1*5a02ffc3SAndrew Turner /*
2*5a02ffc3SAndrew Turner * Double-precision SVE hypot(x) function.
3*5a02ffc3SAndrew Turner *
4*5a02ffc3SAndrew Turner * Copyright (c) 2023, Arm Limited.
5*5a02ffc3SAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*5a02ffc3SAndrew Turner */
7*5a02ffc3SAndrew Turner
8*5a02ffc3SAndrew Turner #include "sv_math.h"
9*5a02ffc3SAndrew Turner #include "pl_sig.h"
10*5a02ffc3SAndrew Turner #include "pl_test.h"
11*5a02ffc3SAndrew Turner
12*5a02ffc3SAndrew Turner static const struct data
13*5a02ffc3SAndrew Turner {
14*5a02ffc3SAndrew Turner uint64_t tiny_bound, thres;
15*5a02ffc3SAndrew Turner } data = {
16*5a02ffc3SAndrew Turner .tiny_bound = 0x0c80000000000000, /* asuint (0x1p-102). */
17*5a02ffc3SAndrew Turner .thres = 0x7300000000000000, /* asuint (inf) - tiny_bound. */
18*5a02ffc3SAndrew Turner };
19*5a02ffc3SAndrew Turner
20*5a02ffc3SAndrew Turner static svfloat64_t NOINLINE
special_case(svfloat64_t sqsum,svfloat64_t x,svfloat64_t y,svbool_t pg,svbool_t special)21*5a02ffc3SAndrew Turner special_case (svfloat64_t sqsum, svfloat64_t x, svfloat64_t y, svbool_t pg,
22*5a02ffc3SAndrew Turner svbool_t special)
23*5a02ffc3SAndrew Turner {
24*5a02ffc3SAndrew Turner return sv_call2_f64 (hypot, x, y, svsqrt_x (pg, sqsum), special);
25*5a02ffc3SAndrew Turner }
26*5a02ffc3SAndrew Turner
27*5a02ffc3SAndrew Turner /* SVE implementation of double-precision hypot.
28*5a02ffc3SAndrew Turner Maximum error observed is 1.21 ULP:
29*5a02ffc3SAndrew Turner _ZGVsMxvv_hypot (-0x1.6a22d0412cdd3p+352, 0x1.d3d89bd66fb1ap+330)
30*5a02ffc3SAndrew Turner got 0x1.6a22d0412cfp+352
31*5a02ffc3SAndrew Turner want 0x1.6a22d0412cf01p+352. */
SV_NAME_D2(hypot)32*5a02ffc3SAndrew Turner svfloat64_t SV_NAME_D2 (hypot) (svfloat64_t x, svfloat64_t y, svbool_t pg)
33*5a02ffc3SAndrew Turner {
34*5a02ffc3SAndrew Turner const struct data *d = ptr_barrier (&data);
35*5a02ffc3SAndrew Turner
36*5a02ffc3SAndrew Turner svfloat64_t sqsum = svmla_x (pg, svmul_x (pg, x, x), y, y);
37*5a02ffc3SAndrew Turner
38*5a02ffc3SAndrew Turner svbool_t special = svcmpge (
39*5a02ffc3SAndrew Turner pg, svsub_x (pg, svreinterpret_u64 (sqsum), d->tiny_bound), d->thres);
40*5a02ffc3SAndrew Turner
41*5a02ffc3SAndrew Turner if (unlikely (svptest_any (pg, special)))
42*5a02ffc3SAndrew Turner return special_case (sqsum, x, y, pg, special);
43*5a02ffc3SAndrew Turner return svsqrt_x (pg, sqsum);
44*5a02ffc3SAndrew Turner }
45*5a02ffc3SAndrew Turner
46*5a02ffc3SAndrew Turner PL_SIG (SV, D, 2, hypot, -10.0, 10.0)
47*5a02ffc3SAndrew Turner PL_TEST_ULP (SV_NAME_D2 (hypot), 0.71)
48*5a02ffc3SAndrew Turner PL_TEST_INTERVAL2 (SV_NAME_D2 (hypot), 0, inf, 0, inf, 10000)
49*5a02ffc3SAndrew Turner PL_TEST_INTERVAL2 (SV_NAME_D2 (hypot), 0, inf, -0, -inf, 10000)
50*5a02ffc3SAndrew Turner PL_TEST_INTERVAL2 (SV_NAME_D2 (hypot), -0, -inf, 0, inf, 10000)
51*5a02ffc3SAndrew Turner PL_TEST_INTERVAL2 (SV_NAME_D2 (hypot), -0, -inf, -0, -inf, 10000)
52