xref: /aosp_15_r20/external/arm-optimized-routines/math/tools/exp.sollya (revision 412f47f9e737e10ed5cc46ec6a8d7fa2264f8a14)
1*412f47f9SXin Li// polynomial for approximating e^x
2*412f47f9SXin Li//
3*412f47f9SXin Li// Copyright (c) 2019, Arm Limited.
4*412f47f9SXin Li// SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
5*412f47f9SXin Li
6*412f47f9SXin Lideg = 5; // poly degree
7*412f47f9SXin LiN = 128; // table entries
8*412f47f9SXin Lib = log(2)/(2*N);  // interval
9*412f47f9SXin Lib = b + b*0x1p-16; // increase interval for non-nearest rounding (TOINT_NARROW)
10*412f47f9SXin Lia = -b;
11*412f47f9SXin Li
12*412f47f9SXin Li// find polynomial with minimal abs error
13*412f47f9SXin Li
14*412f47f9SXin Li// return p that minimizes |exp(x) - poly(x) - x^d*p(x)|
15*412f47f9SXin Liapprox = proc(poly,d) {
16*412f47f9SXin Li  return remez(exp(x)-poly(x), deg-d, [a;b], x^d, 1e-10);
17*412f47f9SXin Li};
18*412f47f9SXin Li
19*412f47f9SXin Li// first 2 coeffs are fixed, iteratively find optimal double prec coeffs
20*412f47f9SXin Lipoly = 1 + x;
21*412f47f9SXin Lifor i from 2 to deg do {
22*412f47f9SXin Li  p = roundcoefficients(approx(poly,i), [|D ...|]);
23*412f47f9SXin Li  poly = poly + x^i*coeff(p,0);
24*412f47f9SXin Li};
25*412f47f9SXin Li
26*412f47f9SXin Lidisplay = hexadecimal;
27*412f47f9SXin Liprint("rel error:", accurateinfnorm(1-poly(x)/exp(x), [a;b], 30));
28*412f47f9SXin Liprint("abs error:", accurateinfnorm(exp(x)-poly(x), [a;b], 30));
29*412f47f9SXin Liprint("in [",a,b,"]");
30*412f47f9SXin Li// double interval error for non-nearest rounding
31*412f47f9SXin Liprint("rel2 error:", accurateinfnorm(1-poly(x)/exp(x), [2*a;2*b], 30));
32*412f47f9SXin Liprint("abs2 error:", accurateinfnorm(exp(x)-poly(x), [2*a;2*b], 30));
33*412f47f9SXin Liprint("in [",2*a,2*b,"]");
34*412f47f9SXin Liprint("coeffs:");
35*412f47f9SXin Lifor i from 0 to deg do coeff(poly,i);
36