xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up24x9-minmax-fp32-wasmsimd-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-wasmsimd-mul16.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 
16 
xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16(
18     size_t channels,
19     size_t output_width,
20     const uint8_t** input,
21     const void* weights,
22     uint8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const uint8_t* zero,
27     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   const v128_t vkernel_zero_point = wasm_u32x4_load16x4(params->fp32_wasmsimd.kernel_zero_point);
33   do {
34     const uint8_t* i0 = input[0];
35     assert(i0 != NULL);
36     if XNN_UNPREDICTABLE(i0 != zero) {
37       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38     }
39     const uint8_t* i1 = input[1];
40     assert(i1 != NULL);
41     if XNN_UNPREDICTABLE(i1 != zero) {
42       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43     }
44     const uint8_t* i2 = input[2];
45     assert(i2 != NULL);
46     if XNN_UNPREDICTABLE(i2 != zero) {
47       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48     }
49     const uint8_t* i3 = input[3];
50     assert(i3 != NULL);
51     if XNN_UNPREDICTABLE(i3 != zero) {
52       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53     }
54     const uint8_t* i4 = input[4];
55     assert(i4 != NULL);
56     if XNN_UNPREDICTABLE(i4 != zero) {
57       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58     }
59     const uint8_t* i5 = input[5];
60     assert(i5 != NULL);
61     if XNN_UNPREDICTABLE(i5 != zero) {
62       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63     }
64     const uint8_t* i6 = input[6];
65     assert(i6 != NULL);
66     if XNN_UNPREDICTABLE(i6 != zero) {
67       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68     }
69     const uint8_t* i7 = input[7];
70     assert(i7 != NULL);
71     if XNN_UNPREDICTABLE(i7 != zero) {
72       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73     }
74     const uint8_t* i8 = input[8];
75     assert(i8 != NULL);
76     if XNN_UNPREDICTABLE(i8 != zero) {
77       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78     }
79     input = (const uint8_t**) ((uintptr_t) input + input_stride);
80 
81     size_t c = channels;
82     const void* w = weights;
83     for (; c >= 24; c -= 24) {
84       v128_t vacc0123 = wasm_v128_load(w);
85       v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
86       v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
87       v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
88       v128_t vaccGHIJ = wasm_v128_load((const void*) ((uintptr_t) w + 16 * sizeof(int32_t)));
89       v128_t vaccKLMN = wasm_v128_load((const void*) ((uintptr_t) w + 20 * sizeof(int32_t)));
90 
91 
92       const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
93       const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
94       const v128_t vi0x89ABCDEF = wasm_u16x8_load8x8(i0 + 8);
95       const v128_t vk0x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
96       const v128_t vi0xGHIJKLMN = wasm_u16x8_load8x8(i0 + 16);
97       const v128_t vk0xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
98       i0 += 24;
99 
100       v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
101       v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
102       v128_t vprodGHIJKLMN = wasm_i16x8_mul(vi0xGHIJKLMN, vk0xGHIJKLMN);
103 
104       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
105       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
106       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
107       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
108       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
109       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
110 
111       const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
112       const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
113       const v128_t vi1x89ABCDEF = wasm_u16x8_load8x8(i1 + 8);
114       const v128_t vk1x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
115       const v128_t vi1xGHIJKLMN = wasm_u16x8_load8x8(i1 + 16);
116       const v128_t vk1xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
117       v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
118       v128_t vsumx89ABCDEF = wasm_i16x8_add(vi0x89ABCDEF, vi1x89ABCDEF);
119       v128_t vsumxGHIJKLMN = wasm_i16x8_add(vi0xGHIJKLMN, vi1xGHIJKLMN);
120       i1 += 24;
121 
122       vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
123       vprod89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF);
124       vprodGHIJKLMN = wasm_i16x8_mul(vi1xGHIJKLMN, vk1xGHIJKLMN);
125 
126       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
127       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
128       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
129       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
130       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
131       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
132 
133       const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
134       const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
135       const v128_t vi2x89ABCDEF = wasm_u16x8_load8x8(i2 + 8);
136       const v128_t vk2x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
137       const v128_t vi2xGHIJKLMN = wasm_u16x8_load8x8(i2 + 16);
138       const v128_t vk2xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
139       vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
140       vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi2x89ABCDEF);
141       vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi2xGHIJKLMN);
142       i2 += 24;
143 
144       vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
145       vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
146       vprodGHIJKLMN = wasm_i16x8_mul(vi2xGHIJKLMN, vk2xGHIJKLMN);
147 
148       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
149       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
150       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
151       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
152       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
153       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
154 
155       const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
156       const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
157       const v128_t vi3x89ABCDEF = wasm_u16x8_load8x8(i3 + 8);
158       const v128_t vk3x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
159       const v128_t vi3xGHIJKLMN = wasm_u16x8_load8x8(i3 + 16);
160       const v128_t vk3xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
161       vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
162       vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi3x89ABCDEF);
163       vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi3xGHIJKLMN);
164       i3 += 24;
165 
166       vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
167       vprod89ABCDEF = wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF);
168       vprodGHIJKLMN = wasm_i16x8_mul(vi3xGHIJKLMN, vk3xGHIJKLMN);
169 
170       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
171       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
172       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
173       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
174       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
175       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
176 
177       const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
178       const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
179       const v128_t vi4x89ABCDEF = wasm_u16x8_load8x8(i4 + 8);
180       const v128_t vk4x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
181       const v128_t vi4xGHIJKLMN = wasm_u16x8_load8x8(i4 + 16);
182       const v128_t vk4xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
183       vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
184       vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi4x89ABCDEF);
185       vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi4xGHIJKLMN);
186       i4 += 24;
187 
188       vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
189       vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
190       vprodGHIJKLMN = wasm_i16x8_mul(vi4xGHIJKLMN, vk4xGHIJKLMN);
191 
192       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
193       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
194       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
195       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
196       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
197       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
198 
199       const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
200       const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
201       const v128_t vi5x89ABCDEF = wasm_u16x8_load8x8(i5 + 8);
202       const v128_t vk5x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
203       const v128_t vi5xGHIJKLMN = wasm_u16x8_load8x8(i5 + 16);
204       const v128_t vk5xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
205       vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
206       vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi5x89ABCDEF);
207       vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi5xGHIJKLMN);
208       i5 += 24;
209 
210       vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
211       vprod89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF);
212       vprodGHIJKLMN = wasm_i16x8_mul(vi5xGHIJKLMN, vk5xGHIJKLMN);
213 
214       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
215       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
216       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
217       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
218       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
219       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
220 
221       const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
222       const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(uint8_t)));
223       const v128_t vi6x89ABCDEF = wasm_u16x8_load8x8(i6 + 8);
224       const v128_t vk6x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(uint8_t)));
225       const v128_t vi6xGHIJKLMN = wasm_u16x8_load8x8(i6 + 16);
226       const v128_t vk6xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(uint8_t)));
227       vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
228       vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi6x89ABCDEF);
229       vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi6xGHIJKLMN);
230       i6 += 24;
231 
232       vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
233       vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
234       vprodGHIJKLMN = wasm_i16x8_mul(vi6xGHIJKLMN, vk6xGHIJKLMN);
235 
236       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
237       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
238       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
239       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
240       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
241       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
242 
243       const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
244       const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(uint8_t)));
245       const v128_t vi7x89ABCDEF = wasm_u16x8_load8x8(i7 + 8);
246       const v128_t vk7x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(uint8_t)));
247       const v128_t vi7xGHIJKLMN = wasm_u16x8_load8x8(i7 + 16);
248       const v128_t vk7xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(uint8_t)));
249       vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
250       vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi7x89ABCDEF);
251       vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi7xGHIJKLMN);
252       i7 += 24;
253 
254       vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
255       vprod89ABCDEF = wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF);
256       vprodGHIJKLMN = wasm_i16x8_mul(vi7xGHIJKLMN, vk7xGHIJKLMN);
257 
258       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
259       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
260       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
261       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
262       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
263       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
264 
265       const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
266       const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(uint8_t)));
267       const v128_t vi8x89ABCDEF = wasm_u16x8_load8x8(i8 + 8);
268       const v128_t vk8x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(uint8_t)));
269       const v128_t vi8xGHIJKLMN = wasm_u16x8_load8x8(i8 + 16);
270       const v128_t vk8xGHIJKLMN = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(uint8_t)));
271       vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
272       vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi8x89ABCDEF);
273       vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi8xGHIJKLMN);
274       i8 += 24;
275 
276       vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
277       vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
278       vprodGHIJKLMN = wasm_i16x8_mul(vi8xGHIJKLMN, vk8xGHIJKLMN);
279 
280       vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
281       vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
282       vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
283       vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
284       vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_u32x4_extend_low_u16x8(vprodGHIJKLMN));
285       vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_u32x4_extend_high_u16x8(vprodGHIJKLMN));
286 
287       vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
288       vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
289       vacc89AB = wasm_i32x4_sub(vacc89AB, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx89ABCDEF), vkernel_zero_point));
290       vaccCDEF = wasm_i32x4_sub(vaccCDEF, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx89ABCDEF), vkernel_zero_point));
291       vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumxGHIJKLMN), vkernel_zero_point));
292       vaccKLMN = wasm_i32x4_sub(vaccKLMN, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumxGHIJKLMN), vkernel_zero_point));
293 
294       w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(uint8_t));
295 
296       vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
297       vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
298       vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
299       vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
300       vaccGHIJ = wasm_f32x4_convert_i32x4(vaccGHIJ);
301       vaccKLMN = wasm_f32x4_convert_i32x4(vaccKLMN);
302 
303       const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
304       vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
305       vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
306       vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
307       vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
308       vaccGHIJ = wasm_f32x4_mul(vaccGHIJ, vscale);
309       vaccKLMN = wasm_f32x4_mul(vaccKLMN, vscale);
310 
311       const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
312       vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
313       vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
314       vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
315       vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
316       vaccGHIJ = wasm_f32x4_add(vaccGHIJ, vmagic_bias);
317       vaccKLMN = wasm_f32x4_add(vaccKLMN, vmagic_bias);
318 
319       const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
320       vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
321       vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
322       vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
323       vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
324       vaccGHIJ = wasm_i32x4_max(vaccGHIJ, vmagic_min);
325       vaccKLMN = wasm_i32x4_max(vaccKLMN, vmagic_min);
326 
327       const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
328       vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
329       vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
330       vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
331       vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
332       vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, vmagic_bias_less_output_zero_point);
333       vaccKLMN = wasm_i32x4_sub(vaccKLMN, vmagic_bias_less_output_zero_point);
334 
335       v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
336       v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
337       v128_t voutGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
338 
339       v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
340       v128_t voutGHIJKLMNGHIJKLMN = wasm_u8x16_narrow_i16x8(voutGHIJKLMN, voutGHIJKLMN);
341 
342       const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
343       vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
344       voutGHIJKLMNGHIJKLMN = wasm_u8x16_min(voutGHIJKLMNGHIJKLMN, voutput_max);
345 
346       wasm_v128_store(output, vout0123456789ABCDEF);
347       *((double*) (output + 16)) = wasm_f64x2_extract_lane(voutGHIJKLMNGHIJKLMN, 0);
348       output += 24;
349     }
350     if XNN_UNLIKELY(c != 0) {
351       const uint8_t* k = (const uint8_t*) ((uintptr_t) w + 24 * sizeof(int32_t));
352       do {
353         v128_t vacc0123 = wasm_v128_load(w);
354         v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
355 
356 
357         const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
358         const v128_t vk0x01234567 = wasm_u16x8_load8x8(k);
359         i0 += 8;
360 
361         v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
362 
363         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
364         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
365 
366         const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
367         const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) (k + 24));
368         v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
369         i1 += 8;
370 
371         vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
372 
373         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
374         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
375 
376         const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
377         const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) (k + 48));
378         vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
379         i2 += 8;
380 
381         vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
382 
383         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
384         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
385 
386         const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
387         const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) (k + 72));
388         vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
389         i3 += 8;
390 
391         vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
392 
393         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
394         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
395 
396         const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
397         const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) (k + 96));
398         vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
399         i4 += 8;
400 
401         vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
402 
403         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
404         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
405 
406         const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
407         const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) (k + 120));
408         vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
409         i5 += 8;
410 
411         vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
412 
413         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
414         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
415 
416         const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
417         const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) (k + 144));
418         vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
419         i6 += 8;
420 
421         vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
422 
423         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
424         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
425 
426         const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
427         const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) (k + 168));
428         vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
429         i7 += 8;
430 
431         vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
432 
433         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
434         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
435 
436         const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
437         const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) (k + 192));
438         vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
439         i8 += 8;
440 
441         vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
442 
443         vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
444         vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
445 
446         k += 8;
447 
448       vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
449       vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
450 
451       vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
452       vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
453 
454       const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
455       vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
456       vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
457 
458       const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
459       vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
460       vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
461 
462       const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
463       vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
464       vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
465 
466       const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
467       vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
468       vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
469 
470       v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
471       v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
472 
473       const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
474       vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
475 
476       w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
477 
478       if XNN_LIKELY(c >= 8) {
479         *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
480         output += 8;
481         c -= 8;
482       } else {
483         if (c & 4) {
484           *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
485           vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
486           output += 4;
487         }
488         uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
489         if (c & 2) {
490           *((uint16_t*) output) = (uint16_t) vout0123;
491           vout0123 >>= 16;
492           output += 2;
493         }
494         if (c & 1) {
495           *output = (uint8_t) vout0123;
496           output += 1;
497         }
498         c = 0;
499       }
500       } while (c != 0);
501     }
502 
503     output = (uint8_t*) ((uintptr_t) output + output_increment);
504   } while (--output_width != 0);
505 }
506