public inbox for glibc-cvs@sourceware.org
help / color / mirror / Atom feed
* [glibc] aarch64: Add vector implementations of acos routines
@ 2023-11-13 14:25 Szabolcs Nagy
0 siblings, 0 replies; only message in thread
From: Szabolcs Nagy @ 2023-11-13 14:25 UTC (permalink / raw)
To: glibc-cvs
https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=b5d23367a8c1e92b26bd48cce984e790208cdac1
commit b5d23367a8c1e92b26bd48cce984e790208cdac1
Author: Joe Ramsay <Joe.Ramsay@arm.com>
Date: Fri Nov 3 12:12:20 2023 +0000
aarch64: Add vector implementations of acos routines
Diff:
---
sysdeps/aarch64/fpu/Makefile | 3 +-
sysdeps/aarch64/fpu/Versions | 4 +
sysdeps/aarch64/fpu/acos_advsimd.c | 122 +++++++++++++++++++++
sysdeps/aarch64/fpu/acos_sve.c | 93 ++++++++++++++++
sysdeps/aarch64/fpu/acosf_advsimd.c | 113 +++++++++++++++++++
sysdeps/aarch64/fpu/acosf_sve.c | 86 +++++++++++++++
sysdeps/aarch64/fpu/bits/math-vector.h | 4 +
sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c | 1 +
sysdeps/aarch64/fpu/test-double-sve-wrappers.c | 1 +
sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c | 1 +
sysdeps/aarch64/fpu/test-float-sve-wrappers.c | 1 +
sysdeps/aarch64/libm-test-ulps | 8 ++
sysdeps/unix/sysv/linux/aarch64/libmvec.abilist | 4 +
13 files changed, 440 insertions(+), 1 deletion(-)
diff --git a/sysdeps/aarch64/fpu/Makefile b/sysdeps/aarch64/fpu/Makefile
index d7c0bd2ed5..606fdd804f 100644
--- a/sysdeps/aarch64/fpu/Makefile
+++ b/sysdeps/aarch64/fpu/Makefile
@@ -1,4 +1,5 @@
-libmvec-supported-funcs = asin \
+libmvec-supported-funcs = acos \
+ asin \
cos \
exp \
exp10 \
diff --git a/sysdeps/aarch64/fpu/Versions b/sysdeps/aarch64/fpu/Versions
index 0f365a1e2e..1037cd92bd 100644
--- a/sysdeps/aarch64/fpu/Versions
+++ b/sysdeps/aarch64/fpu/Versions
@@ -18,6 +18,10 @@ libmvec {
_ZGVsMxv_sinf;
}
GLIBC_2.39 {
+ _ZGVnN4v_acosf;
+ _ZGVnN2v_acos;
+ _ZGVsMxv_acosf;
+ _ZGVsMxv_acos;
_ZGVnN4v_asinf;
_ZGVnN2v_asin;
_ZGVsMxv_asinf;
diff --git a/sysdeps/aarch64/fpu/acos_advsimd.c b/sysdeps/aarch64/fpu/acos_advsimd.c
new file mode 100644
index 0000000000..3121cf66b1
--- /dev/null
+++ b/sysdeps/aarch64/fpu/acos_advsimd.c
@@ -0,0 +1,122 @@
+/* Double-precision AdvSIMD inverse cos
+
+ Copyright (C) 2023 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[12];
+ float64x2_t pi, pi_over_2;
+ uint64x2_t abs_mask;
+} data = {
+ /* Polynomial approximation of (asin(sqrt(x)) - sqrt(x)) / (x * sqrt(x))
+ on [ 0x1p-106, 0x1p-2 ], relative error: 0x1.c3d8e169p-57. */
+ .poly = { V2 (0x1.555555555554ep-3), V2 (0x1.3333333337233p-4),
+ V2 (0x1.6db6db67f6d9fp-5), V2 (0x1.f1c71fbd29fbbp-6),
+ V2 (0x1.6e8b264d467d6p-6), V2 (0x1.1c5997c357e9dp-6),
+ V2 (0x1.c86a22cd9389dp-7), V2 (0x1.856073c22ebbep-7),
+ V2 (0x1.fd1151acb6bedp-8), V2 (0x1.087182f799c1dp-6),
+ V2 (-0x1.6602748120927p-7), V2 (0x1.cfa0dd1f9478p-6), },
+ .pi = V2 (0x1.921fb54442d18p+1),
+ .pi_over_2 = V2 (0x1.921fb54442d18p+0),
+ .abs_mask = V2 (0x7fffffffffffffff),
+};
+
+#define AllMask v_u64 (0xffffffffffffffff)
+#define Oneu (0x3ff0000000000000)
+#define Small (0x3e50000000000000) /* 2^-53. */
+
+#if WANT_SIMD_EXCEPT
+static float64x2_t VPCS_ATTR NOINLINE
+special_case (float64x2_t x, float64x2_t y, uint64x2_t special)
+{
+ return v_call_f64 (acos, x, y, special);
+}
+#endif
+
+/* Double-precision implementation of vector acos(x).
+
+ For |x| < Small, approximate acos(x) by pi/2 - x. Small = 2^-53 for correct
+ rounding.
+ If WANT_SIMD_EXCEPT = 0, Small = 0 and we proceed with the following
+ approximation.
+
+ For |x| in [Small, 0.5], use an order 11 polynomial P such that the final
+ approximation of asin is an odd polynomial:
+
+ acos(x) ~ pi/2 - (x + x^3 P(x^2)).
+
+ The largest observed error in this region is 1.18 ulps,
+ _ZGVnN2v_acos (0x1.fbab0a7c460f6p-2) got 0x1.0d54d1985c068p+0
+ want 0x1.0d54d1985c069p+0.
+
+ For |x| in [0.5, 1.0], use same approximation with a change of variable
+
+ acos(x) = y + y * z * P(z), with z = (1-x)/2 and y = sqrt(z).
+
+ The largest observed error in this region is 1.52 ulps,
+ _ZGVnN2v_acos (0x1.23d362722f591p-1) got 0x1.edbbedf8a7d6ep-1
+ want 0x1.edbbedf8a7d6cp-1. */
+float64x2_t VPCS_ATTR V_NAME_D1 (acos) (float64x2_t x)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ float64x2_t ax = vabsq_f64 (x);
+
+#if WANT_SIMD_EXCEPT
+ /* A single comparison for One, Small and QNaN. */
+ uint64x2_t special
+ = vcgtq_u64 (vsubq_u64 (vreinterpretq_u64_f64 (ax), v_u64 (Small)),
+ v_u64 (Oneu - Small));
+ if (__glibc_unlikely (v_any_u64 (special)))
+ return special_case (x, x, AllMask);
+#endif
+
+ uint64x2_t a_le_half = vcleq_f64 (ax, v_f64 (0.5));
+
+ /* Evaluate polynomial Q(x) = z + z * z2 * P(z2) with
+ z2 = x ^ 2 and z = |x| , if |x| < 0.5
+ z2 = (1 - |x|) / 2 and z = sqrt(z2), if |x| >= 0.5. */
+ float64x2_t z2 = vbslq_f64 (a_le_half, vmulq_f64 (x, x),
+ vfmaq_f64 (v_f64 (0.5), v_f64 (-0.5), ax));
+ float64x2_t z = vbslq_f64 (a_le_half, ax, vsqrtq_f64 (z2));
+
+ /* Use a single polynomial approximation P for both intervals. */
+ float64x2_t z4 = vmulq_f64 (z2, z2);
+ float64x2_t z8 = vmulq_f64 (z4, z4);
+ float64x2_t z16 = vmulq_f64 (z8, z8);
+ float64x2_t p = v_estrin_11_f64 (z2, z4, z8, z16, d->poly);
+
+ /* Finalize polynomial: z + z * z2 * P(z2). */
+ p = vfmaq_f64 (z, vmulq_f64 (z, z2), p);
+
+ /* acos(|x|) = pi/2 - sign(x) * Q(|x|), for |x| < 0.5
+ = 2 Q(|x|) , for 0.5 < x < 1.0
+ = pi - 2 Q(|x|) , for -1.0 < x < -0.5. */
+ float64x2_t y = vbslq_f64 (d->abs_mask, p, x);
+
+ uint64x2_t is_neg = vcltzq_f64 (x);
+ float64x2_t off = vreinterpretq_f64_u64 (
+ vandq_u64 (is_neg, vreinterpretq_u64_f64 (d->pi)));
+ float64x2_t mul = vbslq_f64 (a_le_half, v_f64 (-1.0), v_f64 (2.0));
+ float64x2_t add = vbslq_f64 (a_le_half, d->pi_over_2, off);
+
+ return vfmaq_f64 (add, mul, y);
+}
diff --git a/sysdeps/aarch64/fpu/acos_sve.c b/sysdeps/aarch64/fpu/acos_sve.c
new file mode 100644
index 0000000000..1138a04e73
--- /dev/null
+++ b/sysdeps/aarch64/fpu/acos_sve.c
@@ -0,0 +1,93 @@
+/* Double-precision SVE inverse cos
+
+ Copyright (C) 2023 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
+{
+ float64_t poly[12];
+ float64_t pi, pi_over_2;
+} data = {
+ /* Polynomial approximation of (asin(sqrt(x)) - sqrt(x)) / (x * sqrt(x))
+ on [ 0x1p-106, 0x1p-2 ], relative error: 0x1.c3d8e169p-57. */
+ .poly = { 0x1.555555555554ep-3, 0x1.3333333337233p-4, 0x1.6db6db67f6d9fp-5,
+ 0x1.f1c71fbd29fbbp-6, 0x1.6e8b264d467d6p-6, 0x1.1c5997c357e9dp-6,
+ 0x1.c86a22cd9389dp-7, 0x1.856073c22ebbep-7, 0x1.fd1151acb6bedp-8,
+ 0x1.087182f799c1dp-6, -0x1.6602748120927p-7, 0x1.cfa0dd1f9478p-6, },
+ .pi = 0x1.921fb54442d18p+1,
+ .pi_over_2 = 0x1.921fb54442d18p+0,
+};
+
+/* Double-precision SVE implementation of vector acos(x).
+
+ For |x| in [0, 0.5], use an order 11 polynomial P such that the final
+ approximation of asin is an odd polynomial:
+
+ acos(x) ~ pi/2 - (x + x^3 P(x^2)).
+
+ The largest observed error in this region is 1.18 ulps,
+ _ZGVsMxv_acos (0x1.fbc5fe28ee9e3p-2) got 0x1.0d4d0f55667f6p+0
+ want 0x1.0d4d0f55667f7p+0.
+
+ For |x| in [0.5, 1.0], use same approximation with a change of variable
+
+ acos(x) = y + y * z * P(z), with z = (1-x)/2 and y = sqrt(z).
+
+ The largest observed error in this region is 1.52 ulps,
+ _ZGVsMxv_acos (0x1.24024271a500ap-1) got 0x1.ed82df4243f0dp-1
+ want 0x1.ed82df4243f0bp-1. */
+svfloat64_t SV_NAME_D1 (acos) (svfloat64_t x, const svbool_t pg)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ svuint64_t sign = svand_x (pg, svreinterpret_u64 (x), 0x8000000000000000);
+ svfloat64_t ax = svabs_x (pg, x);
+
+ svbool_t a_gt_half = svacgt (pg, x, 0.5);
+
+ /* Evaluate polynomial Q(x) = z + z * z2 * P(z2) with
+ z2 = x ^ 2 and z = |x| , if |x| < 0.5
+ z2 = (1 - |x|) / 2 and z = sqrt(z2), if |x| >= 0.5. */
+ svfloat64_t z2 = svsel (a_gt_half, svmls_x (pg, sv_f64 (0.5), ax, 0.5),
+ svmul_x (pg, x, x));
+ svfloat64_t z = svsqrt_m (ax, a_gt_half, z2);
+
+ /* Use a single polynomial approximation P for both intervals. */
+ svfloat64_t z4 = svmul_x (pg, z2, z2);
+ svfloat64_t z8 = svmul_x (pg, z4, z4);
+ svfloat64_t z16 = svmul_x (pg, z8, z8);
+ svfloat64_t p = sv_estrin_11_f64_x (pg, z2, z4, z8, z16, d->poly);
+
+ /* Finalize polynomial: z + z * z2 * P(z2). */
+ p = svmla_x (pg, z, svmul_x (pg, z, z2), p);
+
+ /* acos(|x|) = pi/2 - sign(x) * Q(|x|), for |x| < 0.5
+ = 2 Q(|x|) , for 0.5 < x < 1.0
+ = pi - 2 Q(|x|) , for -1.0 < x < -0.5. */
+ svfloat64_t y
+ = svreinterpret_f64 (svorr_x (pg, svreinterpret_u64 (p), sign));
+
+ svbool_t is_neg = svcmplt (pg, x, 0.0);
+ svfloat64_t off = svdup_f64_z (is_neg, d->pi);
+ svfloat64_t mul = svsel (a_gt_half, sv_f64 (2.0), sv_f64 (-1.0));
+ svfloat64_t add = svsel (a_gt_half, off, sv_f64 (d->pi_over_2));
+
+ return svmla_x (pg, add, mul, y);
+}
diff --git a/sysdeps/aarch64/fpu/acosf_advsimd.c b/sysdeps/aarch64/fpu/acosf_advsimd.c
new file mode 100644
index 0000000000..7d39e9b805
--- /dev/null
+++ b/sysdeps/aarch64/fpu/acosf_advsimd.c
@@ -0,0 +1,113 @@
+/* Single-precision AdvSIMD inverse cos
+
+ Copyright (C) 2023 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[5];
+ float32x4_t pi_over_2f, pif;
+} data = {
+ /* Polynomial approximation of (asin(sqrt(x)) - sqrt(x)) / (x * sqrt(x)) on
+ [ 0x1p-24 0x1p-2 ] order = 4 rel error: 0x1.00a23bbp-29 . */
+ .poly = { V4 (0x1.55555ep-3), V4 (0x1.33261ap-4), V4 (0x1.70d7dcp-5),
+ V4 (0x1.b059dp-6), V4 (0x1.3af7d8p-5) },
+ .pi_over_2f = V4 (0x1.921fb6p+0f),
+ .pif = V4 (0x1.921fb6p+1f),
+};
+
+#define AbsMask 0x7fffffff
+#define Half 0x3f000000
+#define One 0x3f800000
+#define Small 0x32800000 /* 2^-26. */
+
+#if WANT_SIMD_EXCEPT
+static float32x4_t VPCS_ATTR NOINLINE
+special_case (float32x4_t x, float32x4_t y, uint32x4_t special)
+{
+ return v_call_f32 (acosf, x, y, special);
+}
+#endif
+
+/* Single-precision implementation of vector acos(x).
+
+ For |x| < Small, approximate acos(x) by pi/2 - x. Small = 2^-26 for correct
+ rounding.
+ If WANT_SIMD_EXCEPT = 0, Small = 0 and we proceed with the following
+ approximation.
+
+ For |x| in [Small, 0.5], use order 4 polynomial P such that the final
+ approximation of asin is an odd polynomial:
+
+ acos(x) ~ pi/2 - (x + x^3 P(x^2)).
+
+ The largest observed error in this region is 1.26 ulps,
+ _ZGVnN4v_acosf (0x1.843bfcp-2) got 0x1.2e934cp+0 want 0x1.2e934ap+0.
+
+ For |x| in [0.5, 1.0], use same approximation with a change of variable
+
+ acos(x) = y + y * z * P(z), with z = (1-x)/2 and y = sqrt(z).
+
+ The largest observed error in this region is 1.32 ulps,
+ _ZGVnN4v_acosf (0x1.15ba56p-1) got 0x1.feb33p-1
+ want 0x1.feb32ep-1. */
+float32x4_t VPCS_ATTR V_NAME_F1 (acos) (float32x4_t x)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ uint32x4_t ix = vreinterpretq_u32_f32 (x);
+ uint32x4_t ia = vandq_u32 (ix, v_u32 (AbsMask));
+
+#if WANT_SIMD_EXCEPT
+ /* A single comparison for One, Small and QNaN. */
+ uint32x4_t special
+ = vcgtq_u32 (vsubq_u32 (ia, v_u32 (Small)), v_u32 (One - Small));
+ if (__glibc_unlikely (v_any_u32 (special)))
+ return special_case (x, x, v_u32 (0xffffffff));
+#endif
+
+ float32x4_t ax = vreinterpretq_f32_u32 (ia);
+ uint32x4_t a_le_half = vcleq_u32 (ia, v_u32 (Half));
+
+ /* Evaluate polynomial Q(x) = z + z * z2 * P(z2) with
+ z2 = x ^ 2 and z = |x| , if |x| < 0.5
+ z2 = (1 - |x|) / 2 and z = sqrt(z2), if |x| >= 0.5. */
+ float32x4_t z2 = vbslq_f32 (a_le_half, vmulq_f32 (x, x),
+ vfmsq_n_f32 (v_f32 (0.5), ax, 0.5));
+ float32x4_t z = vbslq_f32 (a_le_half, ax, vsqrtq_f32 (z2));
+
+ /* Use a single polynomial approximation P for both intervals. */
+ float32x4_t p = v_horner_4_f32 (z2, d->poly);
+ /* Finalize polynomial: z + z * z2 * P(z2). */
+ p = vfmaq_f32 (z, vmulq_f32 (z, z2), p);
+
+ /* acos(|x|) = pi/2 - sign(x) * Q(|x|), for |x| < 0.5
+ = 2 Q(|x|) , for 0.5 < x < 1.0
+ = pi - 2 Q(|x|) , for -1.0 < x < -0.5. */
+ float32x4_t y = vbslq_f32 (v_u32 (AbsMask), p, x);
+
+ uint32x4_t is_neg = vcltzq_f32 (x);
+ float32x4_t off = vreinterpretq_f32_u32 (
+ vandq_u32 (vreinterpretq_u32_f32 (d->pif), is_neg));
+ float32x4_t mul = vbslq_f32 (a_le_half, v_f32 (-1.0), v_f32 (2.0));
+ float32x4_t add = vbslq_f32 (a_le_half, d->pi_over_2f, off);
+
+ return vfmaq_f32 (add, mul, y);
+}
diff --git a/sysdeps/aarch64/fpu/acosf_sve.c b/sysdeps/aarch64/fpu/acosf_sve.c
new file mode 100644
index 0000000000..44253fa999
--- /dev/null
+++ b/sysdeps/aarch64/fpu/acosf_sve.c
@@ -0,0 +1,86 @@
+/* Single-precision SVE inverse cos
+
+ Copyright (C) 2023 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
+{
+ float32_t poly[5];
+ float32_t pi, pi_over_2;
+} data = {
+ /* Polynomial approximation of (asin(sqrt(x)) - sqrt(x)) / (x * sqrt(x)) on
+ [ 0x1p-24 0x1p-2 ] order = 4 rel error: 0x1.00a23bbp-29 . */
+ .poly = { 0x1.55555ep-3, 0x1.33261ap-4, 0x1.70d7dcp-5, 0x1.b059dp-6,
+ 0x1.3af7d8p-5, },
+ .pi = 0x1.921fb6p+1f,
+ .pi_over_2 = 0x1.921fb6p+0f,
+};
+
+/* Single-precision SVE implementation of vector acos(x).
+
+ For |x| in [0, 0.5], use order 4 polynomial P such that the final
+ approximation of asin is an odd polynomial:
+
+ acos(x) ~ pi/2 - (x + x^3 P(x^2)).
+
+ The largest observed error in this region is 1.16 ulps,
+ _ZGVsMxv_acosf(0x1.ffbeccp-2) got 0x1.0c27f8p+0
+ want 0x1.0c27f6p+0.
+
+ For |x| in [0.5, 1.0], use same approximation with a change of variable
+
+ acos(x) = y + y * z * P(z), with z = (1-x)/2 and y = sqrt(z).
+
+ The largest observed error in this region is 1.32 ulps,
+ _ZGVsMxv_acosf (0x1.15ba56p-1) got 0x1.feb33p-1
+ want 0x1.feb32ep-1. */
+svfloat32_t SV_NAME_F1 (acos) (svfloat32_t x, const svbool_t pg)
+{
+ const struct data *d = ptr_barrier (&data);
+
+ svuint32_t sign = svand_x (pg, svreinterpret_u32 (x), 0x80000000);
+ svfloat32_t ax = svabs_x (pg, x);
+ svbool_t a_gt_half = svacgt (pg, x, 0.5);
+
+ /* Evaluate polynomial Q(x) = z + z * z2 * P(z2) with
+ z2 = x ^ 2 and z = |x| , if |x| < 0.5
+ z2 = (1 - |x|) / 2 and z = sqrt(z2), if |x| >= 0.5. */
+ svfloat32_t z2 = svsel (a_gt_half, svmls_x (pg, sv_f32 (0.5), ax, 0.5),
+ svmul_x (pg, x, x));
+ svfloat32_t z = svsqrt_m (ax, a_gt_half, z2);
+
+ /* Use a single polynomial approximation P for both intervals. */
+ svfloat32_t p = sv_horner_4_f32_x (pg, z2, d->poly);
+ /* Finalize polynomial: z + z * z2 * P(z2). */
+ p = svmla_x (pg, z, svmul_x (pg, z, z2), p);
+
+ /* acos(|x|) = pi/2 - sign(x) * Q(|x|), for |x| < 0.5
+ = 2 Q(|x|) , for 0.5 < x < 1.0
+ = pi - 2 Q(|x|) , for -1.0 < x < -0.5. */
+ svfloat32_t y
+ = svreinterpret_f32 (svorr_x (pg, svreinterpret_u32 (p), sign));
+
+ svbool_t is_neg = svcmplt (pg, x, 0.0);
+ svfloat32_t off = svdup_f32_z (is_neg, d->pi);
+ svfloat32_t mul = svsel (a_gt_half, sv_f32 (2.0), sv_f32 (-1.0));
+ svfloat32_t add = svsel (a_gt_half, off, sv_f32 (d->pi_over_2));
+
+ return svmla_x (pg, add, mul, y);
+}
diff --git a/sysdeps/aarch64/fpu/bits/math-vector.h b/sysdeps/aarch64/fpu/bits/math-vector.h
index 03778faf96..f313993d70 100644
--- a/sysdeps/aarch64/fpu/bits/math-vector.h
+++ b/sysdeps/aarch64/fpu/bits/math-vector.h
@@ -49,6 +49,7 @@ typedef __SVBool_t __sv_bool_t;
# define __vpcs __attribute__ ((__aarch64_vector_pcs__))
+__vpcs __f32x4_t _ZGVnN4v_acosf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_asinf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_cosf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_expf (__f32x4_t);
@@ -60,6 +61,7 @@ __vpcs __f32x4_t _ZGVnN4v_log2f (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_sinf (__f32x4_t);
__vpcs __f32x4_t _ZGVnN4v_tanf (__f32x4_t);
+__vpcs __f64x2_t _ZGVnN2v_acos (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_asin (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_cos (__f64x2_t);
__vpcs __f64x2_t _ZGVnN2v_exp (__f64x2_t);
@@ -76,6 +78,7 @@ __vpcs __f64x2_t _ZGVnN2v_tan (__f64x2_t);
#ifdef __SVE_VEC_MATH_SUPPORTED
+__sv_f32_t _ZGVsMxv_acosf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_asinf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_cosf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_expf (__sv_f32_t, __sv_bool_t);
@@ -87,6 +90,7 @@ __sv_f32_t _ZGVsMxv_log2f (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_sinf (__sv_f32_t, __sv_bool_t);
__sv_f32_t _ZGVsMxv_tanf (__sv_f32_t, __sv_bool_t);
+__sv_f64_t _ZGVsMxv_acos (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_asin (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_cos (__sv_f64_t, __sv_bool_t);
__sv_f64_t _ZGVsMxv_exp (__sv_f64_t, __sv_bool_t);
diff --git a/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c b/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
index b5ccd6b1cc..5a0cbf743b 100644
--- a/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
@@ -23,6 +23,7 @@
#define VEC_TYPE float64x2_t
+VPCS_VECTOR_WRAPPER (acos_advsimd, _ZGVnN2v_acos)
VPCS_VECTOR_WRAPPER (asin_advsimd, _ZGVnN2v_asin)
VPCS_VECTOR_WRAPPER (cos_advsimd, _ZGVnN2v_cos)
VPCS_VECTOR_WRAPPER (exp_advsimd, _ZGVnN2v_exp)
diff --git a/sysdeps/aarch64/fpu/test-double-sve-wrappers.c b/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
index fc3b20f421..bd89ff6133 100644
--- a/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
@@ -32,6 +32,7 @@
return svlastb_f64 (svptrue_b64 (), mr); \
}
+SVE_VECTOR_WRAPPER (acos_sve, _ZGVsMxv_acos)
SVE_VECTOR_WRAPPER (asin_sve, _ZGVsMxv_asin)
SVE_VECTOR_WRAPPER (cos_sve, _ZGVsMxv_cos)
SVE_VECTOR_WRAPPER (exp_sve, _ZGVsMxv_exp)
diff --git a/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c b/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
index 0a36aa91f5..3fafca7557 100644
--- a/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
@@ -23,6 +23,7 @@
#define VEC_TYPE float32x4_t
+VPCS_VECTOR_WRAPPER (acosf_advsimd, _ZGVnN4v_acosf)
VPCS_VECTOR_WRAPPER (asinf_advsimd, _ZGVnN4v_asinf)
VPCS_VECTOR_WRAPPER (cosf_advsimd, _ZGVnN4v_cosf)
VPCS_VECTOR_WRAPPER (expf_advsimd, _ZGVnN4v_expf)
diff --git a/sysdeps/aarch64/fpu/test-float-sve-wrappers.c b/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
index f7e4882c7a..b4ec9f777b 100644
--- a/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
@@ -32,6 +32,7 @@
return svlastb_f32 (svptrue_b32 (), mr); \
}
+SVE_VECTOR_WRAPPER (acosf_sve, _ZGVsMxv_acosf)
SVE_VECTOR_WRAPPER (asinf_sve, _ZGVsMxv_asinf)
SVE_VECTOR_WRAPPER (cosf_sve, _ZGVsMxv_cosf)
SVE_VECTOR_WRAPPER (expf_sve, _ZGVsMxv_expf)
diff --git a/sysdeps/aarch64/libm-test-ulps b/sysdeps/aarch64/libm-test-ulps
index 1edc0fc343..c2b6f21b9d 100644
--- a/sysdeps/aarch64/libm-test-ulps
+++ b/sysdeps/aarch64/libm-test-ulps
@@ -6,11 +6,19 @@ double: 1
float: 1
ldouble: 1
+Function: "acos_advsimd":
+double: 1
+float: 1
+
Function: "acos_downward":
double: 1
float: 1
ldouble: 1
+Function: "acos_sve":
+double: 1
+float: 1
+
Function: "acos_towardzero":
double: 1
float: 1
diff --git a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
index 6431c3fe65..f79eaaf241 100644
--- a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
+++ b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
@@ -14,18 +14,22 @@ GLIBC_2.38 _ZGVsMxv_log F
GLIBC_2.38 _ZGVsMxv_logf F
GLIBC_2.38 _ZGVsMxv_sin F
GLIBC_2.38 _ZGVsMxv_sinf F
+GLIBC_2.39 _ZGVnN2v_acos F
GLIBC_2.39 _ZGVnN2v_asin F
GLIBC_2.39 _ZGVnN2v_exp10 F
GLIBC_2.39 _ZGVnN2v_exp2 F
GLIBC_2.39 _ZGVnN2v_log10 F
GLIBC_2.39 _ZGVnN2v_log2 F
GLIBC_2.39 _ZGVnN2v_tan F
+GLIBC_2.39 _ZGVnN4v_acosf F
GLIBC_2.39 _ZGVnN4v_asinf F
GLIBC_2.39 _ZGVnN4v_exp10f F
GLIBC_2.39 _ZGVnN4v_exp2f F
GLIBC_2.39 _ZGVnN4v_log10f F
GLIBC_2.39 _ZGVnN4v_log2f F
GLIBC_2.39 _ZGVnN4v_tanf F
+GLIBC_2.39 _ZGVsMxv_acos F
+GLIBC_2.39 _ZGVsMxv_acosf F
GLIBC_2.39 _ZGVsMxv_asin F
GLIBC_2.39 _ZGVsMxv_asinf F
GLIBC_2.39 _ZGVsMxv_exp10 F
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2023-11-13 14:25 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-11-13 14:25 [glibc] aarch64: Add vector implementations of acos routines Szabolcs Nagy
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).