1*f3087befSAndrew Turner /* 2*f3087befSAndrew Turner * Single-precision vector log(1+x) function. 3*f3087befSAndrew Turner * 4*f3087befSAndrew Turner * Copyright (c) 2022-2024, 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 "v_math.h" 9*f3087befSAndrew Turner #include "test_sig.h" 10*f3087befSAndrew Turner #include "test_defs.h" 11*f3087befSAndrew Turner #include "v_log1pf_inline.h" 12*f3087befSAndrew Turner 13*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT 14*f3087befSAndrew Turner 15*f3087befSAndrew Turner const static struct data 16*f3087befSAndrew Turner { 17*f3087befSAndrew Turner uint32x4_t minus_one, thresh; 18*f3087befSAndrew Turner struct v_log1pf_data d; 19*f3087befSAndrew Turner } data = { 20*f3087befSAndrew Turner .d = V_LOG1PF_CONSTANTS_TABLE, 21*f3087befSAndrew Turner .thresh = V4 (0x4b800000), /* asuint32(INFINITY) - TinyBound. */ 22*f3087befSAndrew Turner .minus_one = V4 (0xbf800000), 23*f3087befSAndrew Turner }; 24*f3087befSAndrew Turner 25*f3087befSAndrew Turner /* asuint32(0x1p-23). ulp=0.5 at 0x1p-23. */ 26*f3087befSAndrew Turner # define TinyBound v_u32 (0x34000000) 27*f3087befSAndrew Turner 28*f3087befSAndrew Turner static float32x4_t NOINLINE VPCS_ATTR 29*f3087befSAndrew Turner special_case (float32x4_t x, uint32x4_t cmp, const struct data *d) 30*f3087befSAndrew Turner { 31*f3087befSAndrew Turner /* Side-step special lanes so fenv exceptions are not triggered 32*f3087befSAndrew Turner inadvertently. */ 33*f3087befSAndrew Turner float32x4_t x_nospecial = v_zerofy_f32 (x, cmp); 34*f3087befSAndrew Turner return v_call_f32 (log1pf, x, log1pf_inline (x_nospecial, &d->d), cmp); 35*f3087befSAndrew Turner } 36*f3087befSAndrew Turner 37*f3087befSAndrew Turner /* Vector log1pf approximation using polynomial on reduced interval. Worst-case 38*f3087befSAndrew Turner error is 1.69 ULP: 39*f3087befSAndrew Turner _ZGVnN4v_log1pf(0x1.04418ap-2) got 0x1.cfcbd8p-3 40*f3087befSAndrew Turner want 0x1.cfcbdcp-3. */ 41*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (log1p) (float32x4_t x) 42*f3087befSAndrew Turner { 43*f3087befSAndrew Turner const struct data *d = ptr_barrier (&data); 44*f3087befSAndrew Turner uint32x4_t ix = vreinterpretq_u32_f32 (x); 45*f3087befSAndrew Turner uint32x4_t ia = vreinterpretq_u32_f32 (vabsq_f32 (x)); 46*f3087befSAndrew Turner 47*f3087befSAndrew Turner uint32x4_t special_cases 48*f3087befSAndrew Turner = vorrq_u32 (vcgeq_u32 (vsubq_u32 (ia, TinyBound), d->thresh), 49*f3087befSAndrew Turner vcgeq_u32 (ix, d->minus_one)); 50*f3087befSAndrew Turner 51*f3087befSAndrew Turner if (unlikely (v_any_u32 (special_cases))) 52*f3087befSAndrew Turner return special_case (x, special_cases, d); 53*f3087befSAndrew Turner 54*f3087befSAndrew Turner return log1pf_inline (x, &d->d); 55*f3087befSAndrew Turner } 56*f3087befSAndrew Turner 57*f3087befSAndrew Turner #else 58*f3087befSAndrew Turner 59*f3087befSAndrew Turner const static struct v_log1pf_data data = V_LOG1PF_CONSTANTS_TABLE; 60*f3087befSAndrew Turner 61*f3087befSAndrew Turner static float32x4_t NOINLINE VPCS_ATTR 62*f3087befSAndrew Turner special_case (float32x4_t x, uint32x4_t cmp) 63*f3087befSAndrew Turner { 64*f3087befSAndrew Turner return v_call_f32 (log1pf, x, log1pf_inline (x, ptr_barrier (&data)), cmp); 65*f3087befSAndrew Turner } 66*f3087befSAndrew Turner 67*f3087befSAndrew Turner /* Vector log1pf approximation using polynomial on reduced interval. Worst-case 68*f3087befSAndrew Turner error is 1.63 ULP: 69*f3087befSAndrew Turner _ZGVnN4v_log1pf(0x1.216d12p-2) got 0x1.fdcb12p-3 70*f3087befSAndrew Turner want 0x1.fdcb16p-3. */ 71*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (log1p) (float32x4_t x) 72*f3087befSAndrew Turner { 73*f3087befSAndrew Turner uint32x4_t special_cases = vornq_u32 (vcleq_f32 (x, v_f32 (-1)), 74*f3087befSAndrew Turner vcaleq_f32 (x, v_f32 (0x1p127f))); 75*f3087befSAndrew Turner 76*f3087befSAndrew Turner if (unlikely (v_any_u32 (special_cases))) 77*f3087befSAndrew Turner return special_case (x, special_cases); 78*f3087befSAndrew Turner 79*f3087befSAndrew Turner return log1pf_inline (x, ptr_barrier (&data)); 80*f3087befSAndrew Turner } 81*f3087befSAndrew Turner 82*f3087befSAndrew Turner #endif 83*f3087befSAndrew Turner 84*f3087befSAndrew Turner HALF_WIDTH_ALIAS_F1 (log1p) 85*f3087befSAndrew Turner 86*f3087befSAndrew Turner TEST_SIG (V, F, 1, log1p, -0.9, 10.0) 87*f3087befSAndrew Turner TEST_ULP (V_NAME_F1 (log1p), 1.20) 88*f3087befSAndrew Turner TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (log1p), WANT_SIMD_EXCEPT) 89*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (log1p), 0.0, 0x1p-23, 30000) 90*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (log1p), 0x1p-23, 1, 50000) 91*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log1p), 1, inf, 50000) 92*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log1p), -1.0, -inf, 1000) 93