/* * Core approximation for double-precision SVE sincospi * * Copyright (c) 2024, Arm Limited. * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception */ #include "sv_math.h" #include "sv_poly_f64.h" static const struct sv_sincospi_data { double c0, c2, c4, c6, c8; double c1, c3, c5, c7, c9; double range_val; } sv_sincospi_data = { /* Polynomial coefficients generated using Remez algorithm, see sinpi.sollya for details. */ .c0 = 0x1.921fb54442d184p1, .c1 = -0x1.4abbce625be53p2, .c2 = 0x1.466bc6775ab16p1, .c3 = -0x1.32d2cce62dc33p-1, .c4 = 0x1.507834891188ep-4, .c5 = -0x1.e30750a28c88ep-8, .c6 = 0x1.e8f48308acda4p-12, .c7 = -0x1.6fc0032b3c29fp-16, .c8 = 0x1.af86ae521260bp-21, .c9 = -0x1.012a9870eeb7dp-25, /* Exclusive upper bound for a signed integer. */ .range_val = 0x1p63 }; /* Double-precision vector function allowing calculation of both sinpi and cospi in one function call, using shared argument reduction and polynomials. Worst-case error for sin is 3.09 ULP: _ZGVsMxvl8l8_sincospi_sin(0x1.7a41deb4b21e1p+14) got 0x1.fd54d0b327cf1p-1 want 0x1.fd54d0b327cf4p-1. Worst-case error for cos is 3.16 ULP: _ZGVsMxvl8l8_sincospi_cos(-0x1.11e3c7e284adep-5) got 0x1.fd2da484ff3ffp-1 want 0x1.fd2da484ff402p-1. */ static inline svfloat64x2_t sv_sincospi_inline (svbool_t pg, svfloat64_t x, const struct sv_sincospi_data *d) { const svbool_t pt = svptrue_b64 (); /* r = x - rint(x). */ /* pt hints unpredicated instruction. */ svfloat64_t rx = svrinta_x (pg, x); svfloat64_t sr = svsub_x (pt, x, rx); /* cospi(x) = sinpi(0.5 - abs(x)) for values -1/2 .. 1/2. */ svfloat64_t cr = svsubr_x (pg, svabs_x (pg, sr), 0.5); /* Pairwise Horner approximation for y = sin(r * pi). */ /* pt hints unpredicated instruction. */ svfloat64_t sr2 = svmul_x (pt, sr, sr); svfloat64_t cr2 = svmul_x (pt, cr, cr); svfloat64_t sr4 = svmul_x (pt, sr2, sr2); svfloat64_t cr4 = svmul_x (pt, cr2, cr2); /* If rint(x) is odd, the sign of the result should be inverted for sinpi and re-introduced for cospi. cmp filters rxs that saturate to max sint. */ svbool_t cmp = svaclt (pg, x, d->range_val); svuint64_t odd = svlsl_x (pt, svreinterpret_u64 (svcvt_s64_z (pg, rx)), 63); sr = svreinterpret_f64 (sveor_x (pt, svreinterpret_u64 (sr), odd)); cr = svreinterpret_f64 (sveor_m (cmp, svreinterpret_u64 (cr), odd)); svfloat64_t sinpix = svmul_x ( pt, sv_lw_pw_horner_9_f64_x (pg, sr2, sr4, &(d->c0), &(d->c1)), sr); svfloat64_t cospix = svmul_x ( pt, sv_lw_pw_horner_9_f64_x (pg, cr2, cr4, &(d->c0), &(d->c1)), cr); return svcreate2 (sinpix, cospix); }