Vector variant of the new C23 sinpi. New tests pass on AArch64.
#define __DECL_SIMD_tanf32x
#define __DECL_SIMD_tanf64x
#define __DECL_SIMD_tanf128x
+
+#define __DECL_SIMD_sinpi
+#define __DECL_SIMD_sinpif
+#define __DECL_SIMD_sinpil
+#define __DECL_SIMD_sinpif16
+#define __DECL_SIMD_sinpif32
+#define __DECL_SIMD_sinpif64
+#define __DECL_SIMD_sinpif128
+#define __DECL_SIMD_sinpif32x
+#define __DECL_SIMD_sinpif64x
+#define __DECL_SIMD_sinpif128x
#endif
/* Cosine of pi * X. */
__MATHCALL (cospi,, (_Mdouble_ __x));
/* Sine of pi * X. */
-__MATHCALL (sinpi,, (_Mdouble_ __x));
+__MATHCALL_VEC (sinpi,, (_Mdouble_ __x));
/* Tangent of pi * X. */
__MATHCALL (tanpi,, (_Mdouble_ __x));
#endif
pow \
sin \
sinh \
+ sinpi \
tan \
tanh
_ZGVnN4v_logp1f;
_ZGVsMxv_logp1;
_ZGVsMxv_logp1f;
+ _ZGVnN2v_sinpi;
+ _ZGVnN2v_sinpif;
+ _ZGVnN4v_sinpif;
+ _ZGVsMxv_sinpi;
+ _ZGVsMxv_sinpif;
}
}
libmvec_hidden_proto (V_NAME_F2(pow));
libmvec_hidden_proto (V_NAME_F1(sin));
libmvec_hidden_proto (V_NAME_F1(sinh));
+libmvec_hidden_proto (V_NAME_F1(sinpi));
libmvec_hidden_proto (V_NAME_F1(tan));
libmvec_hidden_proto (V_NAME_F1(tanh));
libmvec_hidden_proto (V_NAME_F2(atan2));
# define __DECL_SIMD_sinh __DECL_SIMD_aarch64
# undef __DECL_SIMD_sinhf
# define __DECL_SIMD_sinhf __DECL_SIMD_aarch64
+# undef __DECL_SIMD_sinpi
+# define __DECL_SIMD_sinpi __DECL_SIMD_aarch64
+# undef __DECL_SIMD_sinpif
+# define __DECL_SIMD_sinpif __DECL_SIMD_aarch64
# undef __DECL_SIMD_tan
# define __DECL_SIMD_tan __DECL_SIMD_aarch64
# undef __DECL_SIMD_tanf
__vpcs __f32x4_t _ZGVnN4vv_powf (__f32x4_t, __f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_sinf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_sinhf (__f32x4_t);
+__vpcs __f32x4_t _ZGVnN4v_sinpif (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_tanf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_tanhf (__f32x4_t);
__vpcs __f64x2_t _ZGVnN2vv_pow (__f64x2_t, __f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_sin (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_sinh (__f64x2_t);
+__vpcs __f64x2_t _ZGVnN2v_sinpi (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_tan (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_tanh (__f64x2_t);
__sv_f32_t _ZGVsMxvv_powf (__sv_f32_t, __sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_sinf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_sinhf (__sv_f32_t, __sv_bool_t);
+__sv_f32_t _ZGVsMxv_sinpif (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_tanf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_tanhf (__sv_f32_t, __sv_bool_t);
__sv_f64_t _ZGVsMxvv_pow (__sv_f64_t, __sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_sin (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_sinh (__sv_f64_t, __sv_bool_t);
+__sv_f64_t _ZGVsMxv_sinpi (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_tan (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_tanh (__sv_f64_t, __sv_bool_t);
--- /dev/null
+/* Double-precision (Advanced SIMD) sinpi function
+
+ Copyright (C) 2024 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include "v_math.h"
+#include "poly_advsimd_f64.h"
+
+static const struct data
+{
+ float64x2_t poly[10];
+} data = {
+ /* Polynomial coefficients generated using Remez algorithm,
+ see sinpi.sollya for details. */
+ .poly = { V2 (0x1.921fb54442d184p1), V2 (-0x1.4abbce625be53p2),
+ V2 (0x1.466bc6775ab16p1), V2 (-0x1.32d2cce62dc33p-1),
+ V2 (0x1.507834891188ep-4), V2 (-0x1.e30750a28c88ep-8),
+ V2 (0x1.e8f48308acda4p-12), V2 (-0x1.6fc0032b3c29fp-16),
+ V2 (0x1.af86ae521260bp-21), V2 (-0x1.012a9870eeb7dp-25) },
+};
+
+#if WANT_SIMD_EXCEPT
+# define TinyBound v_u64 (0x3bf0000000000000) /* asuint64(0x1p-64). */
+/* asuint64(0x1p64) - TinyBound. */
+# define Thresh v_u64 (0x07f0000000000000)
+
+static float64x2_t VPCS_ATTR NOINLINE
+special_case (float64x2_t x, float64x2_t y, uint64x2_t odd, uint64x2_t cmp)
+{
+ /* Fall back to scalar code. */
+ y = vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), odd));
+ return v_call_f64 (sinpi, x, y, cmp);
+}
+#endif
+
+/* Approximation for vector double-precision sinpi(x).
+ Maximum Error 3.05 ULP:
+ _ZGVnN2v_sinpi(0x1.d32750db30b4ap-2) got 0x1.fb295878301c7p-1
+ want 0x1.fb295878301cap-1. */
+float64x2_t VPCS_ATTR V_NAME_D1 (sinpi) (float64x2_t x)
+{
+ const struct data *d = ptr_barrier (&data);
+
+#if WANT_SIMD_EXCEPT
+ uint64x2_t ir = vreinterpretq_u64_f64 (vabsq_f64 (x));
+ uint64x2_t cmp = vcgeq_u64 (vsubq_u64 (ir, TinyBound), Thresh);
+
+ /* When WANT_SIMD_EXCEPT = 1, special lanes should be set to 0
+ to avoid them under/overflowing and throwing exceptions. */
+ float64x2_t r = v_zerofy_f64 (x, cmp);
+#else
+ float64x2_t r = x;
+#endif
+
+ /* If r is odd, the sign of the result should be inverted. */
+ uint64x2_t odd
+ = vshlq_n_u64 (vreinterpretq_u64_s64 (vcvtaq_s64_f64 (r)), 63);
+
+ /* r = x - rint(x). Range reduction to -1/2 .. 1/2. */
+ r = vsubq_f64 (r, vrndaq_f64 (r));
+
+ /* y = sin(r). */
+ float64x2_t r2 = vmulq_f64 (r, r);
+ float64x2_t r4 = vmulq_f64 (r2, r2);
+ float64x2_t y = vmulq_f64 (v_pw_horner_9_f64 (r2, r4, d->poly), r);
+
+#if WANT_SIMD_EXCEPT
+ if (__glibc_unlikely (v_any_u64 (cmp)))
+ return special_case (x, y, odd, cmp);
+#endif
+
+ return vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), odd));
+}
--- /dev/null
+/* Double-precision (SVE) sinpi function
+
+ Copyright (C) 2024 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include "sv_math.h"
+#include "poly_sve_f64.h"
+
+static const struct data
+{
+ double poly[10], range_val;
+} data = {
+ /* Polynomial coefficients generated using Remez algorithm,
+ see sinpi.sollya for details. */
+ .poly = { 0x1.921fb54442d184p1, -0x1.4abbce625be53p2, 0x1.466bc6775ab16p1,
+ -0x1.32d2cce62dc33p-1, 0x1.507834891188ep-4, -0x1.e30750a28c88ep-8,
+ 0x1.e8f48308acda4p-12, -0x1.6fc0032b3c29fp-16,
+ 0x1.af86ae521260bp-21, -0x1.012a9870eeb7dp-25 },
+ .range_val = 0x1p63,
+};
+
+/* A fast SVE implementation of sinpi.
+ Maximum error 3.10 ULP:
+ _ZGVsMxv_sinpi(0x1.df1a14f1b235p-2) got 0x1.fd64f541606cp-1
+ want 0x1.fd64f541606c3p-1. */
+svfloat64_t SV_NAME_D1 (sinpi) (svfloat64_t x, const svbool_t pg)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ /* range reduction into -1/2 .. 1/2)
+ with n = rint(x) and r = r - n. */
+ svfloat64_t n = svrinta_x (pg, x);
+ svfloat64_t r = svsub_x (pg, x, n);
+
+ /* Result should be negated based on if n is odd or not. */
+ svbool_t cmp = svaclt (pg, x, d->range_val);
+ svuint64_t intn = svreinterpret_u64 (svcvt_s64_z (pg, n));
+ svuint64_t sign = svlsl_z (cmp, intn, 63);
+
+ /* y = sin(r). */
+ svfloat64_t r2 = svmul_x (pg, r, r);
+ svfloat64_t r4 = svmul_x (pg, r2, r2);
+ svfloat64_t y = sv_pw_horner_9_f64_x (pg, r2, r4, d->poly);
+ y = svmul_x (pg, y, r);
+
+ return svreinterpret_f64 (sveor_x (pg, svreinterpret_u64 (y), sign));
+}
--- /dev/null
+/* Single-precision (Advanced SIMD) sinpi function
+
+ Copyright (C) 2024 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include "v_math.h"
+#include "poly_advsimd_f32.h"
+
+static const struct data
+{
+ float32x4_t poly[6];
+} data = {
+ /* Taylor series coefficents for sin(pi * x). */
+ .poly = { V4 (0x1.921fb6p1f), V4 (-0x1.4abbcep2f), V4 (0x1.466bc6p1f),
+ V4 (-0x1.32d2ccp-1f), V4 (0x1.50783p-4f), V4 (-0x1.e30750p-8f) },
+};
+
+#if WANT_SIMD_EXCEPT
+# define TinyBound v_u32 (0x30000000) /* asuint32(0x1p-31f). */
+# define Thresh v_u32 (0x1f000000) /* asuint32(0x1p31f) - TinyBound. */
+
+static float32x4_t VPCS_ATTR NOINLINE
+special_case (float32x4_t x, float32x4_t y, uint32x4_t odd, uint32x4_t cmp)
+{
+ /* Fall back to scalar code. */
+ y = vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd));
+ return v_call_f32 (sinpif, x, y, cmp);
+}
+#endif
+
+/* Approximation for vector single-precision sinpi(x)
+ Maximum Error 3.03 ULP:
+ _ZGVnN4v_sinpif(0x1.c597ccp-2) got 0x1.f7cd56p-1
+ want 0x1.f7cd5p-1. */
+float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (sinpi) (float32x4_t x)
+{
+ const struct data *d = ptr_barrier (&data);
+
+#if WANT_SIMD_EXCEPT
+ uint32x4_t ir = vreinterpretq_u32_f32 (vabsq_f32 (x));
+ uint32x4_t cmp = vcgeq_u32 (vsubq_u32 (ir, TinyBound), Thresh);
+
+ /* When WANT_SIMD_EXCEPT = 1, special lanes should be set to 0
+ to avoid them under/overflowing and throwing exceptions. */
+ float32x4_t r = v_zerofy_f32 (x, cmp);
+#else
+ float32x4_t r = x;
+#endif
+
+ /* If r is odd, the sign of the result should be inverted. */
+ uint32x4_t odd
+ = vshlq_n_u32 (vreinterpretq_u32_s32 (vcvtaq_s32_f32 (r)), 31);
+
+ /* r = x - rint(x). Range reduction to -1/2 .. 1/2. */
+ r = vsubq_f32 (r, vrndaq_f32 (r));
+
+ /* Pairwise Horner approximation for y = sin(r * pi). */
+ float32x4_t r2 = vmulq_f32 (r, r);
+ float32x4_t r4 = vmulq_f32 (r2, r2);
+ float32x4_t y = vmulq_f32 (v_pw_horner_5_f32 (r2, r4, d->poly), r);
+
+#if WANT_SIMD_EXCEPT
+ if (__glibc_unlikely (v_any_u32 (cmp)))
+ return special_case (x, y, odd, cmp);
+#endif
+
+ return vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd));
+}
+
+libmvec_hidden_def (V_NAME_F1 (sinpi))
+HALF_WIDTH_ALIAS_F1 (sinpi)
--- /dev/null
+/* Single-precision (SVE) sinpi function
+
+ Copyright (C) 2024 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include "sv_math.h"
+#include "poly_sve_f32.h"
+
+static const struct data
+{
+ float poly[6], range_val;
+} data = {
+ /* Taylor series coefficents for sin(pi * x). */
+ .poly = { 0x1.921fb6p1f, -0x1.4abbcep2f, 0x1.466bc6p1f, -0x1.32d2ccp-1f,
+ 0x1.50783p-4f, -0x1.e30750p-8f },
+ .range_val = 0x1p31,
+};
+
+/* A fast SVE implementation of sinpif.
+ Maximum error 2.48 ULP:
+ _ZGVsMxv_sinpif(0x1.d062b6p-2) got 0x1.fa8c06p-1
+ want 0x1.fa8c02p-1. */
+svfloat32_t SV_NAME_F1 (sinpi) (svfloat32_t x, const svbool_t pg)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ /* range reduction into -1/2 .. 1/2
+ with n = rint(x) and r = r - n. */
+ svfloat32_t n = svrinta_x (pg, x);
+ svfloat32_t r = svsub_x (pg, x, n);
+
+ /* Result should be negated based on if n is odd or not. */
+ svbool_t cmp = svaclt (pg, x, d->range_val);
+ svuint32_t intn = svreinterpret_u32 (svcvt_s32_z (pg, n));
+ svuint32_t sign = svlsl_z (cmp, intn, 31);
+
+ /* y = sin(r). */
+ svfloat32_t r2 = svmul_x (pg, r, r);
+ svfloat32_t y = sv_horner_5_f32_x (pg, r2, d->poly);
+ y = svmul_x (pg, y, r);
+
+ return svreinterpret_f32 (sveor_x (pg, svreinterpret_u32 (y), sign));
+}
VPCS_VECTOR_WRAPPER_ff (pow_advsimd, _ZGVnN2vv_pow)
VPCS_VECTOR_WRAPPER (sin_advsimd, _ZGVnN2v_sin)
VPCS_VECTOR_WRAPPER (sinh_advsimd, _ZGVnN2v_sinh)
+VPCS_VECTOR_WRAPPER (sinpi_advsimd, _ZGVnN2v_sinpi)
VPCS_VECTOR_WRAPPER (tan_advsimd, _ZGVnN2v_tan)
VPCS_VECTOR_WRAPPER (tanh_advsimd, _ZGVnN2v_tanh)
SVE_VECTOR_WRAPPER_ff (pow_sve, _ZGVsMxvv_pow)
SVE_VECTOR_WRAPPER (sin_sve, _ZGVsMxv_sin)
SVE_VECTOR_WRAPPER (sinh_sve, _ZGVsMxv_sinh)
+SVE_VECTOR_WRAPPER (sinpi_sve, _ZGVsMxv_sinpi)
SVE_VECTOR_WRAPPER (tan_sve, _ZGVsMxv_tan)
SVE_VECTOR_WRAPPER (tanh_sve, _ZGVsMxv_tanh)
VPCS_VECTOR_WRAPPER_ff (powf_advsimd, _ZGVnN4vv_powf)
VPCS_VECTOR_WRAPPER (sinf_advsimd, _ZGVnN4v_sinf)
VPCS_VECTOR_WRAPPER (sinhf_advsimd, _ZGVnN4v_sinhf)
+VPCS_VECTOR_WRAPPER (sinpif_advsimd, _ZGVnN4v_sinpif)
VPCS_VECTOR_WRAPPER (tanf_advsimd, _ZGVnN4v_tanf)
VPCS_VECTOR_WRAPPER (tanhf_advsimd, _ZGVnN4v_tanhf)
SVE_VECTOR_WRAPPER_ff (powf_sve, _ZGVsMxvv_powf)
SVE_VECTOR_WRAPPER (sinf_sve, _ZGVsMxv_sinf)
SVE_VECTOR_WRAPPER (sinhf_sve, _ZGVsMxv_sinhf)
+SVE_VECTOR_WRAPPER (sinpif_sve, _ZGVsMxv_sinpif)
SVE_VECTOR_WRAPPER (tanf_sve, _ZGVsMxv_tanf)
SVE_VECTOR_WRAPPER (tanhf_sve, _ZGVsMxv_tanhf)
float: 2
ldouble: 2
+Function: "sinpi_advsimd":
+double: 2
+float: 2
+
Function: "sinpi_downward":
double: 2
float: 2
ldouble: 2
+Function: "sinpi_sve":
+double: 2
+float: 2
+
Function: "sinpi_towardzero":
double: 2
float: 1
GLIBC_2.40 _ZGVsMxvv_powf F
GLIBC_2.41 _ZGVnN2v_logp1 F
GLIBC_2.41 _ZGVnN2v_logp1f F
+GLIBC_2.41 _ZGVnN2v_sinpi F
+GLIBC_2.41 _ZGVnN2v_sinpif F
GLIBC_2.41 _ZGVnN4v_logp1f F
+GLIBC_2.41 _ZGVnN4v_sinpif F
GLIBC_2.41 _ZGVsMxv_logp1 F
GLIBC_2.41 _ZGVsMxv_logp1f F
+GLIBC_2.41 _ZGVsMxv_sinpi F
+GLIBC_2.41 _ZGVsMxv_sinpif F
projects / thirdparty / glibc.git / commitdiff
