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_arm_loadsplat_2x4_acc3(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_2x4_acc3(
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 
83   float* o0 = output;
84   float* o1 = (float*) ((uintptr_t) o0 + input_width);
85 
86   size_t output_height = input_height;
87   do {
88     if XNN_UNPREDICTABLE(output_height < 2) {
89       i3 = zero;
90       o1 = o0;
91     }
92     if XNN_UNPREDICTABLE(output_height < 3) {
93       i4 = zero;
94     }
95     if XNN_UNPREDICTABLE(output_height < 4) {
96       i5 = zero;
97     }
98 
99     v128_t vi0x0123 = vzero;
100     v128_t vi1x0123 = vzero;
101     v128_t vi2x0123 = vzero;
102     v128_t vi3x0123 = vzero;
103     v128_t vi4x0123 = vzero;
104     v128_t vi5x0123 = vzero;
105 
106     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
107     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
108     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
109     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
110     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
111     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
112 
113     size_t w = input_width;
114     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
115       v128_t vo0p0 = vbias;
116       v128_t vo1p0 = vbias;
117 
118       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
119       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
120       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
121       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
122       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
123       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
124 
125       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
126       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
127 
128       v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12);
129       v128_t vo1p2 = wasm_f32x4_mul(vi2x4567, vk12);
130 
131       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x4567, vk22));
132       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x4567, vk22));
133 
134       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x4567, vk32));
135       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi4x4567, vk32));
136 
137       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
138       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
139 
140       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
141       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
142       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
143       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
144       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
145       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
146 
147       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
148       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
149 
150       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x3456, vk11));
151       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x3456, vk11));
152 
153       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21));
154       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk21));
155 
156       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x3456, vk31));
157       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x3456, vk31));
158 
159       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x3456, vk41));
160       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi5x3456, vk41));
161 
162       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
163       vi0x0123 = vi0x4567;
164       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
165       vi1x0123 = vi1x4567;
166       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
167       vi2x0123 = vi2x4567;
168       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
169       vi3x0123 = vi3x4567;
170       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
171       vi4x0123 = vi4x4567;
172       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
173       vi5x0123 = vi5x4567;
174 
175       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
176       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
177 
178       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
179       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
180 
181       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20));
182       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3x2345, vk20));
183 
184       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
185       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, vk30));
186 
187       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x2345, vk40));
188       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x2345, vk40));
189 
190       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
191       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
192       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
193       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
194       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
195       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
196 
197       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03));
198       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi1x5678, vk03));
199 
200       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
201       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
202 
203       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
204       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
205 
206       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x5678, vk33));
207       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi4x5678, vk33));
208 
209       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43));
210       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43));
211 
212       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
213       vi0x4567 = vi0x89AB;
214       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
215       vi1x4567 = vi1x89AB;
216       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
217       vi2x4567 = vi2x89AB;
218       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
219       vi3x4567 = vi3x89AB;
220       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
221       vi4x4567 = vi4x89AB;
222       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
223       vi5x4567 = vi5x89AB;
224 
225       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x6789, vk04));
226       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x6789, vk04));
227 
228       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, vk14));
229       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, vk14));
230 
231       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
232       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
233 
234       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
235       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
236 
237       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x6789, vk44));
238       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi5x6789, vk44));
239 
240       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
241       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
242       vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
243       vo1p0 = wasm_f32x4_add(vo1p0, vo1p2);
244 
245       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
246       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
247       vo0 = wasm_f32x4_min(vo0, vmax);
248       vo1 = wasm_f32x4_min(vo1, vmax);
249 
250       wasm_v128_store(o1, vo1); o1 += 4;
251       wasm_v128_store(o0, vo0); o0 += 4;
252     }
253     // Always process the last block of 5..8 pixels.
254     if XNN_LIKELY(w > 4 * sizeof(float)) {
255       v128_t vo0p0 = vbias;
256       v128_t vo1p0 = vbias;
257 
258       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
259       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
260       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
261       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
262       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
263       v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
264 
265       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
266       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
267       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
268       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
269       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
270       vi5x89AB = wasm_v128_and(vmask, vi5x89AB);
271 
272       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
273       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
274 
275       v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12);
276       v128_t vo1p2 = wasm_f32x4_mul(vi2x4567, vk12);
277 
278       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x4567, vk22));
279       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x4567, vk22));
280 
281       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x4567, vk32));
282       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi4x4567, vk32));
283 
284       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
285       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
286 
287       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
288       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
289       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
290       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
291       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
292       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
293 
294       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
295       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
296 
297       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x3456, vk11));
298       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x3456, vk11));
299 
300       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21));
301       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk21));
302 
303       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x3456, vk31));
304       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x3456, vk31));
305 
306       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x3456, vk41));
307       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi5x3456, vk41));
308 
309       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
310       vi0x0123 = vi0x4567;
311       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
312       vi1x0123 = vi1x4567;
313       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
314       vi2x0123 = vi2x4567;
315       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
316       vi3x0123 = vi3x4567;
317       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
318       vi4x0123 = vi4x4567;
319       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
320       vi5x0123 = vi5x4567;
321 
322       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
323       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
324 
325       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
326       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
327 
328       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20));
329       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3x2345, vk20));
330 
331       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
332       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, vk30));
333 
334       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x2345, vk40));
335       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x2345, vk40));
336 
337       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
338       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
339       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
340       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
341       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
342       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
343 
344       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03));
345       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi1x5678, vk03));
346 
347       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
348       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
349 
350       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
351       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
352 
353       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x5678, vk33));
354       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi4x5678, vk33));
355 
356       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43));
357       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43));
358 
359       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
360       vi0x4567 = vi0x89AB;
361       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
362       vi1x4567 = vi1x89AB;
363       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
364       vi2x4567 = vi2x89AB;
365       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
366       vi3x4567 = vi3x89AB;
367       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
368       vi4x4567 = vi4x89AB;
369       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
370       vi5x4567 = vi5x89AB;
371 
372       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x6789, vk04));
373       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x6789, vk04));
374 
375       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, vk14));
376       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, vk14));
377 
378       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
379       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
380 
381       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
382       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
383 
384       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x6789, vk44));
385       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi5x6789, vk44));
386 
387       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
388       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
389       vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
390       vo1p0 = wasm_f32x4_add(vo1p0, vo1p2);
391 
392       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
393       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
394       vo0 = wasm_f32x4_min(vo0, vmax);
395       vo1 = wasm_f32x4_min(vo1, vmax);
396 
397       wasm_v128_store(o1, vo1); o1 += 4;
398       wasm_v128_store(o0, vo0); o0 += 4;
399 
400       w -= 4 * sizeof(float);
401     }
402     assert(w >= 1 * sizeof(float));
403     assert(w <= 4 * sizeof(float));
404     {
405       v128_t vo0p0 = vbias;
406       v128_t vo1p0 = vbias;
407 
408       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
409       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
410       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
411       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
412       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
413       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
414 
415       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
416       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
417 
418       v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12);
419       v128_t vo1p2 = wasm_f32x4_mul(vi2x4567, vk12);
420 
421       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x4567, vk22));
422       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x4567, vk22));
423 
424       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x4567, vk32));
425       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi4x4567, vk32));
426 
427       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
428       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
429 
430       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
431       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
432       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
433       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
434       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
435       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
436 
437       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
438       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
439 
440       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x3456, vk11));
441       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x3456, vk11));
442 
443       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk21));
444       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk21));
445 
446       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x3456, vk31));
447       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x3456, vk31));
448 
449       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x3456, vk41));
450       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi5x3456, vk41));
451 
452       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
453       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
454       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
455       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
456       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
457       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
458 
459       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
460       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
461 
462       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
463       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
464 
465       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20));
466       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3x2345, vk20));
467 
468       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
469       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, vk30));
470 
471       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x2345, vk40));
472       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x2345, vk40));
473 
474       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
475       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
476       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
477       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
478       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
479       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
480 
481       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03));
482       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi1x5678, vk03));
483 
484       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
485       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
486 
487       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
488       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
489 
490       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x5678, vk33));
491       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi4x5678, vk33));
492 
493       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43));
494       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43));
495 
496       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
497       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
498       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
499       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
500       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
501       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4);
502 
503       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x6789, vk04));
504       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x6789, vk04));
505 
506       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, vk14));
507       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, vk14));
508 
509       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
510       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
511 
512       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
513       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
514 
515       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x6789, vk44));
516       vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi5x6789, vk44));
517 
518       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
519       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
520       vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
521       vo1p0 = wasm_f32x4_add(vo1p0, vo1p2);
522 
523       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
524       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
525       vo0 = wasm_f32x4_min(vo0, vmax);
526       vo1 = wasm_f32x4_min(vo1, vmax);
527 
528       if XNN_LIKELY(w & (4 * sizeof(float))) {
529         wasm_v128_store(o1, vo1); o1 += 4;
530         wasm_v128_store(o0, vo0); o0 += 4;
531       } else {
532         if (w & (2 * sizeof(float))) {
533           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
534           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
535 
536           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
537           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
538         }
539         if (w & (1 * sizeof(float))) {
540           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
541           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
542         }
543       }
544     }
545 
546     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
547     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
548     i2 = (const float*) ((uintptr_t) i1 + input_width);
549     i3 = (const float*) ((uintptr_t) i2 + input_width);
550     i4 = (const float*) ((uintptr_t) i3 + input_width);
551     i5 = (const float*) ((uintptr_t) i4 + input_width);
552 
553     o0 = o1;
554     o1 = (float*) ((uintptr_t) o0 + input_width);
555 
556     output_height = doz(output_height, 2);
557   } while (output_height != 0);
558 }
559