1*f3087befSAndrew Turner /* 2*f3087befSAndrew Turner * Single-precision vector cosh(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_expf_inline.h" 9*f3087befSAndrew Turner #include "v_math.h" 10*f3087befSAndrew Turner #include "test_sig.h" 11*f3087befSAndrew Turner #include "test_defs.h" 12*f3087befSAndrew Turner 13*f3087befSAndrew Turner static const struct data 14*f3087befSAndrew Turner { 15*f3087befSAndrew Turner struct v_expf_data expf_consts; 16*f3087befSAndrew Turner uint32x4_t tiny_bound; 17*f3087befSAndrew Turner float32x4_t bound; 18*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT 19*f3087befSAndrew Turner uint32x4_t special_bound; 20*f3087befSAndrew Turner #endif 21*f3087befSAndrew Turner } data = { 22*f3087befSAndrew Turner .expf_consts = V_EXPF_DATA, 23*f3087befSAndrew Turner .tiny_bound = V4 (0x20000000), /* 0x1p-63: Round to 1 below this. */ 24*f3087befSAndrew Turner /* 0x1.5a92d8p+6: expf overflows above this, so have to use special case. */ 25*f3087befSAndrew Turner .bound = V4 (0x1.5a92d8p+6), 26*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT 27*f3087befSAndrew Turner .special_bound = V4 (0x42ad496c), 28*f3087befSAndrew Turner #endif 29*f3087befSAndrew Turner }; 30*f3087befSAndrew Turner 31*f3087befSAndrew Turner #if !WANT_SIMD_EXCEPT 32*f3087befSAndrew Turner static float32x4_t NOINLINE VPCS_ATTR 33*f3087befSAndrew Turner special_case (float32x4_t x, float32x4_t half_t, float32x4_t half_over_t, 34*f3087befSAndrew Turner uint32x4_t special) 35*f3087befSAndrew Turner { 36*f3087befSAndrew Turner return v_call_f32 (coshf, x, vaddq_f32 (half_t, half_over_t), special); 37*f3087befSAndrew Turner } 38*f3087befSAndrew Turner #endif 39*f3087befSAndrew Turner 40*f3087befSAndrew Turner /* Single-precision vector cosh, using vector expf. 41*f3087befSAndrew Turner Maximum error is 2.38 ULP: 42*f3087befSAndrew Turner _ZGVnN4v_coshf (0x1.e8001ep+1) got 0x1.6a491ep+4 43*f3087befSAndrew Turner want 0x1.6a4922p+4. */ 44*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (cosh) (float32x4_t x) 45*f3087befSAndrew Turner { 46*f3087befSAndrew Turner const struct data *d = ptr_barrier (&data); 47*f3087befSAndrew Turner 48*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT 49*f3087befSAndrew Turner /* If fp exceptions are to be triggered correctly, fall back to the scalar 50*f3087befSAndrew Turner variant for all inputs if any input is a special value or above the bound 51*f3087befSAndrew Turner at which expf overflows. */ 52*f3087befSAndrew Turner float32x4_t ax = vabsq_f32 (x); 53*f3087befSAndrew Turner uint32x4_t iax = vreinterpretq_u32_f32 (ax); 54*f3087befSAndrew Turner uint32x4_t special = vcgeq_u32 (iax, d->special_bound); 55*f3087befSAndrew Turner if (unlikely (v_any_u32 (special))) 56*f3087befSAndrew Turner return v_call_f32 (coshf, x, x, v_u32 (-1)); 57*f3087befSAndrew Turner 58*f3087befSAndrew Turner uint32x4_t tiny = vcleq_u32 (iax, d->tiny_bound); 59*f3087befSAndrew Turner /* If any input is tiny, avoid underflow exception by fixing tiny lanes of 60*f3087befSAndrew Turner input to 0, which will generate no exceptions. */ 61*f3087befSAndrew Turner if (unlikely (v_any_u32 (tiny))) 62*f3087befSAndrew Turner ax = v_zerofy_f32 (ax, tiny); 63*f3087befSAndrew Turner float32x4_t t = v_expf_inline (ax, &d->expf_consts); 64*f3087befSAndrew Turner #else 65*f3087befSAndrew Turner uint32x4_t special = vcageq_f32 (x, d->bound); 66*f3087befSAndrew Turner float32x4_t t = v_expf_inline (x, &d->expf_consts); 67*f3087befSAndrew Turner #endif 68*f3087befSAndrew Turner 69*f3087befSAndrew Turner /* Calculate cosh by exp(x) / 2 + exp(-x) / 2. */ 70*f3087befSAndrew Turner float32x4_t half_t = vmulq_n_f32 (t, 0.5); 71*f3087befSAndrew Turner float32x4_t half_over_t = vdivq_f32 (v_f32 (0.5), t); 72*f3087befSAndrew Turner 73*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT 74*f3087befSAndrew Turner if (unlikely (v_any_u32 (tiny))) 75*f3087befSAndrew Turner return vbslq_f32 (tiny, v_f32 (1), vaddq_f32 (half_t, half_over_t)); 76*f3087befSAndrew Turner #else 77*f3087befSAndrew Turner if (unlikely (v_any_u32 (special))) 78*f3087befSAndrew Turner return special_case (x, half_t, half_over_t, special); 79*f3087befSAndrew Turner #endif 80*f3087befSAndrew Turner 81*f3087befSAndrew Turner return vaddq_f32 (half_t, half_over_t); 82*f3087befSAndrew Turner } 83*f3087befSAndrew Turner 84*f3087befSAndrew Turner HALF_WIDTH_ALIAS_F1 (cosh) 85*f3087befSAndrew Turner 86*f3087befSAndrew Turner TEST_SIG (V, F, 1, cosh, -10.0, 10.0) 87*f3087befSAndrew Turner TEST_ULP (V_NAME_F1 (cosh), 1.89) 88*f3087befSAndrew Turner TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (cosh), WANT_SIMD_EXCEPT) 89*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (cosh), 0, 0x1p-63, 100) 90*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (cosh), 0x1p-63, 1, 1000) 91*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (cosh), 1, 0x1.5a92d8p+6, 80000) 92*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (cosh), 0x1.5a92d8p+6, inf, 2000) 93