xref: /freebsd/contrib/arm-optimized-routines/pl/math/atan2f_3u.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 /*
2  * Single-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 "atanf_common.h"
11 #include "math_config.h"
12 #include "pl_sig.h"
13 #include "pl_test.h"
14 
15 #define Pi (0x1.921fb6p+1f)
16 #define PiOver2 (0x1.921fb6p+0f)
17 #define PiOver4 (0x1.921fb6p-1f)
18 #define SignMask (0x80000000)
19 
20 /* We calculate atan2f by P(n/d), where n and d are similar to the input
21    arguments, and P is a polynomial. The polynomial may underflow.
22    POLY_UFLOW_BOUND is the lower bound of the difference in exponents of n and d
23    for which P underflows, and is used to special-case such inputs.  */
24 #define POLY_UFLOW_BOUND 24
25 
26 static inline int32_t
27 biased_exponent (float f)
28 {
29   uint32_t fi = asuint (f);
30   int32_t ex = (int32_t) ((fi & 0x7f800000) >> 23);
31   if (unlikely (ex == 0))
32     {
33       /* Subnormal case - we still need to get the exponent right for subnormal
34 	 numbers as division may take us back inside the normal range.  */
35       return ex - __builtin_clz (fi << 9);
36     }
37   return ex;
38 }
39 
40 /* Fast implementation of scalar atan2f. Largest observed error is
41    2.88ulps in [99.0, 101.0] x [99.0, 101.0]:
42    atan2f(0x1.9332d8p+6, 0x1.8cb6c4p+6) got 0x1.964646p-1
43 				       want 0x1.964640p-1.  */
44 float
45 atan2f (float y, float x)
46 {
47   uint32_t ix = asuint (x);
48   uint32_t iy = asuint (y);
49 
50   uint32_t sign_x = ix & SignMask;
51   uint32_t sign_y = iy & SignMask;
52 
53   uint32_t iax = ix & ~SignMask;
54   uint32_t iay = iy & ~SignMask;
55 
56   /* x or y is NaN.  */
57   if ((iax > 0x7f800000) || (iay > 0x7f800000))
58     return x + y;
59 
60   /* m = 2 * sign(x) + sign(y).  */
61   uint32_t m = ((iy >> 31) & 1) | ((ix >> 30) & 2);
62 
63   /* The following follows glibc ieee754 implementation, except
64      that we do not use +-tiny shifts (non-nearest rounding mode).  */
65 
66   int32_t exp_diff = biased_exponent (x) - biased_exponent (y);
67 
68   /* Special case for (x, y) either on or very close to the x axis. Either y =
69      0, or y is tiny and x is huge (difference in exponents >=
70      POLY_UFLOW_BOUND). In the second case, we only want to use this special
71      case when x is negative (i.e. quadrants 2 or 3).  */
72   if (unlikely (iay == 0 || (exp_diff >= POLY_UFLOW_BOUND && m >= 2)))
73     {
74       switch (m)
75 	{
76 	case 0:
77 	case 1:
78 	  return y; /* atan(+-0,+anything)=+-0.  */
79 	case 2:
80 	  return Pi; /* atan(+0,-anything) = pi.  */
81 	case 3:
82 	  return -Pi; /* atan(-0,-anything) =-pi.  */
83 	}
84     }
85   /* Special case for (x, y) either on or very close to the y axis. Either x =
86      0, or x is tiny and y is huge (difference in exponents >=
87      POLY_UFLOW_BOUND).  */
88   if (unlikely (iax == 0 || exp_diff <= -POLY_UFLOW_BOUND))
89     return sign_y ? -PiOver2 : PiOver2;
90 
91   /* x is INF.  */
92   if (iax == 0x7f800000)
93     {
94       if (iay == 0x7f800000)
95 	{
96 	  switch (m)
97 	    {
98 	    case 0:
99 	      return PiOver4; /* atan(+INF,+INF).  */
100 	    case 1:
101 	      return -PiOver4; /* atan(-INF,+INF).  */
102 	    case 2:
103 	      return 3.0f * PiOver4; /* atan(+INF,-INF).  */
104 	    case 3:
105 	      return -3.0f * PiOver4; /* atan(-INF,-INF).  */
106 	    }
107 	}
108       else
109 	{
110 	  switch (m)
111 	    {
112 	    case 0:
113 	      return 0.0f; /* atan(+...,+INF).  */
114 	    case 1:
115 	      return -0.0f; /* atan(-...,+INF).  */
116 	    case 2:
117 	      return Pi; /* atan(+...,-INF).  */
118 	    case 3:
119 	      return -Pi; /* atan(-...,-INF).  */
120 	    }
121 	}
122     }
123   /* y is INF.  */
124   if (iay == 0x7f800000)
125     return sign_y ? -PiOver2 : PiOver2;
126 
127   uint32_t sign_xy = sign_x ^ sign_y;
128 
129   float ax = asfloat (iax);
130   float ay = asfloat (iay);
131 
132   bool pred_aygtax = (ay > ax);
133 
134   /* Set up z for call to atanf.  */
135   float n = pred_aygtax ? -ax : ay;
136   float d = pred_aygtax ? ay : ax;
137   float z = n / d;
138 
139   float ret;
140   if (unlikely (m < 2 && exp_diff >= POLY_UFLOW_BOUND))
141     {
142       /* If (x, y) is very close to x axis and x is positive, the polynomial
143 	 will underflow and evaluate to z.  */
144       ret = z;
145     }
146   else
147     {
148       /* Work out the correct shift.  */
149       float shift = sign_x ? -2.0f : 0.0f;
150       shift = pred_aygtax ? shift + 1.0f : shift;
151       shift *= PiOver2;
152 
153       ret = eval_poly (z, z, shift);
154     }
155 
156   /* Account for the sign of x and y.  */
157   return asfloat (asuint (ret) ^ sign_xy);
158 }
159 
160 /* Arity of 2 means no mathbench entry emitted. See test/mathbench_funcs.h.  */
161 PL_SIG (S, F, 2, atan2)
162 PL_TEST_ULP (atan2f, 2.4)
163 PL_TEST_INTERVAL (atan2f, -10.0, 10.0, 50000)
164 PL_TEST_INTERVAL (atan2f, -1.0, 1.0, 40000)
165 PL_TEST_INTERVAL (atan2f, 0.0, 1.0, 40000)
166 PL_TEST_INTERVAL (atan2f, 1.0, 100.0, 40000)
167 PL_TEST_INTERVAL (atan2f, 1e6, 1e32, 40000)
168