diff options
Diffstat (limited to 'sysdeps/aarch64/fpu/atanhf_advsimd.c')
| -rw-r--r-- | sysdeps/aarch64/fpu/atanhf_advsimd.c | 26 |
1 files changed, 18 insertions, 8 deletions
diff --git a/sysdeps/aarch64/fpu/atanhf_advsimd.c b/sysdeps/aarch64/fpu/atanhf_advsimd.c index ae488f7b54..818b6c92ad 100644 --- a/sysdeps/aarch64/fpu/atanhf_advsimd.c +++ b/sysdeps/aarch64/fpu/atanhf_advsimd.c @@ -40,15 +40,17 @@ const static struct data #define Half v_u32 (0x3f000000) static float32x4_t NOINLINE VPCS_ATTR -special_case (float32x4_t x, float32x4_t y, uint32x4_t special) +special_case (float32x4_t x, float32x4_t halfsign, float32x4_t y, + uint32x4_t special) { - return v_call_f32 (atanhf, x, y, special); + return v_call_f32 (atanhf, vbslq_f32 (AbsMask, x, halfsign), + vmulq_f32 (halfsign, y), special); } /* Approximation for vector single-precision atanh(x) using modified log1p. - The maximum error is 3.08 ULP: - __v_atanhf(0x1.ff215p-5) got 0x1.ffcb7cp-5 - want 0x1.ffcb82p-5. */ + The maximum error is 2.93 ULP: + _ZGVnN4v_atanhf(0x1.f43d7p-5) got 0x1.f4dcfep-5 + want 0x1.f4dcf8p-5. */ VPCS_ATTR float32x4_t NOINLINE V_NAME_F1 (atanh) (float32x4_t x) { const struct data *d = ptr_barrier (&data); @@ -68,11 +70,19 @@ VPCS_ATTR float32x4_t NOINLINE V_NAME_F1 (atanh) (float32x4_t x) uint32x4_t special = vcgeq_u32 (iax, d->one); #endif - float32x4_t y = vdivq_f32 (vaddq_f32 (ax, ax), vsubq_f32 (v_f32 (1), ax)); - y = log1pf_inline (y, d->log1pf_consts); + float32x4_t y = vdivq_f32 (vaddq_f32 (ax, ax), + vsubq_f32 (vreinterpretq_f32_u32 (d->one), ax)); + y = log1pf_inline (y, &d->log1pf_consts); + /* If exceptions not required, pass ax to special-case for shorter dependency + chain. If exceptions are required ax will have been zerofied, so have to + pass x. */ if (__glibc_unlikely (v_any_u32 (special))) - return special_case (x, vmulq_f32 (halfsign, y), special); +#if WANT_SIMD_EXCEPT + return special_case (x, halfsign, y, special); +#else + return special_case (ax, halfsign, y, special); +#endif return vmulq_f32 (halfsign, y); } libmvec_hidden_def (V_NAME_F1 (atanh)) |