• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5p2-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_5x5p2__wasmsimd_arm_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_5x5p2__wasmsimd_arm_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)]) XNN_OOB_READS
28 {
29   assert(input_height != 0);
30   assert(input_width != 0);
31   assert(input_width % sizeof(float) == 0);
32   assert(padding_top == 2);
33 
34   const v128_t vmask = wasm_v128_load(params->scalar.mask);
35   const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
36   const v128_t vmin = wasm_v128_load32_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 vw89AB = wasm_v128_load(weights + 8);
41   const v128_t vwCDEF = wasm_v128_load(weights + 12);
42   const v128_t vwGHIJ = wasm_v128_load(weights + 16);
43   const v128_t vwKLMN = wasm_v128_load(weights + 20);
44   const v128_t vwOP = wasm_v128_load64_splat(weights + 24);
45 
46   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
47 
48   const float* i0 = zero;
49   const float* i1 = zero;
50   const float* i2 = input;
51   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
52   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
53   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
54   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
55   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
56 
57   float* o0 = output;
58   float* o1 = (float*) ((uintptr_t) o0 + input_width);
59   float* o2 = (float*) ((uintptr_t) o1 + input_width);
60   float* o3 = (float*) ((uintptr_t) o2 + input_width);
61 
62   size_t output_height = input_height;
63   do {
64     if XNN_UNPREDICTABLE(output_height < 2) {
65       i3 = zero;
66       o1 = o0;
67     }
68     if XNN_UNPREDICTABLE(output_height < 3) {
69       i4 = zero;
70       o2 = o1;
71     }
72     if XNN_UNPREDICTABLE(output_height < 4) {
73       i5 = zero;
74       o3 = o2;
75     }
76     if XNN_UNPREDICTABLE(output_height < 5) {
77       i6 = zero;
78     }
79     if XNN_UNPREDICTABLE(output_height < 6) {
80       i7 = zero;
81     }
82 
83     v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
84     v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
85     v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
86     v128_t vi3x0123 = wasm_f32x4_const_splat(0.0f);
87     v128_t vi4x0123 = wasm_f32x4_const_splat(0.0f);
88     v128_t vi5x0123 = wasm_f32x4_const_splat(0.0f);
89     v128_t vi6x0123 = wasm_f32x4_const_splat(0.0f);
90     v128_t vi7x0123 = wasm_f32x4_const_splat(0.0f);
91 
92     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
93     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
94     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
95     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
96     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
97     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
98     v128_t vi6x4567 = wasm_v128_load(i6); i6 += 4;
99     v128_t vi7x4567 = wasm_v128_load(i7); i7 += 4;
100 
101     size_t w = input_width;
102     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
103       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
104       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
105       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
106       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
107 
108       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
109       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
110       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
111       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
112       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
113       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
114       const v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
115       const v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
116 
117       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
118       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
119       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
120       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
121 
122       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
123       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
124       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
125       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
126 
127       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
128       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
129       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
130       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
131 
132       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
133       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
134       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
135       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
136 
137       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
138       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
139       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
140       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
141 
142       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
143       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
144       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
145       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
146       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
147       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
148       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
149       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
150 
151       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
152       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
153       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
154       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
155 
156       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
157       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
158       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
159       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
160 
161       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
162       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
163       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
164       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
165 
166       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
167       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
168       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
169       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
170 
171       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
172       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
173       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
174       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
175 
176       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
177       vi0x0123 = vi0x4567;
178       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
179       vi1x0123 = vi1x4567;
180       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
181       vi2x0123 = vi2x4567;
182       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
183       vi3x0123 = vi3x4567;
184       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
185       vi4x0123 = vi4x4567;
186       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
187       vi5x0123 = vi5x4567;
188       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
189       vi6x0123 = vi6x4567;
190       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
191       vi7x0123 = vi7x4567;
192 
193       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
194       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
195       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
196       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
197 
198       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
199       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
200       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
201       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
202 
203       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
204       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
205       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
206       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
207 
208       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
209       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
210       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
211       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
212 
213       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
214       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
215       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
216       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
217 
218       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
219       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
220       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
221       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
222       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
223       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
224       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
225       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
226 
227       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
228       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
229       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
230       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
231 
232       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
233       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
234       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
235       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
236 
237       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
238       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
239       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
240       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
241 
242       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
243       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
244       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
245       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
246 
247       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
248       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
249       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
250       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
251 
252       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
253       vi0x4567 = vi0x89AB;
254       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
255       vi1x4567 = vi1x89AB;
256       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
257       vi2x4567 = vi2x89AB;
258       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
259       vi3x4567 = vi3x89AB;
260       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
261       vi4x4567 = vi4x89AB;
262       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
263       vi5x4567 = vi5x89AB;
264       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
265       vi6x4567 = vi6x89AB;
266       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
267       vi7x4567 = vi7x89AB;
268 
269       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
270       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
271       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
272       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
273 
274       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
275       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
276       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
277       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
278 
279       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
280       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
281       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
282       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
283 
284       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
285       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
286       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
287       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
288 
289       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
290       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
291       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
292       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
293 
294 
295       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
296       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
297       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
298       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
299       vo0 = wasm_f32x4_min(vo0, vmax);
300       vo1 = wasm_f32x4_min(vo1, vmax);
301       vo2 = wasm_f32x4_min(vo2, vmax);
302       vo3 = wasm_f32x4_min(vo3, vmax);
303 
304       wasm_v128_store(o3, vo3); o3 += 4;
305       wasm_v128_store(o2, vo2); o2 += 4;
306       wasm_v128_store(o1, vo1); o1 += 4;
307       wasm_v128_store(o0, vo0); o0 += 4;
308     }
309     // Always process the last block of 5..8 pixels.
310     if XNN_LIKELY(w > 4 * sizeof(float)) {
311       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
312       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
313       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
314       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
315 
316       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
317       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
318       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
319       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
320       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
321       v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
322       v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
323       v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
324 
325       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
326       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
327       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
328       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
329       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
330       vi5x89AB = wasm_v128_and(vmask, vi5x89AB);
331       vi6x89AB = wasm_v128_and(vmask, vi6x89AB);
332       vi7x89AB = wasm_v128_and(vmask, vi7x89AB);
333 
334       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
335       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
336       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
337       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
338 
339       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
340       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
341       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
342       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
343 
344       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
345       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
346       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
347       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
348 
349       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
350       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
351       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
352       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
353 
354       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
355       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
356       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
357       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
358 
359       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
360       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
361       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
362       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
363       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
364       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
365       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
366       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
367 
368       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
369       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
370       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
371       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
372 
373       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
374       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
375       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
376       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
377 
378       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
379       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
380       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
381       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
382 
383       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
384       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
385       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
386       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
387 
388       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
389       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
390       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
391       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
392 
393       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
394       vi0x0123 = vi0x4567;
395       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
396       vi1x0123 = vi1x4567;
397       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
398       vi2x0123 = vi2x4567;
399       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
400       vi3x0123 = vi3x4567;
401       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
402       vi4x0123 = vi4x4567;
403       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
404       vi5x0123 = vi5x4567;
405       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
406       vi6x0123 = vi6x4567;
407       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
408       vi7x0123 = vi7x4567;
409 
410       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
411       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
412       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
413       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
414 
415       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
416       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
417       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
418       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
419 
420       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
421       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
422       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
423       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
424 
425       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
426       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
427       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
428       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
429 
430       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
431       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
432       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
433       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
434 
435       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
436       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
437       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
438       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
439       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
440       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
441       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
442       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
443 
444       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
445       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
446       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
447       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
448 
449       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
450       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
451       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
452       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
453 
454       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
455       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
456       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
457       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
458 
459       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
460       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
461       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
462       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
463 
464       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
465       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
466       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
467       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
468 
469       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
470       vi0x4567 = vi0x89AB;
471       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
472       vi1x4567 = vi1x89AB;
473       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
474       vi2x4567 = vi2x89AB;
475       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
476       vi3x4567 = vi3x89AB;
477       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
478       vi4x4567 = vi4x89AB;
479       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
480       vi5x4567 = vi5x89AB;
481       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
482       vi6x4567 = vi6x89AB;
483       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
484       vi7x4567 = vi7x89AB;
485 
486       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
487       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
488       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
489       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
490 
491       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
492       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
493       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
494       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
495 
496       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
497       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
498       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
499       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
500 
501       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
502       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
503       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
504       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
505 
506       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
507       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
508       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
509       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
510 
511 
512       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
513       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
514       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
515       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
516       vo0 = wasm_f32x4_min(vo0, vmax);
517       vo1 = wasm_f32x4_min(vo1, vmax);
518       vo2 = wasm_f32x4_min(vo2, vmax);
519       vo3 = wasm_f32x4_min(vo3, vmax);
520 
521       wasm_v128_store(o3, vo3); o3 += 4;
522       wasm_v128_store(o2, vo2); o2 += 4;
523       wasm_v128_store(o1, vo1); o1 += 4;
524       wasm_v128_store(o0, vo0); o0 += 4;
525 
526       w -= 4 * sizeof(float);
527     }
528     assert(w >= 1 * sizeof(float));
529     assert(w <= 4 * sizeof(float));
530     {
531       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
532       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
533       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
534       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
535 
536       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
537       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
538       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
539       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
540       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
541       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
542       vi6x4567 = wasm_v128_and(vmask, vi6x4567);
543       vi7x4567 = wasm_v128_and(vmask, vi7x4567);
544 
545       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
546       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
547       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
548       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
549 
550       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
551       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
552       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
553       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
554 
555       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
556       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
557       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
558       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
559 
560       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
561       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
562       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
563       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
564 
565       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
566       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
567       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
568       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
569 
570       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
571       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
572       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
573       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
574       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
575       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
576       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
577       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
578 
579       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
580       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
581       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
582       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
583 
584       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
585       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
586       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
587       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
588 
589       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
590       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
591       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
592       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
593 
594       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
595       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
596       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
597       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
598 
599       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
600       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
601       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
602       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
603 
604       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
605       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
606       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
607       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
608       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
609       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
610       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
611       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
612 
613       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
614       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
615       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
616       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
617 
618       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
619       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
620       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
621       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
622 
623       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
624       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
625       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
626       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
627 
628       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
629       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
630       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
631       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
632 
633       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
634       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
635       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
636       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
637 
638       const v128_t vzero = wasm_f32x4_const_splat(0.0f);
639       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
640       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
641       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
642       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
643       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
644       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
645       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4);
646       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4);
647 
648       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
649       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
650       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
651       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
652 
653       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
654       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
655       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
656       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
657 
658       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
659       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
660       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
661       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
662 
663       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
664       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
665       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
666       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
667 
668       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
669       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
670       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
671       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
672 
673       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
674       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
675       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
676       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
677       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
678       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4);
679       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4);
680       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x5678, vzero, 1, 2, 3, 4);
681 
682       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
683       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
684       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
685       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
686 
687       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
688       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
689       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
690       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
691 
692       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
693       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
694       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
695       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
696 
697       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
698       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
699       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
700       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
701 
702       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
703       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
704       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
705       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
706 
707 
708       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
709       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
710       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
711       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
712       vo0 = wasm_f32x4_min(vo0, vmax);
713       vo1 = wasm_f32x4_min(vo1, vmax);
714       vo2 = wasm_f32x4_min(vo2, vmax);
715       vo3 = wasm_f32x4_min(vo3, vmax);
716 
717       if XNN_LIKELY(w & (4 * sizeof(float))) {
718         wasm_v128_store(o3, vo3); o3 += 4;
719         wasm_v128_store(o2, vo2); o2 += 4;
720         wasm_v128_store(o1, vo1); o1 += 4;
721         wasm_v128_store(o0, vo0); o0 += 4;
722       } else {
723         if (w & (2 * sizeof(float))) {
724           *((double*) o3) = wasm_f64x2_extract_lane(vo3, 0); o3 += 2;
725           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
726           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
727           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
728 
729           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
730           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
731           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
732           vo3 = wasm_v32x4_shuffle(vo3, vo3, 2, 3, 0, 1);
733         }
734         if (w & (1 * sizeof(float))) {
735           *o3 = wasm_f32x4_extract_lane(vo3, 0); o3 += 1;
736           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
737           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
738           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
739         }
740       }
741     }
742 
743     i0 = (const float*) ((uintptr_t) i4 - input_decrement);
744     i1 = (const float*) ((uintptr_t) i5 - input_decrement);
745     i2 = (const float*) ((uintptr_t) i1 + input_width);
746     i3 = (const float*) ((uintptr_t) i2 + input_width);
747     i4 = (const float*) ((uintptr_t) i3 + input_width);
748     i5 = (const float*) ((uintptr_t) i4 + input_width);
749     i6 = (const float*) ((uintptr_t) i5 + input_width);
750     i7 = (const float*) ((uintptr_t) i6 + input_width);
751 
752     o0 = o3;
753     o1 = (float*) ((uintptr_t) o0 + input_width);
754     o2 = (float*) ((uintptr_t) o1 + input_width);
755     o3 = (float*) ((uintptr_t) o2 + input_width);
756 
757     output_height = doz(output_height, 4);
758   } while (output_height != 0);
759 }
760