xref: /freebsd/contrib/arm-optimized-routines/pl/math/acoshf_2u8.c (revision 02e9120893770924227138ba49df1edb3896112a)
1 /*
2  * Single-precision acosh(x) function.
3  *
4  * Copyright (c) 2022-2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "math_config.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 
12 #define Ln2 (0x1.62e4p-1f)
13 #define MinusZero 0x80000000
14 #define SquareLim 0x5f800000 /* asuint(0x1p64).  */
15 #define Two 0x40000000
16 
17 /* Single-precision log from math/.  */
18 float
19 optr_aor_log_f32 (float);
20 
21 /* Single-precision log(1+x) from pl/math.  */
22 float
23 log1pf (float);
24 
25 /* acoshf approximation using a variety of approaches on different intervals:
26 
27    x >= 2^64: We cannot square x without overflow. For huge x, sqrt(x*x - 1) is
28    close enough to x that we can calculate the result by ln(2x) == ln(x) +
29    ln(2). The greatest error in the region is 0.94 ULP:
30    acoshf(0x1.15f706p+92) got 0x1.022e14p+6 want 0x1.022e16p+6.
31 
32    x > 2: Calculate the result directly using definition of asinh(x) = ln(x +
33    sqrt(x*x - 1)). Greatest error in this region is 1.30 ULP:
34    acoshf(0x1.249d8p+1) got 0x1.77e1aep+0 want 0x1.77e1bp+0.
35 
36    0 <= x <= 2: Calculate the result using log1p. For x < 1, acosh(x) is
37    undefined. For 1 <= x <= 2, the greatest error is 2.78 ULP:
38    acoshf(0x1.07887p+0) got 0x1.ef9e9cp-3 want 0x1.ef9ea2p-3.  */
39 float
40 acoshf (float x)
41 {
42   uint32_t ix = asuint (x);
43 
44   if (unlikely (ix >= MinusZero))
45     return __math_invalidf (x);
46 
47   if (unlikely (ix >= SquareLim))
48     return optr_aor_log_f32 (x) + Ln2;
49 
50   if (ix > Two)
51     return optr_aor_log_f32 (x + sqrtf (x * x - 1));
52 
53   float xm1 = x - 1;
54   return log1pf (xm1 + sqrtf (2 * xm1 + xm1 * xm1));
55 }
56 
57 PL_SIG (S, F, 1, acosh, 1.0, 10.0)
58 PL_TEST_ULP (acoshf, 2.30)
59 PL_TEST_INTERVAL (acoshf, 0, 1, 100)
60 PL_TEST_INTERVAL (acoshf, 1, 2, 10000)
61 PL_TEST_INTERVAL (acoshf, 2, 0x1p64, 100000)
62 PL_TEST_INTERVAL (acoshf, 0x1p64, inf, 100000)
63 PL_TEST_INTERVAL (acoshf, -0, -inf, 10000)
64