diff options
Diffstat (limited to 'math/aarch64/advsimd/cbrt.c')
| -rw-r--r-- | math/aarch64/advsimd/cbrt.c | 127 |
1 files changed, 127 insertions, 0 deletions
diff --git a/math/aarch64/advsimd/cbrt.c b/math/aarch64/advsimd/cbrt.c new file mode 100644 index 000000000000..8e72e5b566fc --- /dev/null +++ b/math/aarch64/advsimd/cbrt.c @@ -0,0 +1,127 @@ +/* + * Double-precision vector cbrt(x) function. + * + * Copyright (c) 2022-2024, Arm Limited. + * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception + */ + +#include "v_math.h" +#include "test_sig.h" +#include "test_defs.h" +#include "v_poly_f64.h" + +const static struct data +{ + float64x2_t poly[4], one_third, shift; + int64x2_t exp_bias; + uint64x2_t abs_mask, tiny_bound; + uint32x4_t thresh; + double table[5]; +} data = { + .shift = V2 (0x1.8p52), + .poly = { /* Generated with fpminimax in [0.5, 1]. */ + V2 (0x1.c14e8ee44767p-2), V2 (0x1.dd2d3f99e4c0ep-1), + V2 (-0x1.08e83026b7e74p-1), V2 (0x1.2c74eaa3ba428p-3) }, + .exp_bias = V2 (1022), + .abs_mask = V2(0x7fffffffffffffff), + .tiny_bound = V2(0x0010000000000000), /* Smallest normal. */ + .thresh = V4(0x7fe00000), /* asuint64 (infinity) - tiny_bound. */ + .one_third = V2(0x1.5555555555555p-2), + .table = { /* table[i] = 2^((i - 2) / 3). */ + 0x1.428a2f98d728bp-1, 0x1.965fea53d6e3dp-1, 0x1p0, + 0x1.428a2f98d728bp0, 0x1.965fea53d6e3dp0 } +}; + +#define MantissaMask v_u64 (0x000fffffffffffff) + +static float64x2_t NOINLINE VPCS_ATTR +special_case (float64x2_t x, float64x2_t y, uint32x2_t special) +{ + return v_call_f64 (cbrt, x, y, vmovl_u32 (special)); +} + +/* Approximation for double-precision vector cbrt(x), using low-order + polynomial and two Newton iterations. + + The vector version of frexp does not handle subnormals + correctly. As a result these need to be handled by the scalar + fallback, where accuracy may be worse than that of the vector code + path. + + Greatest observed error in the normal range is 1.79 ULP. Errors repeat + according to the exponent, for instance an error observed for double value + m * 2^e will be observed for any input m * 2^(e + 3*i), where i is an + integer. + _ZGVnN2v_cbrt (0x1.fffff403f0bc6p+1) got 0x1.965fe72821e9bp+0 + want 0x1.965fe72821e99p+0. */ +VPCS_ATTR float64x2_t V_NAME_D1 (cbrt) (float64x2_t x) +{ + const struct data *d = ptr_barrier (&data); + uint64x2_t iax = vreinterpretq_u64_f64 (vabsq_f64 (x)); + + /* Subnormal, +/-0 and special values. */ + uint32x2_t special + = vcge_u32 (vsubhn_u64 (iax, d->tiny_bound), vget_low_u32 (d->thresh)); + + /* Decompose |x| into m * 2^e, where m is in [0.5, 1.0]. This is a vector + version of frexp, which gets subnormal values wrong - these have to be + special-cased as a result. */ + float64x2_t m = vbslq_f64 (MantissaMask, x, v_f64 (0.5)); + int64x2_t exp_bias = d->exp_bias; + uint64x2_t ia12 = vshrq_n_u64 (iax, 52); + int64x2_t e = vsubq_s64 (vreinterpretq_s64_u64 (ia12), exp_bias); + + /* Calculate rough approximation for cbrt(m) in [0.5, 1.0], starting point + for Newton iterations. */ + float64x2_t p = v_pairwise_poly_3_f64 (m, vmulq_f64 (m, m), d->poly); + float64x2_t one_third = d->one_third; + /* Two iterations of Newton's method for iteratively approximating cbrt. */ + float64x2_t m_by_3 = vmulq_f64 (m, one_third); + float64x2_t two_thirds = vaddq_f64 (one_third, one_third); + float64x2_t a + = vfmaq_f64 (vdivq_f64 (m_by_3, vmulq_f64 (p, p)), two_thirds, p); + a = vfmaq_f64 (vdivq_f64 (m_by_3, vmulq_f64 (a, a)), two_thirds, a); + + /* Assemble the result by the following: + + cbrt(x) = cbrt(m) * 2 ^ (e / 3). + + We can get 2 ^ round(e / 3) using ldexp and integer divide, but since e is + not necessarily a multiple of 3 we lose some information. + + Let q = 2 ^ round(e / 3), then t = 2 ^ (e / 3) / q. + + Then we know t = 2 ^ (i / 3), where i is the remainder from e / 3, which + is an integer in [-2, 2], and can be looked up in the table T. Hence the + result is assembled as: + + cbrt(x) = cbrt(m) * t * 2 ^ round(e / 3) * sign. */ + + float64x2_t ef = vcvtq_f64_s64 (e); + float64x2_t eb3f = vrndnq_f64 (vmulq_f64 (ef, one_third)); + int64x2_t em3 = vcvtq_s64_f64 (vfmsq_f64 (ef, eb3f, v_f64 (3))); + int64x2_t ey = vcvtq_s64_f64 (eb3f); + + float64x2_t my = (float64x2_t){ d->table[em3[0] + 2], d->table[em3[1] + 2] }; + my = vmulq_f64 (my, a); + + /* Vector version of ldexp. */ + float64x2_t y = vreinterpretq_f64_s64 ( + vshlq_n_s64 (vaddq_s64 (ey, vaddq_s64 (exp_bias, v_s64 (1))), 52)); + y = vmulq_f64 (y, my); + + if (unlikely (v_any_u32h (special))) + return special_case (x, vbslq_f64 (d->abs_mask, y, x), special); + + /* Copy sign. */ + return vbslq_f64 (d->abs_mask, y, x); +} + +/* Worse-case ULP error assumes that scalar fallback is GLIBC 2.40 cbrt, which + has ULP error of 3.67 at 0x1.7a337e1ba1ec2p-257 [1]. Largest observed error + in the vector path is 1.79 ULP. + [1] Innocente, V., & Zimmermann, P. (2024). Accuracy of Mathematical + Functions in Single, Double, Double Extended, and Quadruple Precision. */ +TEST_ULP (V_NAME_D1 (cbrt), 3.17) +TEST_SIG (V, D, 1, cbrt, -10.0, 10.0) +TEST_SYM_INTERVAL (V_NAME_D1 (cbrt), 0, inf, 1000000) |