xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/experimental/tanf_3u3.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Single-precision scalar tan(x) function.
3  *
4  * Copyright (c) 2021-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 #include "math_config.h"
8 #include "test_sig.h"
9 #include "test_defs.h"
10 #include "poly_scalar_f32.h"
11 
12 /* Useful constants.  */
13 #define NegPio2_1 (-0x1.921fb6p+0f)
14 #define NegPio2_2 (0x1.777a5cp-25f)
15 #define NegPio2_3 (0x1.ee59dap-50f)
16 /* Reduced from 0x1p20 to 0x1p17 to ensure 3.5ulps.  */
17 #define RangeVal (0x1p17f)
18 #define InvPio2 ((0x1.45f306p-1f))
19 #define Shift (0x1.8p+23f)
20 #define AbsMask (0x7fffffff)
21 #define Pio4 (0x1.921fb6p-1)
22 /* 2PI * 2^-64.  */
23 #define Pio2p63 (0x1.921FB54442D18p-62)
24 
25 static inline float
eval_P(float z)26 eval_P (float z)
27 {
28   return pw_horner_5_f32 (z, z * z, __tanf_poly_data.poly_tan);
29 }
30 
31 static inline float
eval_Q(float z)32 eval_Q (float z)
33 {
34   return pairwise_poly_3_f32 (z, z * z, __tanf_poly_data.poly_cotan);
35 }
36 
37 /* Reduction of the input argument x using Cody-Waite approach, such that x = r
38    + n * pi/2 with r lives in [-pi/4, pi/4] and n is a signed integer.  */
39 static inline float
reduce(float x,int32_t * in)40 reduce (float x, int32_t *in)
41 {
42   /* n = rint(x/(pi/2)).  */
43   float r = x;
44   float q = fmaf (InvPio2, r, Shift);
45   float n = q - Shift;
46   /* There is no rounding here, n is representable by a signed integer.  */
47   *in = (int32_t) n;
48   /* r = x - n * (pi/2)  (range reduction into -pi/4 .. pi/4).  */
49   r = fmaf (NegPio2_1, n, r);
50   r = fmaf (NegPio2_2, n, r);
51   r = fmaf (NegPio2_3, n, r);
52   return r;
53 }
54 
55 /* Reduce the range of XI to a multiple of PI/2 using fast integer arithmetic.
56    XI is a reinterpreted float and must be >= 2.0f (the sign bit is ignored).
57    Return the modulo between -PI/4 and PI/4 and store the quadrant in NP.
58    Reduction uses a table of 4/PI with 192 bits of precision.  A 32x96->128 bit
59    multiply computes the exact 2.62-bit fixed-point modulo.  Since the result
60    can have at most 29 leading zeros after the binary point, the double
61    precision result is accurate to 33 bits.  */
62 static inline double
reduce_large(uint32_t xi,int * np)63 reduce_large (uint32_t xi, int *np)
64 {
65   const uint32_t *arr = &__inv_pio4[(xi >> 26) & 15];
66   int shift = (xi >> 23) & 7;
67   uint64_t n, res0, res1, res2;
68 
69   xi = (xi & 0xffffff) | 0x800000;
70   xi <<= shift;
71 
72   res0 = xi * arr[0];
73   res1 = (uint64_t) xi * arr[4];
74   res2 = (uint64_t) xi * arr[8];
75   res0 = (res2 >> 32) | (res0 << 32);
76   res0 += res1;
77 
78   n = (res0 + (1ULL << 61)) >> 62;
79   res0 -= n << 62;
80   double x = (int64_t) res0;
81   *np = n;
82   return x * Pio2p63;
83 }
84 
85 /* Top 12 bits of the float representation with the sign bit cleared.  */
86 static inline uint32_t
top12(float x)87 top12 (float x)
88 {
89   return (asuint (x) >> 20);
90 }
91 
92 /* Fast single-precision tan implementation.
93    Maximum ULP error: 3.293ulps.
94    tanf(0x1.c849eap+16) got -0x1.fe8d98p-1 want -0x1.fe8d9ep-1.  */
95 float
tanf(float x)96 tanf (float x)
97 {
98   /* Get top words.  */
99   uint32_t ix = asuint (x);
100   uint32_t ia = ix & AbsMask;
101   uint32_t ia12 = ia >> 20;
102 
103   /* Dispatch between no reduction (small numbers), fast reduction and
104      slow large numbers reduction. The reduction step determines r float
105      (|r| < pi/4) and n signed integer such that x = r + n * pi/2.  */
106   int32_t n;
107   float r;
108   if (ia12 < top12 (Pio4))
109     {
110       /* Optimize small values.  */
111       if (unlikely (ia12 < top12 (0x1p-12f)))
112 	{
113 	  if (unlikely (ia12 < top12 (0x1p-126f)))
114 	    /* Force underflow for tiny x.  */
115 	    force_eval_float (x * x);
116 	  return x;
117 	}
118 
119       /* tan (x) ~= x + x^3 * P(x^2).  */
120       float x2 = x * x;
121       float y = eval_P (x2);
122       return fmaf (x2, x * y, x);
123     }
124   /* Similar to other trigonometric routines, fast inaccurate reduction is
125      performed for values of x from pi/4 up to RangeVal. In order to keep
126      errors below 3.5ulps, we set the value of RangeVal to 2^17. This might
127      differ for other trigonometric routines. Above this value more advanced
128      but slower reduction techniques need to be implemented to reach a similar
129      accuracy.  */
130   else if (ia12 < top12 (RangeVal))
131     {
132       /* Fast inaccurate reduction.  */
133       r = reduce (x, &n);
134     }
135   else if (ia12 < 0x7f8)
136     {
137       /* Slow accurate reduction.  */
138       uint32_t sign = ix & ~AbsMask;
139       double dar = reduce_large (ia, &n);
140       float ar = (float) dar;
141       r = asfloat (asuint (ar) ^ sign);
142     }
143   else
144     {
145       /* tan(Inf or NaN) is NaN.  */
146       return __math_invalidf (x);
147     }
148 
149   /* If x lives in an interval where |tan(x)|
150      - is finite then use an approximation of tangent in the form
151        tan(r) ~ r + r^3 * P(r^2) = r + r * r^2 * P(r^2).
152      - grows to infinity then use an approximation of cotangent in the form
153        cotan(z) ~ 1/z + z * Q(z^2), where the reciprocal can be computed early.
154        Using symmetries of tangent and the identity tan(r) = cotan(pi/2 - r),
155        we only need to change the sign of r to obtain tan(x) from cotan(r).
156      This 2-interval approach requires 2 different sets of coefficients P and
157      Q, where Q is a lower order polynomial than P.  */
158 
159   /* Determine if x lives in an interval where |tan(x)| grows to infinity.  */
160   uint32_t alt = (uint32_t) n & 1;
161 
162   /* Perform additional reduction if required.  */
163   float z = alt ? -r : r;
164 
165   /* Prepare backward transformation.  */
166   float z2 = r * r;
167   float offset = alt ? 1.0f / z : z;
168   float scale = alt ? z : z * z2;
169 
170   /* Evaluate polynomial approximation of tan or cotan.  */
171   float p = alt ? eval_Q (z2) : eval_P (z2);
172 
173   /* A unified way of assembling the result on both interval types.  */
174   return fmaf (scale, p, offset);
175 }
176 
177 TEST_SIG (S, F, 1, tan, -3.1, 3.1)
178 TEST_ULP (tanf, 2.80)
179 TEST_INTERVAL (tanf, 0, 0xffff0000, 10000)
180 TEST_SYM_INTERVAL (tanf, 0x1p-127, 0x1p-14, 50000)
181 TEST_SYM_INTERVAL (tanf, 0x1p-14, 0.7, 50000)
182 TEST_SYM_INTERVAL (tanf, 0.7, 1.5, 50000)
183 TEST_SYM_INTERVAL (tanf, 1.5, 0x1p17, 50000)
184 TEST_SYM_INTERVAL (tanf, 0x1p17, 0x1p54, 50000)
185 TEST_SYM_INTERVAL (tanf, 0x1p54, inf, 50000)
186