Vector variant of the new C23 cospi. New tests pass on AArch64.
#define __DECL_SIMD_sinpif32x
#define __DECL_SIMD_sinpif64x
#define __DECL_SIMD_sinpif128x
+
+#define __DECL_SIMD_cospi
+#define __DECL_SIMD_cospif
+#define __DECL_SIMD_cospil
+#define __DECL_SIMD_cospif16
+#define __DECL_SIMD_cospif32
+#define __DECL_SIMD_cospif64
+#define __DECL_SIMD_cospif128
+#define __DECL_SIMD_cospif32x
+#define __DECL_SIMD_cospif64x
+#define __DECL_SIMD_cospif128x
#endif
__MATHCALL (atan2pi,, (_Mdouble_ __y, _Mdouble_ __x));
/* Cosine of pi * X. */
-__MATHCALL (cospi,, (_Mdouble_ __x));
+__MATHCALL_VEC (cospi,, (_Mdouble_ __x));
/* Sine of pi * X. */
__MATHCALL_VEC (sinpi,, (_Mdouble_ __x));
/* Tangent of pi * X. */
cbrt \
cos \
cosh \
+ cospi \
erf \
erfc \
exp \
_ZGVsMxv_tanhf;
}
GLIBC_2.41 {
+ _ZGVnN2v_cospi;
+ _ZGVnN2v_cospif;
+ _ZGVnN4v_cospif;
+ _ZGVsMxv_cospi;
+ _ZGVsMxv_cospif;
_ZGVnN2v_logp1;
_ZGVnN2v_logp1f;
_ZGVnN4v_logp1f;
libmvec_hidden_proto (V_NAME_F1(cbrt));
libmvec_hidden_proto (V_NAME_F1(cos));
libmvec_hidden_proto (V_NAME_F1(cosh));
+libmvec_hidden_proto (V_NAME_F1(cospi));
libmvec_hidden_proto (V_NAME_F1(erf));
libmvec_hidden_proto (V_NAME_F1(erfc));
libmvec_hidden_proto (V_NAME_F1(exp10));
# define __DECL_SIMD_cosh __DECL_SIMD_aarch64
# undef __DECL_SIMD_coshf
# define __DECL_SIMD_coshf __DECL_SIMD_aarch64
+# undef __DECL_SIMD_cospi
+# define __DECL_SIMD_cospi __DECL_SIMD_aarch64
+# undef __DECL_SIMD_cospif
+# define __DECL_SIMD_cospif __DECL_SIMD_aarch64
# undef __DECL_SIMD_erf
# define __DECL_SIMD_erf __DECL_SIMD_aarch64
# undef __DECL_SIMD_erff
__vpcs __f32x4_t _ZGVnN4v_cbrtf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_cosf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_coshf (__f32x4_t);
+__vpcs __f32x4_t _ZGVnN4v_cospif (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_erff (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_erfcf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_expf (__f32x4_t);
__vpcs __f64x2_t _ZGVnN2v_cbrt (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_cos (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_cosh (__f64x2_t);
+__vpcs __f64x2_t _ZGVnN2v_cospi (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_erf (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_erfc (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_exp (__f64x2_t);
__sv_f32_t _ZGVsMxv_cbrtf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_cosf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_coshf (__sv_f32_t, __sv_bool_t);
+__sv_f32_t _ZGVsMxv_cospif (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_erff (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_erfcf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_expf (__sv_f32_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_cbrt (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_cos (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_cosh (__sv_f64_t, __sv_bool_t);
+__sv_f64_t _ZGVsMxv_cospi (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_erf (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_erfc (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_exp (__sv_f64_t, __sv_bool_t);
--- /dev/null
+/* Double-precision (Advanced SIMD) cospi 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];
+ float64x2_t range_val;
+} 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) },
+ .range_val = V2 (0x1p63),
+};
+
+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 (cospi, x, y, cmp);
+}
+
+/* Approximation for vector double-precision cospi(x).
+ Maximum Error 3.06 ULP:
+ _ZGVnN2v_cospi(0x1.7dd4c0b03cc66p-5) got 0x1.fa854babfb6bep-1
+ want 0x1.fa854babfb6c1p-1. */
+float64x2_t VPCS_ATTR V_NAME_D1 (cospi) (float64x2_t x)
+{
+ const struct data *d = ptr_barrier (&data);
+
+#if WANT_SIMD_EXCEPT
+ float64x2_t r = vabsq_f64 (x);
+ uint64x2_t cmp = vcaleq_f64 (v_f64 (0x1p64), x);
+
+ /* When WANT_SIMD_EXCEPT = 1, special lanes should be zero'd
+ to avoid them overflowing and throwing exceptions. */
+ r = v_zerofy_f64 (r, cmp);
+ uint64x2_t odd = vshlq_n_u64 (vcvtnq_u64_f64 (r), 63);
+
+#else
+ float64x2_t r = x;
+ uint64x2_t cmp = vcageq_f64 (r, d->range_val);
+ uint64x2_t odd
+ = vshlq_n_u64 (vreinterpretq_u64_s64 (vcvtaq_s64_f64 (r)), 63);
+
+#endif
+
+ r = vsubq_f64 (r, vrndaq_f64 (r));
+
+ /* cospi(x) = sinpi(0.5 - abs(x)) for values -1/2 .. 1/2. */
+ r = vsubq_f64 (v_f64 (0.5), vabsq_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);
+
+ /* Fallback to scalar. */
+ if (__glibc_unlikely (v_any_u64 (cmp)))
+ return special_case (x, y, odd, cmp);
+
+ /* Reintroduce the sign bit for inputs which round to odd. */
+ return vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), odd));
+}
--- /dev/null
+/* Double-precision (SVE) cospi 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];
+ double 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 = 0x1p53,
+};
+
+/* A fast SVE implementation of cospi.
+ Maximum error 3.20 ULP:
+ _ZGVsMxv_cospi(0x1.f18ba32c63159p-6) got 0x1.fdabf595f9763p-1
+ want 0x1.fdabf595f9766p-1. */
+svfloat64_t SV_NAME_D1 (cospi) (svfloat64_t x, const svbool_t pg)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ /* Using cospi(x) = sinpi(0.5 - x)
+ range reduction and offset into sinpi range -1/2 .. 1/2
+ r = 0.5 - |x - rint(x)|. */
+ svfloat64_t n = svrinta_x (pg, x);
+ svfloat64_t r = svsub_x (pg, x, n);
+ r = svsub_x (pg, sv_f64 (0.5), svabs_x (pg, r));
+
+ /* Result should be negated based on if n is odd or not.
+ If ax >= 2^53, the result will always be positive. */
+ 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) cospi 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];
+ float32x4_t range_val;
+} 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) },
+ .range_val = V4 (0x1p31f),
+};
+
+static float32x4_t VPCS_ATTR NOINLINE
+special_case (float32x4_t x, float32x4_t y, uint32x4_t odd, uint32x4_t cmp)
+{
+ y = vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd));
+ return v_call_f32 (cospif, x, y, cmp);
+}
+
+/* Approximation for vector single-precision cospi(x)
+ Maximum Error: 3.17 ULP:
+ _ZGVnN4v_cospif(0x1.d341a8p-5) got 0x1.f7cd56p-1
+ want 0x1.f7cd5p-1. */
+float32x4_t VPCS_ATTR V_NAME_F1 (cospi) (float32x4_t x)
+{
+ const struct data *d = ptr_barrier (&data);
+
+#if WANT_SIMD_EXCEPT
+ float32x4_t r = vabsq_f32 (x);
+ uint32x4_t cmp = vcaleq_f32 (v_f32 (0x1p32f), x);
+
+ /* When WANT_SIMD_EXCEPT = 1, special lanes should be zero'd
+ to avoid them overflowing and throwing exceptions. */
+ r = v_zerofy_f32 (r, cmp);
+ uint32x4_t odd = vshlq_n_u32 (vcvtnq_u32_f32 (r), 31);
+
+#else
+ float32x4_t r = x;
+ uint32x4_t cmp = vcageq_f32 (r, d->range_val);
+
+ uint32x4_t odd
+ = vshlq_n_u32 (vreinterpretq_u32_s32 (vcvtaq_s32_f32 (r)), 31);
+
+#endif
+
+ /* r = x - rint(x). */
+ r = vsubq_f32 (r, vrndaq_f32 (r));
+
+ /* cospi(x) = sinpi(0.5 - abs(x)) for values -1/2 .. 1/2. */
+ r = vsubq_f32 (v_f32 (0.5f), vabsq_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);
+
+ /* Fallback to scalar. */
+ if (__glibc_unlikely (v_any_u32 (cmp)))
+ return special_case (x, y, odd, cmp);
+
+ /* Reintroduce the sign bit for inputs which round to odd. */
+ return vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd));
+}
+
+libmvec_hidden_def (V_NAME_F1 (cospi))
+HALF_WIDTH_ALIAS_F1 (cospi)
--- /dev/null
+/* Single-precision (SVE) cospi 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];
+ float 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 = 0x1p31f,
+};
+
+/* A fast SVE implementation of cospif.
+ Maximum error: 2.60 ULP:
+ _ZGVsMxv_cospif(+/-0x1.cae664p-4) got 0x1.e09c9ep-1
+ want 0x1.e09c98p-1. */
+svfloat32_t SV_NAME_F1 (cospi) (svfloat32_t x, const svbool_t pg)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ /* Using cospi(x) = sinpi(0.5 - x)
+ range reduction and offset into sinpi range -1/2 .. 1/2
+ r = 0.5 - |x - rint(x)|. */
+ svfloat32_t n = svrinta_x (pg, x);
+ svfloat32_t r = svsub_x (pg, x, n);
+ r = svsub_x (pg, sv_f32 (0.5f), svabs_x (pg, r));
+
+ /* Result should be negated based on if n is odd or not.
+ If ax >= 2^31, the result will always be positive. */
+ svbool_t cmp = svaclt (pg, x, d->range_val);
+ svuint32_t intn = svreinterpret_u32 (svcvt_s32_x (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 (cbrt_advsimd, _ZGVnN2v_cbrt)
VPCS_VECTOR_WRAPPER (cos_advsimd, _ZGVnN2v_cos)
VPCS_VECTOR_WRAPPER (cosh_advsimd, _ZGVnN2v_cosh)
+VPCS_VECTOR_WRAPPER (cospi_advsimd, _ZGVnN2v_cospi)
VPCS_VECTOR_WRAPPER (erf_advsimd, _ZGVnN2v_erf)
VPCS_VECTOR_WRAPPER (erfc_advsimd, _ZGVnN2v_erfc)
VPCS_VECTOR_WRAPPER (exp_advsimd, _ZGVnN2v_exp)
SVE_VECTOR_WRAPPER (cbrt_sve, _ZGVsMxv_cbrt)
SVE_VECTOR_WRAPPER (cos_sve, _ZGVsMxv_cos)
SVE_VECTOR_WRAPPER (cosh_sve, _ZGVsMxv_cosh)
+SVE_VECTOR_WRAPPER (cospi_sve, _ZGVsMxv_cospi)
SVE_VECTOR_WRAPPER (erf_sve, _ZGVsMxv_erf)
SVE_VECTOR_WRAPPER (erfc_sve, _ZGVsMxv_erfc)
SVE_VECTOR_WRAPPER (exp_sve, _ZGVsMxv_exp)
VPCS_VECTOR_WRAPPER (cbrtf_advsimd, _ZGVnN4v_cbrtf)
VPCS_VECTOR_WRAPPER (cosf_advsimd, _ZGVnN4v_cosf)
VPCS_VECTOR_WRAPPER (coshf_advsimd, _ZGVnN4v_coshf)
+VPCS_VECTOR_WRAPPER (cospif_advsimd, _ZGVnN4v_cospif)
VPCS_VECTOR_WRAPPER (erff_advsimd, _ZGVnN4v_erff)
VPCS_VECTOR_WRAPPER (erfcf_advsimd, _ZGVnN4v_erfcf)
VPCS_VECTOR_WRAPPER (expf_advsimd, _ZGVnN4v_expf)
SVE_VECTOR_WRAPPER (cbrtf_sve, _ZGVsMxv_cbrtf)
SVE_VECTOR_WRAPPER (cosf_sve, _ZGVsMxv_cosf)
SVE_VECTOR_WRAPPER (coshf_sve, _ZGVsMxv_coshf)
+SVE_VECTOR_WRAPPER (cospif_sve, _ZGVsMxv_cospif)
SVE_VECTOR_WRAPPER (erff_sve, _ZGVsMxv_erff)
SVE_VECTOR_WRAPPER (erfcf_sve, _ZGVsMxv_erfcf)
SVE_VECTOR_WRAPPER (expf_sve, _ZGVsMxv_expf)
float: 2
ldouble: 2
+Function: "cospi_advsimd":
+double: 2
+float: 1
+
Function: "cospi_downward":
double: 1
float: 2
ldouble: 2
+Function: "cospi_sve":
+double: 2
+float: 1
+
Function: "cospi_towardzero":
double: 1
float: 1
GLIBC_2.40 _ZGVsMxvv_hypotf F
GLIBC_2.40 _ZGVsMxvv_pow F
GLIBC_2.40 _ZGVsMxvv_powf F
+GLIBC_2.41 _ZGVnN2v_cospi F
+GLIBC_2.41 _ZGVnN2v_cospif 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_cospif F
GLIBC_2.41 _ZGVnN4v_logp1f F
GLIBC_2.41 _ZGVnN4v_sinpif F
+GLIBC_2.41 _ZGVsMxv_cospi F
+GLIBC_2.41 _ZGVsMxv_cospif F
GLIBC_2.41 _ZGVsMxv_logp1 F
GLIBC_2.41 _ZGVsMxv_logp1f F
GLIBC_2.41 _ZGVsMxv_sinpi F
projects / thirdparty / glibc.git / commitdiff
