1 /*
2  * Double-precision vector atan(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 "v_math.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 #include "poly_advsimd_f64.h"
12 
13 static const struct data
14 {
15   float64x2_t pi_over_2;
16   float64x2_t poly[20];
17 } data = {
18   /* Coefficients of polynomial P such that atan(x)~x+x*P(x^2) on
19 	      [2**-1022, 1.0].  */
20   .poly = { V2 (-0x1.5555555555555p-2),	 V2 (0x1.99999999996c1p-3),
21 	    V2 (-0x1.2492492478f88p-3),	 V2 (0x1.c71c71bc3951cp-4),
22 	    V2 (-0x1.745d160a7e368p-4),	 V2 (0x1.3b139b6a88ba1p-4),
23 	    V2 (-0x1.11100ee084227p-4),	 V2 (0x1.e1d0f9696f63bp-5),
24 	    V2 (-0x1.aebfe7b418581p-5),	 V2 (0x1.842dbe9b0d916p-5),
25 	    V2 (-0x1.5d30140ae5e99p-5),	 V2 (0x1.338e31eb2fbbcp-5),
26 	    V2 (-0x1.00e6eece7de8p-5),	 V2 (0x1.860897b29e5efp-6),
27 	    V2 (-0x1.0051381722a59p-6),	 V2 (0x1.14e9dc19a4a4ep-7),
28 	    V2 (-0x1.d0062b42fe3bfp-9),	 V2 (0x1.17739e210171ap-10),
29 	    V2 (-0x1.ab24da7be7402p-13), V2 (0x1.358851160a528p-16), },
30   .pi_over_2 = V2 (0x1.921fb54442d18p+0),
31 };
32 
33 #define SignMask v_u64 (0x8000000000000000)
34 #define TinyBound 0x3e10000000000000 /* asuint64(0x1p-30).  */
35 #define BigBound 0x4340000000000000  /* asuint64(0x1p53).  */
36 
37 /* Fast implementation of vector atan.
38    Based on atan(x) ~ shift + z + z^3 * P(z^2) with reduction to [0,1] using
39    z=1/x and shift = pi/2. Maximum observed error is 2.27 ulps:
40    _ZGVnN2v_atan (0x1.0005af27c23e9p+0) got 0x1.9225645bdd7c1p-1
41 				       want 0x1.9225645bdd7c3p-1.  */
42 float64x2_t VPCS_ATTR V_NAME_D1 (atan) (float64x2_t x)
43 {
44   const struct data *d = ptr_barrier (&data);
45 
46   /* Small cases, infs and nans are supported by our approximation technique,
47      but do not set fenv flags correctly. Only trigger special case if we need
48      fenv.  */
49   uint64x2_t ix = vreinterpretq_u64_f64 (x);
50   uint64x2_t sign = vandq_u64 (ix, SignMask);
51 
52 #if WANT_SIMD_EXCEPT
53   uint64x2_t ia12 = vandq_u64 (ix, v_u64 (0x7ff0000000000000));
54   uint64x2_t special = vcgtq_u64 (vsubq_u64 (ia12, v_u64 (TinyBound)),
55 				  v_u64 (BigBound - TinyBound));
56   /* If any lane is special, fall back to the scalar routine for all lanes.  */
57   if (unlikely (v_any_u64 (special)))
58     return v_call_f64 (atan, x, v_f64 (0), v_u64 (-1));
59 #endif
60 
61   /* Argument reduction:
62      y := arctan(x) for x < 1
63      y := pi/2 + arctan(-1/x) for x > 1
64      Hence, use z=-1/a if x>=1, otherwise z=a.  */
65   uint64x2_t red = vcagtq_f64 (x, v_f64 (1.0));
66   /* Avoid dependency in abs(x) in division (and comparison).  */
67   float64x2_t z = vbslq_f64 (red, vdivq_f64 (v_f64 (1.0), x), x);
68   float64x2_t shift = vreinterpretq_f64_u64 (
69       vandq_u64 (red, vreinterpretq_u64_f64 (d->pi_over_2)));
70   /* Use absolute value only when needed (odd powers of z).  */
71   float64x2_t az = vbslq_f64 (
72       SignMask, vreinterpretq_f64_u64 (vandq_u64 (SignMask, red)), z);
73 
74   /* Calculate the polynomial approximation.
75      Use split Estrin scheme for P(z^2) with deg(P)=19. Use split instead of
76      full scheme to avoid underflow in x^16.
77      The order 19 polynomial P approximates
78      (atan(sqrt(x))-sqrt(x))/x^(3/2).  */
79   float64x2_t z2 = vmulq_f64 (z, z);
80   float64x2_t x2 = vmulq_f64 (z2, z2);
81   float64x2_t x4 = vmulq_f64 (x2, x2);
82   float64x2_t x8 = vmulq_f64 (x4, x4);
83   float64x2_t y
84       = vfmaq_f64 (v_estrin_7_f64 (z2, x2, x4, d->poly),
85 		   v_estrin_11_f64 (z2, x2, x4, x8, d->poly + 8), x8);
86 
87   /* Finalize. y = shift + z + z^3 * P(z^2).  */
88   y = vfmaq_f64 (az, y, vmulq_f64 (z2, az));
89   y = vaddq_f64 (y, shift);
90 
91   /* y = atan(x) if x>0, -atan(-x) otherwise.  */
92   y = vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), sign));
93   return y;
94 }
95 
96 PL_SIG (V, D, 1, atan, -10.0, 10.0)
97 PL_TEST_ULP (V_NAME_D1 (atan), 1.78)
98 PL_TEST_EXPECT_FENV (V_NAME_D1 (atan), WANT_SIMD_EXCEPT)
99 PL_TEST_INTERVAL (V_NAME_D1 (atan), 0, 0x1p-30, 10000)
100 PL_TEST_INTERVAL (V_NAME_D1 (atan), -0, -0x1p-30, 1000)
101 PL_TEST_INTERVAL (V_NAME_D1 (atan), 0x1p-30, 0x1p53, 900000)
102 PL_TEST_INTERVAL (V_NAME_D1 (atan), -0x1p-30, -0x1p53, 90000)
103 PL_TEST_INTERVAL (V_NAME_D1 (atan), 0x1p53, inf, 10000)
104 PL_TEST_INTERVAL (V_NAME_D1 (atan), -0x1p53, -inf, 1000)
105