1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5s2p2-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_5x5s2p2__wasmsimd_arm_splat_3x4_acc2(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2(
20     size_t input_height,
21     size_t input_width,
22     const float* input,
23     const float* weights,
24     const float* zero,
25     float* output,
26     uint32_t padding_top,
27     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(input_height != 0);
30   assert(input_width != 0);
31   assert(input_width % sizeof(float) == 0);
32   assert(padding_top >= 1);
33   assert(padding_top <= 2);
34 
35   const v128_t vmask_even = wasm_v128_load(params->scalar.mask_even);
36   const v128_t vmask_odd  = wasm_v128_load(params->scalar.mask_odd);
37   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
38   const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
39 
40   const v128_t vw0123 = wasm_v128_load(weights);
41   const v128_t vw4567 = wasm_v128_load(weights + 4);
42   const v128_t vw89AB = wasm_v128_load(weights + 8);
43   const v128_t vwCDEF = wasm_v128_load(weights + 12);
44   const v128_t vwGHIJ = wasm_v128_load(weights + 16);
45   const v128_t vwKLMN = wasm_v128_load(weights + 20);
46   const v128_t vwOP = wasm_v64x2_load_splat(weights + 24);
47 
48   const v128_t vzero = wasm_f32x4_splat(0.0f);
49 
50   const uint32_t padding_top_less_1 = padding_top - 1;
51   const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(float));
52 
53   const float* i0 = zero;
54   const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
55   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
56   if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
57     i1 = zero;
58   }
59   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
60   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
61   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
62   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
63   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
64   const float* i8 = (const float*) ((uintptr_t) i7 + input_width);
65 
66   const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
67 
68   float* o0 = output;
69   float* o1 = (float*) ((uintptr_t) o0 + output_width);
70   float* o2 = (float*) ((uintptr_t) o1 + output_width);
71 
72   size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
73   size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
74   do {
75     if XNN_UNPREDICTABLE(padded_input_height < 6) {
76       i3 = zero;
77     }
78     if XNN_UNPREDICTABLE(padded_input_height < 7) {
79       i4 = zero;
80       o1 = o0;
81     }
82     if XNN_UNPREDICTABLE(padded_input_height < 8) {
83       i5 = zero;
84     }
85     if XNN_UNPREDICTABLE(padded_input_height < 9) {
86       i6 = zero;
87       o2 = o1;
88     }
89     if XNN_UNPREDICTABLE(padded_input_height < 10) {
90       i7 = zero;
91     }
92     if XNN_UNPREDICTABLE(padded_input_height < 11) {
93       i8 = zero;
94     }
95 
96     v128_t vi0x0246 = vzero;
97     v128_t vi1x0246 = vzero;
98     v128_t vi2x0246 = vzero;
99     v128_t vi3x0246 = vzero;
100     v128_t vi4x0246 = vzero;
101     v128_t vi5x0246 = vzero;
102     v128_t vi6x0246 = vzero;
103     v128_t vi7x0246 = vzero;
104     v128_t vi8x0246 = vzero;
105 
106     v128_t vi0x1357 = vzero;
107     v128_t vi1x1357 = vzero;
108     v128_t vi2x1357 = vzero;
109     v128_t vi3x1357 = vzero;
110     v128_t vi4x1357 = vzero;
111     v128_t vi5x1357 = vzero;
112     v128_t vi6x1357 = vzero;
113     v128_t vi7x1357 = vzero;
114     v128_t vi8x1357 = vzero;
115 
116     const v128_t vi0x89AB = wasm_v128_load(i0);
117     const v128_t vi0xCDEF = wasm_v128_load(i0 + 4);
118     i0 += 8;
119     const v128_t vi1x89AB = wasm_v128_load(i1);
120     const v128_t vi1xCDEF = wasm_v128_load(i1 + 4);
121     i1 += 8;
122     const v128_t vi2x89AB = wasm_v128_load(i2);
123     const v128_t vi2xCDEF = wasm_v128_load(i2 + 4);
124     i2 += 8;
125     const v128_t vi3x89AB = wasm_v128_load(i3);
126     const v128_t vi3xCDEF = wasm_v128_load(i3 + 4);
127     i3 += 8;
128     const v128_t vi4x89AB = wasm_v128_load(i4);
129     const v128_t vi4xCDEF = wasm_v128_load(i4 + 4);
130     i4 += 8;
131     const v128_t vi5x89AB = wasm_v128_load(i5);
132     const v128_t vi5xCDEF = wasm_v128_load(i5 + 4);
133     i5 += 8;
134     const v128_t vi6x89AB = wasm_v128_load(i6);
135     const v128_t vi6xCDEF = wasm_v128_load(i6 + 4);
136     i6 += 8;
137     const v128_t vi7x89AB = wasm_v128_load(i7);
138     const v128_t vi7xCDEF = wasm_v128_load(i7 + 4);
139     i7 += 8;
140     const v128_t vi8x89AB = wasm_v128_load(i8);
141     const v128_t vi8xCDEF = wasm_v128_load(i8 + 4);
142     i8 += 8;
143 
144     v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6);
145     v128_t vi0x9BDF = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 1, 3, 5, 7);
146     v128_t vi1x8ACE = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 0, 2, 4, 6);
147     v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7);
148     v128_t vi2x8ACE = wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 0, 2, 4, 6);
149     v128_t vi2x9BDF = wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 1, 3, 5, 7);
150     v128_t vi3x8ACE = wasm_v32x4_shuffle(vi3x89AB, vi3xCDEF, 0, 2, 4, 6);
151     v128_t vi3x9BDF = wasm_v32x4_shuffle(vi3x89AB, vi3xCDEF, 1, 3, 5, 7);
152     v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6);
153     v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7);
154     v128_t vi5x8ACE = wasm_v32x4_shuffle(vi5x89AB, vi5xCDEF, 0, 2, 4, 6);
155     v128_t vi5x9BDF = wasm_v32x4_shuffle(vi5x89AB, vi5xCDEF, 1, 3, 5, 7);
156     v128_t vi6x8ACE = wasm_v32x4_shuffle(vi6x89AB, vi6xCDEF, 0, 2, 4, 6);
157     v128_t vi6x9BDF = wasm_v32x4_shuffle(vi6x89AB, vi6xCDEF, 1, 3, 5, 7);
158     v128_t vi7x8ACE = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 0, 2, 4, 6);
159     v128_t vi7x9BDF = wasm_v32x4_shuffle(vi7x89AB, vi7xCDEF, 1, 3, 5, 7);
160     v128_t vi8x8ACE = wasm_v32x4_shuffle(vi8x89AB, vi8xCDEF, 0, 2, 4, 6);
161     v128_t vi8x9BDF = wasm_v32x4_shuffle(vi8x89AB, vi8xCDEF, 1, 3, 5, 7);
162 
163     size_t w = input_width;
164     for (; w > 8 * sizeof(float); w -= 8 * sizeof(float)) {
165       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
166       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
167       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
168 
169       v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
170       v128_t vo1p1 = wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
171       v128_t vo2p1 = wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
172 
173       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
174       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
175       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
176 
177       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
178       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
179       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
180 
181       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
182       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
183       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi7x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
184 
185       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
186       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi6x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
187       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi8x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
188 
189       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
190       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
191       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
192 
193       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
194       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
195       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
196 
197       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
198       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
199       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
200 
201       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
202       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
203       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
204 
205       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
206       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi6x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
207       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi8x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
208 
209       const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6);
210       vi0x0246 = vi0x8ACE;
211       const v128_t vi1x68AC = wasm_v32x4_shuffle(vi1x0246, vi1x8ACE, 3, 4, 5, 6);
212       vi1x0246 = vi1x8ACE;
213       const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6);
214       vi2x0246 = vi2x8ACE;
215       const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6);
216       vi3x0246 = vi3x8ACE;
217       const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6);
218       vi4x0246 = vi4x8ACE;
219       const v128_t vi5x68AC = wasm_v32x4_shuffle(vi5x0246, vi5x8ACE, 3, 4, 5, 6);
220       vi5x0246 = vi5x8ACE;
221       const v128_t vi6x68AC = wasm_v32x4_shuffle(vi6x0246, vi6x8ACE, 3, 4, 5, 6);
222       vi6x0246 = vi6x8ACE;
223       const v128_t vi7x68AC = wasm_v32x4_shuffle(vi7x0246, vi7x8ACE, 3, 4, 5, 6);
224       vi7x0246 = vi7x8ACE;
225       const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6);
226       vi8x0246 = vi8x8ACE;
227 
228       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
229       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
230       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
231 
232       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
233       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
234       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
235 
236       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
237       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
238       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
239 
240       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
241       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
242       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi7x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
243 
244       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
245       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
246       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi8x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
247 
248       const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6);
249       vi0x1357 = vi0x9BDF;
250       const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6);
251       vi1x1357 = vi1x9BDF;
252       const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6);
253       vi2x1357 = vi2x9BDF;
254       const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6);
255       vi3x1357 = vi3x9BDF;
256       const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6);
257       vi4x1357 = vi4x9BDF;
258       const v128_t vi5x79BD = wasm_v32x4_shuffle(vi5x1357, vi5x9BDF, 3, 4, 5, 6);
259       vi5x1357 = vi5x9BDF;
260       const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6);
261       vi6x1357 = vi6x9BDF;
262       const v128_t vi7x79BD = wasm_v32x4_shuffle(vi7x1357, vi7x9BDF, 3, 4, 5, 6);
263       vi7x1357 = vi7x9BDF;
264       const v128_t vi8x79BD = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6);
265       vi8x1357 = vi8x9BDF;
266 
267       const v128_t vi0xGHIJ = wasm_v128_load(i0);
268       const v128_t vi0xKLMN = wasm_v128_load(i0 + 4);
269       i0 += 8;
270       const v128_t vi1xGHIJ = wasm_v128_load(i1);
271       const v128_t vi1xKLMN = wasm_v128_load(i1 + 4);
272       i1 += 8;
273       const v128_t vi2xGHIJ = wasm_v128_load(i2);
274       const v128_t vi2xKLMN = wasm_v128_load(i2 + 4);
275       i2 += 8;
276       const v128_t vi3xGHIJ = wasm_v128_load(i3);
277       const v128_t vi3xKLMN = wasm_v128_load(i3 + 4);
278       i3 += 8;
279       const v128_t vi4xGHIJ = wasm_v128_load(i4);
280       const v128_t vi4xKLMN = wasm_v128_load(i4 + 4);
281       i4 += 8;
282       const v128_t vi5xGHIJ = wasm_v128_load(i5);
283       const v128_t vi5xKLMN = wasm_v128_load(i5 + 4);
284       i5 += 8;
285       const v128_t vi6xGHIJ = wasm_v128_load(i6);
286       const v128_t vi6xKLMN = wasm_v128_load(i6 + 4);
287       i6 += 8;
288       const v128_t vi7xGHIJ = wasm_v128_load(i7);
289       const v128_t vi7xKLMN = wasm_v128_load(i7 + 4);
290       i7 += 8;
291       const v128_t vi8xGHIJ = wasm_v128_load(i8);
292       const v128_t vi8xKLMN = wasm_v128_load(i8 + 4);
293       i8 += 8;
294 
295       const v128_t vi0xGIKM = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 0, 2, 4, 6);
296       const v128_t vi0xHJLN = wasm_v32x4_shuffle(vi0xGHIJ, vi0xKLMN, 1, 3, 5, 7);
297       const v128_t vi1xGIKM = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 0, 2, 4, 6);
298       const v128_t vi1xHJLN = wasm_v32x4_shuffle(vi1xGHIJ, vi1xKLMN, 1, 3, 5, 7);
299       const v128_t vi2xGIKM = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 0, 2, 4, 6);
300       const v128_t vi2xHJLN = wasm_v32x4_shuffle(vi2xGHIJ, vi2xKLMN, 1, 3, 5, 7);
301       const v128_t vi3xGIKM = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 0, 2, 4, 6);
302       const v128_t vi3xHJLN = wasm_v32x4_shuffle(vi3xGHIJ, vi3xKLMN, 1, 3, 5, 7);
303       const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6);
304       const v128_t vi4xHJLN = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 1, 3, 5, 7);
305       const v128_t vi5xGIKM = wasm_v32x4_shuffle(vi5xGHIJ, vi5xKLMN, 0, 2, 4, 6);
306       const v128_t vi5xHJLN = wasm_v32x4_shuffle(vi5xGHIJ, vi5xKLMN, 1, 3, 5, 7);
307       const v128_t vi6xGIKM = wasm_v32x4_shuffle(vi6xGHIJ, vi6xKLMN, 0, 2, 4, 6);
308       const v128_t vi6xHJLN = wasm_v32x4_shuffle(vi6xGHIJ, vi6xKLMN, 1, 3, 5, 7);
309       const v128_t vi7xGIKM = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 0, 2, 4, 6);
310       const v128_t vi7xHJLN = wasm_v32x4_shuffle(vi7xGHIJ, vi7xKLMN, 1, 3, 5, 7);
311       const v128_t vi8xGIKM = wasm_v32x4_shuffle(vi8xGHIJ, vi8xKLMN, 0, 2, 4, 6);
312       const v128_t vi8xHJLN = wasm_v32x4_shuffle(vi8xGHIJ, vi8xKLMN, 1, 3, 5, 7);
313 
314       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
315       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
316       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
317 
318       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
319       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
320       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
321 
322       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
323       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
324       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
325 
326       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
327       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
328       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
329 
330       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
331       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi6x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
332       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi8x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
333 
334       const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vi0xGIKM, 1, 2, 3, 4);
335       vi0x8ACE = vi0xGIKM;
336       vi0x9BDF = vi0xHJLN;
337       const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4);
338       vi1x8ACE = vi1xGIKM;
339       vi1x9BDF = vi1xHJLN;
340       const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4);
341       vi2x8ACE = vi2xGIKM;
342       vi2x9BDF = vi2xHJLN;
343       const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4);
344       vi3x8ACE = vi3xGIKM;
345       vi3x9BDF = vi3xHJLN;
346       const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4);
347       vi4x8ACE = vi4xGIKM;
348       vi4x9BDF = vi4xHJLN;
349       const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4);
350       vi5x8ACE = vi5xGIKM;
351       vi5x9BDF = vi5xHJLN;
352       const v128_t vi6xACEG = wasm_v32x4_shuffle(vi6x8ACE, vi6xGIKM, 1, 2, 3, 4);
353       vi6x8ACE = vi6xGIKM;
354       vi6x9BDF = vi6xHJLN;
355       const v128_t vi7xACEG = wasm_v32x4_shuffle(vi7x8ACE, vi7xGIKM, 1, 2, 3, 4);
356       vi7x8ACE = vi7xGIKM;
357       vi7x9BDF = vi7xHJLN;
358       const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vi8xGIKM, 1, 2, 3, 4);
359       vi8x8ACE = vi8xGIKM;
360       vi8x9BDF = vi8xHJLN;
361 
362       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
363       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
364       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
365 
366       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
367       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
368       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
369 
370       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
371       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
372       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
373 
374       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
375       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
376       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi7xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
377 
378       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
379       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi6xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
380       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi8xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
381 
382       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
383       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
384       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
385 
386       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
387       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
388       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
389       vo0 = wasm_f32x4_min(vo0, vmax);
390       vo1 = wasm_f32x4_min(vo1, vmax);
391       vo2 = wasm_f32x4_min(vo2, vmax);
392 
393       wasm_v128_store(o2, vo2); o2 += 4;
394       wasm_v128_store(o1, vo1); o1 += 4;
395       wasm_v128_store(o0, vo0); o0 += 4;
396     }
397     // Last block has 1-8 pixels to process.
398     assert(w <= 8 * sizeof(float));
399     assert(w >= 1 * sizeof(float));
400     {
401       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
402       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
403       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
404 
405       vi0x8ACE = wasm_v128_and(vmask_even, vi0x8ACE);
406       vi1x8ACE = wasm_v128_and(vmask_even, vi1x8ACE);
407       vi2x8ACE = wasm_v128_and(vmask_even, vi2x8ACE);
408       vi3x8ACE = wasm_v128_and(vmask_even, vi3x8ACE);
409       vi4x8ACE = wasm_v128_and(vmask_even, vi4x8ACE);
410       vi5x8ACE = wasm_v128_and(vmask_even, vi5x8ACE);
411       vi6x8ACE = wasm_v128_and(vmask_even, vi6x8ACE);
412       vi7x8ACE = wasm_v128_and(vmask_even, vi7x8ACE);
413       vi8x8ACE = wasm_v128_and(vmask_even, vi8x8ACE);
414 
415       vi0x9BDF = wasm_v128_and(vmask_odd, vi0x9BDF);
416       vi1x9BDF = wasm_v128_and(vmask_odd, vi1x9BDF);
417       vi2x9BDF = wasm_v128_and(vmask_odd, vi2x9BDF);
418       vi3x9BDF = wasm_v128_and(vmask_odd, vi3x9BDF);
419       vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF);
420       vi5x9BDF = wasm_v128_and(vmask_odd, vi5x9BDF);
421       vi6x9BDF = wasm_v128_and(vmask_odd, vi6x9BDF);
422       vi7x9BDF = wasm_v128_and(vmask_odd, vi7x9BDF);
423       vi8x9BDF = wasm_v128_and(vmask_odd, vi8x9BDF);
424 
425       v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
426       v128_t vo1p1 = wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
427       v128_t vo2p1 = wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
428 
429       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
430       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
431       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
432 
433       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
434       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
435       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
436 
437       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
438       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
439       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi7x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
440 
441       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
442       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi6x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
443       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi8x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
444 
445       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
446       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
447       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
448 
449       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
450       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
451       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
452 
453       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
454       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
455       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
456 
457       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
458       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
459       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
460 
461       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
462       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi6x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
463       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi8x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
464 
465       const v128_t vi0x68AC = wasm_v32x4_shuffle(vi0x0246, vi0x8ACE, 3, 4, 5, 6);
466       const v128_t vi1x68AC = wasm_v32x4_shuffle(vi1x0246, vi1x8ACE, 3, 4, 5, 6);
467       const v128_t vi2x68AC = wasm_v32x4_shuffle(vi2x0246, vi2x8ACE, 3, 4, 5, 6);
468       const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6);
469       const v128_t vi4x68AC = wasm_v32x4_shuffle(vi4x0246, vi4x8ACE, 3, 4, 5, 6);
470       const v128_t vi5x68AC = wasm_v32x4_shuffle(vi5x0246, vi5x8ACE, 3, 4, 5, 6);
471       const v128_t vi6x68AC = wasm_v32x4_shuffle(vi6x0246, vi6x8ACE, 3, 4, 5, 6);
472       const v128_t vi7x68AC = wasm_v32x4_shuffle(vi7x0246, vi7x8ACE, 3, 4, 5, 6);
473       const v128_t vi8x68AC = wasm_v32x4_shuffle(vi8x0246, vi8x8ACE, 3, 4, 5, 6);
474 
475       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
476       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
477       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
478 
479       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
480       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
481       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
482 
483       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
484       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
485       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
486 
487       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
488       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
489       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi7x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
490 
491       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
492       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
493       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi8x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
494 
495       const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6);
496       const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6);
497       const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6);
498       const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6);
499       const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6);
500       const v128_t vi5x79BD = wasm_v32x4_shuffle(vi5x1357, vi5x9BDF, 3, 4, 5, 6);
501       const v128_t vi6x79BD = wasm_v32x4_shuffle(vi6x1357, vi6x9BDF, 3, 4, 5, 6);
502       const v128_t vi7x79BD = wasm_v32x4_shuffle(vi7x1357, vi7x9BDF, 3, 4, 5, 6);
503       const v128_t vi8x79BD = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6);
504 
505       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
506       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
507       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
508 
509       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
510       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
511       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
512 
513       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
514       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
515       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
516 
517       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
518       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
519       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
520 
521       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
522       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi6x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
523       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi8x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
524 
525       const v128_t vi0xACEG = wasm_v32x4_shuffle(vi0x8ACE, vzero, 1, 2, 3, 4);
526       const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4);
527       const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4);
528       const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4);
529       const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4);
530       const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4);
531       const v128_t vi6xACEG = wasm_v32x4_shuffle(vi6x8ACE, vzero, 1, 2, 3, 4);
532       const v128_t vi7xACEG = wasm_v32x4_shuffle(vi7x8ACE, vzero, 1, 2, 3, 4);
533       const v128_t vi8xACEG = wasm_v32x4_shuffle(vi8x8ACE, vzero, 1, 2, 3, 4);
534 
535       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
536       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
537       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
538 
539       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
540       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
541       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
542 
543       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
544       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
545       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
546 
547       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
548       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
549       vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi7xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
550 
551       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
552       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi6xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
553       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi8xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
554 
555       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
556       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
557       vo2p0 = wasm_f32x4_add(vo2p0, vo2p1);
558 
559       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
560       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
561       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
562       vo0 = wasm_f32x4_min(vo0, vmax);
563       vo1 = wasm_f32x4_min(vo1, vmax);
564       vo2 = wasm_f32x4_min(vo2, vmax);
565 
566       size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
567       if XNN_LIKELY(w_tmp >= 4) {
568         wasm_v128_store(o2, vo2); o2 += 4;
569         wasm_v128_store(o1, vo1); o1 += 4;
570         wasm_v128_store(o0, vo0); o0 += 4;
571       } else {
572         if (w_tmp & 2) {
573           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
574           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
575           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
576 
577           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
578           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
579           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
580         }
581         if (w_tmp & 1) {
582           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
583           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
584           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
585         }
586       }
587     }
588 
589     i0 = (const float*) ((uintptr_t) i6 - input_decrement);
590     i1 = (const float*) ((uintptr_t) i7 - input_decrement);
591     i2 = (const float*) ((uintptr_t) i8 - input_decrement);
592     i3 = (const float*) ((uintptr_t) i2 + input_width);
593     i4 = (const float*) ((uintptr_t) i3 + input_width);
594     i5 = (const float*) ((uintptr_t) i4 + input_width);
595     i6 = (const float*) ((uintptr_t) i5 + input_width);
596     i7 = (const float*) ((uintptr_t) i6 + input_width);
597     i8 = (const float*) ((uintptr_t) i7 + input_width);
598 
599     o0 = o2;
600     o1 = (float*) ((uintptr_t) o0 + output_width);
601     o2 = (float*) ((uintptr_t) o1 + output_width);
602 
603     output_height = doz(output_height, 3);
604     padded_input_height = doz(padded_input_height, 6);
605   } while (output_height != 0);
606 }
607