xref: /aosp_15_r20/external/XNNPACK/src/f32-vrelu/wasm_shr_x2.S (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker# Copyright 2020 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker#
3*4bdc9457SAndroid Build Coastguard Worker# This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker# LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker
6*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/assembly.h>
7*4bdc9457SAndroid Build Coastguard Worker
8*4bdc9457SAndroid Build Coastguard Worker# void xnn_f32_vrelu_ukernel__wasm32_shr_x2(
9*4bdc9457SAndroid Build Coastguard Worker#     size_t n,             0
10*4bdc9457SAndroid Build Coastguard Worker#     const float* x,       1
11*4bdc9457SAndroid Build Coastguard Worker#     float* y,             2
12*4bdc9457SAndroid Build Coastguard Worker#     const union params)   3 unused
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker# locals
15*4bdc9457SAndroid Build Coastguard Worker#     float value0          4
16*4bdc9457SAndroid Build Coastguard Worker#     float value1          5
17*4bdc9457SAndroid Build Coastguard Worker#     float mask0           6
18*4bdc9457SAndroid Build Coastguard Worker#     float mask1           7
19*4bdc9457SAndroid Build Coastguard Worker
20*4bdc9457SAndroid Build Coastguard Worker
21*4bdc9457SAndroid Build Coastguard WorkerBEGIN_FUNCTION  xnn_f32_vrelu_ukernel__wasm32_shr_x2
22*4bdc9457SAndroid Build Coastguard Worker    .functype   xnn_f32_vrelu_ukernel__wasm32_shr_x2 (i32, i32, i32, i32) -> ()
23*4bdc9457SAndroid Build Coastguard Worker    .local      i32, i32, i32, i32
24*4bdc9457SAndroid Build Coastguard Worker
25*4bdc9457SAndroid Build Coastguard Worker    local.get    0
26*4bdc9457SAndroid Build Coastguard Worker    i32.const    8       # count >= 8
27*4bdc9457SAndroid Build Coastguard Worker    i32.ge_s
28*4bdc9457SAndroid Build Coastguard Worker    if
29*4bdc9457SAndroid Build Coastguard Worker      loop
30*4bdc9457SAndroid Build Coastguard Worker        local.get    1        # src
31*4bdc9457SAndroid Build Coastguard Worker        i32.load     0        # load float from src
32*4bdc9457SAndroid Build Coastguard Worker        local.set    4
33*4bdc9457SAndroid Build Coastguard Worker        local.get    1        # src
34*4bdc9457SAndroid Build Coastguard Worker        i32.load     4        # load 2nd float from src + 4
35*4bdc9457SAndroid Build Coastguard Worker        local.set    5
36*4bdc9457SAndroid Build Coastguard Worker
37*4bdc9457SAndroid Build Coastguard Worker        local.get    1        # src += 8
38*4bdc9457SAndroid Build Coastguard Worker        i32.const    8
39*4bdc9457SAndroid Build Coastguard Worker        i32.add
40*4bdc9457SAndroid Build Coastguard Worker        local.set    1
41*4bdc9457SAndroid Build Coastguard Worker
42*4bdc9457SAndroid Build Coastguard Worker        local.get    4        # (v >> 31) - 1) & v
43*4bdc9457SAndroid Build Coastguard Worker        i32.const    31
44*4bdc9457SAndroid Build Coastguard Worker        i32.shr_u
45*4bdc9457SAndroid Build Coastguard Worker        local.set    6
46*4bdc9457SAndroid Build Coastguard Worker        local.get    5        # 2nd mask
47*4bdc9457SAndroid Build Coastguard Worker        i32.const    31
48*4bdc9457SAndroid Build Coastguard Worker        i32.shr_u
49*4bdc9457SAndroid Build Coastguard Worker        local.set    7
50*4bdc9457SAndroid Build Coastguard Worker
51*4bdc9457SAndroid Build Coastguard Worker        local.get    6
52*4bdc9457SAndroid Build Coastguard Worker        i32.const    -1
53*4bdc9457SAndroid Build Coastguard Worker        i32.add
54*4bdc9457SAndroid Build Coastguard Worker        local.set    6
55*4bdc9457SAndroid Build Coastguard Worker        local.get    7
56*4bdc9457SAndroid Build Coastguard Worker        i32.const    -1
57*4bdc9457SAndroid Build Coastguard Worker        i32.add
58*4bdc9457SAndroid Build Coastguard Worker        local.set    7
59*4bdc9457SAndroid Build Coastguard Worker
60*4bdc9457SAndroid Build Coastguard Worker        local.get    4
61*4bdc9457SAndroid Build Coastguard Worker        local.get    6
62*4bdc9457SAndroid Build Coastguard Worker        i32.and
63*4bdc9457SAndroid Build Coastguard Worker        local.set    4
64*4bdc9457SAndroid Build Coastguard Worker        local.get    5
65*4bdc9457SAndroid Build Coastguard Worker        local.get    7
66*4bdc9457SAndroid Build Coastguard Worker        i32.and
67*4bdc9457SAndroid Build Coastguard Worker        local.set    5
68*4bdc9457SAndroid Build Coastguard Worker
69*4bdc9457SAndroid Build Coastguard Worker        local.get    2        # dst
70*4bdc9457SAndroid Build Coastguard Worker        local.get    4
71*4bdc9457SAndroid Build Coastguard Worker        i32.store    0        # store float
72*4bdc9457SAndroid Build Coastguard Worker        local.get    2        # dst
73*4bdc9457SAndroid Build Coastguard Worker        local.get    5
74*4bdc9457SAndroid Build Coastguard Worker        i32.store    4        # store 2nd float
75*4bdc9457SAndroid Build Coastguard Worker
76*4bdc9457SAndroid Build Coastguard Worker        local.get    2        # dst += 8
77*4bdc9457SAndroid Build Coastguard Worker        i32.const    8
78*4bdc9457SAndroid Build Coastguard Worker        i32.add
79*4bdc9457SAndroid Build Coastguard Worker        local.set    2
80*4bdc9457SAndroid Build Coastguard Worker
81*4bdc9457SAndroid Build Coastguard Worker        local.get    0
82*4bdc9457SAndroid Build Coastguard Worker        i32.const    -8
83*4bdc9457SAndroid Build Coastguard Worker        i32.add              # count -= 8
84*4bdc9457SAndroid Build Coastguard Worker        local.set    0
85*4bdc9457SAndroid Build Coastguard Worker
86*4bdc9457SAndroid Build Coastguard Worker        local.get    0
87*4bdc9457SAndroid Build Coastguard Worker        i32.const    8       # count >= 8
88*4bdc9457SAndroid Build Coastguard Worker        i32.ge_s
89*4bdc9457SAndroid Build Coastguard Worker        br_if        0       # loop
90*4bdc9457SAndroid Build Coastguard Worker      end_loop
91*4bdc9457SAndroid Build Coastguard Worker    end_if
92*4bdc9457SAndroid Build Coastguard Worker
93*4bdc9457SAndroid Build Coastguard Worker    local.get    0
94*4bdc9457SAndroid Build Coastguard Worker    i32.const    4       # if count >= 4
95*4bdc9457SAndroid Build Coastguard Worker    i32.ge_s
96*4bdc9457SAndroid Build Coastguard Worker    if
97*4bdc9457SAndroid Build Coastguard Worker      local.get    1        # src
98*4bdc9457SAndroid Build Coastguard Worker      i32.load     0        # load float from src
99*4bdc9457SAndroid Build Coastguard Worker      local.set    4
100*4bdc9457SAndroid Build Coastguard Worker
101*4bdc9457SAndroid Build Coastguard Worker      local.get    4        # (v >> 31) - 1) & v
102*4bdc9457SAndroid Build Coastguard Worker      i32.const    31
103*4bdc9457SAndroid Build Coastguard Worker      i32.shr_u
104*4bdc9457SAndroid Build Coastguard Worker      local.set    5
105*4bdc9457SAndroid Build Coastguard Worker
106*4bdc9457SAndroid Build Coastguard Worker      local.get    5
107*4bdc9457SAndroid Build Coastguard Worker      i32.const    -1
108*4bdc9457SAndroid Build Coastguard Worker      i32.add
109*4bdc9457SAndroid Build Coastguard Worker      local.set    5
110*4bdc9457SAndroid Build Coastguard Worker
111*4bdc9457SAndroid Build Coastguard Worker      local.get    4
112*4bdc9457SAndroid Build Coastguard Worker      local.get    5
113*4bdc9457SAndroid Build Coastguard Worker      i32.and
114*4bdc9457SAndroid Build Coastguard Worker      local.set    4
115*4bdc9457SAndroid Build Coastguard Worker
116*4bdc9457SAndroid Build Coastguard Worker      local.get    2        # dst
117*4bdc9457SAndroid Build Coastguard Worker      local.get    4
118*4bdc9457SAndroid Build Coastguard Worker      i32.store    0        # store float
119*4bdc9457SAndroid Build Coastguard Worker    end_if
120*4bdc9457SAndroid Build Coastguard WorkerEND_FUNCTION xnn_f32_vrelu_ukernel__wasm32_shr_x2
121