1*f3087befSAndrew Turner /* 2*f3087befSAndrew Turner * Single-precision vector cos function. 3*f3087befSAndrew Turner * 4*f3087befSAndrew Turner * Copyright (c) 2019-2024, Arm Limited. 5*f3087befSAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception 6*f3087befSAndrew Turner */ 7*f3087befSAndrew Turner 8*f3087befSAndrew Turner #include "mathlib.h" 9*f3087befSAndrew Turner #include "v_math.h" 10*f3087befSAndrew Turner #include "test_defs.h" 11*f3087befSAndrew Turner #include "test_sig.h" 12*f3087befSAndrew Turner 13*f3087befSAndrew Turner static const struct data 14*f3087befSAndrew Turner { 15*f3087befSAndrew Turner float32x4_t poly[4]; 16*f3087befSAndrew Turner float32x4_t range_val, inv_pi, pi_1, pi_2, pi_3; 17*f3087befSAndrew Turner } data = { 18*f3087befSAndrew Turner /* 1.886 ulp error. */ 19*f3087befSAndrew Turner .poly = { V4 (-0x1.555548p-3f), V4 (0x1.110df4p-7f), V4 (-0x1.9f42eap-13f), 20*f3087befSAndrew Turner V4 (0x1.5b2e76p-19f) }, 21*f3087befSAndrew Turner 22*f3087befSAndrew Turner .pi_1 = V4 (0x1.921fb6p+1f), 23*f3087befSAndrew Turner .pi_2 = V4 (-0x1.777a5cp-24f), 24*f3087befSAndrew Turner .pi_3 = V4 (-0x1.ee59dap-49f), 25*f3087befSAndrew Turner 26*f3087befSAndrew Turner .inv_pi = V4 (0x1.45f306p-2f), 27*f3087befSAndrew Turner .range_val = V4 (0x1p20f) 28*f3087befSAndrew Turner }; 29*f3087befSAndrew Turner 30*f3087befSAndrew Turner #define C(i) d->poly[i] 31*f3087befSAndrew Turner 32*f3087befSAndrew Turner static float32x4_t VPCS_ATTR NOINLINE 33*f3087befSAndrew Turner special_case (float32x4_t x, float32x4_t y, uint32x4_t odd, uint32x4_t cmp) 34*f3087befSAndrew Turner { 35*f3087befSAndrew Turner /* Fall back to scalar code. */ 36*f3087befSAndrew Turner y = vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd)); 37*f3087befSAndrew Turner return v_call_f32 (cosf, x, y, cmp); 38*f3087befSAndrew Turner } 39*f3087befSAndrew Turner 40*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (cos) (float32x4_t x) 41*f3087befSAndrew Turner { 42*f3087befSAndrew Turner const struct data *d = ptr_barrier (&data); 43*f3087befSAndrew Turner float32x4_t n, r, r2, r3, y; 44*f3087befSAndrew Turner uint32x4_t odd, cmp; 45*f3087befSAndrew Turner 46*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT 47*f3087befSAndrew Turner r = vabsq_f32 (x); 48*f3087befSAndrew Turner cmp = vcgeq_u32 (vreinterpretq_u32_f32 (r), 49*f3087befSAndrew Turner vreinterpretq_u32_f32 (d->range_val)); 50*f3087befSAndrew Turner if (unlikely (v_any_u32 (cmp))) 51*f3087befSAndrew Turner /* If fenv exceptions are to be triggered correctly, set any special lanes 52*f3087befSAndrew Turner to 1 (which is neutral w.r.t. fenv). These lanes will be fixed by 53*f3087befSAndrew Turner special-case handler later. */ 54*f3087befSAndrew Turner r = vbslq_f32 (cmp, v_f32 (1.0f), r); 55*f3087befSAndrew Turner #else 56*f3087befSAndrew Turner cmp = vcageq_f32 (x, d->range_val); 57*f3087befSAndrew Turner r = x; 58*f3087befSAndrew Turner #endif 59*f3087befSAndrew Turner 60*f3087befSAndrew Turner /* n = rint((|x|+pi/2)/pi) - 0.5. */ 61*f3087befSAndrew Turner n = vrndaq_f32 (vfmaq_f32 (v_f32 (0.5), r, d->inv_pi)); 62*f3087befSAndrew Turner odd = vshlq_n_u32 (vreinterpretq_u32_s32 (vcvtq_s32_f32 (n)), 31); 63*f3087befSAndrew Turner n = vsubq_f32 (n, v_f32 (0.5f)); 64*f3087befSAndrew Turner 65*f3087befSAndrew Turner /* r = |x| - n*pi (range reduction into -pi/2 .. pi/2). */ 66*f3087befSAndrew Turner r = vfmsq_f32 (r, d->pi_1, n); 67*f3087befSAndrew Turner r = vfmsq_f32 (r, d->pi_2, n); 68*f3087befSAndrew Turner r = vfmsq_f32 (r, d->pi_3, n); 69*f3087befSAndrew Turner 70*f3087befSAndrew Turner /* y = sin(r). */ 71*f3087befSAndrew Turner r2 = vmulq_f32 (r, r); 72*f3087befSAndrew Turner r3 = vmulq_f32 (r2, r); 73*f3087befSAndrew Turner y = vfmaq_f32 (C (2), C (3), r2); 74*f3087befSAndrew Turner y = vfmaq_f32 (C (1), y, r2); 75*f3087befSAndrew Turner y = vfmaq_f32 (C (0), y, r2); 76*f3087befSAndrew Turner y = vfmaq_f32 (r, y, r3); 77*f3087befSAndrew Turner 78*f3087befSAndrew Turner if (unlikely (v_any_u32 (cmp))) 79*f3087befSAndrew Turner return special_case (x, y, odd, cmp); 80*f3087befSAndrew Turner return vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd)); 81*f3087befSAndrew Turner } 82*f3087befSAndrew Turner 83*f3087befSAndrew Turner HALF_WIDTH_ALIAS_F1 (cos) 84*f3087befSAndrew Turner 85*f3087befSAndrew Turner TEST_SIG (V, F, 1, cos, -3.1, 3.1) 86*f3087befSAndrew Turner TEST_ULP (V_NAME_F1 (cos), 1.4) 87*f3087befSAndrew Turner TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (cos), WANT_SIMD_EXCEPT) 88*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (cos), 0, 0x1p20, 500000) 89*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (cos), 0x1p20, inf, 10000) 90