1*412f47f9SXin Li// polynomial used for __v_log2f(x) 2*412f47f9SXin Li// 3*412f47f9SXin Li// Copyright (c) 2022-2023, Arm Limited. 4*412f47f9SXin Li// SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception 5*412f47f9SXin Li 6*412f47f9SXin Lideg = 9; // poly degree 7*412f47f9SXin Lia = -1/3; 8*412f47f9SXin Lib = 1/3; 9*412f47f9SXin Li 10*412f47f9SXin Liln2 = evaluate(log(2),0); 11*412f47f9SXin Liinvln2 = single(1/ln2); 12*412f47f9SXin Li 13*412f47f9SXin Li// find log2(1+x)/x polynomial with minimal relative error 14*412f47f9SXin Li// (minimal relative error polynomial for log2(1+x) is the same * x) 15*412f47f9SXin Lideg = deg-1; // because of /x 16*412f47f9SXin Li 17*412f47f9SXin Li// f = log2(1+x)/x; using taylor series 18*412f47f9SXin Lif = 0; 19*412f47f9SXin Lifor i from 0 to 60 do { f = f + (-x)^i/(i+1); }; 20*412f47f9SXin Lif = f * invln2; 21*412f47f9SXin Li 22*412f47f9SXin Li// return p that minimizes |f(x) - poly(x) - x^d*p(x)|/|f(x)| 23*412f47f9SXin Liapprox = proc(poly,d) { 24*412f47f9SXin Li return remez(1 - poly(x)/f(x), deg-d, [a;b], x^d/f(x), 1e-10); 25*412f47f9SXin Li}; 26*412f47f9SXin Li 27*412f47f9SXin Li// first coeff is fixed, iteratively find optimal double prec coeffs 28*412f47f9SXin Lipoly = invln2; 29*412f47f9SXin Lifor i from 1 to deg do { 30*412f47f9SXin Li p = roundcoefficients(approx(poly,i), [|SG ...|]); 31*412f47f9SXin Li poly = poly + x^i*coeff(p,0); 32*412f47f9SXin Li}; 33*412f47f9SXin Li 34*412f47f9SXin Lidisplay = hexadecimal; 35*412f47f9SXin Liprint("rel error:", accurateinfnorm(1-poly(x)/f(x), [a;b], 30)); 36*412f47f9SXin Liprint("in [",a,b,"]"); 37*412f47f9SXin Liprint("coeffs:"); 38*412f47f9SXin Lifor i from 0 to deg do coeff(poly,i); 39