1.d906bcp-1, -0x1.e9f416p-1, -0x1.f6297ep-1, -0x1.fd88dap-1,
-0x1p0, -0x1.fd88dap-1, -0x1.f6297cp-1, -0x1.e9f416p-1,
-0x1.d906bcp-1, -0x1.c38b3p-1, -0x1.a9b662p-1, -0x1.8bc806p-1,
-0x1.6a09e6p-1, -0x1.44cf32p-1, -0x1.1c73b4p-1, -0x1.e2b5d4p-2,
-0x1.87de2ap-2, -0x1.294062p-2, -0x1.8f8b84p-3, -0x1.917a6cp-4};
static LIBC_INLINE int32_t range_reduction(float x, float &y) {
float kf = fputil::nearest_integer(x * 32);
y = fputil::multiply_add<float>(x, 32.0, -kf);
return static_cast<int32_t>(kf);
}
LLVM_LIBC_FUNCTION(float16, sinpif16, (float16 x)) {
using FPBits = typename fputil::FPBits<float16>;
FPBits xbits(x);
uint16_t x_u = xbits.uintval();
uint16_t x_abs = x_u & 0x7fff;
// Range reduction:
// For |x| > 1/32, we perform range reduction as follows:
// Find k and y such that:
// x = (k + y) * 1/32
// k is an integer
// |y| < 0.5
//
// This is done by performing:
// k = round(x * 32)
// y = x * 32 - k
//
// Once k and y are computed, we then deduce the answer by the sine of sum
// formula:
// sin(x * pi) = sin((k + y) * pi/32)
// = sin(k * pi/32) * cos(y * pi/32) + sin (y * pi/32) * cos (k *
// pi/32)
// The values of sin(k * pi/32) and cos (k * pi/32) for k = 0...63 are
// precomputed and stored using a vector of 64 single precision floats. sin(y
// * pi/32) and cos(y * pi/32) are computed using degree-9 chebyshev
// polynomials generated by Sollya.
// For signed zeros
if (LIBC_UNLIKELY(x_abs == 0U))
return x;
// Numbers greater or equal to 2^10 are integers, or infinity, or NaN
if (LIBC_UNLIKELY(x_abs >= 0x6400)) {
// Check for NaN or infinity values
if (LIBC_UNLIKELY(x_abs >= 0x7c00)) {
// If value is equal to infinity
if (x_abs == 0x7c00) {
fputil::set_errno_if_required(EDOM);
fputil::raise_except_if_required(FE_INVALID);
}
return x + FPBits::quiet_nan().get_val();
}
return FPBits::zero(xbits.sign()).get_val();
}
float f32 = x;
float y;
int32_t k = range_reduction(f32, y);
float sin_k = SIN_K_PI_OVER_32[k & 63];
float cos_k = SIN_K_PI_OVER_32[(k + 16) & 63];
// Recall;
// sin(x * pi/32) = sin((k + y) * pi/32)
// = sin(y * pi/32) * cos(k * pi/32) + cos(y * pi/32) * sin(k *
// pi/32) Recall, after range reduction, -0.5 <= y <= 0.5. For very small
// values of y, calculating sin(y * p/32) can be inaccurate. Generating a
// polynomial for sin(y * p/32)/y instead significantly reduces the relative
// errors.
float ysq = y * y;
// Degree-6 minimax even polynomial for sin(y*pi/32)/y generated by Sollya
// with: > Q = fpminimax(sin(y*pi/32)/y, [|0, 2, 4, 6|], [|SG...|], [0, 0.5]);
float sin_y = y * fputil::polyeval(ysq, 0x1.921fb6p-4f, -0x1.4aeabcp-13f,
0x1.a03354p-21f, -0x1.ad02d2p-20f);
// Note that cosm1_y = cos(y*pi/32) - 1 = cos_y - 1
// Derivation:
// sin(x * pi) = sin((k + y) * pi/32)
// = sin_y * cos_k + cos_y * sin_k
// = cos_k * sin_y + sin_k * (1 + cos_y - 1)
// Degree-6 minimax even polynomial for cos(y*pi/32) generated by Sollya with:
// > P = fpminimax(cos(y*pi/32), [|0, 2, 4, 6|],[|1, SG...|], [0, 0.5]);
float cosm1_y = ysq * fputil::polyeval(ysq, -0x1.3bd3ccp-8f, 0x1.03a61ap-18f,
0x1.a6f7a2p-29f);
if (LIBC_UNLIKELY(sin_y == 0 && sin_k == 0))
return FPBits::zero(xbits.sign()).get_val();
// Since, cosm1_y = cos_y - 1, therefore:
// sin(x * pi) = cos_k * sin_y + sin_k + (cosm1_y * sin_k)
return fputil::cast<float16>(fputil::multiply_add(
sin_y, cos_k, fputil::multiply_add(cosm1_y, sin_k, sin_k)));
}
} // namespace LIBC_NAMESPACE_DECL

21
libc/src/math/sinpif16.h Normal file
View File

@@ -0,0 +1,21 @@
//===-- Implementation header for sinpif16 ---------------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache Licese 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_MATH_SINPIF16_H
#define LLVM_LIBC_SRC_MATH_SINPIF16_H
#include "src/__support/macros/config.h"
#include "src/__support/macros/properties/types.h"
namespace LIBC_NAMESPACE_DECL {
float16 sinpif16(float16 x);
} // namespace LIBC_NAMESPACE_DECL
#endif // LLVM_LIBC_SRC_MATH_SINPIF16_H

11
libc/test/src/math/CMakeLists.txt
View File

@@ -90,6 +90,17 @@ add_fp_unittest(
libc.src.__support.FPUtil.fp_bits
)
add_fp_unittest(
sinpif16_test
NEED_MPFR
SUITE
libc-math-unittests
SRCS
sinpif16_test.cpp
DEPENDS
libc.src.math.sinpif16
)
add_fp_unittest(
sin_test
NEED_MPFR

40
libc/test/src/math/sinpif16_test.cpp Normal file
View File

@@ -0,0 +1,40 @@
//===-- Exhaustive test for sinpif16 --------------------------------------===//
//
// 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
//
//===---------------------------------------------------------------------===//
#include "src/math/sinpif16.h"
#include "test/UnitTest/FPMatcher.h"
#include "test/UnitTest/Test.h"
#include "utils/MPFRWrapper/MPFRUtils.h"
using LlvmLibcSinpif16Test = LIBC_NAMESPACE::testing::FPTest<float16>;
namespace mpfr = LIBC_NAMESPACE::testing::mpfr;
// Range: [0, Inf]
static constexpr uint16_t POS_START = 0x0000U;
static constexpr uint16_t POS_STOP = 0x7c00U;
// Range: [-Inf, 0]
static constexpr uint16_t NEG_START = 0x8000U;
static constexpr uint16_t NEG_STOP = 0xfc00U;
TEST_F(LlvmLibcSinpif16Test, PositiveRange) {
for (uint16_t v = POS_START; v <= POS_STOP; ++v) {
float16 x = FPBits(v).get_val();
EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sinpi, x,
LIBC_NAMESPACE::sinpif16(x), 0.5);
}
}
TEST_F(LlvmLibcSinpif16Test, NegativeRange) {
for (uint16_t v = NEG_START; v <= NEG_STOP; ++v) {
float16 x = FPBits(v).get_val();
EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sinpi, x,
LIBC_NAMESPACE::sinpif16(x), 0.5);
}
}

11
libc/test/src/math/smoke/CMakeLists.txt
View File

@@ -51,6 +51,17 @@ add_fp_unittest(
libc.src.__support.FPUtil.fp_bits
)
add_fp_unittest(
sinpif16_test
SUITE
libc-math-smoke-tests
SRCS
sinpif16_test.cpp
DEPENDS
libc.src.errno.errno
libc.src.math.sinpif16
)
add_fp_unittest(
sincosf_test
SUITE

42
libc/test/src/math/smoke/sinpif16_test.cpp Normal file
View File

@@ -0,0 +1,42 @@
//===-- Unittests for sinpif16 --------------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "src/errno/libc_errno.h"
#include "src/math/sinpif16.h"
#include "test/UnitTest/FPMatcher.h"
#include "test/UnitTest/Test.h"
using LlvmLibcSinpif16Test = LIBC_NAMESPACE::testing::FPTest<float16>;
TEST_F(LlvmLibcSinpif16Test, SpecialNumbers) {
LIBC_NAMESPACE::libc_errno = 0;
EXPECT_FP_EQ(aNaN, LIBC_NAMESPACE::sinpif16(aNaN));
EXPECT_MATH_ERRNO(0);
EXPECT_FP_EQ(zero, LIBC_NAMESPACE::sinpif16(zero));
EXPECT_MATH_ERRNO(0);
EXPECT_FP_EQ(neg_zero, LIBC_NAMESPACE::sinpif16(neg_zero));
EXPECT_MATH_ERRNO(0);
EXPECT_FP_EQ(aNaN, LIBC_NAMESPACE::sinpif16(inf));
EXPECT_MATH_ERRNO(EDOM);
EXPECT_FP_EQ(aNaN, LIBC_NAMESPACE::sinpif16(neg_inf));
EXPECT_MATH_ERRNO(EDOM);
}
TEST_F(LlvmLibcSinpif16Test, Integers) {
EXPECT_FP_EQ(neg_zero, LIBC_NAMESPACE::sinpif16(-0x420));
EXPECT_FP_EQ(neg_zero, LIBC_NAMESPACE::sinpif16(-0x1p+10));
EXPECT_FP_EQ(neg_zero, LIBC_NAMESPACE::sinpif16(-0x1.4p+14));
EXPECT_FP_EQ(zero, LIBC_NAMESPACE::sinpif16(0x420));
EXPECT_FP_EQ(zero, LIBC_NAMESPACE::sinpif16(0x1.cp+15));
EXPECT_FP_EQ(zero, LIBC_NAMESPACE::sinpif16(0x1.cp+7));
}

18
libc/utils/MPFRWrapper/MPFRUtils.cpp
View File

@@ -488,14 +488,28 @@ public:
(MPFR_VERSION_MAJOR == 4 && MPFR_VERSION_MINOR >= 2)
mpfr_sinpi(result.value, value, mpfr_rounding);
return result;
#else
if (mpfr_integer_p(value)) {
mpfr_set_si(result.value, 0, mpfr_rounding);
return result;
}
MPFRNumber value_mul_two(*this);
mpfr_mul_si(value_mul_two.value, value, 2, MPFR_RNDN);
if (mpfr_integer_p(value_mul_two.value)) {
auto d = mpfr_get_si(value, MPFR_RNDD);
mpfr_set_si(result.value, (d & 1) ? -1 : 1, mpfr_rounding);
return result;
}
MPFRNumber value_pi(0.0, 1280);
mpfr_const_pi(value_pi.value, MPFR_RNDN);
mpfr_mul(value_pi.value, value_pi.value, value, MPFR_RNDN);
mpfr_sin(result.value, value_pi.value, mpfr_rounding);
#endif
return result;
#endif
}
MPFRNumber sinh() const {