• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5p2-wasmsimd-loadsplat.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <wasm_simd128.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16 
17 
18 
xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc5(
20     size_t input_height,
21     size_t input_width,
22     const float* input,
23     const float* weights,
24     const float* zero,
25     float* output,
26     uint32_t padding_top,
27     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(input_height != 0);
30   assert(input_width != 0);
31   assert(input_width % sizeof(float) == 0);
32   assert(padding_top == 2);
33 
34   const v128_t vmask = wasm_v128_load(params->scalar.mask);
35   const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
36   const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
37 
38   const v128_t vw0123 = wasm_v128_load(weights);
39   const v128_t vw4567 = wasm_v128_load(weights + 4);
40   const v128_t vw89AB = wasm_v128_load(weights + 8);
41   const v128_t vwCDEF = wasm_v128_load(weights + 12);
42   const v128_t vwGHIJ = wasm_v128_load(weights + 16);
43   const v128_t vwKLMN = wasm_v128_load(weights + 20);
44   const v128_t vwOP = wasm_v128_load64_splat(weights + 24);
45   const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
46   const v128_t vk00 =  wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
47   const v128_t vk01 =  wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
48   const v128_t vk02 =  wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3);
49   const v128_t vk03 =  wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0);
50   const v128_t vk04 =  wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1);
51   const v128_t vk10 =  wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);
52   const v128_t vk11 =  wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);
53   const v128_t vk12 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0);
54   const v128_t vk13 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1);
55   const v128_t vk14 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2);
56   const v128_t vk20 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3);
57   const v128_t vk21 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0);
58   const v128_t vk22 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1);
59   const v128_t vk23 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2);
60   const v128_t vk24 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3);
61   const v128_t vk30 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0);
62   const v128_t vk31 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1);
63   const v128_t vk32 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2);
64   const v128_t vk33 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3);
65   const v128_t vk34 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0);
66   const v128_t vk40 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1);
67   const v128_t vk41 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2);
68   const v128_t vk42 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3);
69   const v128_t vk43 =  wasm_v32x4_shuffle(vwOP,   vwOP,   0, 0, 0, 0);
70   const v128_t vk44 =  wasm_v32x4_shuffle(vwOP,   vwOP,   1, 1, 1, 1);
71 
72   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
73 
74   const float* i0 = zero;
75   const float* i1 = zero;
76   const float* i2 = input;
77   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
78   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
79 
80   float* o0 = output;
81 
82   size_t output_height = input_height;
83   do {
84     if XNN_UNPREDICTABLE(output_height < 2) {
85       i3 = zero;
86     }
87     if XNN_UNPREDICTABLE(output_height < 3) {
88       i4 = zero;
89     }
90 
91     v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
92     v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
93     v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
94     v128_t vi3x0123 = wasm_f32x4_const_splat(0.0f);
95     v128_t vi4x0123 = wasm_f32x4_const_splat(0.0f);
96 
97     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
98     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
99     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
100     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
101     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
102 
103     size_t w = input_width;
104     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
105       v128_t vo0p0 = vbias;
106 
107       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
108       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
109       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
110       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
111       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
112 
113       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
114 
115       v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12);
116 
117       v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, vk22);
118 
119       v128_t vo0p4 = wasm_f32x4_mul(vi3x4567, vk32);
120 
121       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x4567, vk42));
122 
123       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
124       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
125       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
126       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
127       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
128 
129       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x3456, vk01));
130 
131       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x3456, vk11));
132 
133       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x3456, vk21));
134 
135       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
136 
137       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
138 
139       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
140       vi0x0123 = vi0x4567;
141       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
142       vi1x0123 = vi1x4567;
143       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
144       vi2x0123 = vi2x4567;
145       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
146       vi3x0123 = vi3x4567;
147       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
148       vi4x0123 = vi4x4567;
149 
150       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x2345, vk00));
151 
152       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x2345, vk10));
153 
154       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x2345, vk20));
155 
156       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
157 
158       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x2345, vk40));
159 
160       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
161       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
162       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
163       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
164       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
165 
166       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03));
167 
168       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x5678, vk13));
169 
170       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x5678, vk23));
171 
172       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
173 
174       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
175 
176       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
177       vi0x4567 = vi0x89AB;
178       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
179       vi1x4567 = vi1x89AB;
180       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
181       vi2x4567 = vi2x89AB;
182       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
183       vi3x4567 = vi3x89AB;
184       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
185       vi4x4567 = vi4x89AB;
186 
187       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x6789, vk04));
188 
189       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x6789, vk14));
190 
191       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x6789, vk24));
192 
193       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34));
194 
195       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x6789, vk44));
196 
197       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
198       vo0p2 = wasm_f32x4_add(vo0p2, vo0p3);
199       vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
200       vo0p0 = wasm_f32x4_add(vo0p0, vo0p4);
201 
202       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
203       vo0 = wasm_f32x4_pmin(vmax, vo0);
204 
205       wasm_v128_store(o0, vo0); o0 += 4;
206     }
207     // Always process the last block of 5..8 pixels.
208     if XNN_LIKELY(w > 4 * sizeof(float)) {
209       v128_t vo0p0 = vbias;
210 
211       v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
212       v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
213       v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
214       v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
215       v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
216 
217       vi0x89AB = wasm_v128_and(vmask, vi0x89AB);
218       vi1x89AB = wasm_v128_and(vmask, vi1x89AB);
219       vi2x89AB = wasm_v128_and(vmask, vi2x89AB);
220       vi3x89AB = wasm_v128_and(vmask, vi3x89AB);
221       vi4x89AB = wasm_v128_and(vmask, vi4x89AB);
222 
223       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
224 
225       v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12);
226 
227       v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, vk22);
228 
229       v128_t vo0p4 = wasm_f32x4_mul(vi3x4567, vk32);
230 
231       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x4567, vk42));
232 
233       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
234       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
235       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
236       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
237       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
238 
239       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x3456, vk01));
240 
241       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x3456, vk11));
242 
243       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x3456, vk21));
244 
245       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
246 
247       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
248 
249       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
250       vi0x0123 = vi0x4567;
251       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
252       vi1x0123 = vi1x4567;
253       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
254       vi2x0123 = vi2x4567;
255       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
256       vi3x0123 = vi3x4567;
257       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
258       vi4x0123 = vi4x4567;
259 
260       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x2345, vk00));
261 
262       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x2345, vk10));
263 
264       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x2345, vk20));
265 
266       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
267 
268       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x2345, vk40));
269 
270       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
271       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
272       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
273       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
274       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
275 
276       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03));
277 
278       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x5678, vk13));
279 
280       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x5678, vk23));
281 
282       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
283 
284       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
285 
286       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 2, 3, 4, 5);
287       vi0x4567 = vi0x89AB;
288       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 2, 3, 4, 5);
289       vi1x4567 = vi1x89AB;
290       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 2, 3, 4, 5);
291       vi2x4567 = vi2x89AB;
292       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 2, 3, 4, 5);
293       vi3x4567 = vi3x89AB;
294       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);
295       vi4x4567 = vi4x89AB;
296 
297       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x6789, vk04));
298 
299       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x6789, vk14));
300 
301       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x6789, vk24));
302 
303       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34));
304 
305       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x6789, vk44));
306 
307       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
308       vo0p2 = wasm_f32x4_add(vo0p2, vo0p3);
309       vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
310       vo0p0 = wasm_f32x4_add(vo0p0, vo0p4);
311 
312       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
313       vo0 = wasm_f32x4_pmin(vmax, vo0);
314 
315       wasm_v128_store(o0, vo0); o0 += 4;
316 
317       w -= 4 * sizeof(float);
318     }
319     assert(w >= 1 * sizeof(float));
320     assert(w <= 4 * sizeof(float));
321     {
322       v128_t vo0p0 = vbias;
323 
324       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
325       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
326       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
327       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
328       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
329 
330       v128_t vo0p1 = wasm_f32x4_mul(vi0x4567, vk02);
331 
332       v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12);
333 
334       v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, vk22);
335 
336       v128_t vo0p4 = wasm_f32x4_mul(vi3x4567, vk32);
337 
338       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x4567, vk42));
339 
340       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
341       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
342       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
343       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
344       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
345 
346       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x3456, vk01));
347 
348       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x3456, vk11));
349 
350       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x3456, vk21));
351 
352       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, vk31));
353 
354       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, vk41));
355 
356       const v128_t vi0x2345 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 2, 3, 4, 5);
357       const v128_t vi1x2345 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 2, 3, 4, 5);
358       const v128_t vi2x2345 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 2, 3, 4, 5);
359       const v128_t vi3x2345 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 2, 3, 4, 5);
360       const v128_t vi4x2345 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 2, 3, 4, 5);
361 
362       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x2345, vk00));
363 
364       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x2345, vk10));
365 
366       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x2345, vk20));
367 
368       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, vk30));
369 
370       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x2345, vk40));
371 
372       const v128_t vzero = wasm_f32x4_const_splat(0.0f);
373       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
374       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
375       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
376       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
377       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
378 
379       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03));
380 
381       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x5678, vk13));
382 
383       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x5678, vk23));
384 
385       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, vk33));
386 
387       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43));
388 
389       const v128_t vi0x6789 = wasm_v32x4_shuffle(vi0x5678, vzero, 1, 2, 3, 4);
390       const v128_t vi1x6789 = wasm_v32x4_shuffle(vi1x5678, vzero, 1, 2, 3, 4);
391       const v128_t vi2x6789 = wasm_v32x4_shuffle(vi2x5678, vzero, 1, 2, 3, 4);
392       const v128_t vi3x6789 = wasm_v32x4_shuffle(vi3x5678, vzero, 1, 2, 3, 4);
393       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);
394 
395       vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x6789, vk04));
396 
397       vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x6789, vk14));
398 
399       vo0p4 = wasm_f32x4_add(vo0p4, wasm_f32x4_mul(vi2x6789, vk24));
400 
401       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x6789, vk34));
402 
403       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x6789, vk44));
404 
405       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
406       vo0p2 = wasm_f32x4_add(vo0p2, vo0p3);
407       vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
408       vo0p0 = wasm_f32x4_add(vo0p0, vo0p4);
409 
410       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
411       vo0 = wasm_f32x4_pmin(vmax, vo0);
412 
413       if XNN_LIKELY(w & (4 * sizeof(float))) {
414         wasm_v128_store(o0, vo0); o0 += 4;
415       } else {
416         if (w & (2 * sizeof(float))) {
417           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
418 
419           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
420         }
421         if (w & (1 * sizeof(float))) {
422           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
423         }
424       }
425     }
426 
427     i0 = (const float*) ((uintptr_t) i1 - input_decrement);
428     i1 = (const float*) ((uintptr_t) i2 - input_decrement);
429     i2 = (const float*) ((uintptr_t) i1 + input_width);
430     i3 = (const float*) ((uintptr_t) i2 + input_width);
431     i4 = (const float*) ((uintptr_t) i3 + input_width);
432 
433 
434   } while (--output_height != 0);
435 }
436