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_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_arm_loadsplat_2x4_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 
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       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
129       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
130 
131       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
132       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
133 
134       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, vk32));
135       vo1p1 = wasm_f32x4_add(vo1p1, 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       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
151       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
152 
153       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk21));
154       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk21));
155 
156       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
157       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
158 
159       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
160       vo1p1 = wasm_f32x4_add(vo1p1, 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       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
182       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
183 
184       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, vk30));
185       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, vk30));
186 
187       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
188       vo1p0 = wasm_f32x4_add(vo1p0, 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       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03));
198       vo1p1 = wasm_f32x4_add(vo1p1, 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       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
207       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
208 
209       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
210       vo1p1 = wasm_f32x4_add(vo1p1, 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       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
226       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
227 
228       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, vk14));
229       vo1p1 = wasm_f32x4_add(vo1p1, 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       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
238       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
239 
240       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
241       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
242 
243       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
244       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
245       vo0 = wasm_f32x4_min(vo0, vmax);
246       vo1 = wasm_f32x4_min(vo1, vmax);
247 
248       wasm_v128_store(o1, vo1); o1 += 4;
249       wasm_v128_store(o0, vo0); o0 += 4;
250     }
251     // Always process the last block of 5..8 pixels.
252     if XNN_LIKELY(w > 4 * sizeof(float)) {
253       v128_t vo0p0 = vbias;
254       v128_t vo1p0 = vbias;
255 
256       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
257       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
258       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
259       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
260       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
261       v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
262 
263       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
264       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
265       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
266       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
267       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
268       vi5x89AB = wasm_v128_and(vmask, vi5x89AB);
269 
270       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
271       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
272 
273       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
274       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
275 
276       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
277       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
278 
279       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, vk32));
280       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, vk32));
281 
282       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
283       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
284 
285       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
286       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
287       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
288       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
289       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
290       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
291 
292       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
293       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
294 
295       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
296       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
297 
298       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk21));
299       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk21));
300 
301       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
302       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
303 
304       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
305       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, vk41));
306 
307       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
308       vi0x0123 = vi0x4567;
309       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
310       vi1x0123 = vi1x4567;
311       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
312       vi2x0123 = vi2x4567;
313       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
314       vi3x0123 = vi3x4567;
315       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
316       vi4x0123 = vi4x4567;
317       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
318       vi5x0123 = vi5x4567;
319 
320       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
321       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
322 
323       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
324       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
325 
326       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
327       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
328 
329       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, vk30));
330       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, vk30));
331 
332       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
333       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
334 
335       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
336       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
337       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
338       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
339       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
340       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
341 
342       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03));
343       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03));
344 
345       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
346       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
347 
348       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
349       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
350 
351       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
352       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
353 
354       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
355       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43));
356 
357       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
358       vi0x4567 = vi0x89AB;
359       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
360       vi1x4567 = vi1x89AB;
361       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
362       vi2x4567 = vi2x89AB;
363       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
364       vi3x4567 = vi3x89AB;
365       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
366       vi4x4567 = vi4x89AB;
367       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5);
368       vi5x4567 = vi5x89AB;
369 
370       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
371       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
372 
373       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, vk14));
374       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14));
375 
376       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
377       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
378 
379       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
380       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
381 
382       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
383       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
384 
385       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
386       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
387 
388       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
389       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
390       vo0 = wasm_f32x4_min(vo0, vmax);
391       vo1 = wasm_f32x4_min(vo1, vmax);
392 
393       wasm_v128_store(o1, vo1); o1 += 4;
394       wasm_v128_store(o0, vo0); o0 += 4;
395 
396       w -= 4 * sizeof(float);
397     }
398     assert(w >= 1 * sizeof(float));
399     assert(w <= 4 * sizeof(float));
400     {
401       v128_t vo0p0 = vbias;
402       v128_t vo1p0 = vbias;
403 
404       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
405       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
406       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
407       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
408       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
409       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
410 
411       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
412       v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, vk02);
413 
414       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk12));
415       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk12));
416 
417       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk22));
418       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk22));
419 
420       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, vk32));
421       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, vk32));
422 
423       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, vk42));
424       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, vk42));
425 
426       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
427       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
428       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
429       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
430       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
431       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
432 
433       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk01));
434       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk01));
435 
436       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk11));
437       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk11));
438 
439       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk21));
440       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk21));
441 
442       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
443       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, vk31));
444 
445       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
446       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, vk41));
447 
448       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
449       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
450       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
451       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
452       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
453       const v128_t vi5x2345 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 2, 3, 4, 5);
454 
455       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x2345, vk00));
456       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x2345, vk00));
457 
458       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x2345, vk10));
459       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x2345, vk10));
460 
461       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, vk20));
462       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, vk20));
463 
464       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, vk30));
465       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, vk30));
466 
467       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, vk40));
468       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, vk40));
469 
470       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
471       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
472       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
473       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
474       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
475       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
476 
477       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03));
478       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03));
479 
480       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk13));
481       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk13));
482 
483       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23));
484       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23));
485 
486       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
487       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, vk33));
488 
489       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
490       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43));
491 
492       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
493       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
494       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
495       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
496       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
497       const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4);
498 
499       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x6789, vk04));
500       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x6789, vk04));
501 
502       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, vk14));
503       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, vk14));
504 
505       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x6789, vk24));
506       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x6789, vk24));
507 
508       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, vk34));
509       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34));
510 
511       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));
512       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));
513 
514       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
515       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
516 
517       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
518       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
519       vo0 = wasm_f32x4_min(vo0, vmax);
520       vo1 = wasm_f32x4_min(vo1, vmax);
521 
522       if XNN_LIKELY(w & (4 * sizeof(float))) {
523         wasm_v128_store(o1, vo1); o1 += 4;
524         wasm_v128_store(o0, vo0); o0 += 4;
525       } else {
526         if (w & (2 * sizeof(float))) {
527           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
528           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
529 
530           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
531           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
532         }
533         if (w & (1 * sizeof(float))) {
534           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
535           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
536         }
537       }
538     }
539 
540     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
541     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
542     i2 = (const float*) ((uintptr_t) i1 + input_width);
543     i3 = (const float*) ((uintptr_t) i2 + input_width);
544     i4 = (const float*) ((uintptr_t) i3 + input_width);
545     i5 = (const float*) ((uintptr_t) i4 + input_width);
546 
547     o0 = o1;
548     o1 = (float*) ((uintptr_t) o0 + input_width);
549 
550     output_height = doz(output_height, 2);
551   } while (output_height != 0);
552 }
553