xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/atan2.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Double-precision vector atan2(x) function.
3  *
4  * Copyright (c) 2021-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "v_math.h"
9 #include "test_sig.h"
10 #include "test_defs.h"
11 
12 static const struct data
13 {
14   float64x2_t c0, c2, c4, c6, c8, c10, c12, c14, c16, c18;
15   float64x2_t pi_over_2;
16   double c1, c3, c5, c7, c9, c11, c13, c15, c17, c19;
17   uint64x2_t zeroinfnan, minustwo;
18 } data = {
19   /* Coefficients of polynomial P such that atan(x)~x+x*P(x^2) on
20 	      [2**-1022, 1.0].  */
21   .c0 = V2 (-0x1.5555555555555p-2),
22   .c1 = 0x1.99999999996c1p-3,
23   .c2 = V2 (-0x1.2492492478f88p-3),
24   .c3 = 0x1.c71c71bc3951cp-4,
25   .c4 = V2 (-0x1.745d160a7e368p-4),
26   .c5 = 0x1.3b139b6a88ba1p-4,
27   .c6 = V2 (-0x1.11100ee084227p-4),
28   .c7 = 0x1.e1d0f9696f63bp-5,
29   .c8 = V2 (-0x1.aebfe7b418581p-5),
30   .c9 = 0x1.842dbe9b0d916p-5,
31   .c10 = V2 (-0x1.5d30140ae5e99p-5),
32   .c11 = 0x1.338e31eb2fbbcp-5,
33   .c12 = V2 (-0x1.00e6eece7de8p-5),
34   .c13 = 0x1.860897b29e5efp-6,
35   .c14 = V2 (-0x1.0051381722a59p-6),
36   .c15 = 0x1.14e9dc19a4a4ep-7,
37   .c16 = V2 (-0x1.d0062b42fe3bfp-9),
38   .c17 = 0x1.17739e210171ap-10,
39   .c18 = V2 (-0x1.ab24da7be7402p-13),
40   .c19 = 0x1.358851160a528p-16,
41   .pi_over_2 = V2 (0x1.921fb54442d18p+0),
42   .zeroinfnan = V2 (2 * 0x7ff0000000000000ul - 1),
43   .minustwo = V2 (0xc000000000000000),
44 };
45 
46 #define SignMask v_u64 (0x8000000000000000)
47 
48 /* Special cases i.e. 0, infinity, NaN (fall back to scalar calls).  */
49 static float64x2_t VPCS_ATTR NOINLINE
special_case(float64x2_t y,float64x2_t x,float64x2_t ret,uint64x2_t sign_xy,uint64x2_t cmp)50 special_case (float64x2_t y, float64x2_t x, float64x2_t ret,
51 	      uint64x2_t sign_xy, uint64x2_t cmp)
52 {
53   /* Account for the sign of x and y.  */
54   ret = vreinterpretq_f64_u64 (
55       veorq_u64 (vreinterpretq_u64_f64 (ret), sign_xy));
56   return v_call2_f64 (atan2, y, x, ret, cmp);
57 }
58 
59 /* Returns 1 if input is the bit representation of 0, infinity or nan.  */
60 static inline uint64x2_t
zeroinfnan(uint64x2_t i,const struct data * d)61 zeroinfnan (uint64x2_t i, const struct data *d)
62 {
63   /* (2 * i - 1) >= (2 * asuint64 (INFINITY) - 1).  */
64   return vcgeq_u64 (vsubq_u64 (vaddq_u64 (i, i), v_u64 (1)), d->zeroinfnan);
65 }
66 
67 /* Fast implementation of vector atan2.
68    Maximum observed error is 2.8 ulps:
69    _ZGVnN2vv_atan2 (0x1.9651a429a859ap+5, 0x1.953075f4ee26p+5)
70 	got 0x1.92d628ab678ccp-1
71        want 0x1.92d628ab678cfp-1.  */
V_NAME_D2(atan2)72 float64x2_t VPCS_ATTR V_NAME_D2 (atan2) (float64x2_t y, float64x2_t x)
73 {
74   const struct data *d = ptr_barrier (&data);
75 
76   uint64x2_t ix = vreinterpretq_u64_f64 (x);
77   uint64x2_t iy = vreinterpretq_u64_f64 (y);
78 
79   uint64x2_t special_cases
80       = vorrq_u64 (zeroinfnan (ix, d), zeroinfnan (iy, d));
81 
82   uint64x2_t sign_x = vandq_u64 (ix, SignMask);
83   uint64x2_t sign_y = vandq_u64 (iy, SignMask);
84   uint64x2_t sign_xy = veorq_u64 (sign_x, sign_y);
85 
86   float64x2_t ax = vabsq_f64 (x);
87   float64x2_t ay = vabsq_f64 (y);
88 
89   uint64x2_t pred_xlt0 = vcltzq_f64 (x);
90   uint64x2_t pred_aygtax = vcagtq_f64 (y, x);
91 
92   /* Set up z for call to atan.  */
93   float64x2_t n = vbslq_f64 (pred_aygtax, vnegq_f64 (ax), ay);
94   float64x2_t q = vbslq_f64 (pred_aygtax, ay, ax);
95   float64x2_t z = vdivq_f64 (n, q);
96 
97   /* Work out the correct shift.  */
98   float64x2_t shift
99       = vreinterpretq_f64_u64 (vandq_u64 (pred_xlt0, d->minustwo));
100   shift = vbslq_f64 (pred_aygtax, vaddq_f64 (shift, v_f64 (1.0)), shift);
101   shift = vmulq_f64 (shift, d->pi_over_2);
102 
103   /* Calculate the polynomial approximation.
104      Use split Estrin scheme for P(z^2) with deg(P)=19. Use split instead of
105      full scheme to avoid underflow in x^16.
106      The order 19 polynomial P approximates
107      (atan(sqrt(x))-sqrt(x))/x^(3/2).  */
108   float64x2_t z2 = vmulq_f64 (z, z);
109   float64x2_t x2 = vmulq_f64 (z2, z2);
110   float64x2_t x4 = vmulq_f64 (x2, x2);
111   float64x2_t x8 = vmulq_f64 (x4, x4);
112 
113   float64x2_t c13 = vld1q_f64 (&d->c1);
114   float64x2_t c57 = vld1q_f64 (&d->c5);
115   float64x2_t c911 = vld1q_f64 (&d->c9);
116   float64x2_t c1315 = vld1q_f64 (&d->c13);
117   float64x2_t c1719 = vld1q_f64 (&d->c17);
118 
119   /* estrin_7.  */
120   float64x2_t p01 = vfmaq_laneq_f64 (d->c0, z2, c13, 0);
121   float64x2_t p23 = vfmaq_laneq_f64 (d->c2, z2, c13, 1);
122   float64x2_t p03 = vfmaq_f64 (p01, x2, p23);
123 
124   float64x2_t p45 = vfmaq_laneq_f64 (d->c4, z2, c57, 0);
125   float64x2_t p67 = vfmaq_laneq_f64 (d->c6, z2, c57, 1);
126   float64x2_t p47 = vfmaq_f64 (p45, x2, p67);
127 
128   float64x2_t p07 = vfmaq_f64 (p03, x4, p47);
129 
130   /* estrin_11.  */
131   float64x2_t p89 = vfmaq_laneq_f64 (d->c8, z2, c911, 0);
132   float64x2_t p1011 = vfmaq_laneq_f64 (d->c10, z2, c911, 1);
133   float64x2_t p811 = vfmaq_f64 (p89, x2, p1011);
134 
135   float64x2_t p1213 = vfmaq_laneq_f64 (d->c12, z2, c1315, 0);
136   float64x2_t p1415 = vfmaq_laneq_f64 (d->c14, z2, c1315, 1);
137   float64x2_t p1215 = vfmaq_f64 (p1213, x2, p1415);
138 
139   float64x2_t p1617 = vfmaq_laneq_f64 (d->c16, z2, c1719, 0);
140   float64x2_t p1819 = vfmaq_laneq_f64 (d->c18, z2, c1719, 1);
141   float64x2_t p1619 = vfmaq_f64 (p1617, x2, p1819);
142 
143   float64x2_t p815 = vfmaq_f64 (p811, x4, p1215);
144   float64x2_t p819 = vfmaq_f64 (p815, x8, p1619);
145 
146   float64x2_t ret = vfmaq_f64 (p07, p819, x8);
147 
148   /* Finalize. y = shift + z + z^3 * P(z^2).  */
149   ret = vfmaq_f64 (z, ret, vmulq_f64 (z2, z));
150   ret = vaddq_f64 (ret, shift);
151 
152   if (unlikely (v_any_u64 (special_cases)))
153     return special_case (y, x, ret, sign_xy, special_cases);
154 
155   /* Account for the sign of x and y.  */
156   ret = vreinterpretq_f64_u64 (
157       veorq_u64 (vreinterpretq_u64_f64 (ret), sign_xy));
158 
159   return ret;
160 }
161 
162 /* Arity of 2 means no mathbench entry emitted. See test/mathbench_funcs.h.  */
163 TEST_SIG (V, D, 2, atan2)
164 // TODO tighten this once __v_atan2 is fixed
165 TEST_ULP (V_NAME_D2 (atan2), 2.9)
166 TEST_DISABLE_FENV (V_NAME_D2 (atan2))
167 TEST_INTERVAL (V_NAME_D2 (atan2), -10.0, 10.0, 50000)
168 TEST_INTERVAL (V_NAME_D2 (atan2), -1.0, 1.0, 40000)
169 TEST_INTERVAL (V_NAME_D2 (atan2), 0.0, 1.0, 40000)
170 TEST_INTERVAL (V_NAME_D2 (atan2), 1.0, 100.0, 40000)
171 TEST_INTERVAL (V_NAME_D2 (atan2), 1e6, 1e32, 40000)
172