1*f3087befSAndrew Turner /* 2*f3087befSAndrew Turner * Single-precision vector log function - inline version 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 "v_math.h" 9*f3087befSAndrew Turner 10*f3087befSAndrew Turner struct v_logf_data 11*f3087befSAndrew Turner { 12*f3087befSAndrew Turner float32x4_t poly[7]; 13*f3087befSAndrew Turner float32x4_t ln2; 14*f3087befSAndrew Turner uint32x4_t off, mantissa_mask; 15*f3087befSAndrew Turner }; 16*f3087befSAndrew Turner 17*f3087befSAndrew Turner #define V_LOGF_CONSTANTS \ 18*f3087befSAndrew Turner { \ 19*f3087befSAndrew Turner .poly \ 20*f3087befSAndrew Turner = { V4 (-0x1.3e737cp-3f), V4 (0x1.5a9aa2p-3f), V4 (-0x1.4f9934p-3f), \ 21*f3087befSAndrew Turner V4 (0x1.961348p-3f), V4 (-0x1.00187cp-2f), V4 (0x1.555d7cp-2f), \ 22*f3087befSAndrew Turner V4 (-0x1.ffffc8p-2f) }, \ 23*f3087befSAndrew Turner .ln2 = V4 (0x1.62e43p-1f), .off = V4 (0x3f2aaaab), \ 24*f3087befSAndrew Turner .mantissa_mask = V4 (0x007fffff) \ 25*f3087befSAndrew Turner } 26*f3087befSAndrew Turner 27*f3087befSAndrew Turner #define P(i) d->poly[7 - i] 28*f3087befSAndrew Turner 29*f3087befSAndrew Turner static inline float32x4_t 30*f3087befSAndrew Turner v_logf_inline (float32x4_t x, const struct v_logf_data *d) 31*f3087befSAndrew Turner { 32*f3087befSAndrew Turner float32x4_t n, p, q, r, r2, y; 33*f3087befSAndrew Turner uint32x4_t u; 34*f3087befSAndrew Turner 35*f3087befSAndrew Turner u = vreinterpretq_u32_f32 (x); 36*f3087befSAndrew Turner 37*f3087befSAndrew Turner /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3. */ 38*f3087befSAndrew Turner u = vsubq_u32 (u, d->off); 39*f3087befSAndrew Turner n = vcvtq_f32_s32 ( 40*f3087befSAndrew Turner vshrq_n_s32 (vreinterpretq_s32_u32 (u), 23)); /* signextend. */ 41*f3087befSAndrew Turner u = vandq_u32 (u, d->mantissa_mask); 42*f3087befSAndrew Turner u = vaddq_u32 (u, d->off); 43*f3087befSAndrew Turner r = vsubq_f32 (vreinterpretq_f32_u32 (u), v_f32 (1.0f)); 44*f3087befSAndrew Turner 45*f3087befSAndrew Turner /* y = log(1+r) + n*ln2. */ 46*f3087befSAndrew Turner r2 = vmulq_f32 (r, r); 47*f3087befSAndrew Turner /* n*ln2 + r + r2*(P1 + r*P2 + r2*(P3 + r*P4 + r2*(P5 + r*P6 + r2*P7))). */ 48*f3087befSAndrew Turner p = vfmaq_f32 (P (5), P (6), r); 49*f3087befSAndrew Turner q = vfmaq_f32 (P (3), P (4), r); 50*f3087befSAndrew Turner y = vfmaq_f32 (P (1), P (2), r); 51*f3087befSAndrew Turner p = vfmaq_f32 (p, P (7), r2); 52*f3087befSAndrew Turner q = vfmaq_f32 (q, p, r2); 53*f3087befSAndrew Turner y = vfmaq_f32 (y, q, r2); 54*f3087befSAndrew Turner p = vfmaq_f32 (r, d->ln2, n); 55*f3087befSAndrew Turner 56*f3087befSAndrew Turner return vfmaq_f32 (p, y, r2); 57*f3087befSAndrew Turner } 58*f3087befSAndrew Turner 59*f3087befSAndrew Turner #undef P 60