diff options
Diffstat (limited to 'pl/math/v_sinh_3u.c')
-rw-r--r-- | pl/math/v_sinh_3u.c | 120 |
1 files changed, 72 insertions, 48 deletions
diff --git a/pl/math/v_sinh_3u.c b/pl/math/v_sinh_3u.c index 57ec66ecc282..a644f54b4a0f 100644 --- a/pl/math/v_sinh_3u.c +++ b/pl/math/v_sinh_3u.c @@ -6,47 +6,73 @@ */ #include "v_math.h" -#include "estrin.h" +#include "poly_advsimd_f64.h" #include "pl_sig.h" #include "pl_test.h" -#define AbsMask 0x7fffffffffffffff -#define Half 0x3fe0000000000000 -#define BigBound \ - 0x4080000000000000 /* 2^9. expm1 helper overflows for large input. */ -#define TinyBound \ - 0x3e50000000000000 /* 2^-26, below which sinh(x) rounds to x. */ -#define InvLn2 v_f64 (0x1.71547652b82fep0) -#define MLn2hi v_f64 (-0x1.62e42fefa39efp-1) -#define MLn2lo v_f64 (-0x1.abc9e3b39803fp-56) -#define Shift v_f64 (0x1.8p52) -#define One 0x3ff0000000000000 -#define C(i) v_f64 (__expm1_poly[i]) +static const struct data +{ + float64x2_t poly[11]; + float64x2_t inv_ln2, m_ln2, shift; + uint64x2_t halff; + int64x2_t onef; +#if WANT_SIMD_EXCEPT + uint64x2_t tiny_bound, thresh; +#else + uint64x2_t large_bound; +#endif +} data = { + /* Generated using Remez, deg=12 in [-log(2)/2, log(2)/2]. */ + .poly = { V2 (0x1p-1), V2 (0x1.5555555555559p-3), V2 (0x1.555555555554bp-5), + V2 (0x1.111111110f663p-7), V2 (0x1.6c16c16c1b5f3p-10), + V2 (0x1.a01a01affa35dp-13), V2 (0x1.a01a018b4ecbbp-16), + V2 (0x1.71ddf82db5bb4p-19), V2 (0x1.27e517fc0d54bp-22), + V2 (0x1.af5eedae67435p-26), V2 (0x1.1f143d060a28ap-29), }, -#if V_SUPPORTED + .inv_ln2 = V2 (0x1.71547652b82fep0), + .m_ln2 = (float64x2_t) {-0x1.62e42fefa39efp-1, -0x1.abc9e3b39803fp-56}, + .shift = V2 (0x1.8p52), -static inline v_f64_t -expm1_inline (v_f64_t x) + .halff = V2 (0x3fe0000000000000), + .onef = V2 (0x3ff0000000000000), +#if WANT_SIMD_EXCEPT + /* 2^-26, below which sinh(x) rounds to x. */ + .tiny_bound = V2 (0x3e50000000000000), + /* asuint(large_bound) - asuint(tiny_bound). */ + .thresh = V2 (0x0230000000000000), +#else +/* 2^9. expm1 helper overflows for large input. */ + .large_bound = V2 (0x4080000000000000), +#endif +}; + +static inline float64x2_t +expm1_inline (float64x2_t x) { + const struct data *d = ptr_barrier (&data); + /* Reduce argument: exp(x) - 1 = 2^i * (expm1(f) + 1) - 1 where i = round(x / ln2) and f = x - i * ln2 (f in [-ln2/2, ln2/2]). */ - v_f64_t j = v_fma_f64 (InvLn2, x, Shift) - Shift; - v_s64_t i = v_to_s64_f64 (j); - v_f64_t f = v_fma_f64 (j, MLn2hi, x); - f = v_fma_f64 (j, MLn2lo, f); + float64x2_t j = vsubq_f64 (vfmaq_f64 (d->shift, d->inv_ln2, x), d->shift); + int64x2_t i = vcvtq_s64_f64 (j); + float64x2_t f = vfmaq_laneq_f64 (x, j, d->m_ln2, 0); + f = vfmaq_laneq_f64 (f, j, d->m_ln2, 1); /* Approximate expm1(f) using polynomial. */ - v_f64_t f2 = f * f, f4 = f2 * f2, f8 = f4 * f4; - v_f64_t p = v_fma_f64 (f2, ESTRIN_10 (f, f2, f4, f8, C), f); + float64x2_t f2 = vmulq_f64 (f, f); + float64x2_t f4 = vmulq_f64 (f2, f2); + float64x2_t f8 = vmulq_f64 (f4, f4); + float64x2_t p = vfmaq_f64 (f, f2, v_estrin_10_f64 (f, f2, f4, f8, d->poly)); /* t = 2^i. */ - v_f64_t t = v_as_f64_u64 (v_as_u64_s64 (i << 52) + One); + float64x2_t t = vreinterpretq_f64_u64 ( + vreinterpretq_u64_s64 (vaddq_s64 (vshlq_n_s64 (i, 52), d->onef))); /* expm1(x) ~= p * t + (t - 1). */ - return v_fma_f64 (p, t, t - 1); + return vfmaq_f64 (vsubq_f64 (t, v_f64 (1.0)), p, t); } -static NOINLINE VPCS_ATTR v_f64_t -special_case (v_f64_t x) +static float64x2_t NOINLINE VPCS_ATTR +special_case (float64x2_t x) { return v_call_f64 (sinh, x, x, v_u64 (-1)); } @@ -54,20 +80,22 @@ special_case (v_f64_t x) /* Approximation for vector double-precision sinh(x) using expm1. sinh(x) = (exp(x) - exp(-x)) / 2. The greatest observed error is 2.57 ULP: - sinh(0x1.9fb1d49d1d58bp-2) got 0x1.ab34e59d678dcp-2 - want 0x1.ab34e59d678d9p-2. */ -VPCS_ATTR v_f64_t V_NAME (sinh) (v_f64_t x) + _ZGVnN2v_sinh (0x1.9fb1d49d1d58bp-2) got 0x1.ab34e59d678dcp-2 + want 0x1.ab34e59d678d9p-2. */ +float64x2_t VPCS_ATTR V_NAME_D1 (sinh) (float64x2_t x) { - v_u64_t ix = v_as_u64_f64 (x); - v_u64_t iax = ix & AbsMask; - v_f64_t ax = v_as_f64_u64 (iax); - v_u64_t sign = ix & ~AbsMask; - v_f64_t halfsign = v_as_f64_u64 (sign | Half); + const struct data *d = ptr_barrier (&data); + + float64x2_t ax = vabsq_f64 (x); + uint64x2_t sign + = veorq_u64 (vreinterpretq_u64_f64 (x), vreinterpretq_u64_f64 (ax)); + float64x2_t halfsign = vreinterpretq_f64_u64 (vorrq_u64 (sign, d->halff)); #if WANT_SIMD_EXCEPT - v_u64_t special = v_cond_u64 ((iax - TinyBound) >= (BigBound - TinyBound)); + uint64x2_t special = vcgeq_u64 ( + vsubq_u64 (vreinterpretq_u64_f64 (ax), d->tiny_bound), d->thresh); #else - v_u64_t special = v_cond_u64 (iax >= BigBound); + uint64x2_t special = vcgeq_u64 (vreinterpretq_u64_f64 (ax), d->large_bound); #endif /* Fall back to scalar variant for all lanes if any of them are special. */ @@ -77,18 +105,14 @@ VPCS_ATTR v_f64_t V_NAME (sinh) (v_f64_t x) /* Up to the point that expm1 overflows, we can use it to calculate sinh using a slight rearrangement of the definition of sinh. This allows us to retain acceptable accuracy for very small inputs. */ - v_f64_t t = expm1_inline (ax); - return (t + t / (t + 1)) * halfsign; + float64x2_t t = expm1_inline (ax); + t = vaddq_f64 (t, vdivq_f64 (t, vaddq_f64 (t, v_f64 (1.0)))); + return vmulq_f64 (t, halfsign); } -VPCS_ALIAS PL_SIG (V, D, 1, sinh, -10.0, 10.0) -PL_TEST_ULP (V_NAME (sinh), 2.08) -PL_TEST_EXPECT_FENV (V_NAME (sinh), WANT_SIMD_EXCEPT) -PL_TEST_INTERVAL (V_NAME (sinh), 0, TinyBound, 1000) -PL_TEST_INTERVAL (V_NAME (sinh), -0, -TinyBound, 1000) -PL_TEST_INTERVAL (V_NAME (sinh), TinyBound, BigBound, 500000) -PL_TEST_INTERVAL (V_NAME (sinh), -TinyBound, -BigBound, 500000) -PL_TEST_INTERVAL (V_NAME (sinh), BigBound, inf, 1000) -PL_TEST_INTERVAL (V_NAME (sinh), -BigBound, -inf, 1000) -#endif +PL_TEST_ULP (V_NAME_D1 (sinh), 2.08) +PL_TEST_EXPECT_FENV (V_NAME_D1 (sinh), WANT_SIMD_EXCEPT) +PL_TEST_SYM_INTERVAL (V_NAME_D1 (sinh), 0, 0x1p-26, 1000) +PL_TEST_SYM_INTERVAL (V_NAME_D1 (sinh), 0x1p-26, 0x1p9, 500000) +PL_TEST_SYM_INTERVAL (V_NAME_D1 (sinh), 0x1p9, inf, 1000) |