diff options
Diffstat (limited to 'sysdeps/aarch64/fpu/tanhf_advsimd.c')
-rw-r--r-- | sysdeps/aarch64/fpu/tanhf_advsimd.c | 21 |
1 files changed, 14 insertions, 7 deletions
diff --git a/sysdeps/aarch64/fpu/tanhf_advsimd.c b/sysdeps/aarch64/fpu/tanhf_advsimd.c index 50defd6ef0..3ced9b7a41 100644 --- a/sysdeps/aarch64/fpu/tanhf_advsimd.c +++ b/sysdeps/aarch64/fpu/tanhf_advsimd.c @@ -28,13 +28,16 @@ static const struct data /* 0x1.205966p+3, above which tanhf rounds to 1 (or -1 for negative). */ .boring_bound = V4 (0x41102cb3), .large_bound = V4 (0x7f800000), - .onef = V4 (0x3f800000), }; static float32x4_t NOINLINE VPCS_ATTR -special_case (float32x4_t x, float32x4_t y, uint32x4_t special) +special_case (float32x4_t x, uint32x4_t is_boring, float32x4_t boring, + float32x4_t q, uint32x4_t special) { - return v_call_f32 (tanhf, x, y, special); + return v_call_f32 ( + tanhf, x, + vbslq_f32 (is_boring, boring, vdivq_f32 (q, vaddq_f32 (q, v_f32 (2.0)))), + special); } /* Approximation for single-precision vector tanh(x), using a simplified @@ -50,7 +53,9 @@ float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (tanh) (float32x4_t x) uint32x4_t iax = vreinterpretq_u32_f32 (ax); uint32x4_t sign = veorq_u32 (ix, iax); uint32x4_t is_boring = vcgtq_u32 (iax, d->boring_bound); - float32x4_t boring = vreinterpretq_f32_u32 (vorrq_u32 (sign, d->onef)); + /* expm1 exponent bias is 1.0f reinterpreted to int. */ + float32x4_t boring = vreinterpretq_f32_u32 (vorrq_u32 ( + sign, vreinterpretq_u32_s32 (d->expm1f_consts.exponent_bias))); #if WANT_SIMD_EXCEPT /* If fp exceptions are to be triggered properly, set all special and boring @@ -66,10 +71,12 @@ float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (tanh) (float32x4_t x) /* tanh(x) = (e^2x - 1) / (e^2x + 1). */ float32x4_t q = expm1f_inline (vmulq_n_f32 (x, 2), &d->expm1f_consts); - float32x4_t y = vdivq_f32 (q, vaddq_f32 (q, v_f32 (2.0))); + if (__glibc_unlikely (v_any_u32 (special))) - return special_case (vreinterpretq_f32_u32 (ix), - vbslq_f32 (is_boring, boring, y), special); + return special_case (vreinterpretq_f32_u32 (ix), is_boring, boring, q, + special); + + float32x4_t y = vdivq_f32 (q, vaddq_f32 (q, v_f32 (2.0))); return vbslq_f32 (is_boring, boring, y); } libmvec_hidden_def (V_NAME_F1 (tanh)) |