diff options
Diffstat (limited to 'sysdeps/aarch64/fpu/acoshf_sve.c')
| -rw-r--r-- | sysdeps/aarch64/fpu/acoshf_sve.c | 21 |
1 files changed, 12 insertions, 9 deletions
diff --git a/sysdeps/aarch64/fpu/acoshf_sve.c b/sysdeps/aarch64/fpu/acoshf_sve.c index 2110894e62..491365e24d 100644 --- a/sysdeps/aarch64/fpu/acoshf_sve.c +++ b/sysdeps/aarch64/fpu/acoshf_sve.c @@ -17,23 +17,26 @@ License along with the GNU C Library; if not, see <https://www.gnu.org/licenses/>. */ +#include "sv_math.h" +#include "sv_log1pf_inline.h" + #define One 0x3f800000 #define Thres 0x20000000 /* asuint(0x1p64) - One. */ -#include "sv_log1pf_inline.h" - static svfloat32_t NOINLINE -special_case (svfloat32_t x, svfloat32_t y, svbool_t special) +special_case (svfloat32_t xm1, svfloat32_t tmp, svbool_t special) { + svfloat32_t x = svadd_x (svptrue_b32 (), xm1, 1.0f); + svfloat32_t y = sv_log1pf_inline (tmp, svptrue_b32 ()); return sv_call_f32 (acoshf, x, y, special); } /* Single-precision SVE acosh(x) routine. Implements the same algorithm as vector acoshf and log1p. - Maximum error is 2.78 ULPs: - SV_NAME_F1 (acosh) (0x1.01e996p+0) got 0x1.f45b42p-4 - want 0x1.f45b3cp-4. */ + Maximum error is 2.47 ULPs: + SV_NAME_F1 (acosh) (0x1.01ca76p+0) got 0x1.e435a6p-4 + want 0x1.e435a2p-4. */ svfloat32_t SV_NAME_F1 (acosh) (svfloat32_t x, const svbool_t pg) { svuint32_t ix = svreinterpret_u32 (x); @@ -41,9 +44,9 @@ svfloat32_t SV_NAME_F1 (acosh) (svfloat32_t x, const svbool_t pg) svfloat32_t xm1 = svsub_x (pg, x, 1.0f); svfloat32_t u = svmul_x (pg, xm1, svadd_x (pg, x, 1.0f)); - svfloat32_t y = sv_log1pf_inline (svadd_x (pg, xm1, svsqrt_x (pg, u)), pg); + svfloat32_t tmp = svadd_x (pg, xm1, svsqrt_x (pg, u)); if (__glibc_unlikely (svptest_any (pg, special))) - return special_case (x, y, special); - return y; + return special_case (xm1, tmp, special); + return sv_log1pf_inline (tmp, pg); } |