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_5x4(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_5x4(
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   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
58   const float* i8 = (const float*) ((uintptr_t) i7 + input_width);
59 
60   float* o0 = output;
61   float* o1 = (float*) ((uintptr_t) o0 + input_width);
62   float* o2 = (float*) ((uintptr_t) o1 + input_width);
63   float* o3 = (float*) ((uintptr_t) o2 + input_width);
64   float* o4 = (float*) ((uintptr_t) o3 + input_width);
65 
66   size_t output_height = input_height;
67   do {
68     if XNN_UNPREDICTABLE(output_height < 2) {
69       i3 = zero;
70       o1 = o0;
71     }
72     if XNN_UNPREDICTABLE(output_height < 3) {
73       i4 = zero;
74       o2 = o1;
75     }
76     if XNN_UNPREDICTABLE(output_height < 4) {
77       i5 = zero;
78       o3 = o2;
79     }
80     if XNN_UNPREDICTABLE(output_height < 5) {
81       i6 = zero;
82       o4 = o3;
83     }
84     if XNN_UNPREDICTABLE(output_height < 6) {
85       i7 = zero;
86     }
87     if XNN_UNPREDICTABLE(output_height < 7) {
88       i8 = zero;
89     }
90 
91     v128_t vi0x0123 = vzero;
92     v128_t vi1x0123 = vzero;
93     v128_t vi2x0123 = vzero;
94     v128_t vi3x0123 = vzero;
95     v128_t vi4x0123 = vzero;
96     v128_t vi5x0123 = vzero;
97     v128_t vi6x0123 = vzero;
98     v128_t vi7x0123 = vzero;
99     v128_t vi8x0123 = vzero;
100 
101     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
102     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
103     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
104     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
105     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
106     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
107     v128_t vi6x4567 = wasm_v128_load(i6); i6 += 4;
108     v128_t vi7x4567 = wasm_v128_load(i7); i7 += 4;
109     v128_t vi8x4567 = wasm_v128_load(i8); i8 += 4;
110 
111     size_t w = input_width;
112     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
113       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
114       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
115       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
116       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
117       v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
118 
119       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
120       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
121       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
122       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
123       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
124       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
125       const v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
126       const v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
127       const v128_t vi8x89AB = wasm_v128_load(i8); i8 += 4;
128 
129       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
130       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
131       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
132       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
133       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
134 
135       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
136       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
137       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
138       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
139       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
140 
141       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
142       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
143       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
144       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
145       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
146 
147       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
148       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
149       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
150       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
151       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
152 
153       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
154       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
155       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
156       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
157       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
158 
159       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
160       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
161       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
162       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
163       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
164       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
165       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
166       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
167       const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6);
168 
169       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
170       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
171       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
172       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
173       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
174 
175       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
176       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
177       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
178       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
179       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
180 
181       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
182       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
183       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
184       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
185       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
186 
187       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
188       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
189       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
190       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
191       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
192 
193       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
194       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
195       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
196       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
197       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
198 
199       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
200       vi0x0123 = vi0x4567;
201       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
202       vi1x0123 = vi1x4567;
203       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
204       vi2x0123 = vi2x4567;
205       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
206       vi3x0123 = vi3x4567;
207       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
208       vi4x0123 = vi4x4567;
209       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
210       vi5x0123 = vi5x4567;
211       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
212       vi6x0123 = vi6x4567;
213       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
214       vi7x0123 = vi7x4567;
215       const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5);
216       vi8x0123 = vi8x4567;
217 
218       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
219       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
220       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
221       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
222       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
223 
224       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
225       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
226       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
227       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
228       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
229 
230       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
231       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
232       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
233       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
234       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
235 
236       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
237       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
238       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
239       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
240       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
241 
242       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
243       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
244       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
245       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
246       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
247 
248       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
249       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
250       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
251       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
252       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
253       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
254       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
255       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
256       const v128_t vi8x5678 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 1, 2, 3, 4);
257 
258       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
259       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
260       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
261       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
262       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
263 
264       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
265       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
266       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
267       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
268       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
269 
270       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
271       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
272       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
273       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
274       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
275 
276       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
277       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
278       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
279       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
280       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
281 
282       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
283       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
284       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
285       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
286       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
287 
288       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
289       vi0x4567 = vi0x89AB;
290       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
291       vi1x4567 = vi1x89AB;
292       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
293       vi2x4567 = vi2x89AB;
294       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
295       vi3x4567 = vi3x89AB;
296       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
297       vi4x4567 = vi4x89AB;
298       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
299       vi5x4567 = vi5x89AB;
300       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
301       vi6x4567 = vi6x89AB;
302       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
303       vi7x4567 = vi7x89AB;
304       const v128_t vi8x6789 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 2, 3, 4, 5);
305       vi8x4567 = vi8x89AB;
306 
307       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
308       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
309       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
310       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
311       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
312 
313       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
314       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
315       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
316       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
317       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
318 
319       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
320       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
321       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
322       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
323       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
324 
325       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
326       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
327       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
328       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
329       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
330 
331       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
332       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
333       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
334       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
335       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
336 
337 
338       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
339       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
340       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
341       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
342       v128_t vo4 = wasm_f32x4_max(vo4p0, vmin);
343       vo0 = wasm_f32x4_min(vo0, vmax);
344       vo1 = wasm_f32x4_min(vo1, vmax);
345       vo2 = wasm_f32x4_min(vo2, vmax);
346       vo3 = wasm_f32x4_min(vo3, vmax);
347       vo4 = wasm_f32x4_min(vo4, vmax);
348 
349       wasm_v128_store(o4, vo4); o4 += 4;
350       wasm_v128_store(o3, vo3); o3 += 4;
351       wasm_v128_store(o2, vo2); o2 += 4;
352       wasm_v128_store(o1, vo1); o1 += 4;
353       wasm_v128_store(o0, vo0); o0 += 4;
354     }
355     // Always process the last block of 5..8 pixels.
356     if XNN_LIKELY(w > 4 * sizeof(float)) {
357       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
358       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
359       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
360       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
361       v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
362 
363       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
364       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
365       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
366       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
367       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
368       v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
369       v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
370       v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
371       v128_t vi8x89AB = wasm_v128_load(i8); i8 += 4;
372 
373       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
374       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
375       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
376       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
377       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
378       vi5x89AB = wasm_v128_and(vmask, vi5x89AB);
379       vi6x89AB = wasm_v128_and(vmask, vi6x89AB);
380       vi7x89AB = wasm_v128_and(vmask, vi7x89AB);
381       vi8x89AB = wasm_v128_and(vmask, vi8x89AB);
382 
383       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
384       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
385       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
386       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
387       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
388 
389       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
390       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
391       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
392       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
393       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
394 
395       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
396       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
397       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
398       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
399       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
400 
401       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
402       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
403       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
404       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
405       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
406 
407       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
408       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
409       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
410       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
411       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
412 
413       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
414       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
415       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
416       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
417       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
418       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
419       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
420       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
421       const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6);
422 
423       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
424       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
425       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
426       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
427       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
428 
429       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
430       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
431       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
432       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
433       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
434 
435       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
436       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
437       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
438       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
439       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
440 
441       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
442       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
443       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
444       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
445       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
446 
447       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
448       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
449       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
450       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
451       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
452 
453       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
454       vi0x0123 = vi0x4567;
455       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
456       vi1x0123 = vi1x4567;
457       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
458       vi2x0123 = vi2x4567;
459       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
460       vi3x0123 = vi3x4567;
461       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
462       vi4x0123 = vi4x4567;
463       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
464       vi5x0123 = vi5x4567;
465       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
466       vi6x0123 = vi6x4567;
467       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
468       vi7x0123 = vi7x4567;
469       const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5);
470       vi8x0123 = vi8x4567;
471 
472       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
473       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
474       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
475       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
476       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
477 
478       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
479       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
480       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
481       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
482       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
483 
484       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
485       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
486       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
487       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
488       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
489 
490       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
491       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
492       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
493       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
494       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
495 
496       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
497       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
498       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
499       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
500       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
501 
502       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
503       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
504       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
505       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
506       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
507       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
508       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
509       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
510       const v128_t vi8x5678 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 1, 2, 3, 4);
511 
512       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
513       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
514       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
515       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
516       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
517 
518       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
519       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
520       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
521       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
522       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
523 
524       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
525       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
526       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
527       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
528       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
529 
530       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
531       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
532       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
533       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
534       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
535 
536       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
537       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
538       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
539       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
540       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
541 
542       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
543       vi0x4567 = vi0x89AB;
544       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
545       vi1x4567 = vi1x89AB;
546       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
547       vi2x4567 = vi2x89AB;
548       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
549       vi3x4567 = vi3x89AB;
550       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
551       vi4x4567 = vi4x89AB;
552       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
553       vi5x4567 = vi5x89AB;
554       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
555       vi6x4567 = vi6x89AB;
556       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
557       vi7x4567 = vi7x89AB;
558       const v128_t vi8x6789 = wasm_v32x4_shuffle(vi8x4567, vi8x89AB, 2, 3, 4, 5);
559       vi8x4567 = vi8x89AB;
560 
561       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
562       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
563       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
564       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
565       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
566 
567       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
568       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
569       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
570       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
571       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
572 
573       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
574       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
575       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
576       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
577       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
578 
579       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
580       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
581       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
582       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
583       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
584 
585       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
586       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
587       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
588       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
589       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
590 
591 
592       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
593       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
594       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
595       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
596       v128_t vo4 = wasm_f32x4_max(vo4p0, vmin);
597       vo0 = wasm_f32x4_min(vo0, vmax);
598       vo1 = wasm_f32x4_min(vo1, vmax);
599       vo2 = wasm_f32x4_min(vo2, vmax);
600       vo3 = wasm_f32x4_min(vo3, vmax);
601       vo4 = wasm_f32x4_min(vo4, vmax);
602 
603       wasm_v128_store(o4, vo4); o4 += 4;
604       wasm_v128_store(o3, vo3); o3 += 4;
605       wasm_v128_store(o2, vo2); o2 += 4;
606       wasm_v128_store(o1, vo1); o1 += 4;
607       wasm_v128_store(o0, vo0); o0 += 4;
608 
609       w -= 4 * sizeof(float);
610     }
611     assert(w >= 1 * sizeof(float));
612     assert(w <= 4 * sizeof(float));
613     {
614       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
615       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
616       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
617       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
618       v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
619 
620       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
621       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
622       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
623       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
624       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
625       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
626       vi6x4567 = wasm_v128_and(vmask, vi6x4567);
627       vi7x4567 = wasm_v128_and(vmask, vi7x4567);
628       vi8x4567 = wasm_v128_and(vmask, vi8x4567);
629 
630       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
631       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
632       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
633       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
634       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
635 
636       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
637       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
638       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
639       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
640       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
641 
642       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
643       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
644       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
645       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
646       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
647 
648       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
649       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
650       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
651       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
652       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
653 
654       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
655       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
656       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
657       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
658       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
659 
660       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
661       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
662       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
663       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
664       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
665       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
666       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
667       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
668       const v128_t vi8x3456 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 3, 4, 5, 6);
669 
670       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
671       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
672       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
673       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
674       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
675 
676       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
677       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
678       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
679       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
680       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
681 
682       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
683       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
684       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
685       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
686       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
687 
688       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
689       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
690       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
691       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
692       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
693 
694       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
695       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
696       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
697       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
698       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
699 
700       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
701       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
702       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
703       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
704       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
705       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
706       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
707       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
708       const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5);
709 
710       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
711       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
712       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
713       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
714       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
715 
716       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
717       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
718       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
719       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
720       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
721 
722       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
723       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
724       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
725       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
726       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
727 
728       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
729       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
730       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
731       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
732       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
733 
734       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
735       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
736       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
737       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
738       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
739 
740       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
741       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
742       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
743       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
744       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
745       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
746       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4);
747       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4);
748       const v128_t vi8x5678 = wasm_v32x4_shuffle(vi8x4567, vzero, 1, 2, 3, 4);
749 
750       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
751       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
752       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
753       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
754       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
755 
756       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
757       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
758       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
759       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
760       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
761 
762       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
763       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
764       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
765       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
766       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
767 
768       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
769       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
770       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
771       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
772       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
773 
774       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
775       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
776       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
777       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
778       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x5678, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
779 
780       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
781       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
782       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
783       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
784       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
785       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4);
786       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4);
787       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x5678, vzero, 1, 2, 3, 4);
788       const v128_t vi8x6789 = wasm_v32x4_shuffle(vi8x5678, vzero, 1, 2, 3, 4);
789 
790       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
791       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
792       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
793       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
794       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
795 
796       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
797       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
798       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
799       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
800       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
801 
802       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
803       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
804       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
805       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
806       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
807 
808       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
809       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
810       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
811       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
812       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
813 
814       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
815       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
816       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
817       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
818       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
819 
820 
821       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
822       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
823       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
824       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
825       v128_t vo4 = wasm_f32x4_max(vo4p0, vmin);
826       vo0 = wasm_f32x4_min(vo0, vmax);
827       vo1 = wasm_f32x4_min(vo1, vmax);
828       vo2 = wasm_f32x4_min(vo2, vmax);
829       vo3 = wasm_f32x4_min(vo3, vmax);
830       vo4 = wasm_f32x4_min(vo4, vmax);
831 
832       if XNN_LIKELY(w & (4 * sizeof(float))) {
833         wasm_v128_store(o4, vo4); o4 += 4;
834         wasm_v128_store(o3, vo3); o3 += 4;
835         wasm_v128_store(o2, vo2); o2 += 4;
836         wasm_v128_store(o1, vo1); o1 += 4;
837         wasm_v128_store(o0, vo0); o0 += 4;
838       } else {
839         if (w & (2 * sizeof(float))) {
840           *((double*) o4) = wasm_f64x2_extract_lane(vo4, 0); o4 += 2;
841           *((double*) o3) = wasm_f64x2_extract_lane(vo3, 0); o3 += 2;
842           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
843           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
844           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
845 
846           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
847           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
848           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
849           vo3 = wasm_v32x4_shuffle(vo3, vo3, 2, 3, 0, 1);
850           vo4 = wasm_v32x4_shuffle(vo4, vo4, 2, 3, 0, 1);
851         }
852         if (w & (1 * sizeof(float))) {
853           *o4 = wasm_f32x4_extract_lane(vo4, 0); o4 += 1;
854           *o3 = wasm_f32x4_extract_lane(vo3, 0); o3 += 1;
855           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
856           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
857           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
858         }
859       }
860     }
861 
862     i0 = (const float*) ((uintptr_t) i5 - input_decrement);
863     i1 = (const float*) ((uintptr_t) i6 - input_decrement);
864     i2 = (const float*) ((uintptr_t) i1 + input_width);
865     i3 = (const float*) ((uintptr_t) i2 + input_width);
866     i4 = (const float*) ((uintptr_t) i3 + input_width);
867     i5 = (const float*) ((uintptr_t) i4 + input_width);
868     i6 = (const float*) ((uintptr_t) i5 + input_width);
869     i7 = (const float*) ((uintptr_t) i6 + input_width);
870     i8 = (const float*) ((uintptr_t) i7 + input_width);
871 
872     o0 = o4;
873     o1 = (float*) ((uintptr_t) o0 + input_width);
874     o2 = (float*) ((uintptr_t) o1 + input_width);
875     o3 = (float*) ((uintptr_t) o2 + input_width);
876     o4 = (float*) ((uintptr_t) o3 + input_width);
877 
878     output_height = doz(output_height, 5);
879   } while (output_height != 0);
880 }
881