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_acc2(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2(
20     size_t input_height,
21     size_t input_width,
22     const float* input,
23     const float* weights,
24     const float* zero,
25     float* output,
26     uint32_t padding_top,
27     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(input_height != 0);
30   assert(input_width != 0);
31   assert(input_width % sizeof(float) == 0);
32   assert(padding_top == 2);
33 
34   const v128_t vmask = wasm_v128_load(params->scalar.mask);
35   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
36   const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
37 
38   const v128_t vw0123 = wasm_v128_load(weights);
39   const v128_t vw4567 = wasm_v128_load(weights + 4);
40   const v128_t vw89AB = wasm_v128_load(weights + 8);
41   const v128_t vwCDEF = wasm_v128_load(weights + 12);
42   const v128_t vwGHIJ = wasm_v128_load(weights + 16);
43   const v128_t vwKLMN = wasm_v128_load(weights + 20);
44   const v128_t vwOP = wasm_v64x2_load_splat(weights + 24);
45   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 v128_t vzero = wasm_f32x4_splat(0.0f);
73 
74   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
75 
76   const float* i0 = zero;
77   const float* i1 = zero;
78   const float* i2 = input;
79   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
80   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
81   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
82   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
83   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
84 
85   float* o0 = output;
86   float* o1 = (float*) ((uintptr_t) o0 + input_width);
87   float* o2 = (float*) ((uintptr_t) o1 + input_width);
88   float* o3 = (float*) ((uintptr_t) o2 + input_width);
89 
90   size_t output_height = input_height;
91   do {
92     if XNN_UNPREDICTABLE(output_height < 2) {
93       i3 = zero;
94       o1 = o0;
95     }
96     if XNN_UNPREDICTABLE(output_height < 3) {
97       i4 = zero;
98       o2 = o1;
99     }
100     if XNN_UNPREDICTABLE(output_height < 4) {
101       i5 = zero;
102       o3 = o2;
103     }
104     if XNN_UNPREDICTABLE(output_height < 5) {
105       i6 = zero;
106     }
107     if XNN_UNPREDICTABLE(output_height < 6) {
108       i7 = zero;
109     }
110 
111     v128_t vi0x0123 = vzero;
112     v128_t vi1x0123 = vzero;
113     v128_t vi2x0123 = vzero;
114     v128_t vi3x0123 = vzero;
115     v128_t vi4x0123 = vzero;
116     v128_t vi5x0123 = vzero;
117     v128_t vi6x0123 = vzero;
118     v128_t vi7x0123 = vzero;
119 
120     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
121     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
122     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
123     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
124     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
125     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
126     v128_t vi6x4567 = wasm_v128_load(i6); i6 += 4;
127     v128_t vi7x4567 = wasm_v128_load(i7); i7 += 4;
128 
129     size_t w = input_width;
130     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
131       v128_t vo0p0 = vbias;
132       v128_t vo1p0 = vbias;
133       v128_t vo2p0 = vbias;
134       v128_t vo3p0 = vbias;
135 
136       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
137       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
138       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
139       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
140       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
141       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
142       const v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
143       const v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
144 
145       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
146       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
147       v128_t vo2p1 = wasm_f32x4_mul(vi2x4567, vk02);
148       v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02);
149 
150       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
151       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
152       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));
153       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk12));
154 
155       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
156       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
157       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));
158       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk22));
159 
160       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, vk32));
161       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, vk32));
162       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, vk32));
163       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x4567, vk32));
164 
165       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
166       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
167       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));
168       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, vk42));
169 
170       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
171       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
172       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
173       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
174       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
175       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
176       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
177       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
178 
179       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
180       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
181       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x3456, vk01));
182       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi3x3456, vk01));
183 
184       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
185       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
186       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));
187       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk11));
188 
189       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk21));
190       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk21));
191       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x3456, vk21));
192       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x3456, vk21));
193 
194       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
195       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
196       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));
197       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, vk31));
198 
199       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
200       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, vk41));
201       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x3456, vk41));
202       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x3456, vk41));
203 
204       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
205       vi0x0123 = vi0x4567;
206       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
207       vi1x0123 = vi1x4567;
208       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
209       vi2x0123 = vi2x4567;
210       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
211       vi3x0123 = vi3x4567;
212       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
213       vi4x0123 = vi4x4567;
214       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
215       vi5x0123 = vi5x4567;
216       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
217       vi6x0123 = vi6x4567;
218       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
219       vi7x0123 = vi7x4567;
220 
221       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
222       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
223       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));
224       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, vk00));
225 
226       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
227       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
228       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x2345, vk10));
229       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi4x2345, vk10));
230 
231       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
232       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
233       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));
234       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, vk20));
235 
236       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, vk30));
237       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, vk30));
238       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, vk30));
239       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x2345, vk30));
240 
241       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
242       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
243       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));
244       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, vk40));
245 
246       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
247       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
248       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
249       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
250       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
251       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
252       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
253       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
254 
255       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03));
256       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03));
257       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, vk03));
258       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi3x5678, vk03));
259 
260       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
261       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
262       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));
263       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk13));
264 
265       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
266       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
267       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, vk23));
268       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x5678, vk23));
269 
270       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
271       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
272       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));
273       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, vk33));
274 
275       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
276       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43));
277       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, vk43));
278       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x5678, vk43));
279 
280       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
281       vi0x4567 = vi0x89AB;
282       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
283       vi1x4567 = vi1x89AB;
284       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
285       vi2x4567 = vi2x89AB;
286       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
287       vi3x4567 = vi3x89AB;
288       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
289       vi4x4567 = vi4x89AB;
290       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
291       vi5x4567 = vi5x89AB;
292       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
293       vi6x4567 = vi6x89AB;
294       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
295       vi7x4567 = vi7x89AB;
296 
297       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
298       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
299       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));
300       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, vk04));
301 
302       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, vk14));
303       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14));
304       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x6789, vk14));
305       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi4x6789, vk14));
306 
307       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
308       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
309       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));
310       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24));
311 
312       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
313       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
314       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34));
315       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x6789, vk34));
316 
317       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
318       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
319       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));
320       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));
321 
322       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
323       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
324       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
325       vo3p0 = wasm_f32x4_add(vo3p0, vo3p1);
326 
327       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
328       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
329       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
330       v128_t vo3 = wasm_v128_bitselect(vmin, vo3p0, wasm_f32x4_lt(vo3p0, vmin));
331       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
332       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
333       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));
334       vo3 = wasm_v128_bitselect(vo3, vmax, wasm_f32x4_le(vo3, vmax));
335 
336       wasm_v128_store(o3, vo3); o3 += 4;
337       wasm_v128_store(o2, vo2); o2 += 4;
338       wasm_v128_store(o1, vo1); o1 += 4;
339       wasm_v128_store(o0, vo0); o0 += 4;
340     }
341     // Always process the last block of 5..8 pixels.
342     if XNN_LIKELY(w > 4 * sizeof(float)) {
343       v128_t vo0p0 = vbias;
344       v128_t vo1p0 = vbias;
345       v128_t vo2p0 = vbias;
346       v128_t vo3p0 = vbias;
347 
348       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
349       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
350       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
351       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
352       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
353       v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
354       v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
355       v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
356 
357       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
358       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
359       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
360       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
361       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
362       vi5x89AB = wasm_v128_and(vmask, vi5x89AB);
363       vi6x89AB = wasm_v128_and(vmask, vi6x89AB);
364       vi7x89AB = wasm_v128_and(vmask, vi7x89AB);
365 
366       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
367       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
368       v128_t vo2p1 = wasm_f32x4_mul(vi2x4567, vk02);
369       v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02);
370 
371       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
372       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
373       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));
374       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk12));
375 
376       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
377       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
378       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));
379       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk22));
380 
381       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, vk32));
382       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, vk32));
383       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, vk32));
384       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x4567, vk32));
385 
386       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
387       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
388       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));
389       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, vk42));
390 
391       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
392       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
393       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
394       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
395       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
396       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
397       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
398       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
399 
400       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
401       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
402       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x3456, vk01));
403       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi3x3456, vk01));
404 
405       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
406       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
407       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));
408       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk11));
409 
410       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk21));
411       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk21));
412       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x3456, vk21));
413       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x3456, vk21));
414 
415       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
416       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
417       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));
418       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, vk31));
419 
420       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
421       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, vk41));
422       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x3456, vk41));
423       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x3456, vk41));
424 
425       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
426       vi0x0123 = vi0x4567;
427       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
428       vi1x0123 = vi1x4567;
429       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
430       vi2x0123 = vi2x4567;
431       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
432       vi3x0123 = vi3x4567;
433       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
434       vi4x0123 = vi4x4567;
435       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
436       vi5x0123 = vi5x4567;
437       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
438       vi6x0123 = vi6x4567;
439       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
440       vi7x0123 = vi7x4567;
441 
442       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
443       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
444       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));
445       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, vk00));
446 
447       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
448       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
449       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x2345, vk10));
450       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi4x2345, vk10));
451 
452       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
453       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
454       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));
455       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, vk20));
456 
457       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, vk30));
458       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, vk30));
459       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, vk30));
460       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x2345, vk30));
461 
462       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
463       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
464       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));
465       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, vk40));
466 
467       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
468       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
469       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
470       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
471       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
472       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
473       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
474       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
475 
476       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03));
477       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03));
478       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, vk03));
479       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi3x5678, vk03));
480 
481       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
482       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
483       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));
484       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk13));
485 
486       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
487       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
488       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, vk23));
489       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x5678, vk23));
490 
491       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
492       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
493       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));
494       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, vk33));
495 
496       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
497       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43));
498       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, vk43));
499       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x5678, vk43));
500 
501       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
502       vi0x4567 = vi0x89AB;
503       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
504       vi1x4567 = vi1x89AB;
505       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
506       vi2x4567 = vi2x89AB;
507       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
508       vi3x4567 = vi3x89AB;
509       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
510       vi4x4567 = vi4x89AB;
511       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
512       vi5x4567 = vi5x89AB;
513       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 2, 3, 4, 5);
514       vi6x4567 = vi6x89AB;
515       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 2, 3, 4, 5);
516       vi7x4567 = vi7x89AB;
517 
518       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
519       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
520       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));
521       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, vk04));
522 
523       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, vk14));
524       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14));
525       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x6789, vk14));
526       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi4x6789, vk14));
527 
528       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
529       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
530       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));
531       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24));
532 
533       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
534       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
535       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34));
536       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x6789, vk34));
537 
538       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
539       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
540       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));
541       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));
542 
543       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
544       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
545       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
546       vo3p0 = wasm_f32x4_add(vo3p0, vo3p1);
547 
548       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
549       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
550       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
551       v128_t vo3 = wasm_v128_bitselect(vmin, vo3p0, wasm_f32x4_lt(vo3p0, vmin));
552       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
553       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
554       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));
555       vo3 = wasm_v128_bitselect(vo3, vmax, wasm_f32x4_le(vo3, vmax));
556 
557       wasm_v128_store(o3, vo3); o3 += 4;
558       wasm_v128_store(o2, vo2); o2 += 4;
559       wasm_v128_store(o1, vo1); o1 += 4;
560       wasm_v128_store(o0, vo0); o0 += 4;
561 
562       w -= 4 * sizeof(float);
563     }
564     assert(w >= 1 * sizeof(float));
565     assert(w <= 4 * sizeof(float));
566     {
567       v128_t vo0p0 = vbias;
568       v128_t vo1p0 = vbias;
569       v128_t vo2p0 = vbias;
570       v128_t vo3p0 = vbias;
571 
572       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
573       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
574       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
575       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
576       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
577       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
578       vi6x4567 = wasm_v128_and(vmask, vi6x4567);
579       vi7x4567 = wasm_v128_and(vmask, vi7x4567);
580 
581       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
582       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
583       v128_t vo2p1 = wasm_f32x4_mul(vi2x4567, vk02);
584       v128_t vo3p1 = wasm_f32x4_mul(vi3x4567, vk02);
585 
586       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
587       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
588       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk12));
589       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk12));
590 
591       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
592       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
593       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk22));
594       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk22));
595 
596       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, vk32));
597       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, vk32));
598       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, vk32));
599       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x4567, vk32));
600 
601       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
602       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
603       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x4567, vk42));
604       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x4567, vk42));
605 
606       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
607       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
608       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
609       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
610       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
611       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
612       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
613       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
614 
615       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
616       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
617       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x3456, vk01));
618       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi3x3456, vk01));
619 
620       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
621       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
622       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk11));
623       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk11));
624 
625       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk21));
626       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk21));
627       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x3456, vk21));
628       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x3456, vk21));
629 
630       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
631       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
632       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31));
633       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, vk31));
634 
635       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
636       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, vk41));
637       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x3456, vk41));
638       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x3456, vk41));
639 
640       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
641       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
642       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
643       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
644       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
645       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
646       const v128_t vi6x2345 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 2, 3, 4, 5);
647       const v128_t vi7x2345 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 2, 3, 4, 5);
648 
649       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
650       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
651       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x2345, vk00));
652       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x2345, vk00));
653 
654       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
655       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
656       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x2345, vk10));
657       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi4x2345, vk10));
658 
659       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
660       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
661       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, vk20));
662       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, vk20));
663 
664       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, vk30));
665       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, vk30));
666       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, vk30));
667       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x2345, vk30));
668 
669       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
670       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
671       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x2345, vk40));
672       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x2345, vk40));
673 
674       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
675       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
676       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
677       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
678       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
679       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
680       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4);
681       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4);
682 
683       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03));
684       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03));
685       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, vk03));
686       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi3x5678, vk03));
687 
688       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
689       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
690       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk13));
691       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk13));
692 
693       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
694       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
695       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, vk23));
696       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x5678, vk23));
697 
698       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
699       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
700       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33));
701       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, vk33));
702 
703       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
704       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43));
705       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, vk43));
706       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x5678, vk43));
707 
708       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
709       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
710       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
711       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
712       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
713       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4);
714       const v128_t vi6x6789 = wasm_v32x4_shuffle(vi6x5678, vzero, 1, 2, 3, 4);
715       const v128_t vi7x6789 = wasm_v32x4_shuffle(vi7x5678, vzero, 1, 2, 3, 4);
716 
717       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
718       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
719       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x6789, vk04));
720       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x6789, vk04));
721 
722       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, vk14));
723       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14));
724       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi3x6789, vk14));
725       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi4x6789, vk14));
726 
727       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
728       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
729       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));
730       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24));
731 
732       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
733       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
734       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34));
735       vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x6789, vk34));
736 
737       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
738       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
739       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));
740       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));
741 
742       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
743       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
744       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
745       vo3p0 = wasm_f32x4_add(vo3p0, vo3p1);
746 
747       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
748       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
749       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
750       v128_t vo3 = wasm_v128_bitselect(vmin, vo3p0, wasm_f32x4_lt(vo3p0, vmin));
751       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
752       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
753       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));
754       vo3 = wasm_v128_bitselect(vo3, vmax, wasm_f32x4_le(vo3, vmax));
755 
756       if XNN_LIKELY(w & (4 * sizeof(float))) {
757         wasm_v128_store(o3, vo3); o3 += 4;
758         wasm_v128_store(o2, vo2); o2 += 4;
759         wasm_v128_store(o1, vo1); o1 += 4;
760         wasm_v128_store(o0, vo0); o0 += 4;
761       } else {
762         if (w & (2 * sizeof(float))) {
763           *((double*) o3) = wasm_f64x2_extract_lane(vo3, 0); o3 += 2;
764           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
765           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
766           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
767 
768           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
769           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
770           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
771           vo3 = wasm_v32x4_shuffle(vo3, vo3, 2, 3, 0, 1);
772         }
773         if (w & (1 * sizeof(float))) {
774           *o3 = wasm_f32x4_extract_lane(vo3, 0); o3 += 1;
775           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
776           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
777           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
778         }
779       }
780     }
781 
782     i0 = (const float*) ((uintptr_t) i4 - input_decrement);
783     i1 = (const float*) ((uintptr_t) i5 - input_decrement);
784     i2 = (const float*) ((uintptr_t) i1 + input_width);
785     i3 = (const float*) ((uintptr_t) i2 + input_width);
786     i4 = (const float*) ((uintptr_t) i3 + input_width);
787     i5 = (const float*) ((uintptr_t) i4 + input_width);
788     i6 = (const float*) ((uintptr_t) i5 + input_width);
789     i7 = (const float*) ((uintptr_t) i6 + input_width);
790 
791     o0 = o3;
792     o1 = (float*) ((uintptr_t) o0 + input_width);
793     o2 = (float*) ((uintptr_t) o1 + input_width);
794     o3 = (float*) ((uintptr_t) o2 + input_width);
795 
796     output_height = doz(output_height, 4);
797   } while (output_height != 0);
798 }
799