xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_hypotf_1u5.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /*
2  * Single-precision vector hypot(x) function.
3  *
4  * Copyright (c) 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 #if WANT_SIMD_EXCEPT
13 static const struct data
14 {
15   uint32x4_t tiny_bound, thres;
16 } data = {
17   .tiny_bound = V4 (0x20000000), /* asuint (0x1p-63).  */
18   .thres = V4 (0x3f000000), /* asuint (0x1p63) - tiny_bound.  */
19 };
20 #else
21 static const struct data
22 {
23   uint32x4_t tiny_bound;
24   uint16x8_t thres;
25 } data = {
26   .tiny_bound = V4 (0x0C800000), /* asuint (0x1p-102).  */
27   .thres = V8 (0x7300), /* asuint (inf) - tiny_bound.  */
28 };
29 #endif
30 
31 static float32x4_t VPCS_ATTR NOINLINE
32 special_case (float32x4_t x, float32x4_t y, float32x4_t sqsum,
33 	      uint16x4_t special)
34 {
35   return v_call2_f32 (hypotf, x, y, vsqrtq_f32 (sqsum), vmovl_u16 (special));
36 }
37 
38 /* Vector implementation of single-precision hypot.
39    Maximum error observed is 1.21 ULP:
40    _ZGVnN4vv_hypotf (0x1.6a419cp-13, 0x1.82a852p-22) got 0x1.6a41d2p-13
41 						    want 0x1.6a41dp-13.  */
42 #if WANT_SIMD_EXCEPT
43 
44 float32x4_t VPCS_ATTR V_NAME_F2 (hypot) (float32x4_t x, float32x4_t y)
45 {
46   const struct data *d = ptr_barrier (&data);
47 
48   float32x4_t ax = vabsq_f32 (x);
49   float32x4_t ay = vabsq_f32 (y);
50 
51   uint32x4_t ix = vreinterpretq_u32_f32 (ax);
52   uint32x4_t iy = vreinterpretq_u32_f32 (ay);
53 
54   /* Extreme values, NaNs, and infinities should be handled by the scalar
55      fallback for correct flag handling.  */
56   uint32x4_t specialx = vcgeq_u32 (vsubq_u32 (ix, d->tiny_bound), d->thres);
57   uint32x4_t specialy = vcgeq_u32 (vsubq_u32 (iy, d->tiny_bound), d->thres);
58   ax = v_zerofy_f32 (ax, specialx);
59   ay = v_zerofy_f32 (ay, specialy);
60   uint16x4_t special = vaddhn_u32 (specialx, specialy);
61 
62   float32x4_t sqsum = vfmaq_f32 (vmulq_f32 (ax, ax), ay, ay);
63 
64   if (unlikely (v_any_u16h (special)))
65     return special_case (x, y, sqsum, special);
66 
67   return vsqrtq_f32 (sqsum);
68 }
69 #else
70 
71 float32x4_t VPCS_ATTR V_NAME_F2 (hypot) (float32x4_t x, float32x4_t y)
72 {
73   const struct data *d = ptr_barrier (&data);
74 
75   float32x4_t sqsum = vfmaq_f32 (vmulq_f32 (x, x), y, y);
76 
77   uint16x4_t special = vcge_u16 (
78       vsubhn_u32 (vreinterpretq_u32_f32 (sqsum), d->tiny_bound),
79       vget_low_u16 (d->thres));
80 
81   if (unlikely (v_any_u16h (special)))
82     return special_case (x, y, sqsum, special);
83 
84   return vsqrtq_f32 (sqsum);
85 }
86 #endif
87 
88 PL_SIG (V, F, 2, hypot, -10.0, 10.0)
89 PL_TEST_ULP (V_NAME_F2 (hypot), 1.21)
90 PL_TEST_EXPECT_FENV (V_NAME_F2 (hypot), WANT_SIMD_EXCEPT)
91 PL_TEST_INTERVAL2 (V_NAME_F2 (hypot), 0, inf, 0, inf, 10000)
92 PL_TEST_INTERVAL2 (V_NAME_F2 (hypot), 0, inf, -0, -inf, 10000)
93 PL_TEST_INTERVAL2 (V_NAME_F2 (hypot), -0, -inf, 0, inf, 10000)
94 PL_TEST_INTERVAL2 (V_NAME_F2 (hypot), -0, -inf, -0, -inf, 10000)
95