xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/experimental/atan2f_3u.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner  * Single-precision scalar atan2(x) function.
3*f3087befSAndrew Turner  *
4*f3087befSAndrew Turner  * Copyright (c) 2021-2024, Arm Limited.
5*f3087befSAndrew Turner  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*f3087befSAndrew Turner  */
7*f3087befSAndrew Turner 
8*f3087befSAndrew Turner #include <stdbool.h>
9*f3087befSAndrew Turner 
10*f3087befSAndrew Turner #include "atanf_common.h"
11*f3087befSAndrew Turner #include "math_config.h"
12*f3087befSAndrew Turner #include "test_sig.h"
13*f3087befSAndrew Turner #include "test_defs.h"
14*f3087befSAndrew Turner 
15*f3087befSAndrew Turner #define Pi (0x1.921fb6p+1f)
16*f3087befSAndrew Turner #define PiOver2 (0x1.921fb6p+0f)
17*f3087befSAndrew Turner #define PiOver4 (0x1.921fb6p-1f)
18*f3087befSAndrew Turner #define SignMask (0x80000000)
19*f3087befSAndrew Turner 
20*f3087befSAndrew Turner /* We calculate atan2f by P(n/d), where n and d are similar to the input
21*f3087befSAndrew Turner    arguments, and P is a polynomial. The polynomial may underflow.
22*f3087befSAndrew Turner    POLY_UFLOW_BOUND is the lower bound of the difference in exponents of n and
23*f3087befSAndrew Turner    d for which P underflows, and is used to special-case such inputs.  */
24*f3087befSAndrew Turner #define POLY_UFLOW_BOUND 24
25*f3087befSAndrew Turner 
26*f3087befSAndrew Turner static inline int32_t
biased_exponent(float f)27*f3087befSAndrew Turner biased_exponent (float f)
28*f3087befSAndrew Turner {
29*f3087befSAndrew Turner   uint32_t fi = asuint (f);
30*f3087befSAndrew Turner   int32_t ex = (int32_t) ((fi & 0x7f800000) >> 23);
31*f3087befSAndrew Turner   if (unlikely (ex == 0))
32*f3087befSAndrew Turner     {
33*f3087befSAndrew Turner       /* Subnormal case - we still need to get the exponent right for subnormal
34*f3087befSAndrew Turner 	 numbers as division may take us back inside the normal range.  */
35*f3087befSAndrew Turner       return ex - __builtin_clz (fi << 9);
36*f3087befSAndrew Turner     }
37*f3087befSAndrew Turner   return ex;
38*f3087befSAndrew Turner }
39*f3087befSAndrew Turner 
40*f3087befSAndrew Turner /* Fast implementation of scalar atan2f. Largest observed error is
41*f3087befSAndrew Turner    2.88ulps in [99.0, 101.0] x [99.0, 101.0]:
42*f3087befSAndrew Turner    atan2f(0x1.9332d8p+6, 0x1.8cb6c4p+6) got 0x1.964646p-1
43*f3087befSAndrew Turner 				       want 0x1.964640p-1.  */
44*f3087befSAndrew Turner float
atan2f(float y,float x)45*f3087befSAndrew Turner atan2f (float y, float x)
46*f3087befSAndrew Turner {
47*f3087befSAndrew Turner   uint32_t ix = asuint (x);
48*f3087befSAndrew Turner   uint32_t iy = asuint (y);
49*f3087befSAndrew Turner 
50*f3087befSAndrew Turner   uint32_t sign_x = ix & SignMask;
51*f3087befSAndrew Turner   uint32_t sign_y = iy & SignMask;
52*f3087befSAndrew Turner 
53*f3087befSAndrew Turner   uint32_t iax = ix & ~SignMask;
54*f3087befSAndrew Turner   uint32_t iay = iy & ~SignMask;
55*f3087befSAndrew Turner 
56*f3087befSAndrew Turner   /* x or y is NaN.  */
57*f3087befSAndrew Turner   if ((iax > 0x7f800000) || (iay > 0x7f800000))
58*f3087befSAndrew Turner     return x + y;
59*f3087befSAndrew Turner 
60*f3087befSAndrew Turner   /* m = 2 * sign(x) + sign(y).  */
61*f3087befSAndrew Turner   uint32_t m = ((iy >> 31) & 1) | ((ix >> 30) & 2);
62*f3087befSAndrew Turner 
63*f3087befSAndrew Turner   /* The following follows glibc ieee754 implementation, except
64*f3087befSAndrew Turner      that we do not use +-tiny shifts (non-nearest rounding mode).  */
65*f3087befSAndrew Turner 
66*f3087befSAndrew Turner   int32_t exp_diff = biased_exponent (x) - biased_exponent (y);
67*f3087befSAndrew Turner 
68*f3087befSAndrew Turner   /* Special case for (x, y) either on or very close to the x axis. Either y =
69*f3087befSAndrew Turner      0, or y is tiny and x is huge (difference in exponents >=
70*f3087befSAndrew Turner      POLY_UFLOW_BOUND). In the second case, we only want to use this special
71*f3087befSAndrew Turner      case when x is negative (i.e. quadrants 2 or 3).  */
72*f3087befSAndrew Turner   if (unlikely (iay == 0 || (exp_diff >= POLY_UFLOW_BOUND && m >= 2)))
73*f3087befSAndrew Turner     {
74*f3087befSAndrew Turner       switch (m)
75*f3087befSAndrew Turner 	{
76*f3087befSAndrew Turner 	case 0:
77*f3087befSAndrew Turner 	case 1:
78*f3087befSAndrew Turner 	  return y; /* atan(+-0,+anything)=+-0.  */
79*f3087befSAndrew Turner 	case 2:
80*f3087befSAndrew Turner 	  return Pi; /* atan(+0,-anything) = pi.  */
81*f3087befSAndrew Turner 	case 3:
82*f3087befSAndrew Turner 	  return -Pi; /* atan(-0,-anything) =-pi.  */
83*f3087befSAndrew Turner 	}
84*f3087befSAndrew Turner     }
85*f3087befSAndrew Turner   /* Special case for (x, y) either on or very close to the y axis. Either x =
86*f3087befSAndrew Turner      0, or x is tiny and y is huge (difference in exponents >=
87*f3087befSAndrew Turner      POLY_UFLOW_BOUND).  */
88*f3087befSAndrew Turner   if (unlikely (iax == 0 || exp_diff <= -POLY_UFLOW_BOUND))
89*f3087befSAndrew Turner     return sign_y ? -PiOver2 : PiOver2;
90*f3087befSAndrew Turner 
91*f3087befSAndrew Turner   /* x is INF.  */
92*f3087befSAndrew Turner   if (iax == 0x7f800000)
93*f3087befSAndrew Turner     {
94*f3087befSAndrew Turner       if (iay == 0x7f800000)
95*f3087befSAndrew Turner 	{
96*f3087befSAndrew Turner 	  switch (m)
97*f3087befSAndrew Turner 	    {
98*f3087befSAndrew Turner 	    case 0:
99*f3087befSAndrew Turner 	      return PiOver4; /* atan(+INF,+INF).  */
100*f3087befSAndrew Turner 	    case 1:
101*f3087befSAndrew Turner 	      return -PiOver4; /* atan(-INF,+INF).  */
102*f3087befSAndrew Turner 	    case 2:
103*f3087befSAndrew Turner 	      return 3.0f * PiOver4; /* atan(+INF,-INF).  */
104*f3087befSAndrew Turner 	    case 3:
105*f3087befSAndrew Turner 	      return -3.0f * PiOver4; /* atan(-INF,-INF).  */
106*f3087befSAndrew Turner 	    }
107*f3087befSAndrew Turner 	}
108*f3087befSAndrew Turner       else
109*f3087befSAndrew Turner 	{
110*f3087befSAndrew Turner 	  switch (m)
111*f3087befSAndrew Turner 	    {
112*f3087befSAndrew Turner 	    case 0:
113*f3087befSAndrew Turner 	      return 0.0f; /* atan(+...,+INF).  */
114*f3087befSAndrew Turner 	    case 1:
115*f3087befSAndrew Turner 	      return -0.0f; /* atan(-...,+INF).  */
116*f3087befSAndrew Turner 	    case 2:
117*f3087befSAndrew Turner 	      return Pi; /* atan(+...,-INF).  */
118*f3087befSAndrew Turner 	    case 3:
119*f3087befSAndrew Turner 	      return -Pi; /* atan(-...,-INF).  */
120*f3087befSAndrew Turner 	    }
121*f3087befSAndrew Turner 	}
122*f3087befSAndrew Turner     }
123*f3087befSAndrew Turner   /* y is INF.  */
124*f3087befSAndrew Turner   if (iay == 0x7f800000)
125*f3087befSAndrew Turner     return sign_y ? -PiOver2 : PiOver2;
126*f3087befSAndrew Turner 
127*f3087befSAndrew Turner   uint32_t sign_xy = sign_x ^ sign_y;
128*f3087befSAndrew Turner 
129*f3087befSAndrew Turner   float ax = asfloat (iax);
130*f3087befSAndrew Turner   float ay = asfloat (iay);
131*f3087befSAndrew Turner 
132*f3087befSAndrew Turner   bool pred_aygtax = (ay > ax);
133*f3087befSAndrew Turner 
134*f3087befSAndrew Turner   /* Set up z for call to atanf.  */
135*f3087befSAndrew Turner   float n = pred_aygtax ? -ax : ay;
136*f3087befSAndrew Turner   float d = pred_aygtax ? ay : ax;
137*f3087befSAndrew Turner   float z = n / d;
138*f3087befSAndrew Turner 
139*f3087befSAndrew Turner   float ret;
140*f3087befSAndrew Turner   if (unlikely (m < 2 && exp_diff >= POLY_UFLOW_BOUND))
141*f3087befSAndrew Turner     {
142*f3087befSAndrew Turner       /* If (x, y) is very close to x axis and x is positive, the polynomial
143*f3087befSAndrew Turner 	 will underflow and evaluate to z.  */
144*f3087befSAndrew Turner       ret = z;
145*f3087befSAndrew Turner     }
146*f3087befSAndrew Turner   else
147*f3087befSAndrew Turner     {
148*f3087befSAndrew Turner       /* Work out the correct shift.  */
149*f3087befSAndrew Turner       float shift = sign_x ? -2.0f : 0.0f;
150*f3087befSAndrew Turner       shift = pred_aygtax ? shift + 1.0f : shift;
151*f3087befSAndrew Turner       shift *= PiOver2;
152*f3087befSAndrew Turner 
153*f3087befSAndrew Turner       ret = eval_poly (z, z, shift);
154*f3087befSAndrew Turner     }
155*f3087befSAndrew Turner 
156*f3087befSAndrew Turner   /* Account for the sign of x and y.  */
157*f3087befSAndrew Turner   return asfloat (asuint (ret) ^ sign_xy);
158*f3087befSAndrew Turner }
159*f3087befSAndrew Turner 
160*f3087befSAndrew Turner /* Arity of 2 means no mathbench entry emitted. See test/mathbench_funcs.h.  */
161*f3087befSAndrew Turner TEST_SIG (S, F, 2, atan2)
162*f3087befSAndrew Turner TEST_ULP (atan2f, 2.4)
163*f3087befSAndrew Turner TEST_INTERVAL (atan2f, -10.0, 10.0, 50000)
164*f3087befSAndrew Turner TEST_INTERVAL (atan2f, -1.0, 1.0, 40000)
165*f3087befSAndrew Turner TEST_INTERVAL (atan2f, 0.0, 1.0, 40000)
166*f3087befSAndrew Turner TEST_INTERVAL (atan2f, 1.0, 100.0, 40000)
167*f3087befSAndrew Turner TEST_INTERVAL (atan2f, 1e6, 1e32, 40000)
168