• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/3x3p1-wasmsimd-splat.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <wasm_simd128.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16 
17 
18 
xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4(
20     size_t input_height,
21     size_t input_width,
22     const float* input,
23     const float* weights,
24     const float* zero,
25     float* output,
26     uint32_t padding_top,
27     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(input_height != 0);
30   assert(input_width != 0);
31   assert(input_width % sizeof(float) == 0);
32   assert(padding_top == 1);
33 
34   const v128_t vmask = wasm_v128_load(params->scalar.mask);
35   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
36   const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
37 
38   const v128_t vw0123 = wasm_v128_load(weights);
39   const v128_t vw4567 = wasm_v128_load(weights + 4);
40   const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);
41 
42   const v128_t vzero = wasm_f32x4_splat(0.0f);
43 
44   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
45 
46   const float* i0 = zero;
47   const float* i1 = input;
48   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
49   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
50   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
51   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
52 
53   float* o0 = output;
54   float* o1 = (float*) ((uintptr_t) o0 + input_width);
55   float* o2 = (float*) ((uintptr_t) o1 + input_width);
56   float* o3 = (float*) ((uintptr_t) o2 + input_width);
57 
58   size_t output_height = input_height;
59   do {
60     if XNN_UNPREDICTABLE(output_height < 2) {
61       i2 = zero;
62       o1 = o0;
63     }
64     if XNN_UNPREDICTABLE(output_height < 3) {
65       i3 = zero;
66       o2 = o1;
67     }
68     if XNN_UNPREDICTABLE(output_height < 4) {
69       i4 = zero;
70       o3 = o2;
71     }
72     if XNN_UNPREDICTABLE(output_height < 5) {
73       i5 = zero;
74     }
75 
76     v128_t vi0x0123 = vzero;
77     v128_t vi1x0123 = vzero;
78     v128_t vi2x0123 = vzero;
79     v128_t vi3x0123 = vzero;
80     v128_t vi4x0123 = vzero;
81     v128_t vi5x0123 = vzero;
82 
83     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
84     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
85     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
86     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
87     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
88     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
89 
90     size_t w = input_width;
91     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
92       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
93       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
94       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
95       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
96 
97       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
98       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
99       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
100       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
101       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
102       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
103 
104       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
105       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
106       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
107       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
108 
109       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
110       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
111       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
112       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
113 
114       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
115       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
116       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
117       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
118 
119       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
120       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
121       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
122       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
123       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
124       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
125 
126       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
127       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
128       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
129       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
130 
131       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
132       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
133       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
134       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
135 
136       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
137       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
138       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
139       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
140 
141       vi0x0123 = vi0x4567;
142       vi1x0123 = vi1x4567;
143       vi2x0123 = vi2x4567;
144       vi3x0123 = vi3x4567;
145       vi4x0123 = vi4x4567;
146       vi5x0123 = vi5x4567;
147 
148       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
149       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
150       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
151       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
152       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
153       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
154 
155       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
156       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
157       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
158       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
159 
160       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
161       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
162       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
163       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
164 
165       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
166       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
167       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
168       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
169 
170       vi0x4567 = vi0x89AB;
171       vi1x4567 = vi1x89AB;
172       vi2x4567 = vi2x89AB;
173       vi3x4567 = vi3x89AB;
174       vi4x4567 = vi4x89AB;
175       vi5x4567 = vi5x89AB;
176 
177 
178       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
179       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
180       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
181       v128_t vo3 = wasm_v128_bitselect(vmin, vo3p0, wasm_f32x4_lt(vo3p0, vmin));
182       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
183       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
184       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));
185       vo3 = wasm_v128_bitselect(vo3, vmax, wasm_f32x4_le(vo3, vmax));
186 
187       wasm_v128_store(o3, vo3); o3 += 4;
188       wasm_v128_store(o2, vo2); o2 += 4;
189       wasm_v128_store(o1, vo1); o1 += 4;
190       wasm_v128_store(o0, vo0); o0 += 4;
191     }
192     // Always process the last block of 1..4 pixels.
193     assert(w >= 1 * sizeof(float));
194     assert(w <= 4 * sizeof(float));
195     {
196       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
197       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
198       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
199       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
200 
201       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
202       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
203       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
204       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
205       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
206       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
207 
208       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
209       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
210       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
211       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
212 
213       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
214       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
215       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
216       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
217 
218       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
219       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
220       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
221       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
222 
223       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
224       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
225       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
226       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
227       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
228       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
229 
230       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
231       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
232       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
233       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
234 
235       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
236       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
237       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
238       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
239 
240       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
241       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
242       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
243       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
244 
245       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
246       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
247       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
248       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
249       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
250       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
251 
252       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
253       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
254       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
255       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
256 
257       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
258       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
259       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
260       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
261 
262       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
263       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
264       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
265       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
266 
267 
268       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
269       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
270       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
271       v128_t vo3 = wasm_v128_bitselect(vmin, vo3p0, wasm_f32x4_lt(vo3p0, vmin));
272       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
273       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
274       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));
275       vo3 = wasm_v128_bitselect(vo3, vmax, wasm_f32x4_le(vo3, vmax));
276 
277       if XNN_LIKELY(w == 4 * sizeof(float)) {
278         wasm_v128_store(o3, vo3); o3 += 4;
279         wasm_v128_store(o2, vo2); o2 += 4;
280         wasm_v128_store(o1, vo1); o1 += 4;
281         wasm_v128_store(o0, vo0); o0 += 4;
282       } else {
283         if (w & (2 * sizeof(float))) {
284           *((double*) o3) = wasm_f64x2_extract_lane(vo3, 0); o3 += 2;
285           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
286           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
287           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
288 
289           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
290           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
291           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
292           vo3 = wasm_v32x4_shuffle(vo3, vo3, 2, 3, 0, 1);
293         }
294         if (w & (1 * sizeof(float))) {
295           *o3 = wasm_f32x4_extract_lane(vo3, 0); o3 += 1;
296           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
297           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
298           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
299         }
300       }
301     }
302 
303     i0 = (const float*) ((uintptr_t) i4 - input_decrement);
304     i1 = (const float*) ((uintptr_t) i5 - input_decrement);
305     i2 = (const float*) ((uintptr_t) i1 + input_width);
306     i3 = (const float*) ((uintptr_t) i2 + input_width);
307     i4 = (const float*) ((uintptr_t) i3 + input_width);
308     i5 = (const float*) ((uintptr_t) i4 + input_width);
309 
310     o0 = o3;
311     o1 = (float*) ((uintptr_t) o0 + input_width);
312     o2 = (float*) ((uintptr_t) o1 + input_width);
313     o3 = (float*) ((uintptr_t) o2 + input_width);
314 
315     output_height = doz(output_height, 4);
316   } while (output_height != 0);
317 }
318