1 /* 2 * Single-precision vector sin function. 3 * 4 * Copyright (c) 2019, Arm Limited. 5 * SPDX-License-Identifier: MIT 6 */ 7 8 #include "mathlib.h" 9 #include "v_math.h" 10 #if V_SUPPORTED 11 12 static const float Poly[] = { 13 /* 1.886 ulp error */ 14 0x1.5b2e76p-19f, 15 -0x1.9f42eap-13f, 16 0x1.110df4p-7f, 17 -0x1.555548p-3f, 18 }; 19 #define Pi1 v_f32 (0x1.921fb6p+1f) 20 #define Pi2 v_f32 (-0x1.777a5cp-24f) 21 #define Pi3 v_f32 (-0x1.ee59dap-49f) 22 #define A3 v_f32 (Poly[3]) 23 #define A5 v_f32 (Poly[2]) 24 #define A7 v_f32 (Poly[1]) 25 #define A9 v_f32 (Poly[0]) 26 #define RangeVal v_f32 (0x1p20f) 27 #define InvPi v_f32 (0x1.45f306p-2f) 28 #define Shift v_f32 (0x1.8p+23f) 29 #define AbsMask v_u32 (0x7fffffff) 30 31 VPCS_ATTR 32 static v_f32_t 33 specialcase (v_f32_t x, v_f32_t y, v_u32_t cmp) 34 { 35 /* Fall back to scalar code. */ 36 return v_call_f32 (sinf, x, y, cmp); 37 } 38 39 VPCS_ATTR 40 v_f32_t 41 V_NAME(sinf) (v_f32_t x) 42 { 43 v_f32_t n, r, r2, y; 44 v_u32_t sign, odd, cmp; 45 46 r = v_as_f32_u32 (v_as_u32_f32 (x) & AbsMask); 47 sign = v_as_u32_f32 (x) & ~AbsMask; 48 cmp = v_cond_u32 (v_as_u32_f32 (r) >= v_as_u32_f32 (RangeVal)); 49 50 /* n = rint(|x|/pi) */ 51 n = v_fma_f32 (InvPi, r, Shift); 52 odd = v_as_u32_f32 (n) << 31; 53 n -= Shift; 54 55 /* r = |x| - n*pi (range reduction into -pi/2 .. pi/2) */ 56 r = v_fma_f32 (-Pi1, n, r); 57 r = v_fma_f32 (-Pi2, n, r); 58 r = v_fma_f32 (-Pi3, n, r); 59 60 /* y = sin(r) */ 61 r2 = r * r; 62 y = v_fma_f32 (A9, r2, A7); 63 y = v_fma_f32 (y, r2, A5); 64 y = v_fma_f32 (y, r2, A3); 65 y = v_fma_f32 (y * r2, r, r); 66 67 /* sign fix */ 68 y = v_as_f32_u32 (v_as_u32_f32 (y) ^ sign ^ odd); 69 70 if (unlikely (v_any_u32 (cmp))) 71 return specialcase (x, y, cmp); 72 return y; 73 } 74 VPCS_ALIAS 75 #endif 76