xref: /freebsd/contrib/arm-optimized-routines/pl/math/log1pf_2u1.c (revision 911f0260390e18cf85f3dbf2c719b593efdc1e3c)
1 /*
2  * Single-precision log(1+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 "hornerf.h"
9 #include "math_config.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 #define Ln2 (0x1.62e43p-1f)
14 #define SignMask (0x80000000)
15 
16 /* Biased exponent of the largest float m for which m^8 underflows.  */
17 #define M8UFLOW_BOUND_BEXP 112
18 /* Biased exponent of the largest float for which we just return x.  */
19 #define TINY_BOUND_BEXP 103
20 
21 #define C(i) __log1pf_data.coeffs[i]
22 
23 static inline float
24 eval_poly (float m, uint32_t e)
25 {
26 #ifdef LOG1PF_2U5
27 
28   /* 2.5 ulp variant. Approximate log(1+m) on [-0.25, 0.5] using
29      slightly modified Estrin scheme (no x^0 term, and x term is just x).  */
30   float p_12 = fmaf (m, C (1), C (0));
31   float p_34 = fmaf (m, C (3), C (2));
32   float p_56 = fmaf (m, C (5), C (4));
33   float p_78 = fmaf (m, C (7), C (6));
34 
35   float m2 = m * m;
36   float p_02 = fmaf (m2, p_12, m);
37   float p_36 = fmaf (m2, p_56, p_34);
38   float p_79 = fmaf (m2, C (8), p_78);
39 
40   float m4 = m2 * m2;
41   float p_06 = fmaf (m4, p_36, p_02);
42 
43   if (unlikely (e < M8UFLOW_BOUND_BEXP))
44     return p_06;
45 
46   float m8 = m4 * m4;
47   return fmaf (m8, p_79, p_06);
48 
49 #elif defined(LOG1PF_1U3)
50 
51   /* 1.3 ulp variant. Approximate log(1+m) on [-0.25, 0.5] using Horner
52      scheme. Our polynomial approximation for log1p has the form
53      x + C1 * x^2 + C2 * x^3 + C3 * x^4 + ...
54      Hence approximation has the form m + m^2 * P(m)
55        where P(x) = C1 + C2 * x + C3 * x^2 + ... .  */
56   return fmaf (m, m * HORNER_8 (m, C), m);
57 
58 #else
59 #error No log1pf approximation exists with the requested precision. Options are 13 or 25.
60 #endif
61 }
62 
63 static inline uint32_t
64 biased_exponent (uint32_t ix)
65 {
66   return (ix & 0x7f800000) >> 23;
67 }
68 
69 /* log1pf approximation using polynomial on reduced interval. Worst-case error
70    when using Estrin is roughly 2.02 ULP:
71    log1pf(0x1.21e13ap-2) got 0x1.fe8028p-3 want 0x1.fe802cp-3.  */
72 float
73 log1pf (float x)
74 {
75   uint32_t ix = asuint (x);
76   uint32_t ia = ix & ~SignMask;
77   uint32_t ia12 = ia >> 20;
78   uint32_t e = biased_exponent (ix);
79 
80   /* Handle special cases first.  */
81   if (unlikely (ia12 >= 0x7f8 || ix >= 0xbf800000 || ix == 0x80000000
82 		|| e <= TINY_BOUND_BEXP))
83     {
84       if (ix == 0xff800000)
85 	{
86 	  /* x == -Inf => log1pf(x) =  NaN.  */
87 	  return NAN;
88 	}
89       if ((ix == 0x7f800000 || e <= TINY_BOUND_BEXP) && ia12 <= 0x7f8)
90 	{
91 	  /* |x| < TinyBound => log1p(x)  =  x.
92 	      x ==       Inf => log1pf(x) = Inf.  */
93 	  return x;
94 	}
95       if (ix == 0xbf800000)
96 	{
97 	  /* x == -1.0 => log1pf(x) = -Inf.  */
98 	  return __math_divzerof (-1);
99 	}
100       if (ia12 >= 0x7f8)
101 	{
102 	  /* x == +/-NaN => log1pf(x) = NaN.  */
103 	  return __math_invalidf (asfloat (ia));
104 	}
105       /* x <    -1.0 => log1pf(x) = NaN.  */
106       return __math_invalidf (x);
107     }
108 
109   /* With x + 1 = t * 2^k (where t = m + 1 and k is chosen such that m
110 			   is in [-0.25, 0.5]):
111      log1p(x) = log(t) + log(2^k) = log1p(m) + k*log(2).
112 
113      We approximate log1p(m) with a polynomial, then scale by
114      k*log(2). Instead of doing this directly, we use an intermediate
115      scale factor s = 4*k*log(2) to ensure the scale is representable
116      as a normalised fp32 number.  */
117 
118   if (ix <= 0x3f000000 || ia <= 0x3e800000)
119     {
120       /* If x is in [-0.25, 0.5] then we can shortcut all the logic
121 	 below, as k = 0 and m = x.  All we need is to return the
122 	 polynomial.  */
123       return eval_poly (x, e);
124     }
125 
126   float m = x + 1.0f;
127 
128   /* k is used scale the input. 0x3f400000 is chosen as we are trying to
129      reduce x to the range [-0.25, 0.5]. Inside this range, k is 0.
130      Outside this range, if k is reinterpreted as (NOT CONVERTED TO) float:
131 	 let k = sign * 2^p      where sign = -1 if x < 0
132 					       1 otherwise
133 	 and p is a negative integer whose magnitude increases with the
134 	 magnitude of x.  */
135   int k = (asuint (m) - 0x3f400000) & 0xff800000;
136 
137   /* By using integer arithmetic, we obtain the necessary scaling by
138      subtracting the unbiased exponent of k from the exponent of x.  */
139   float m_scale = asfloat (asuint (x) - k);
140 
141   /* Scale up to ensure that the scale factor is representable as normalised
142      fp32 number (s in [2**-126,2**26]), and scale m down accordingly.  */
143   float s = asfloat (asuint (4.0f) - k);
144   m_scale = m_scale + fmaf (0.25f, s, -1.0f);
145 
146   float p = eval_poly (m_scale, biased_exponent (asuint (m_scale)));
147 
148   /* The scale factor to be applied back at the end - by multiplying float(k)
149      by 2^-23 we get the unbiased exponent of k.  */
150   float scale_back = (float) k * 0x1.0p-23f;
151 
152   /* Apply the scaling back.  */
153   return fmaf (scale_back, Ln2, p);
154 }
155 
156 PL_SIG (S, F, 1, log1p, -0.9, 10.0)
157 PL_TEST_ULP (log1pf, 1.52)
158 PL_TEST_INTERVAL (log1pf, -10.0, 10.0, 10000)
159 PL_TEST_INTERVAL (log1pf, 0.0, 0x1p-23, 50000)
160 PL_TEST_INTERVAL (log1pf, 0x1p-23, 0.001, 50000)
161 PL_TEST_INTERVAL (log1pf, 0.001, 1.0, 50000)
162 PL_TEST_INTERVAL (log1pf, 0.0, -0x1p-23, 50000)
163 PL_TEST_INTERVAL (log1pf, -0x1p-23, -0.001, 50000)
164 PL_TEST_INTERVAL (log1pf, -0.001, -1.0, 50000)
165 PL_TEST_INTERVAL (log1pf, -1.0, inf, 5000)
166