xref: /freebsd/contrib/arm-optimized-routines/pl/math/atan2_2u5.c (revision a03411e84728e9b267056fd31c7d1d9d1dc1b01e)
1 /*
2  * Double-precision scalar atan2(x) function.
3  *
4  * Copyright (c) 2021-2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include <stdbool.h>
9 
10 #include "atan_common.h"
11 #include "math_config.h"
12 #include "pl_sig.h"
13 #include "pl_test.h"
14 
15 #define Pi (0x1.921fb54442d18p+1)
16 #define PiOver2 (0x1.921fb54442d18p+0)
17 #define PiOver4 (0x1.921fb54442d18p-1)
18 #define SignMask (0x8000000000000000)
19 #define ExpMask (0x7ff0000000000000)
20 
21 /* We calculate atan2 by P(n/d), where n and d are similar to the input
22    arguments, and P is a polynomial. Evaluating P(x) requires calculating x^8,
23    which may underflow if n and d have very different magnitude.
24    POW8_EXP_UFLOW_BOUND is the lower bound of the difference in exponents of n
25    and d for which P underflows, and is used to special-case such inputs.  */
26 #define POW8_EXP_UFLOW_BOUND 62
27 
28 static inline int64_t
29 biased_exponent (double f)
30 {
31   uint64_t fi = asuint64 (f);
32   return (fi & ExpMask) >> 52;
33 }
34 
35 /* Fast implementation of scalar atan2. Largest errors are when y and x are
36    close together. The greatest observed error is 2.28 ULP:
37    atan2(-0x1.5915b1498e82fp+732, 0x1.54d11ef838826p+732)
38    got -0x1.954f42f1fa841p-1 want -0x1.954f42f1fa843p-1.  */
39 double
40 atan2 (double y, double x)
41 {
42   uint64_t ix = asuint64 (x);
43   uint64_t iy = asuint64 (y);
44 
45   uint64_t sign_x = ix & SignMask;
46   uint64_t sign_y = iy & SignMask;
47 
48   uint64_t iax = ix & ~SignMask;
49   uint64_t iay = iy & ~SignMask;
50 
51   bool xisnan = isnan (x);
52   if (unlikely (isnan (y) && !xisnan))
53     return __math_invalid (y);
54   if (unlikely (xisnan))
55     return __math_invalid (x);
56 
57   /* m = 2 * sign(x) + sign(y).  */
58   uint32_t m = ((iy >> 63) & 1) | ((ix >> 62) & 2);
59 
60   int64_t exp_diff = biased_exponent (x) - biased_exponent (y);
61 
62   /* y = 0.  */
63   if (iay == 0)
64     {
65       switch (m)
66 	{
67 	case 0:
68 	case 1:
69 	  return y; /* atan(+-0,+anything)=+-0.  */
70 	case 2:
71 	  return Pi; /* atan(+0,-anything) = pi.  */
72 	case 3:
73 	  return -Pi; /* atan(-0,-anything) =-pi.  */
74 	}
75     }
76   /* Special case for (x, y) either on or very close to the y axis. Either x =
77      0, or y is much larger than x (difference in exponents >=
78      POW8_EXP_UFLOW_BOUND).  */
79   if (unlikely (iax == 0 || exp_diff <= -POW8_EXP_UFLOW_BOUND))
80     return sign_y ? -PiOver2 : PiOver2;
81 
82   /* Special case for either x is INF or (x, y) is very close to x axis and x is
83      negative.  */
84   if (unlikely (iax == 0x7ff0000000000000
85 		|| (exp_diff >= POW8_EXP_UFLOW_BOUND && m >= 2)))
86     {
87       if (iay == 0x7ff0000000000000)
88 	{
89 	  switch (m)
90 	    {
91 	    case 0:
92 	      return PiOver4; /* atan(+INF,+INF).  */
93 	    case 1:
94 	      return -PiOver4; /* atan(-INF,+INF).  */
95 	    case 2:
96 	      return 3.0 * PiOver4; /* atan(+INF,-INF).  */
97 	    case 3:
98 	      return -3.0 * PiOver4; /* atan(-INF,-INF).  */
99 	    }
100 	}
101       else
102 	{
103 	  switch (m)
104 	    {
105 	    case 0:
106 	      return 0.0; /* atan(+...,+INF).  */
107 	    case 1:
108 	      return -0.0; /* atan(-...,+INF).  */
109 	    case 2:
110 	      return Pi; /* atan(+...,-INF).  */
111 	    case 3:
112 	      return -Pi; /* atan(-...,-INF).  */
113 	    }
114 	}
115     }
116   /* y is INF.  */
117   if (iay == 0x7ff0000000000000)
118     return sign_y ? -PiOver2 : PiOver2;
119 
120   uint64_t sign_xy = sign_x ^ sign_y;
121 
122   double ax = asdouble (iax);
123   double ay = asdouble (iay);
124   uint64_t pred_aygtax = (ay > ax);
125 
126   /* Set up z for call to atan.  */
127   double n = pred_aygtax ? -ax : ay;
128   double d = pred_aygtax ? ay : ax;
129   double z = n / d;
130 
131   double ret;
132   if (unlikely (m < 2 && exp_diff >= POW8_EXP_UFLOW_BOUND))
133     {
134       /* If (x, y) is very close to x axis and x is positive, the polynomial
135 	 will underflow and evaluate to z.  */
136       ret = z;
137     }
138   else
139     {
140       /* Work out the correct shift.  */
141       double shift = sign_x ? -2.0 : 0.0;
142       shift = pred_aygtax ? shift + 1.0 : shift;
143       shift *= PiOver2;
144 
145       ret = eval_poly (z, z, shift);
146     }
147 
148   /* Account for the sign of x and y.  */
149   return asdouble (asuint64 (ret) ^ sign_xy);
150 }
151 
152 /* Arity of 2 means no mathbench entry emitted. See test/mathbench_funcs.h.  */
153 PL_SIG (S, D, 2, atan2)
154 PL_TEST_ULP (atan2, 1.78)
155 PL_TEST_INTERVAL (atan2, -10.0, 10.0, 50000)
156 PL_TEST_INTERVAL (atan2, -1.0, 1.0, 40000)
157 PL_TEST_INTERVAL (atan2, 0.0, 1.0, 40000)
158 PL_TEST_INTERVAL (atan2, 1.0, 100.0, 40000)
159 PL_TEST_INTERVAL (atan2, 1e6, 1e32, 40000)
160