xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/acoshf.c (revision dd21556857e8d40f66bf5ad54754d9d52669ebf7)
1 /*
2  * Single-precision vector acosh(x) function.
3  * Copyright (c) 2023-2024, Arm Limited.
4  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
5  */
6 
7 #include "v_math.h"
8 #include "test_sig.h"
9 #include "test_defs.h"
10 #include "v_log1pf_inline.h"
11 
12 #define SquareLim 0x1p64
13 
14 const static struct data
15 {
16   struct v_log1pf_data log1pf_consts;
17   uint32x4_t one;
18 } data = { .log1pf_consts = V_LOG1PF_CONSTANTS_TABLE, .one = V4 (0x3f800000) };
19 
20 #define Thresh vdup_n_u16 (0x2000) /* top(asuint(SquareLim) - asuint(1)).  */
21 
22 static float32x4_t NOINLINE VPCS_ATTR
23 special_case (float32x4_t x, float32x4_t y, uint16x4_t special,
24 	      const struct v_log1pf_data *d)
25 {
26   return v_call_f32 (acoshf, x, log1pf_inline (y, d), vmovl_u16 (special));
27 }
28 
29 /* Vector approximation for single-precision acosh, based on log1p. Maximum
30    error depends on WANT_SIMD_EXCEPT. With SIMD fp exceptions enabled, it
31    is 3.00 ULP:
32    _ZGVnN4v_acoshf(0x1.01df3ap+0) got 0x1.ef0a82p-4
33 				 want 0x1.ef0a7cp-4.
34    With exceptions disabled, we can compute u with a shorter dependency chain,
35    which gives maximum error of 3.22 ULP:
36    _ZGVnN4v_acoshf(0x1.007ef2p+0) got 0x1.fdcdccp-5
37 				 want 0x1.fdcdd2p-5.  */
38 
39 float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (acosh) (float32x4_t x)
40 {
41   const struct data *d = ptr_barrier (&data);
42   uint32x4_t ix = vreinterpretq_u32_f32 (x);
43   uint16x4_t special = vcge_u16 (vsubhn_u32 (ix, d->one), Thresh);
44 
45 #if WANT_SIMD_EXCEPT
46   /* Mask special lanes with 1 to side-step spurious invalid or overflow. Use
47      only xm1 to calculate u, as operating on x will trigger invalid for NaN.
48      Widening sign-extend special predicate in order to mask with it.  */
49   uint32x4_t p
50       = vreinterpretq_u32_s32 (vmovl_s16 (vreinterpret_s16_u16 (special)));
51   float32x4_t xm1 = v_zerofy_f32 (vsubq_f32 (x, v_f32 (1)), p);
52   float32x4_t u = vfmaq_f32 (vaddq_f32 (xm1, xm1), xm1, xm1);
53 #else
54   float32x4_t xm1 = vsubq_f32 (x, vreinterpretq_f32_u32 (d->one));
55   float32x4_t u
56       = vmulq_f32 (xm1, vaddq_f32 (x, vreinterpretq_f32_u32 (d->one)));
57 #endif
58 
59   float32x4_t y = vaddq_f32 (xm1, vsqrtq_f32 (u));
60 
61   if (unlikely (v_any_u16h (special)))
62     return special_case (x, y, special, &d->log1pf_consts);
63   return log1pf_inline (y, &d->log1pf_consts);
64 }
65 
66 HALF_WIDTH_ALIAS_F1 (acosh)
67 
68 TEST_SIG (V, F, 1, acosh, 1.0, 10.0)
69 #if WANT_SIMD_EXCEPT
70 TEST_ULP (V_NAME_F1 (acosh), 2.50)
71 #else
72 TEST_ULP (V_NAME_F1 (acosh), 2.78)
73 #endif
74 TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (acosh), WANT_SIMD_EXCEPT)
75 TEST_INTERVAL (V_NAME_F1 (acosh), 0, 1, 500)
76 TEST_INTERVAL (V_NAME_F1 (acosh), 1, SquareLim, 100000)
77 TEST_INTERVAL (V_NAME_F1 (acosh), SquareLim, inf, 1000)
78 TEST_INTERVAL (V_NAME_F1 (acosh), -0, -inf, 1000)
79