xref: /aosp_15_r20/external/XNNPACK/src/f32-vrelu/wasm_shr_x1.S (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1# Copyright 2020 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5
6#include <xnnpack/assembly.h>
7
8# void xnn_f32_vrelu_ukernel__wasm32_shr_x1(
9#     size_t n,             0
10#     const float* x,       1
11#     float* y,             2
12#     const union params)   3 unused
13
14# locals
15#     float v               4
16#     float mask            5
17
18BEGIN_FUNCTION  xnn_f32_vrelu_ukernel__wasm32_shr_x1
19    .functype   xnn_f32_vrelu_ukernel__wasm32_shr_x1 (i32, i32, i32, i32) -> ()
20    .local      i32, i32  # 4 - value, 5 - mask
21
22    loop
23      local.get    1        # src
24      i32.load     0        # load float from src
25      local.set    4
26
27      local.get    1        # src += 4
28      i32.const    4
29      i32.add
30      local.set    1
31
32      local.get    4        # (v >> 31) - 1) & v
33      i32.const    31
34      i32.shr_u
35      local.set    5
36
37      local.get    5
38      i32.const    -1
39      i32.add
40      local.set    5
41
42      local.get    4
43      local.get    5
44      i32.and
45      local.set    4
46
47      local.get    2        # dst
48      local.get    4
49      i32.store    0        # store float
50
51      local.get    2        # dst += 4
52      i32.const    4
53      i32.add
54      local.set    2
55
56      local.get    0
57      i32.const    -4
58      i32.add              # count -= 4
59      local.set    0
60
61      local.get    0
62      i32.const    0       # count > 0
63      i32.gt_s
64      br_if        0       # loop
65    end_loop
66END_FUNCTION xnn_f32_vrelu_ukernel__wasm32_shr_x1