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_qc8_dwconv_minmax_fp32_ukernel_up24x25__wasmsimd_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__wasmsimd_mul16_add16(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 const int8_t* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
82 }
83 const int8_t* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
87 }
88 const int8_t* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
92 }
93 const int8_t* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
97 }
98 const int8_t* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
102 }
103 const int8_t* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
107 }
108 const int8_t* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
112 }
113 const int8_t* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
117 }
118 const int8_t* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
122 }
123 const int8_t* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
127 }
128 const int8_t* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
132 }
133 const int8_t* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
137 }
138 const int8_t* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
142 }
143 const int8_t* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
147 }
148 const int8_t* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
152 }
153 const int8_t* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const int8_t**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const void* w = weights;
162 for (; c >= 24; c -= 24) {
163 v128_t vacc0123 = wasm_v128_load(w);
164 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
165 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
166 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
167 v128_t vaccGHIJ = wasm_v128_load((const void*) ((uintptr_t) w + 16 * sizeof(int32_t)));
168 v128_t vaccKLMN = wasm_v128_load((const void*) ((uintptr_t) w + 20 * sizeof(int32_t)));
169
170
171 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
172 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
173 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
174 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
175 const v128_t vi0xGHIJKLMN = wasm_i16x8_load8x8(i0 + 16);
176 const v128_t vk0xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
177 i0 += 24;
178
179 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
180 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
181 v128_t vprodGHIJKLMN = wasm_i16x8_mul(vi0xGHIJKLMN, vk0xGHIJKLMN);
182
183
184 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
185 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
186 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
187 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
188 const v128_t vi1xGHIJKLMN = wasm_i16x8_load8x8(i1 + 16);
189 const v128_t vk1xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
190 i1 += 24;
191
192 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
193 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
194 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi1xGHIJKLMN, vk1xGHIJKLMN));
195
196 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
197 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
198 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
199 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
200 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
201 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
202
203 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
204 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
205 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
206 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
207 const v128_t vi2xGHIJKLMN = wasm_i16x8_load8x8(i2 + 16);
208 const v128_t vk2xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
209 i2 += 24;
210
211 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
212 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
213 vprodGHIJKLMN = wasm_i16x8_mul(vi2xGHIJKLMN, vk2xGHIJKLMN);
214
215
216 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
217 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
218 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
219 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
220 const v128_t vi3xGHIJKLMN = wasm_i16x8_load8x8(i3 + 16);
221 const v128_t vk3xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
222 i3 += 24;
223
224 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
225 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
226 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi3xGHIJKLMN, vk3xGHIJKLMN));
227
228 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
229 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
230 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
231 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
232 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
233 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
234
235 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
236 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
237 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
238 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
239 const v128_t vi4xGHIJKLMN = wasm_i16x8_load8x8(i4 + 16);
240 const v128_t vk4xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
241 i4 += 24;
242
243 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
244 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
245 vprodGHIJKLMN = wasm_i16x8_mul(vi4xGHIJKLMN, vk4xGHIJKLMN);
246
247
248 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
249 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
250 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
251 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
252 const v128_t vi5xGHIJKLMN = wasm_i16x8_load8x8(i5 + 16);
253 const v128_t vk5xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
254 i5 += 24;
255
256 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
257 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF));
258 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi5xGHIJKLMN, vk5xGHIJKLMN));
259
260 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
261 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
262 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
263 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
264 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
265 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
266
267 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
268 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
269 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
270 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
271 const v128_t vi6xGHIJKLMN = wasm_i16x8_load8x8(i6 + 16);
272 const v128_t vk6xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
273 i6 += 24;
274
275 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
276 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
277 vprodGHIJKLMN = wasm_i16x8_mul(vi6xGHIJKLMN, vk6xGHIJKLMN);
278
279
280 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
281 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
282 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
283 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
284 const v128_t vi7xGHIJKLMN = wasm_i16x8_load8x8(i7 + 16);
285 const v128_t vk7xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
286 i7 += 24;
287
288 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
289 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF));
290 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi7xGHIJKLMN, vk7xGHIJKLMN));
291
292 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
293 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
294 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
295 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
296 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
297 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
298
299 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
300 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
301 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
302 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
303 const v128_t vi8xGHIJKLMN = wasm_i16x8_load8x8(i8 + 16);
304 const v128_t vk8xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
305 i8 += 24;
306
307 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
308 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
309 vprodGHIJKLMN = wasm_i16x8_mul(vi8xGHIJKLMN, vk8xGHIJKLMN);
310
311
312 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
313 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
314 const v128_t vi9x89ABCDEF = wasm_i16x8_load8x8(i9 + 8);
315 const v128_t vk9x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t)));
316 const v128_t vi9xGHIJKLMN = wasm_i16x8_load8x8(i9 + 16);
317 const v128_t vk9xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t)));
318 i9 += 24;
319
320 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi9x01234567, vk9x01234567));
321 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi9x89ABCDEF, vk9x89ABCDEF));
322 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi9xGHIJKLMN, vk9xGHIJKLMN));
323
324 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
325 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
326 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
327 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
328 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
329 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
330
331 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
332 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t)));
333 const v128_t vi10x89ABCDEF = wasm_i16x8_load8x8(i10 + 8);
334 const v128_t vk10x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t)));
335 const v128_t vi10xGHIJKLMN = wasm_i16x8_load8x8(i10 + 16);
336 const v128_t vk10xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t)));
337 i10 += 24;
338
339 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
340 vprod89ABCDEF = wasm_i16x8_mul(vi10x89ABCDEF, vk10x89ABCDEF);
341 vprodGHIJKLMN = wasm_i16x8_mul(vi10xGHIJKLMN, vk10xGHIJKLMN);
342
343
344 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
345 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t)));
346 const v128_t vi11x89ABCDEF = wasm_i16x8_load8x8(i11 + 8);
347 const v128_t vk11x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t)));
348 const v128_t vi11xGHIJKLMN = wasm_i16x8_load8x8(i11 + 16);
349 const v128_t vk11xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t)));
350 i11 += 24;
351
352 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi11x01234567, vk11x01234567));
353 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi11x89ABCDEF, vk11x89ABCDEF));
354 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi11xGHIJKLMN, vk11xGHIJKLMN));
355
356 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
357 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
358 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
359 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
360 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
361 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
362
363 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
364 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)));
365 const v128_t vi12x89ABCDEF = wasm_i16x8_load8x8(i12 + 8);
366 const v128_t vk12x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t)));
367 const v128_t vi12xGHIJKLMN = wasm_i16x8_load8x8(i12 + 16);
368 const v128_t vk12xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t)));
369 i12 += 24;
370
371 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
372 vprod89ABCDEF = wasm_i16x8_mul(vi12x89ABCDEF, vk12x89ABCDEF);
373 vprodGHIJKLMN = wasm_i16x8_mul(vi12xGHIJKLMN, vk12xGHIJKLMN);
374
375
376 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
377 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t)));
378 const v128_t vi13x89ABCDEF = wasm_i16x8_load8x8(i13 + 8);
379 const v128_t vk13x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t)));
380 const v128_t vi13xGHIJKLMN = wasm_i16x8_load8x8(i13 + 16);
381 const v128_t vk13xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t)));
382 i13 += 24;
383
384 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi13x01234567, vk13x01234567));
385 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi13x89ABCDEF, vk13x89ABCDEF));
386 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi13xGHIJKLMN, vk13xGHIJKLMN));
387
388 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
389 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
390 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
391 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
392 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
393 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
394
395 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
396 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t)));
397 const v128_t vi14x89ABCDEF = wasm_i16x8_load8x8(i14 + 8);
398 const v128_t vk14x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t)));
399 const v128_t vi14xGHIJKLMN = wasm_i16x8_load8x8(i14 + 16);
400 const v128_t vk14xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t)));
401 i14 += 24;
402
403 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
404 vprod89ABCDEF = wasm_i16x8_mul(vi14x89ABCDEF, vk14x89ABCDEF);
405 vprodGHIJKLMN = wasm_i16x8_mul(vi14xGHIJKLMN, vk14xGHIJKLMN);
406
407
408 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
409 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t)));
410 const v128_t vi15x89ABCDEF = wasm_i16x8_load8x8(i15 + 8);
411 const v128_t vk15x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t)));
412 const v128_t vi15xGHIJKLMN = wasm_i16x8_load8x8(i15 + 16);
413 const v128_t vk15xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t)));
414 i15 += 24;
415
416 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi15x01234567, vk15x01234567));
417 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi15x89ABCDEF, vk15x89ABCDEF));
418 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi15xGHIJKLMN, vk15xGHIJKLMN));
419
420 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
421 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
422 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
423 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
424 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
425 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
426
427 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
428 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t)));
429 const v128_t vi16x89ABCDEF = wasm_i16x8_load8x8(i16 + 8);
430 const v128_t vk16x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t)));
431 const v128_t vi16xGHIJKLMN = wasm_i16x8_load8x8(i16 + 16);
432 const v128_t vk16xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t)));
433 i16 += 24;
434
435 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
436 vprod89ABCDEF = wasm_i16x8_mul(vi16x89ABCDEF, vk16x89ABCDEF);
437 vprodGHIJKLMN = wasm_i16x8_mul(vi16xGHIJKLMN, vk16xGHIJKLMN);
438
439
440 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
441 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t)));
442 const v128_t vi17x89ABCDEF = wasm_i16x8_load8x8(i17 + 8);
443 const v128_t vk17x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t)));
444 const v128_t vi17xGHIJKLMN = wasm_i16x8_load8x8(i17 + 16);
445 const v128_t vk17xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t)));
446 i17 += 24;
447
448 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi17x01234567, vk17x01234567));
449 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi17x89ABCDEF, vk17x89ABCDEF));
450 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi17xGHIJKLMN, vk17xGHIJKLMN));
451
452 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
453 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
454 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
455 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
456 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
457 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
458
459 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
460 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t)));
461 const v128_t vi18x89ABCDEF = wasm_i16x8_load8x8(i18 + 8);
462 const v128_t vk18x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t)));
463 const v128_t vi18xGHIJKLMN = wasm_i16x8_load8x8(i18 + 16);
464 const v128_t vk18xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t)));
465 i18 += 24;
466
467 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
468 vprod89ABCDEF = wasm_i16x8_mul(vi18x89ABCDEF, vk18x89ABCDEF);
469 vprodGHIJKLMN = wasm_i16x8_mul(vi18xGHIJKLMN, vk18xGHIJKLMN);
470
471
472 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
473 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t)));
474 const v128_t vi19x89ABCDEF = wasm_i16x8_load8x8(i19 + 8);
475 const v128_t vk19x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t)));
476 const v128_t vi19xGHIJKLMN = wasm_i16x8_load8x8(i19 + 16);
477 const v128_t vk19xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t)));
478 i19 += 24;
479
480 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi19x01234567, vk19x01234567));
481 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi19x89ABCDEF, vk19x89ABCDEF));
482 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi19xGHIJKLMN, vk19xGHIJKLMN));
483
484 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
485 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
486 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
487 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
488 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
489 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
490
491 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
492 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t)));
493 const v128_t vi20x89ABCDEF = wasm_i16x8_load8x8(i20 + 8);
494 const v128_t vk20x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t)));
495 const v128_t vi20xGHIJKLMN = wasm_i16x8_load8x8(i20 + 16);
496 const v128_t vk20xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t)));
497 i20 += 24;
498
499 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
500 vprod89ABCDEF = wasm_i16x8_mul(vi20x89ABCDEF, vk20x89ABCDEF);
501 vprodGHIJKLMN = wasm_i16x8_mul(vi20xGHIJKLMN, vk20xGHIJKLMN);
502
503
504 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
505 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t)));
506 const v128_t vi21x89ABCDEF = wasm_i16x8_load8x8(i21 + 8);
507 const v128_t vk21x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t)));
508 const v128_t vi21xGHIJKLMN = wasm_i16x8_load8x8(i21 + 16);
509 const v128_t vk21xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t)));
510 i21 += 24;
511
512 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi21x01234567, vk21x01234567));
513 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi21x89ABCDEF, vk21x89ABCDEF));
514 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi21xGHIJKLMN, vk21xGHIJKLMN));
515
516 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
517 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
518 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
519 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
520 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
521 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
522
523 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
524 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t)));
525 const v128_t vi22x89ABCDEF = wasm_i16x8_load8x8(i22 + 8);
526 const v128_t vk22x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t)));
527 const v128_t vi22xGHIJKLMN = wasm_i16x8_load8x8(i22 + 16);
528 const v128_t vk22xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t)));
529 i22 += 24;
530
531 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
532 vprod89ABCDEF = wasm_i16x8_mul(vi22x89ABCDEF, vk22x89ABCDEF);
533 vprodGHIJKLMN = wasm_i16x8_mul(vi22xGHIJKLMN, vk22xGHIJKLMN);
534
535
536 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
537 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t)));
538 const v128_t vi23x89ABCDEF = wasm_i16x8_load8x8(i23 + 8);
539 const v128_t vk23x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t)));
540 const v128_t vi23xGHIJKLMN = wasm_i16x8_load8x8(i23 + 16);
541 const v128_t vk23xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t)));
542 i23 += 24;
543
544 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi23x01234567, vk23x01234567));
545 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF));
546 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi23xGHIJKLMN, vk23xGHIJKLMN));
547
548 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
549 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
550 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
551 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
552 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
553 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
554
555 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
556 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t)));
557 const v128_t vi24x89ABCDEF = wasm_i16x8_load8x8(i24 + 8);
558 const v128_t vk24x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t)));
559 const v128_t vi24xGHIJKLMN = wasm_i16x8_load8x8(i24 + 16);
560 const v128_t vk24xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t)));
561 i24 += 24;
562
563 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
564 vprod89ABCDEF = wasm_i16x8_mul(vi24x89ABCDEF, vk24x89ABCDEF);
565 vprodGHIJKLMN = wasm_i16x8_mul(vi24xGHIJKLMN, vk24xGHIJKLMN);
566
567 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
568 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
569 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
570 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
571 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
572 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
573
574
575 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
576
577 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
578 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
579 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
580 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
581 vaccGHIJ = wasm_f32x4_convert_i32x4(vaccGHIJ);
582 vaccKLMN = wasm_f32x4_convert_i32x4(vaccKLMN);
583
584 const v128_t vscale0123 = wasm_v128_load(w);
585 const v128_t vscale4567 = wasm_v128_load((const float*) w + 4);
586 const v128_t vscale89AB = wasm_v128_load((const float*) w + 8);
587 const v128_t vscaleCDEF = wasm_v128_load((const float*) w + 12);
588 const v128_t vscaleGHIJ = wasm_v128_load((const float*) w + 16);
589 const v128_t vscaleKLMN = wasm_v128_load((const float*) w + 20);
590 w = (const void*) ((const float*) w + 24);
591
592 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
593 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
594 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale89AB);
595 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscaleCDEF);
596 vaccGHIJ = wasm_f32x4_mul(vaccGHIJ, vscaleGHIJ);
597 vaccKLMN = wasm_f32x4_mul(vaccKLMN, vscaleKLMN);
598
599 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
600 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
601 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
602 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
603 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
604 vaccGHIJ = wasm_f32x4_add(vaccGHIJ, vmagic_bias);
605 vaccKLMN = wasm_f32x4_add(vaccKLMN, vmagic_bias);
606
607 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
608 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
609 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
610 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
611 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
612 vaccGHIJ = wasm_i32x4_max(vaccGHIJ, vmagic_min);
613 vaccKLMN = wasm_i32x4_max(vaccKLMN, vmagic_min);
614
615 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
616 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
617 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
618 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
619 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
620 vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, vmagic_bias_less_output_zero_point);
621 vaccKLMN = wasm_i32x4_sub(vaccKLMN, vmagic_bias_less_output_zero_point);
622
623 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
624 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
625 v128_t voutGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
626
627 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
628 v128_t voutGHIJKLMNGHIJKLMN = wasm_i8x16_narrow_i16x8(voutGHIJKLMN, voutGHIJKLMN);
629
630 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
631 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
632 voutGHIJKLMNGHIJKLMN = wasm_i8x16_min(voutGHIJKLMNGHIJKLMN, voutput_max);
633
634 wasm_v128_store(output, vout0123456789ABCDEF);
635 *((double*) (output + 16)) = wasm_f64x2_extract_lane(voutGHIJKLMNGHIJKLMN, 0);
636 output += 24;
637 }
638 if XNN_UNLIKELY(c != 0) {
639 const int8_t* k = (const int8_t*) ((uintptr_t) w + 24 * sizeof(int32_t));
640 do {
641 v128_t vacc0123 = wasm_v128_load(w);
642 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
643
644
645 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
646 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
647 i0 += 8;
648
649 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
650
651
652 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
653 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 24));
654 i1 += 8;
655
656 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
657
658 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
659 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
660
661 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
662 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
663 i2 += 8;
664
665 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
666
667
668 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
669 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 72));
670 i3 += 8;
671
672 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
673
674 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
675 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
676
677 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
678 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
679 i4 += 8;
680
681 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
682
683
684 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
685 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 120));
686 i5 += 8;
687
688 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
689
690 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
691 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
692
693 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
694 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 144));
695 i6 += 8;
696
697 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
698
699
700 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
701 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 168));
702 i7 += 8;
703
704 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
705
706 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
707 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
708
709 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
710 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 192));
711 i8 += 8;
712
713 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
714
715
716 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
717 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) (k + 216));
718 i9 += 8;
719
720 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi9x01234567, vk9x01234567));
721
722 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
723 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
724
725 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
726 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) (k + 240));
727 i10 += 8;
728
729 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
730
731
732 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
733 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) (k + 264));
734 i11 += 8;
735
736 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi11x01234567, vk11x01234567));
737
738 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
739 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
740
741 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
742 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) (k + 288));
743 i12 += 8;
744
745 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
746
747
748 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
749 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) (k + 312));
750 i13 += 8;
751
752 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi13x01234567, vk13x01234567));
753
754 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
755 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
756
757 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
758 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) (k + 336));
759 i14 += 8;
760
761 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
762
763
764 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
765 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) (k + 360));
766 i15 += 8;
767
768 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi15x01234567, vk15x01234567));
769
770 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
771 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
772
773 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
774 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) (k + 384));
775 i16 += 8;
776
777 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
778
779
780 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
781 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) (k + 408));
782 i17 += 8;
783
784 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi17x01234567, vk17x01234567));
785
786 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
787 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
788
789 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
790 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) (k + 432));
791 i18 += 8;
792
793 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
794
795
796 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
797 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) (k + 456));
798 i19 += 8;
799
800 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi19x01234567, vk19x01234567));
801
802 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
803 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
804
805 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
806 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) (k + 480));
807 i20 += 8;
808
809 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
810
811
812 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
813 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) (k + 504));
814 i21 += 8;
815
816 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi21x01234567, vk21x01234567));
817
818 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
819 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
820
821 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
822 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) (k + 528));
823 i22 += 8;
824
825 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
826
827
828 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
829 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) (k + 552));
830 i23 += 8;
831
832 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi23x01234567, vk23x01234567));
833
834 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
835 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
836
837 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
838 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) (k + 576));
839 i24 += 8;
840
841 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
842
843 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
844 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
845
846 k += 8;
847
848
849 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
850 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
851
852 const v128_t vscale0123 = wasm_v128_load((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t)));
853 const v128_t vscale4567 = wasm_v128_load((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t) + 4 * sizeof(float)));
854
855 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
856 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
857
858 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
859 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
860 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
861
862 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
863 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
864 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
865
866 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
867 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
868 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
869
870 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
871 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
872
873 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
874 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
875
876 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
877
878 if XNN_LIKELY(c >= 8) {
879 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
880 output += 8;
881 c -= 8;
882 } else {
883 if (c & 4) {
884 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
885 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
886 output += 4;
887 }
888 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
889 if (c & 2) {
890 *((uint16_t*) output) = (uint16_t) vout0123;
891 vout0123 >>= 16;
892 output += 2;
893 }
894 if (c & 1) {
895 *output = (int8_t) vout0123;
896 output += 1;
897 }
898 c = 0;
899 }
900 } while (c != 0);
901 }
902
903 output = (int8_t*) ((uintptr_t) output + output_increment);
904 } while (--output_width != 0);
905 }
906