• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5p2-wasmsimd-loadsplat.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_x86_loadsplat_4x4(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4(
20     size_t input_height,
21     size_t input_width,
22     const float* input,
23     const float* weights,
24     const float* zero,
25     float* output,
26     uint32_t padding_top,
27     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(input_height != 0);
30   assert(input_width != 0);
31   assert(input_width % sizeof(float) == 0);
32   assert(padding_top == 2);
33 
34   const v128_t vmask = wasm_v128_load(params->scalar.mask);
35   const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
36   const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
37 
38   const v128_t vw0123 = wasm_v128_load(weights);
39   const v128_t vw4567 = wasm_v128_load(weights + 4);
40   const v128_t vw89AB = wasm_v128_load(weights + 8);
41   const v128_t vwCDEF = wasm_v128_load(weights + 12);
42   const v128_t vwGHIJ = wasm_v128_load(weights + 16);
43   const v128_t vwKLMN = wasm_v128_load(weights + 20);
44   const v128_t vwOP = wasm_v128_load64_splat(weights + 24);
45   const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
46   const v128_t vk00 =  wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
47   const v128_t vk01 =  wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
48   const v128_t vk02 =  wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3);
49   const v128_t vk03 =  wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0);
50   const v128_t vk04 =  wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1);
51   const v128_t vk10 =  wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);
52   const v128_t vk11 =  wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);
53   const v128_t vk12 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0);
54   const v128_t vk13 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1);
55   const v128_t vk14 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2);
56   const v128_t vk20 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3);
57   const v128_t vk21 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0);
58   const v128_t vk22 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1);
59   const v128_t vk23 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2);
60   const v128_t vk24 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3);
61   const v128_t vk30 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0);
62   const v128_t vk31 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1);
63   const v128_t vk32 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2);
64   const v128_t vk33 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3);
65   const v128_t vk34 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0);
66   const v128_t vk40 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1);
67   const v128_t vk41 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2);
68   const v128_t vk42 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3);
69   const v128_t vk43 =  wasm_v32x4_shuffle(vwOP,   vwOP,   0, 0, 0, 0);
70   const v128_t vk44 =  wasm_v32x4_shuffle(vwOP,   vwOP,   1, 1, 1, 1);
71 
72   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
73 
74   const float* i0 = zero;
75   const float* i1 = zero;
76   const float* i2 = input;
77   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
78   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
79   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
80   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
81   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
82 
83   float* o0 = output;
84   float* o1 = (float*) ((uintptr_t) o0 + input_width);
85   float* o2 = (float*) ((uintptr_t) o1 + input_width);
86   float* o3 = (float*) ((uintptr_t) o2 + input_width);
87 
88   size_t output_height = input_height;
89   do {
90     if XNN_UNPREDICTABLE(output_height < 2) {
91       i3 = zero;
92       o1 = o0;
93     }
94     if XNN_UNPREDICTABLE(output_height < 3) {
95       i4 = zero;
96       o2 = o1;
97     }
98     if XNN_UNPREDICTABLE(output_height < 4) {
99       i5 = zero;
100       o3 = o2;
101     }
102     if XNN_UNPREDICTABLE(output_height < 5) {
103       i6 = zero;
104     }
105     if XNN_UNPREDICTABLE(output_height < 6) {
106       i7 = zero;
107     }
108 
109     v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
110     v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
111     v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
112     v128_t vi3x0123 = wasm_f32x4_const_splat(0.0f);
113     v128_t vi4x0123 = wasm_f32x4_const_splat(0.0f);
114     v128_t vi5x0123 = wasm_f32x4_const_splat(0.0f);
115     v128_t vi6x0123 = wasm_f32x4_const_splat(0.0f);
116     v128_t vi7x0123 = wasm_f32x4_const_splat(0.0f);
117 
118     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
119     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
120     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
121     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
122     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
123     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
124     v128_t vi6x4567 = wasm_v128_load(i6); i6 += 4;
125     v128_t vi7x4567 = wasm_v128_load(i7); i7 += 4;
126 
127     size_t w = input_width;
128     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
129       v128_t vo0p0 = vbias;
130       v128_t vo1p0 = vbias;
131       v128_t vo2p0 = vbias;
132       v128_t vo3p0 = vbias;
133 
134       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
135       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
136       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
137       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
138       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
139       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
140       const v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
141       const v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
142 
143       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, vk02));
144       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, vk02));
145       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, vk02));
146       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, vk02));
147 
148       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
149       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
150       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));
151       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk12));
152 
153       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
154       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
155       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));
156       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk22));
157 
158       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, vk32));
159       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, vk32));
160       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, vk32));
161       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, vk32));
162 
163       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
164       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
165       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));
166       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, vk42));
167 
168       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
169       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
170       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
171       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
172       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
173       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
174       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
175       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
176 
177       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk01));
178       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk01));
179       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk01));
180       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk01));
181 
182       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
183       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
184       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));
185       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk11));
186 
187       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21));
188       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk21));
189       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk21));
190       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk21));
191 
192       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
193       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
194       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));
195       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, vk31));
196 
197       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, vk41));
198       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, vk41));
199       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, vk41));
200       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, vk41));
201 
202       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
203       vi0x0123 = vi0x4567;
204       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
205       vi1x0123 = vi1x4567;
206       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
207       vi2x0123 = vi2x4567;
208       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
209       vi3x0123 = vi3x4567;
210       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
211       vi4x0123 = vi4x4567;
212       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
213       vi5x0123 = vi5x4567;
214       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
215       vi6x0123 = vi6x4567;
216       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
217       vi7x0123 = vi7x4567;
218 
219       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
220       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
221       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));
222       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, vk00));
223 
224       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, vk10));
225       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10));
226       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, vk10));
227       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, vk10));
228 
229       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
230       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
231       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));
232       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, vk20));
233 
234       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
235       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, vk30));
236       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, vk30));
237       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, vk30));
238 
239       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
240       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
241       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));
242       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, vk40));
243 
244       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
245       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
246       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
247       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
248       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
249       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
250       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
251       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
252 
253       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03));
254       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03));
255       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03));
256       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk03));
257 
258       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
259       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
260       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));
261       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk13));
262 
263       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23));
264       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk23));
265       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23));
266       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23));
267 
268       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
269       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
270       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));
271       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, vk33));
272 
273       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43));
274       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43));
275       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43));
276       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, vk43));
277 
278       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
279       vi0x4567 = vi0x89AB;
280       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
281       vi1x4567 = vi1x89AB;
282       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
283       vi2x4567 = vi2x89AB;
284       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
285       vi3x4567 = vi3x89AB;
286       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
287       vi4x4567 = vi4x89AB;
288       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
289       vi5x4567 = vi5x89AB;
290       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
291       vi6x4567 = vi6x89AB;
292       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
293       vi7x4567 = vi7x89AB;
294 
295       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
296       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
297       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));
298       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, vk04));
299 
300       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, vk14));
301       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14));
302       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, vk14));
303       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14));
304 
305       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
306       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
307       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));
308       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24));
309 
310       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34));
311       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34));
312       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34));
313       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, vk34));
314 
315       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
316       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
317       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));
318       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));
319 
320 
321       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
322       v128_t vo1 = wasm_f32x4_pmax(vmin, vo1p0);
323       v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0);
324       v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0);
325       vo0 = wasm_f32x4_pmin(vmax, vo0);
326       vo1 = wasm_f32x4_pmin(vmax, vo1);
327       vo2 = wasm_f32x4_pmin(vmax, vo2);
328       vo3 = wasm_f32x4_pmin(vmax, vo3);
329 
330       wasm_v128_store(o3, vo3); o3 += 4;
331       wasm_v128_store(o2, vo2); o2 += 4;
332       wasm_v128_store(o1, vo1); o1 += 4;
333       wasm_v128_store(o0, vo0); o0 += 4;
334     }
335     // Always process the last block of 5..8 pixels.
336     if XNN_LIKELY(w > 4 * sizeof(float)) {
337       v128_t vo0p0 = vbias;
338       v128_t vo1p0 = vbias;
339       v128_t vo2p0 = vbias;
340       v128_t vo3p0 = vbias;
341 
342       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
343       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
344       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
345       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
346       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
347       v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
348       v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
349       v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
350 
351       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
352       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
353       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
354       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
355       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
356       vi5x89AB = wasm_v128_and(vmask, vi5x89AB);
357       vi6x89AB = wasm_v128_and(vmask, vi6x89AB);
358       vi7x89AB = wasm_v128_and(vmask, vi7x89AB);
359 
360       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, vk02));
361       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, vk02));
362       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, vk02));
363       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, vk02));
364 
365       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
366       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
367       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));
368       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk12));
369 
370       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
371       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
372       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));
373       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk22));
374 
375       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, vk32));
376       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, vk32));
377       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, vk32));
378       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, vk32));
379 
380       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
381       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
382       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));
383       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, vk42));
384 
385       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
386       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
387       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
388       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
389       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
390       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
391       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
392       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
393 
394       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk01));
395       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk01));
396       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk01));
397       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk01));
398 
399       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
400       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
401       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));
402       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk11));
403 
404       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21));
405       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk21));
406       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk21));
407       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk21));
408 
409       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
410       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
411       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));
412       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, vk31));
413 
414       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, vk41));
415       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, vk41));
416       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, vk41));
417       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, vk41));
418 
419       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
420       vi0x0123 = vi0x4567;
421       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
422       vi1x0123 = vi1x4567;
423       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
424       vi2x0123 = vi2x4567;
425       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
426       vi3x0123 = vi3x4567;
427       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
428       vi4x0123 = vi4x4567;
429       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
430       vi5x0123 = vi5x4567;
431       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
432       vi6x0123 = vi6x4567;
433       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
434       vi7x0123 = vi7x4567;
435 
436       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
437       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
438       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));
439       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, vk00));
440 
441       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, vk10));
442       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10));
443       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, vk10));
444       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, vk10));
445 
446       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
447       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
448       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));
449       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, vk20));
450 
451       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
452       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, vk30));
453       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, vk30));
454       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, vk30));
455 
456       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
457       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
458       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));
459       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, vk40));
460 
461       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
462       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
463       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
464       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
465       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
466       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
467       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
468       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
469 
470       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03));
471       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03));
472       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03));
473       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk03));
474 
475       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
476       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
477       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));
478       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk13));
479 
480       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23));
481       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk23));
482       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23));
483       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23));
484 
485       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
486       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
487       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));
488       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, vk33));
489 
490       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43));
491       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43));
492       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43));
493       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, vk43));
494 
495       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
496       vi0x4567 = vi0x89AB;
497       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
498       vi1x4567 = vi1x89AB;
499       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
500       vi2x4567 = vi2x89AB;
501       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
502       vi3x4567 = vi3x89AB;
503       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
504       vi4x4567 = vi4x89AB;
505       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
506       vi5x4567 = vi5x89AB;
507       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
508       vi6x4567 = vi6x89AB;
509       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
510       vi7x4567 = vi7x89AB;
511 
512       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
513       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
514       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));
515       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, vk04));
516 
517       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, vk14));
518       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14));
519       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, vk14));
520       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14));
521 
522       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
523       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
524       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));
525       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24));
526 
527       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34));
528       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34));
529       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34));
530       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, vk34));
531 
532       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
533       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
534       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));
535       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));
536 
537 
538       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
539       v128_t vo1 = wasm_f32x4_pmax(vmin, vo1p0);
540       v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0);
541       v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0);
542       vo0 = wasm_f32x4_pmin(vmax, vo0);
543       vo1 = wasm_f32x4_pmin(vmax, vo1);
544       vo2 = wasm_f32x4_pmin(vmax, vo2);
545       vo3 = wasm_f32x4_pmin(vmax, vo3);
546 
547       wasm_v128_store(o3, vo3); o3 += 4;
548       wasm_v128_store(o2, vo2); o2 += 4;
549       wasm_v128_store(o1, vo1); o1 += 4;
550       wasm_v128_store(o0, vo0); o0 += 4;
551 
552       w -= 4 * sizeof(float);
553     }
554     assert(w >= 1 * sizeof(float));
555     assert(w <= 4 * sizeof(float));
556     {
557       v128_t vo0p0 = vbias;
558       v128_t vo1p0 = vbias;
559       v128_t vo2p0 = vbias;
560       v128_t vo3p0 = vbias;
561 
562       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
563       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
564       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
565       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
566       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
567       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
568       vi6x4567 = wasm_v128_and(vmask, vi6x4567);
569       vi7x4567 = wasm_v128_and(vmask, vi7x4567);
570 
571       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, vk02));
572       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, vk02));
573       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, vk02));
574       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, vk02));
575 
576       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
577       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
578       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));
579       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk12));
580 
581       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
582       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
583       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));
584       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk22));
585 
586       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, vk32));
587       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, vk32));
588       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, vk32));
589       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, vk32));
590 
591       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
592       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
593       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));
594       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, vk42));
595 
596       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
597       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
598       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
599       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
600       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
601       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
602       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
603       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
604 
605       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk01));
606       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk01));
607       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk01));
608       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk01));
609 
610       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
611       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
612       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));
613       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk11));
614 
615       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21));
616       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk21));
617       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk21));
618       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk21));
619 
620       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
621       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
622       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));
623       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, vk31));
624 
625       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x3456, vk41));
626       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, vk41));
627       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x3456, vk41));
628       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x3456, vk41));
629 
630       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
631       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
632       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
633       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
634       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
635       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
636       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
637       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
638 
639       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
640       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
641       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));
642       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, vk00));
643 
644       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x2345, vk10));
645       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x2345, vk10));
646       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x2345, vk10));
647       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x2345, vk10));
648 
649       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
650       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
651       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));
652       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, vk20));
653 
654       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
655       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, vk30));
656       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, vk30));
657       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, vk30));
658 
659       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
660       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
661       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));
662       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, vk40));
663 
664       const v128_t vzero = wasm_f32x4_const_splat(0.0f);
665       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
666       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
667       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
668       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
669       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
670       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
671       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4);
672       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4);
673 
674       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03));
675       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03));
676       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03));
677       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk03));
678 
679       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
680       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
681       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));
682       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk13));
683 
684       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23));
685       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk23));
686       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23));
687       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23));
688 
689       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
690       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
691       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));
692       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, vk33));
693 
694       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43));
695       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43));
696       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43));
697       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, vk43));
698 
699       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
700       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
701       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
702       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
703       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
704       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4);
705       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4);
706       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x5678, vzero, 1, 2, 3, 4);
707 
708       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
709       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
710       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));
711       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, vk04));
712 
713       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, vk14));
714       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, vk14));
715       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x6789, vk14));
716       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14));
717 
718       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
719       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
720       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));
721       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24));
722 
723       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34));
724       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34));
725       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34));
726       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x6789, vk34));
727 
728       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
729       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
730       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));
731       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));
732 
733 
734       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
735       v128_t vo1 = wasm_f32x4_pmax(vmin, vo1p0);
736       v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0);
737       v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0);
738       vo0 = wasm_f32x4_pmin(vmax, vo0);
739       vo1 = wasm_f32x4_pmin(vmax, vo1);
740       vo2 = wasm_f32x4_pmin(vmax, vo2);
741       vo3 = wasm_f32x4_pmin(vmax, vo3);
742 
743       if XNN_LIKELY(w & (4 * sizeof(float))) {
744         wasm_v128_store(o3, vo3); o3 += 4;
745         wasm_v128_store(o2, vo2); o2 += 4;
746         wasm_v128_store(o1, vo1); o1 += 4;
747         wasm_v128_store(o0, vo0); o0 += 4;
748       } else {
749         if (w & (2 * sizeof(float))) {
750           *((double*) o3) = wasm_f64x2_extract_lane(vo3, 0); o3 += 2;
751           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
752           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
753           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
754 
755           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
756           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
757           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
758           vo3 = wasm_v32x4_shuffle(vo3, vo3, 2, 3, 0, 1);
759         }
760         if (w & (1 * sizeof(float))) {
761           *o3 = wasm_f32x4_extract_lane(vo3, 0); o3 += 1;
762           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
763           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
764           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
765         }
766       }
767     }
768 
769     i0 = (const float*) ((uintptr_t) i4 - input_decrement);
770     i1 = (const float*) ((uintptr_t) i5 - input_decrement);
771     i2 = (const float*) ((uintptr_t) i1 + input_width);
772     i3 = (const float*) ((uintptr_t) i2 + input_width);
773     i4 = (const float*) ((uintptr_t) i3 + input_width);
774     i5 = (const float*) ((uintptr_t) i4 + input_width);
775     i6 = (const float*) ((uintptr_t) i5 + input_width);
776     i7 = (const float*) ((uintptr_t) i6 + input_width);
777 
778     o0 = o3;
779     o1 = (float*) ((uintptr_t) o0 + input_width);
780     o2 = (float*) ((uintptr_t) o1 + input_width);
781     o3 = (float*) ((uintptr_t) o2 + input_width);
782 
783     output_height = doz(output_height, 4);
784   } while (output_height != 0);
785 }
786