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_3x4_acc2(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_3x4_acc2(
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 == 2);
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 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_v64x2_load_splat(weights + 24);
45 
46   const v128_t vzero = wasm_f32x4_splat(0.0f);
47 
48   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
49 
50   const float* i0 = zero;
51   const float* i1 = zero;
52   const float* i2 = input;
53   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
54   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
55   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
56   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
57 
58   float* o0 = output;
59   float* o1 = (float*) ((uintptr_t) o0 + input_width);
60   float* o2 = (float*) ((uintptr_t) o1 + 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     }
75     if XNN_UNPREDICTABLE(output_height < 5) {
76       i6 = zero;
77     }
78 
79     v128_t vi0x0123 = vzero;
80     v128_t vi1x0123 = vzero;
81     v128_t vi2x0123 = vzero;
82     v128_t vi3x0123 = vzero;
83     v128_t vi4x0123 = vzero;
84     v128_t vi5x0123 = vzero;
85     v128_t vi6x0123 = vzero;
86 
87     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
88     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
89     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
90     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
91     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
92     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
93     v128_t vi6x4567 = wasm_v128_load(i6); i6 += 4;
94 
95     size_t w = input_width;
96     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
97       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
98       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
99       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
100 
101       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
102       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
103       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
104       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
105       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
106       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
107       const v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
108 
109       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
110       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
111       v128_t vo2p1 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
112 
113       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
114       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
115       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
116 
117       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
118       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
119       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
120 
121       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
122       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
123       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
124 
125       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
126       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
127       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
128 
129       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
130       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
131       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
132       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
133       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
134       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
135       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
136 
137       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
138       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
139       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
140 
141       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
142       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
143       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
144 
145       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
146       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
147       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
148 
149       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
150       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
151       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
152 
153       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
154       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
155       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
156 
157       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
158       vi0x0123 = vi0x4567;
159       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
160       vi1x0123 = vi1x4567;
161       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
162       vi2x0123 = vi2x4567;
163       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
164       vi3x0123 = vi3x4567;
165       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
166       vi4x0123 = vi4x4567;
167       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
168       vi5x0123 = vi5x4567;
169       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
170       vi6x0123 = vi6x4567;
171 
172       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
173       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
174       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
175 
176       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
177       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
178       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
179 
180       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
181       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
182       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
183 
184       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
185       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
186       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
187 
188       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
189       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
190       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
191 
192       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
193       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
194       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
195       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
196       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
197       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
198       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
199 
200       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
201       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
202       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
203 
204       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
205       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
206       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
207 
208       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
209       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
210       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
211 
212       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
213       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
214       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
215 
216       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
217       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
218       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
219 
220       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
221       vi0x4567 = vi0x89AB;
222       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
223       vi1x4567 = vi1x89AB;
224       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
225       vi2x4567 = vi2x89AB;
226       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
227       vi3x4567 = vi3x89AB;
228       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
229       vi4x4567 = vi4x89AB;
230       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
231       vi5x4567 = vi5x89AB;
232       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
233       vi6x4567 = vi6x89AB;
234 
235       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
236       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
237       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
238 
239       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
240       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
241       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
242 
243       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
244       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
245       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
246 
247       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
248       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
249       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
250 
251       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
252       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
253       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
254 
255       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
256       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
257       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
258 
259       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
260       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
261       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
262       vo0 = wasm_f32x4_min(vo0, vmax);
263       vo1 = wasm_f32x4_min(vo1, vmax);
264       vo2 = wasm_f32x4_min(vo2, vmax);
265 
266       wasm_v128_store(o2, vo2); o2 += 4;
267       wasm_v128_store(o1, vo1); o1 += 4;
268       wasm_v128_store(o0, vo0); o0 += 4;
269     }
270     // Always process the last block of 5..8 pixels.
271     if XNN_LIKELY(w > 4 * sizeof(float)) {
272       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
273       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
274       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
275 
276       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
277       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
278       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
279       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
280       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
281       v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
282       v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
283 
284       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
285       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
286       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
287       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
288       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
289       vi5x89AB = wasm_v128_and(vmask, vi5x89AB);
290       vi6x89AB = wasm_v128_and(vmask, vi6x89AB);
291 
292       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
293       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
294       v128_t vo2p1 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
295 
296       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
297       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
298       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
299 
300       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
301       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
302       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
303 
304       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
305       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
306       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
307 
308       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
309       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
310       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
311 
312       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
313       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
314       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
315       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
316       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
317       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
318       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
319 
320       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
321       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
322       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
323 
324       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
325       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
326       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
327 
328       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
329       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
330       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
331 
332       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
333       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
334       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
335 
336       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
337       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
338       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
339 
340       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
341       vi0x0123 = vi0x4567;
342       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
343       vi1x0123 = vi1x4567;
344       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
345       vi2x0123 = vi2x4567;
346       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
347       vi3x0123 = vi3x4567;
348       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
349       vi4x0123 = vi4x4567;
350       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
351       vi5x0123 = vi5x4567;
352       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
353       vi6x0123 = vi6x4567;
354 
355       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
356       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
357       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
358 
359       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
360       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
361       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
362 
363       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
364       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
365       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
366 
367       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
368       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
369       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
370 
371       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
372       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
373       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
374 
375       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
376       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
377       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
378       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
379       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
380       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
381       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
382 
383       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
384       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
385       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
386 
387       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
388       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
389       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
390 
391       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
392       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
393       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
394 
395       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
396       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
397       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
398 
399       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
400       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
401       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
402 
403       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
404       vi0x4567 = vi0x89AB;
405       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
406       vi1x4567 = vi1x89AB;
407       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
408       vi2x4567 = vi2x89AB;
409       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
410       vi3x4567 = vi3x89AB;
411       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
412       vi4x4567 = vi4x89AB;
413       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
414       vi5x4567 = vi5x89AB;
415       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
416       vi6x4567 = vi6x89AB;
417 
418       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
419       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
420       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
421 
422       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
423       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
424       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
425 
426       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
427       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
428       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
429 
430       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
431       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
432       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
433 
434       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
435       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
436       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
437 
438       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
439       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
440       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
441 
442       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
443       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
444       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
445       vo0 = wasm_f32x4_min(vo0, vmax);
446       vo1 = wasm_f32x4_min(vo1, vmax);
447       vo2 = wasm_f32x4_min(vo2, vmax);
448 
449       wasm_v128_store(o2, vo2); o2 += 4;
450       wasm_v128_store(o1, vo1); o1 += 4;
451       wasm_v128_store(o0, vo0); o0 += 4;
452 
453       w -= 4 * sizeof(float);
454     }
455     assert(w >= 1 * sizeof(float));
456     assert(w <= 4 * sizeof(float));
457     {
458       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
459       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
460       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
461 
462       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
463       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
464       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
465       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
466       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
467       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
468       vi6x4567 = wasm_v128_and(vmask, vi6x4567);
469 
470       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
471       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
472       v128_t vo2p1 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
473 
474       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
475       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
476       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
477 
478       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
479       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
480       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
481 
482       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
483       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
484       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
485 
486       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
487       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
488       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
489 
490       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
491       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
492       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
493       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
494       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
495       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
496       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
497 
498       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
499       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
500       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
501 
502       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
503       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
504       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
505 
506       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
507       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
508       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
509 
510       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
511       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
512       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
513 
514       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
515       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
516       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
517 
518       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
519       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
520       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
521       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
522       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
523       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
524       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
525 
526       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
527       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
528       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
529 
530       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
531       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
532       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
533 
534       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
535       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
536       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
537 
538       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
539       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
540       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
541 
542       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
543       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
544       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
545 
546       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
547       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
548       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
549       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
550       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
551       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
552       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4);
553 
554       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
555       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
556       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
557 
558       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
559       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
560       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
561 
562       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
563       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
564       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
565 
566       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
567       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
568       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
569 
570       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
571       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
572       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
573 
574       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
575       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
576       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
577       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
578       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
579       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4);
580       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4);
581 
582       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
583       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
584       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
585 
586       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
587       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
588       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
589 
590       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
591       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
592       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
593 
594       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
595       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
596       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
597 
598       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
599       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
600       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
601 
602       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
603       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
604       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
605 
606       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
607       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
608       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
609       vo0 = wasm_f32x4_min(vo0, vmax);
610       vo1 = wasm_f32x4_min(vo1, vmax);
611       vo2 = wasm_f32x4_min(vo2, vmax);
612 
613       if XNN_LIKELY(w & (4 * sizeof(float))) {
614         wasm_v128_store(o2, vo2); o2 += 4;
615         wasm_v128_store(o1, vo1); o1 += 4;
616         wasm_v128_store(o0, vo0); o0 += 4;
617       } else {
618         if (w & (2 * sizeof(float))) {
619           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
620           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
621           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
622 
623           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
624           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
625           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
626         }
627         if (w & (1 * sizeof(float))) {
628           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
629           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
630           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
631         }
632       }
633     }
634 
635     i0 = (const float*) ((uintptr_t) i3 - input_decrement);
636     i1 = (const float*) ((uintptr_t) i4 - input_decrement);
637     i2 = (const float*) ((uintptr_t) i1 + input_width);
638     i3 = (const float*) ((uintptr_t) i2 + input_width);
639     i4 = (const float*) ((uintptr_t) i3 + input_width);
640     i5 = (const float*) ((uintptr_t) i4 + input_width);
641     i6 = (const float*) ((uintptr_t) i5 + input_width);
642 
643     o0 = o2;
644     o1 = (float*) ((uintptr_t) o0 + input_width);
645     o2 = (float*) ((uintptr_t) o1 + input_width);
646 
647     output_height = doz(output_height, 3);
648   } while (output_height != 0);
649 }
650