1 /* 2 * Single-precision SVE cospi(x) function. 3 * 4 * Copyright (c) 2023, Arm Limited. 5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception 6 */ 7 8 #include "mathlib.h" 9 #include "sv_math.h" 10 #include "pl_sig.h" 11 #include "pl_test.h" 12 #include "poly_sve_f32.h" 13 14 static const struct data 15 { 16 float poly[6]; 17 float range_val; 18 } data = { 19 /* Taylor series coefficents for sin(pi * x). */ 20 .poly = { 0x1.921fb6p1f, -0x1.4abbcep2f, 0x1.466bc6p1f, -0x1.32d2ccp-1f, 21 0x1.50783p-4f, -0x1.e30750p-8f }, 22 .range_val = 0x1p31f, 23 }; 24 25 /* A fast SVE implementation of cospif. 26 Maximum error: 2.60 ULP: 27 _ZGVsMxv_cospif(+/-0x1.cae664p-4) got 0x1.e09c9ep-1 28 want 0x1.e09c98p-1. */ 29 svfloat32_t SV_NAME_F1 (cospi) (svfloat32_t x, const svbool_t pg) 30 { 31 const struct data *d = ptr_barrier (&data); 32 33 /* Using cospi(x) = sinpi(0.5 - x) 34 range reduction and offset into sinpi range -1/2 .. 1/2 35 r = 0.5 - |x - rint(x)|. */ 36 svfloat32_t n = svrinta_x (pg, x); 37 svfloat32_t r = svsub_x (pg, x, n); 38 r = svsub_x (pg, sv_f32 (0.5f), svabs_x (pg, r)); 39 40 /* Result should be negated based on if n is odd or not. 41 If ax >= 2^31, the result will always be positive. */ 42 svbool_t cmp = svaclt (pg, x, d->range_val); 43 svuint32_t intn = svreinterpret_u32 (svcvt_s32_x (pg, n)); 44 svuint32_t sign = svlsl_z (cmp, intn, 31); 45 46 /* y = sin(r). */ 47 svfloat32_t r2 = svmul_x (pg, r, r); 48 svfloat32_t y = sv_horner_5_f32_x (pg, r2, d->poly); 49 y = svmul_x (pg, y, r); 50 51 return svreinterpret_f32 (sveor_x (pg, svreinterpret_u32 (y), sign)); 52 } 53 54 PL_SIG (SV, F, 1, cospi, -0.9, 0.9) 55 PL_TEST_ULP (SV_NAME_F1 (cospi), 2.08) 56 PL_TEST_SYM_INTERVAL (SV_NAME_F1 (cospi), 0, 0x1p-31, 5000) 57 PL_TEST_SYM_INTERVAL (SV_NAME_F1 (cospi), 0x1p-31, 0.5, 10000) 58 PL_TEST_SYM_INTERVAL (SV_NAME_F1 (cospi), 0.5, 0x1p31f, 10000) 59 PL_TEST_SYM_INTERVAL (SV_NAME_F1 (cospi), 0x1p31f, inf, 10000) 60