xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_sinhf_2u3.c (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
1 /*
2  * Single-precision vector sinh(x) function.
3  *
4  * Copyright (c) 2022-2023, 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 
12 #include "v_expm1f_inline.h"
13 
14 static const struct data
15 {
16   struct v_expm1f_data expm1f_consts;
17   uint32x4_t halff;
18 #if WANT_SIMD_EXCEPT
19   uint32x4_t tiny_bound, thresh;
20 #else
21   uint32x4_t oflow_bound;
22 #endif
23 } data = {
24   .expm1f_consts = V_EXPM1F_DATA,
25   .halff = V4 (0x3f000000),
26 #if WANT_SIMD_EXCEPT
27   /* 0x1.6a09e8p-32, below which expm1f underflows.  */
28   .tiny_bound = V4 (0x2fb504f4),
29   /* asuint(oflow_bound) - asuint(tiny_bound).  */
30   .thresh = V4 (0x12fbbbb3),
31 #else
32   /* 0x1.61814ep+6, above which expm1f helper overflows.  */
33   .oflow_bound = V4 (0x42b0c0a7),
34 #endif
35 };
36 
37 static float32x4_t NOINLINE VPCS_ATTR
38 special_case (float32x4_t x, float32x4_t y, uint32x4_t special)
39 {
40   return v_call_f32 (sinhf, x, y, special);
41 }
42 
43 /* Approximation for vector single-precision sinh(x) using expm1.
44    sinh(x) = (exp(x) - exp(-x)) / 2.
45    The maximum error is 2.26 ULP:
46    _ZGVnN4v_sinhf (0x1.e34a9ep-4) got 0x1.e469ep-4
47 				 want 0x1.e469e4p-4.  */
48 float32x4_t VPCS_ATTR V_NAME_F1 (sinh) (float32x4_t x)
49 {
50   const struct data *d = ptr_barrier (&data);
51 
52   uint32x4_t ix = vreinterpretq_u32_f32 (x);
53   float32x4_t ax = vabsq_f32 (x);
54   uint32x4_t iax = vreinterpretq_u32_f32 (ax);
55   uint32x4_t sign = veorq_u32 (ix, iax);
56   float32x4_t halfsign = vreinterpretq_f32_u32 (vorrq_u32 (sign, d->halff));
57 
58 #if WANT_SIMD_EXCEPT
59   uint32x4_t special = vcgeq_u32 (vsubq_u32 (iax, d->tiny_bound), d->thresh);
60   ax = v_zerofy_f32 (ax, special);
61 #else
62   uint32x4_t special = vcgeq_u32 (iax, d->oflow_bound);
63 #endif
64 
65   /* Up to the point that expm1f overflows, we can use it to calculate sinhf
66        using a slight rearrangement of the definition of asinh. This allows us
67      to retain acceptable accuracy for very small inputs.  */
68   float32x4_t t = expm1f_inline (ax, &d->expm1f_consts);
69   t = vaddq_f32 (t, vdivq_f32 (t, vaddq_f32 (t, v_f32 (1.0))));
70 
71   /* Fall back to the scalar variant for any lanes that should trigger an
72      exception.  */
73   if (unlikely (v_any_u32 (special)))
74     return special_case (x, vmulq_f32 (t, halfsign), special);
75 
76   return vmulq_f32 (t, halfsign);
77 }
78 
79 PL_SIG (V, F, 1, sinh, -10.0, 10.0)
80 PL_TEST_ULP (V_NAME_F1 (sinh), 1.76)
81 PL_TEST_EXPECT_FENV (V_NAME_F1 (sinh), WANT_SIMD_EXCEPT)
82 PL_TEST_SYM_INTERVAL (V_NAME_F1 (sinh), 0, 0x2fb504f4, 1000)
83 PL_TEST_SYM_INTERVAL (V_NAME_F1 (sinh), 0x2fb504f4, 0x42b0c0a7, 100000)
84 PL_TEST_SYM_INTERVAL (V_NAME_F1 (sinh), 0x42b0c0a7, inf, 1000)
85