1 /*
2 * Single-precision SVE powi(x, n) function.
3 *
4 * Copyright (c) 2020-2024, Arm Limited.
5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6 */
7
8 #include "sv_math.h"
9
10 /* Optimized single-precision vector powi (float base, integer power).
11 powi is developed for environments in which accuracy is of much less
12 importance than performance, hence we provide no estimate for worst-case
13 error. */
14 svfloat32_t
_ZGVsMxvv_powi(svfloat32_t as,svint32_t ns,svbool_t p)15 _ZGVsMxvv_powi (svfloat32_t as, svint32_t ns, svbool_t p)
16 {
17 /* Compute powi by successive squaring, right to left. */
18 svfloat32_t acc = sv_f32 (1.f);
19 svbool_t want_recip = svcmplt (p, ns, 0);
20 svuint32_t ns_abs = svreinterpret_u32 (svabs_x (p, ns));
21
22 /* We use a max to avoid needing to check whether any lane != 0 on each
23 iteration. */
24 uint32_t max_n = svmaxv (p, ns_abs);
25
26 svfloat32_t c = as;
27 /* Successively square c, and use merging predication (_m) to determine
28 whether or not to perform the multiplication or keep the previous
29 iteration. */
30 while (true)
31 {
32 svbool_t px = svcmpeq (p, svand_x (p, ns_abs, 1), 1);
33 acc = svmul_m (px, acc, c);
34 max_n >>= 1;
35 if (max_n == 0)
36 break;
37
38 ns_abs = svlsr_x (p, ns_abs, 1);
39 c = svmul_x (p, c, c);
40 }
41
42 /* Negative powers are handled by computing the abs(n) version and then
43 taking the reciprocal. */
44 if (svptest_any (want_recip, want_recip))
45 acc = svdivr_m (want_recip, acc, 1.0f);
46
47 return acc;
48 }
49 CLOSE_SVE_ATTR
50