xref: /freebsd/contrib/arm-optimized-routines/pl/math/sv_atan2_2u5.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 /*
2  * Double-precision vector 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 "sv_math.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 
12 #if SV_SUPPORTED
13 
14 #include "sv_atan_common.h"
15 
16 /* Useful constants.  */
17 #define PiOver2 sv_f64 (0x1.921fb54442d18p+0)
18 #define SignMask sv_u64 (0x8000000000000000)
19 
20 /* Special cases i.e. 0, infinity, nan (fall back to scalar calls).  */
21 __attribute__ ((noinline)) static sv_f64_t
22 specialcase (sv_f64_t y, sv_f64_t x, sv_f64_t ret, const svbool_t cmp)
23 {
24   return sv_call2_f64 (atan2, y, x, ret, cmp);
25 }
26 
27 /* Returns a predicate indicating true if the input is the bit representation of
28    0, infinity or nan.  */
29 static inline svbool_t
30 zeroinfnan (sv_u64_t i, const svbool_t pg)
31 {
32   return svcmpge_u64 (pg, svsub_n_u64_x (pg, svlsl_n_u64_x (pg, i, 1), 1),
33 		      sv_u64 (2 * asuint64 (INFINITY) - 1));
34 }
35 
36 /* Fast implementation of SVE atan2. Errors are greatest when y and
37    x are reasonably close together. The greatest observed error is 2.28 ULP:
38    sv_atan2(-0x1.5915b1498e82fp+732, 0x1.54d11ef838826p+732)
39    got -0x1.954f42f1fa841p-1 want -0x1.954f42f1fa843p-1.  */
40 sv_f64_t
41 __sv_atan2_x (sv_f64_t y, sv_f64_t x, const svbool_t pg)
42 {
43   sv_u64_t ix = sv_as_u64_f64 (x);
44   sv_u64_t iy = sv_as_u64_f64 (y);
45 
46   svbool_t cmp_x = zeroinfnan (ix, pg);
47   svbool_t cmp_y = zeroinfnan (iy, pg);
48   svbool_t cmp_xy = svorr_b_z (pg, cmp_x, cmp_y);
49 
50   sv_u64_t sign_x = svand_u64_x (pg, ix, SignMask);
51   sv_u64_t sign_y = svand_u64_x (pg, iy, SignMask);
52   sv_u64_t sign_xy = sveor_u64_x (pg, sign_x, sign_y);
53 
54   sv_f64_t ax = svabs_f64_x (pg, x);
55   sv_f64_t ay = svabs_f64_x (pg, y);
56 
57   svbool_t pred_xlt0 = svcmplt_f64 (pg, x, sv_f64 (0.0));
58   svbool_t pred_aygtax = svcmpgt_f64 (pg, ay, ax);
59 
60   /* Set up z for call to atan.  */
61   sv_f64_t n = svsel_f64 (pred_aygtax, svneg_f64_x (pg, ax), ay);
62   sv_f64_t d = svsel_f64 (pred_aygtax, ay, ax);
63   sv_f64_t z = svdiv_f64_x (pg, n, d);
64 
65   /* Work out the correct shift.  */
66   sv_f64_t shift = svsel_f64 (pred_xlt0, sv_f64 (-2.0), sv_f64 (0.0));
67   shift = svsel_f64 (pred_aygtax, svadd_n_f64_x (pg, shift, 1.0), shift);
68   shift = svmul_f64_x (pg, shift, PiOver2);
69 
70   sv_f64_t ret = __sv_atan_common (pg, pg, z, z, shift);
71 
72   /* Account for the sign of x and y.  */
73   ret = sv_as_f64_u64 (sveor_u64_x (pg, sv_as_u64_f64 (ret), sign_xy));
74 
75   if (unlikely (svptest_any (pg, cmp_xy)))
76     {
77       return specialcase (y, x, ret, cmp_xy);
78     }
79 
80   return ret;
81 }
82 
83 PL_ALIAS (__sv_atan2_x, _ZGVsMxvv_atan2)
84 
85 /* Arity of 2 means no mathbench entry emitted. See test/mathbench_funcs.h.  */
86 PL_SIG (SV, D, 2, atan2)
87 PL_TEST_ULP (__sv_atan2, 1.78)
88 PL_TEST_INTERVAL (__sv_atan2, -10.0, 10.0, 50000)
89 PL_TEST_INTERVAL (__sv_atan2, -1.0, 1.0, 40000)
90 PL_TEST_INTERVAL (__sv_atan2, 0.0, 1.0, 40000)
91 PL_TEST_INTERVAL (__sv_atan2, 1.0, 100.0, 40000)
92 PL_TEST_INTERVAL (__sv_atan2, 1e6, 1e32, 40000)
93 #endif
94