/* * Double-precision SVE asinh(x) function. * * Copyright (c) 2022-2023, Arm Limited. * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception */ #include "sv_math.h" #include "poly_sve_f64.h" #include "pl_sig.h" #include "pl_test.h" #define OneTop sv_u64 (0x3ff) /* top12(asuint64(1.0f)). */ #define HugeBound sv_u64 (0x5fe) /* top12(asuint64(0x1p511)). */ #define TinyBound (0x3e5) /* top12(asuint64(0x1p-26)). */ #define SignMask (0x8000000000000000) /* Constants & data for log. */ #define A(i) __v_log_data.poly[i] #define Ln2 (0x1.62e42fefa39efp-1) #define N (1 << V_LOG_TABLE_BITS) #define OFF (0x3fe6900900000000) static svfloat64_t NOINLINE special_case (svfloat64_t x, svfloat64_t y, svbool_t special) { return sv_call_f64 (asinh, x, y, special); } static inline svfloat64_t __sv_log_inline (svfloat64_t x, const svbool_t pg) { /* Double-precision SVE log, copied from pl/math/sv_log_2u5.c with some cosmetic modification and special-cases removed. See that file for details of the algorithm used. */ svuint64_t ix = svreinterpret_u64 (x); svuint64_t tmp = svsub_x (pg, ix, OFF); svuint64_t i = svand_x (pg, svlsr_x (pg, tmp, (51 - V_LOG_TABLE_BITS)), (N - 1) << 1); svint64_t k = svasr_x (pg, svreinterpret_s64 (tmp), 52); svuint64_t iz = svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52)); svfloat64_t z = svreinterpret_f64 (iz); svfloat64_t invc = svld1_gather_index (pg, &__v_log_data.table[0].invc, i); svfloat64_t logc = svld1_gather_index (pg, &__v_log_data.table[0].logc, i); svfloat64_t r = svmla_x (pg, sv_f64 (-1.0), invc, z); svfloat64_t kd = svcvt_f64_x (pg, k); svfloat64_t hi = svmla_x (pg, svadd_x (pg, logc, r), kd, Ln2); svfloat64_t r2 = svmul_x (pg, r, r); svfloat64_t y = svmla_x (pg, sv_f64 (A (2)), r, A (3)); svfloat64_t p = svmla_x (pg, sv_f64 (A (0)), r, A (1)); y = svmla_x (pg, y, r2, A (4)); y = svmla_x (pg, p, r2, y); y = svmla_x (pg, hi, r2, y); return y; } /* Double-precision implementation of SVE asinh(x). asinh is very sensitive around 1, so it is impractical to devise a single low-cost algorithm which is sufficiently accurate on a wide range of input. Instead we use two different algorithms: asinh(x) = sign(x) * log(|x| + sqrt(x^2 + 1) if |x| >= 1 = sign(x) * (|x| + |x|^3 * P(x^2)) otherwise where log(x) is an optimized log approximation, and P(x) is a polynomial shared with the scalar routine. The greatest observed error 2.51 ULP, in |x| >= 1: _ZGVsMxv_asinh(0x1.170469d024505p+0) got 0x1.e3181c43b0f36p-1 want 0x1.e3181c43b0f39p-1. */ svfloat64_t SV_NAME_D1 (asinh) (svfloat64_t x, const svbool_t pg) { svuint64_t ix = svreinterpret_u64 (x); svuint64_t iax = svbic_x (pg, ix, SignMask); svuint64_t sign = svand_x (pg, ix, SignMask); svfloat64_t ax = svreinterpret_f64 (iax); svuint64_t top12 = svlsr_x (pg, iax, 52); svbool_t ge1 = svcmpge (pg, top12, OneTop); svbool_t special = svcmpge (pg, top12, HugeBound); /* Option 1: |x| >= 1. Compute asinh(x) according by asinh(x) = log(x + sqrt(x^2 + 1)). */ svfloat64_t option_1 = sv_f64 (0); if (likely (svptest_any (pg, ge1))) { svfloat64_t axax = svmul_x (pg, ax, ax); option_1 = __sv_log_inline ( svadd_x (pg, ax, svsqrt_x (pg, svadd_x (pg, axax, 1))), pg); } /* Option 2: |x| < 1. Compute asinh(x) using a polynomial. The largest observed error in this region is 1.51 ULPs: _ZGVsMxv_asinh(0x1.fe12bf8c616a2p-1) got 0x1.c1e649ee2681bp-1 want 0x1.c1e649ee2681dp-1. */ svfloat64_t option_2 = sv_f64 (0); if (likely (svptest_any (pg, svnot_z (pg, ge1)))) { svfloat64_t x2 = svmul_x (pg, ax, ax); svfloat64_t z2 = svmul_x (pg, x2, x2); svfloat64_t z4 = svmul_x (pg, z2, z2); svfloat64_t z8 = svmul_x (pg, z4, z4); svfloat64_t z16 = svmul_x (pg, z8, z8); svfloat64_t p = sv_estrin_17_f64_x (pg, x2, z2, z4, z8, z16, __asinh_data.poly); option_2 = svmla_x (pg, ax, p, svmul_x (pg, x2, ax)); } /* Choose the right option for each lane. */ svfloat64_t y = svsel (ge1, option_1, option_2); /* Apply sign of x to y. */ y = svreinterpret_f64 (sveor_x (pg, svreinterpret_u64 (y), sign)); if (unlikely (svptest_any (pg, special))) return special_case (x, y, special); return y; } PL_SIG (SV, D, 1, asinh, -10.0, 10.0) PL_TEST_ULP (SV_NAME_D1 (asinh), 2.52) /* Test vector asinh 3 times, with control lane < 1, > 1 and special. Ensures the svsel is choosing the right option in all cases. */ #define SV_ASINH_INTERVAL(lo, hi, n) \ PL_TEST_SYM_INTERVAL_C (SV_NAME_D1 (asinh), lo, hi, n, 0.5) \ PL_TEST_SYM_INTERVAL_C (SV_NAME_D1 (asinh), lo, hi, n, 2) \ PL_TEST_SYM_INTERVAL_C (SV_NAME_D1 (asinh), lo, hi, n, 0x1p600) SV_ASINH_INTERVAL (0, 0x1p-26, 50000) SV_ASINH_INTERVAL (0x1p-26, 1, 50000) SV_ASINH_INTERVAL (1, 0x1p511, 50000) SV_ASINH_INTERVAL (0x1p511, inf, 40000)