diff options
| author | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2019-10-14 15:21:28 +0100 |
|---|---|---|
| committer | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2019-11-05 18:38:51 +0000 |
| commit | 69170e15cc2a10d7883e2fb40fb4437052febc13 (patch) | |
| tree | 80a149cd724906606532751fab4ff870ad750930 /math/v_expf_1u.c | |
| parent | 65464ec60ba73d113c8bd1632708014fa82dc9c6 (diff) | |
| download | arm-optimized-routines-69170e15cc2a10d7883e2fb40fb4437052febc13.tar.gz | |
Add vector exp2f
Same design as in expf. Worst-case error of __v_exp2f and __v_exp2f_1u
is 1.96 and 0.88 ulp respectively.
It is not clear if round/convert instructions are better or +- Shift.
For expf the latter, for exp2f the former seems more consistently
faster, but both options are kept in the code for now.
Diffstat (limited to 'math/v_expf_1u.c')
| -rw-r--r-- | math/v_expf_1u.c | 8 |
1 files changed, 8 insertions, 0 deletions
diff --git a/math/v_expf_1u.c b/math/v_expf_1u.c index 37d3d1e..023bd24 100644 --- a/math/v_expf_1u.c +++ b/math/v_expf_1u.c @@ -51,11 +51,19 @@ V_NAME(expf_1u) (v_f32_t x) /* exp(x) = 2^n * poly(r), with poly(r) in [1/sqrt(2),sqrt(2)] x = ln2*n + r, with r in [-ln2/2, ln2/2]. */ +#if 1 z = v_fma_f32 (x, InvLn2, Shift); n = z - Shift; r = v_fma_f32 (n, -Ln2hi, x); r = v_fma_f32 (n, -Ln2lo, r); e = v_as_u32_f32 (z) << 23; +#else + z = x * InvLn2; + n = v_round_f32 (z); + r = v_fma_f32 (n, -Ln2hi, x); + r = v_fma_f32 (n, -Ln2lo, r); + e = v_as_u32_s32 (v_round_s32 (z)) << 23; +#endif scale = v_as_f32_u32 (e + v_u32 (0x3f800000)); absn = v_abs_f32 (n); cmp = v_cond_u32 (absn > v_f32 (126.0f)); |