/* * Helper for single-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_f32.h" const static struct sv_sincospif_data { float c0, c2, c4; float c1, c3, c5; float range_val; } sv_sincospif_data = { /* Taylor series coefficents for sin(pi * x). */ .c0 = 0x1.921fb6p1f, .c1 = -0x1.4abbcep2f, .c2 = 0x1.466bc6p1f, .c3 = -0x1.32d2ccp-1f, .c4 = 0x1.50783p-4f, .c5 = -0x1.e30750p-8f, /* Exclusive upper bound for a signed integer. */ .range_val = 0x1p31f, }; /* Single-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 3.04 ULP: _ZGVsMxvl4l4_sincospif_sin(0x1.b51b8p-2) got 0x1.f28b5ep-1 want 0x1.f28b58p-1. Worst-case error for cos is 3.18 ULP: _ZGVsMxvl4l4_sincospif_cos(0x1.d341a8p-5) got 0x1.f7cd56p-1 want 0x1.f7cd5p-1. */ static inline svfloat32x2_t sv_sincospif_inline (svbool_t pg, svfloat32_t x, const struct sv_sincospif_data *d) { const svbool_t pt = svptrue_b32 (); /* r = x - rint(x). */ svfloat32_t rx = svrinta_x (pg, x); svfloat32_t sr = svsub_x (pt, x, rx); /* cospi(x) = sinpi(0.5 - abs(r)) for values -1/2 .. 1/2. */ svfloat32_t cr = svsubr_x (pt, svabs_x (pg, sr), 0.5f); /* Pairwise Horner approximation for y = sin(r * pi). */ svfloat32_t sr2 = svmul_x (pt, sr, sr); svfloat32_t sr4 = svmul_x (pt, sr2, sr2); svfloat32_t cr2 = svmul_x (pt, cr, cr); svfloat32_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); svuint32_t odd = svlsl_x (pt, svreinterpret_u32 (svcvt_s32_z (pg, rx)), 31); sr = svreinterpret_f32 (sveor_x (pt, svreinterpret_u32 (sr), odd)); cr = svreinterpret_f32 (sveor_m (cmp, svreinterpret_u32 (cr), odd)); svfloat32_t c135 = svld1rq_f32 (svptrue_b32 (), &d->c1); svfloat32_t sp01 = svmla_lane (sv_f32 (d->c0), sr2, c135, 0); svfloat32_t sp23 = svmla_lane (sv_f32 (d->c2), sr2, c135, 1); svfloat32_t sp45 = svmla_lane (sv_f32 (d->c4), sr2, c135, 2); svfloat32_t cp01 = svmla_lane (sv_f32 (d->c0), cr2, c135, 0); svfloat32_t cp23 = svmla_lane (sv_f32 (d->c2), cr2, c135, 1); svfloat32_t cp45 = svmla_lane (sv_f32 (d->c4), cr2, c135, 2); svfloat32_t sp = svmla_x (pg, sp23, sr4, sp45); svfloat32_t cp = svmla_x (pg, cp23, cr4, cp45); sp = svmla_x (pg, sp01, sr4, sp); cp = svmla_x (pg, cp01, cr4, cp); svfloat32_t sinpix = svmul_x (pt, sp, sr); svfloat32_t cospix = svmul_x (pt, cp, cr); return svcreate2 (sinpix, cospix); }