diff options
author | Joe Ramsay <Joe.Ramsay@arm.com> | 2023-11-03 12:12:22 +0000 |
---|---|---|
committer | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2023-11-10 17:07:43 +0000 |
commit | b07038c5d304a7afc312516ce0ff886a57bf3163 (patch) | |
tree | 34d0a8e9e4d170229e4385dd849883c9af5dbee9 /sysdeps/aarch64/fpu/atan2f_sve.c | |
parent | d30c39f80d19d62e8fd750c424ccb7eb06b617e5 (diff) | |
download | glibc-b07038c5d304a7afc312516ce0ff886a57bf3163.tar.gz glibc-b07038c5d304a7afc312516ce0ff886a57bf3163.tar.xz glibc-b07038c5d304a7afc312516ce0ff886a57bf3163.zip |
aarch64: Add vector implementations of atan2 routines
Diffstat (limited to 'sysdeps/aarch64/fpu/atan2f_sve.c')
-rw-r--r-- | sysdeps/aarch64/fpu/atan2f_sve.c | 110 |
1 files changed, 110 insertions, 0 deletions
diff --git a/sysdeps/aarch64/fpu/atan2f_sve.c b/sysdeps/aarch64/fpu/atan2f_sve.c new file mode 100644 index 0000000000..606a62c144 --- /dev/null +++ b/sysdeps/aarch64/fpu/atan2f_sve.c @@ -0,0 +1,110 @@ +/* Single-precision SVE atan2 + + 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[8]; + float32_t pi_over_2; +} data = { + /* Coefficients of polynomial P such that atan(x)~x+x*P(x^2) on + [2**-128, 1.0]. */ + .poly = { -0x1.55555p-2f, 0x1.99935ep-3f, -0x1.24051ep-3f, 0x1.bd7368p-4f, + -0x1.491f0ep-4f, 0x1.93a2c0p-5f, -0x1.4c3c60p-6f, 0x1.01fd88p-8f }, + .pi_over_2 = 0x1.921fb6p+0f, +}; + +#define SignMask sv_u32 (0x80000000) + +/* Special cases i.e. 0, infinity, nan (fall back to scalar calls). */ +static inline svfloat32_t +special_case (svfloat32_t y, svfloat32_t x, svfloat32_t ret, + const svbool_t cmp) +{ + return sv_call2_f32 (atan2f, y, x, ret, cmp); +} + +/* Returns a predicate indicating true if the input is the bit representation + of 0, infinity or nan. */ +static inline svbool_t +zeroinfnan (svuint32_t i, const svbool_t pg) +{ + return svcmpge (pg, svsub_x (pg, svlsl_x (pg, i, 1), 1), + sv_u32 (2 * 0x7f800000lu - 1)); +} + +/* Fast implementation of SVE atan2f based on atan(x) ~ shift + z + z^3 * + P(z^2) with reduction to [0,1] using z=1/x and shift = pi/2. Maximum + observed error is 2.95 ULP: + _ZGVsMxvv_atan2f (0x1.93836cp+6, 0x1.8cae1p+6) got 0x1.967f06p-1 + want 0x1.967f00p-1. */ +svfloat32_t SV_NAME_F2 (atan2) (svfloat32_t y, svfloat32_t x, const svbool_t pg) +{ + const struct data *data_ptr = ptr_barrier (&data); + + svuint32_t ix = svreinterpret_u32 (x); + svuint32_t iy = svreinterpret_u32 (y); + + svbool_t cmp_x = zeroinfnan (ix, pg); + svbool_t cmp_y = zeroinfnan (iy, pg); + svbool_t cmp_xy = svorr_z (pg, cmp_x, cmp_y); + + svuint32_t sign_x = svand_x (pg, ix, SignMask); + svuint32_t sign_y = svand_x (pg, iy, SignMask); + svuint32_t sign_xy = sveor_x (pg, sign_x, sign_y); + + svfloat32_t ax = svabs_x (pg, x); + svfloat32_t ay = svabs_x (pg, y); + + svbool_t pred_xlt0 = svcmplt (pg, x, 0.0); + svbool_t pred_aygtax = svcmpgt (pg, ay, ax); + + /* Set up z for call to atan. */ + svfloat32_t n = svsel (pred_aygtax, svneg_x (pg, ax), ay); + svfloat32_t d = svsel (pred_aygtax, ay, ax); + svfloat32_t z = svdiv_x (pg, n, d); + + /* Work out the correct shift. */ + svfloat32_t shift = svsel (pred_xlt0, sv_f32 (-2.0), sv_f32 (0.0)); + shift = svsel (pred_aygtax, svadd_x (pg, shift, 1.0), shift); + shift = svmul_x (pg, shift, sv_f32 (data_ptr->pi_over_2)); + + /* Use split Estrin scheme for P(z^2) with deg(P)=7. */ + svfloat32_t z2 = svmul_x (pg, z, z); + svfloat32_t z4 = svmul_x (pg, z2, z2); + svfloat32_t z8 = svmul_x (pg, z4, z4); + + svfloat32_t ret = sv_estrin_7_f32_x (pg, z2, z4, z8, data_ptr->poly); + + /* ret = shift + z + z^3 * P(z^2). */ + svfloat32_t z3 = svmul_x (pg, z2, z); + ret = svmla_x (pg, z, z3, ret); + + ret = svadd_m (pg, ret, shift); + + /* Account for the sign of x and y. */ + ret = svreinterpret_f32 (sveor_x (pg, svreinterpret_u32 (ret), sign_xy)); + + if (__glibc_unlikely (svptest_any (pg, cmp_xy))) + return special_case (y, x, ret, cmp_xy); + + return ret; +} |