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(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(
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 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
184 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
185 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
186 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
187 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
188 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
189
190 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
191 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
192 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
193 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
194 const v128_t vi1xGHIJKLMN = wasm_i16x8_load8x8(i1 + 16);
195 const v128_t vk1xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
196 i1 += 24;
197
198 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
199 vprod89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF);
200 vprodGHIJKLMN = wasm_i16x8_mul(vi1xGHIJKLMN, vk1xGHIJKLMN);
201
202 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
203 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
204 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
205 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
206 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
207 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
208
209 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
210 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
211 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
212 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
213 const v128_t vi2xGHIJKLMN = wasm_i16x8_load8x8(i2 + 16);
214 const v128_t vk2xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
215 i2 += 24;
216
217 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
218 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
219 vprodGHIJKLMN = wasm_i16x8_mul(vi2xGHIJKLMN, vk2xGHIJKLMN);
220
221 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
222 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
223 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
224 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
225 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
226 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
227
228 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
229 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
230 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
231 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
232 const v128_t vi3xGHIJKLMN = wasm_i16x8_load8x8(i3 + 16);
233 const v128_t vk3xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
234 i3 += 24;
235
236 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
237 vprod89ABCDEF = wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF);
238 vprodGHIJKLMN = wasm_i16x8_mul(vi3xGHIJKLMN, vk3xGHIJKLMN);
239
240 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
241 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
242 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
243 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
244 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
245 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
246
247 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
248 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
249 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
250 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
251 const v128_t vi4xGHIJKLMN = wasm_i16x8_load8x8(i4 + 16);
252 const v128_t vk4xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
253 i4 += 24;
254
255 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
256 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
257 vprodGHIJKLMN = wasm_i16x8_mul(vi4xGHIJKLMN, vk4xGHIJKLMN);
258
259 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
260 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
261 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
262 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
263 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
264 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
265
266 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
267 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
268 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
269 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
270 const v128_t vi5xGHIJKLMN = wasm_i16x8_load8x8(i5 + 16);
271 const v128_t vk5xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
272 i5 += 24;
273
274 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
275 vprod89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF);
276 vprodGHIJKLMN = wasm_i16x8_mul(vi5xGHIJKLMN, vk5xGHIJKLMN);
277
278 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
279 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
280 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
281 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
282 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
283 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
284
285 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
286 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
287 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
288 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
289 const v128_t vi6xGHIJKLMN = wasm_i16x8_load8x8(i6 + 16);
290 const v128_t vk6xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
291 i6 += 24;
292
293 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
294 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
295 vprodGHIJKLMN = wasm_i16x8_mul(vi6xGHIJKLMN, vk6xGHIJKLMN);
296
297 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
298 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
299 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
300 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
301 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
302 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
303
304 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
305 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
306 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
307 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
308 const v128_t vi7xGHIJKLMN = wasm_i16x8_load8x8(i7 + 16);
309 const v128_t vk7xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
310 i7 += 24;
311
312 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
313 vprod89ABCDEF = wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF);
314 vprodGHIJKLMN = wasm_i16x8_mul(vi7xGHIJKLMN, vk7xGHIJKLMN);
315
316 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
317 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
318 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
319 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
320 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
321 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
322
323 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
324 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
325 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
326 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
327 const v128_t vi8xGHIJKLMN = wasm_i16x8_load8x8(i8 + 16);
328 const v128_t vk8xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
329 i8 += 24;
330
331 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
332 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
333 vprodGHIJKLMN = wasm_i16x8_mul(vi8xGHIJKLMN, vk8xGHIJKLMN);
334
335 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
336 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
337 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
338 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
339 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
340 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
341
342 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
343 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
344 const v128_t vi9x89ABCDEF = wasm_i16x8_load8x8(i9 + 8);
345 const v128_t vk9x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 224 * sizeof(int8_t)));
346 const v128_t vi9xGHIJKLMN = wasm_i16x8_load8x8(i9 + 16);
347 const v128_t vk9xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 232 * sizeof(int8_t)));
348 i9 += 24;
349
350 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
351 vprod89ABCDEF = wasm_i16x8_mul(vi9x89ABCDEF, vk9x89ABCDEF);
352 vprodGHIJKLMN = wasm_i16x8_mul(vi9xGHIJKLMN, vk9xGHIJKLMN);
353
354 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
355 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
356 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
357 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
358 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
359 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
360
361 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
362 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 240 * sizeof(int8_t)));
363 const v128_t vi10x89ABCDEF = wasm_i16x8_load8x8(i10 + 8);
364 const v128_t vk10x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 248 * sizeof(int8_t)));
365 const v128_t vi10xGHIJKLMN = wasm_i16x8_load8x8(i10 + 16);
366 const v128_t vk10xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 256 * sizeof(int8_t)));
367 i10 += 24;
368
369 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
370 vprod89ABCDEF = wasm_i16x8_mul(vi10x89ABCDEF, vk10x89ABCDEF);
371 vprodGHIJKLMN = wasm_i16x8_mul(vi10xGHIJKLMN, vk10xGHIJKLMN);
372
373 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
374 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
375 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
376 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
377 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
378 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
379
380 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
381 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 264 * sizeof(int8_t)));
382 const v128_t vi11x89ABCDEF = wasm_i16x8_load8x8(i11 + 8);
383 const v128_t vk11x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 272 * sizeof(int8_t)));
384 const v128_t vi11xGHIJKLMN = wasm_i16x8_load8x8(i11 + 16);
385 const v128_t vk11xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 280 * sizeof(int8_t)));
386 i11 += 24;
387
388 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
389 vprod89ABCDEF = wasm_i16x8_mul(vi11x89ABCDEF, vk11x89ABCDEF);
390 vprodGHIJKLMN = wasm_i16x8_mul(vi11xGHIJKLMN, vk11xGHIJKLMN);
391
392 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
393 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
394 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
395 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
396 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
397 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
398
399 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
400 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 288 * sizeof(int8_t)));
401 const v128_t vi12x89ABCDEF = wasm_i16x8_load8x8(i12 + 8);
402 const v128_t vk12x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 296 * sizeof(int8_t)));
403 const v128_t vi12xGHIJKLMN = wasm_i16x8_load8x8(i12 + 16);
404 const v128_t vk12xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 304 * sizeof(int8_t)));
405 i12 += 24;
406
407 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
408 vprod89ABCDEF = wasm_i16x8_mul(vi12x89ABCDEF, vk12x89ABCDEF);
409 vprodGHIJKLMN = wasm_i16x8_mul(vi12xGHIJKLMN, vk12xGHIJKLMN);
410
411 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
412 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
413 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
414 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
415 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
416 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
417
418 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
419 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 312 * sizeof(int8_t)));
420 const v128_t vi13x89ABCDEF = wasm_i16x8_load8x8(i13 + 8);
421 const v128_t vk13x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 320 * sizeof(int8_t)));
422 const v128_t vi13xGHIJKLMN = wasm_i16x8_load8x8(i13 + 16);
423 const v128_t vk13xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 328 * sizeof(int8_t)));
424 i13 += 24;
425
426 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
427 vprod89ABCDEF = wasm_i16x8_mul(vi13x89ABCDEF, vk13x89ABCDEF);
428 vprodGHIJKLMN = wasm_i16x8_mul(vi13xGHIJKLMN, vk13xGHIJKLMN);
429
430 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
431 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
432 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
433 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
434 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
435 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
436
437 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
438 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 336 * sizeof(int8_t)));
439 const v128_t vi14x89ABCDEF = wasm_i16x8_load8x8(i14 + 8);
440 const v128_t vk14x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 344 * sizeof(int8_t)));
441 const v128_t vi14xGHIJKLMN = wasm_i16x8_load8x8(i14 + 16);
442 const v128_t vk14xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 352 * sizeof(int8_t)));
443 i14 += 24;
444
445 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
446 vprod89ABCDEF = wasm_i16x8_mul(vi14x89ABCDEF, vk14x89ABCDEF);
447 vprodGHIJKLMN = wasm_i16x8_mul(vi14xGHIJKLMN, vk14xGHIJKLMN);
448
449 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
450 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
451 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
452 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
453 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
454 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
455
456 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
457 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 360 * sizeof(int8_t)));
458 const v128_t vi15x89ABCDEF = wasm_i16x8_load8x8(i15 + 8);
459 const v128_t vk15x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 368 * sizeof(int8_t)));
460 const v128_t vi15xGHIJKLMN = wasm_i16x8_load8x8(i15 + 16);
461 const v128_t vk15xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 376 * sizeof(int8_t)));
462 i15 += 24;
463
464 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
465 vprod89ABCDEF = wasm_i16x8_mul(vi15x89ABCDEF, vk15x89ABCDEF);
466 vprodGHIJKLMN = wasm_i16x8_mul(vi15xGHIJKLMN, vk15xGHIJKLMN);
467
468 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
469 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
470 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
471 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
472 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
473 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
474
475 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
476 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 384 * sizeof(int8_t)));
477 const v128_t vi16x89ABCDEF = wasm_i16x8_load8x8(i16 + 8);
478 const v128_t vk16x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 392 * sizeof(int8_t)));
479 const v128_t vi16xGHIJKLMN = wasm_i16x8_load8x8(i16 + 16);
480 const v128_t vk16xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 400 * sizeof(int8_t)));
481 i16 += 24;
482
483 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
484 vprod89ABCDEF = wasm_i16x8_mul(vi16x89ABCDEF, vk16x89ABCDEF);
485 vprodGHIJKLMN = wasm_i16x8_mul(vi16xGHIJKLMN, vk16xGHIJKLMN);
486
487 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
488 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
489 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
490 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
491 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
492 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
493
494 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
495 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 408 * sizeof(int8_t)));
496 const v128_t vi17x89ABCDEF = wasm_i16x8_load8x8(i17 + 8);
497 const v128_t vk17x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 416 * sizeof(int8_t)));
498 const v128_t vi17xGHIJKLMN = wasm_i16x8_load8x8(i17 + 16);
499 const v128_t vk17xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 424 * sizeof(int8_t)));
500 i17 += 24;
501
502 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
503 vprod89ABCDEF = wasm_i16x8_mul(vi17x89ABCDEF, vk17x89ABCDEF);
504 vprodGHIJKLMN = wasm_i16x8_mul(vi17xGHIJKLMN, vk17xGHIJKLMN);
505
506 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
507 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
508 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
509 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
510 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
511 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
512
513 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
514 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 432 * sizeof(int8_t)));
515 const v128_t vi18x89ABCDEF = wasm_i16x8_load8x8(i18 + 8);
516 const v128_t vk18x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 440 * sizeof(int8_t)));
517 const v128_t vi18xGHIJKLMN = wasm_i16x8_load8x8(i18 + 16);
518 const v128_t vk18xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 448 * sizeof(int8_t)));
519 i18 += 24;
520
521 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
522 vprod89ABCDEF = wasm_i16x8_mul(vi18x89ABCDEF, vk18x89ABCDEF);
523 vprodGHIJKLMN = wasm_i16x8_mul(vi18xGHIJKLMN, vk18xGHIJKLMN);
524
525 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
526 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
527 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
528 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
529 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
530 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
531
532 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
533 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 456 * sizeof(int8_t)));
534 const v128_t vi19x89ABCDEF = wasm_i16x8_load8x8(i19 + 8);
535 const v128_t vk19x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 464 * sizeof(int8_t)));
536 const v128_t vi19xGHIJKLMN = wasm_i16x8_load8x8(i19 + 16);
537 const v128_t vk19xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 472 * sizeof(int8_t)));
538 i19 += 24;
539
540 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
541 vprod89ABCDEF = wasm_i16x8_mul(vi19x89ABCDEF, vk19x89ABCDEF);
542 vprodGHIJKLMN = wasm_i16x8_mul(vi19xGHIJKLMN, vk19xGHIJKLMN);
543
544 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
545 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
546 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
547 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
548 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
549 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
550
551 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
552 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 480 * sizeof(int8_t)));
553 const v128_t vi20x89ABCDEF = wasm_i16x8_load8x8(i20 + 8);
554 const v128_t vk20x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 488 * sizeof(int8_t)));
555 const v128_t vi20xGHIJKLMN = wasm_i16x8_load8x8(i20 + 16);
556 const v128_t vk20xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 496 * sizeof(int8_t)));
557 i20 += 24;
558
559 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
560 vprod89ABCDEF = wasm_i16x8_mul(vi20x89ABCDEF, vk20x89ABCDEF);
561 vprodGHIJKLMN = wasm_i16x8_mul(vi20xGHIJKLMN, vk20xGHIJKLMN);
562
563 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
564 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
565 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
566 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
567 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
568 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
569
570 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
571 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 504 * sizeof(int8_t)));
572 const v128_t vi21x89ABCDEF = wasm_i16x8_load8x8(i21 + 8);
573 const v128_t vk21x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 512 * sizeof(int8_t)));
574 const v128_t vi21xGHIJKLMN = wasm_i16x8_load8x8(i21 + 16);
575 const v128_t vk21xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 520 * sizeof(int8_t)));
576 i21 += 24;
577
578 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
579 vprod89ABCDEF = wasm_i16x8_mul(vi21x89ABCDEF, vk21x89ABCDEF);
580 vprodGHIJKLMN = wasm_i16x8_mul(vi21xGHIJKLMN, vk21xGHIJKLMN);
581
582 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
583 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
584 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
585 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
586 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
587 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
588
589 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
590 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 528 * sizeof(int8_t)));
591 const v128_t vi22x89ABCDEF = wasm_i16x8_load8x8(i22 + 8);
592 const v128_t vk22x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 536 * sizeof(int8_t)));
593 const v128_t vi22xGHIJKLMN = wasm_i16x8_load8x8(i22 + 16);
594 const v128_t vk22xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 544 * sizeof(int8_t)));
595 i22 += 24;
596
597 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
598 vprod89ABCDEF = wasm_i16x8_mul(vi22x89ABCDEF, vk22x89ABCDEF);
599 vprodGHIJKLMN = wasm_i16x8_mul(vi22xGHIJKLMN, vk22xGHIJKLMN);
600
601 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
602 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
603 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
604 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
605 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
606 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
607
608 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
609 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 552 * sizeof(int8_t)));
610 const v128_t vi23x89ABCDEF = wasm_i16x8_load8x8(i23 + 8);
611 const v128_t vk23x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 560 * sizeof(int8_t)));
612 const v128_t vi23xGHIJKLMN = wasm_i16x8_load8x8(i23 + 16);
613 const v128_t vk23xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 568 * sizeof(int8_t)));
614 i23 += 24;
615
616 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
617 vprod89ABCDEF = wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF);
618 vprodGHIJKLMN = wasm_i16x8_mul(vi23xGHIJKLMN, vk23xGHIJKLMN);
619
620 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
621 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
622 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
623 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
624 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
625 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
626
627 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
628 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 576 * sizeof(int8_t)));
629 const v128_t vi24x89ABCDEF = wasm_i16x8_load8x8(i24 + 8);
630 const v128_t vk24x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 584 * sizeof(int8_t)));
631 const v128_t vi24xGHIJKLMN = wasm_i16x8_load8x8(i24 + 16);
632 const v128_t vk24xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 592 * sizeof(int8_t)));
633 i24 += 24;
634
635 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
636 vprod89ABCDEF = wasm_i16x8_mul(vi24x89ABCDEF, vk24x89ABCDEF);
637 vprodGHIJKLMN = wasm_i16x8_mul(vi24xGHIJKLMN, vk24xGHIJKLMN);
638
639 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
640 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
641 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
642 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
643 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
644 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
645
646
647 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t));
648
649 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
650 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
651 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
652 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
653 vaccGHIJ = wasm_f32x4_convert_i32x4(vaccGHIJ);
654 vaccKLMN = wasm_f32x4_convert_i32x4(vaccKLMN);
655
656 const v128_t vscale0123 = wasm_v128_load(w);
657 const v128_t vscale4567 = wasm_v128_load((const float*) w + 4);
658 const v128_t vscale89AB = wasm_v128_load((const float*) w + 8);
659 const v128_t vscaleCDEF = wasm_v128_load((const float*) w + 12);
660 const v128_t vscaleGHIJ = wasm_v128_load((const float*) w + 16);
661 const v128_t vscaleKLMN = wasm_v128_load((const float*) w + 20);
662 w = (const void*) ((const float*) w + 24);
663
664 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
665 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
666 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale89AB);
667 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscaleCDEF);
668 vaccGHIJ = wasm_f32x4_mul(vaccGHIJ, vscaleGHIJ);
669 vaccKLMN = wasm_f32x4_mul(vaccKLMN, vscaleKLMN);
670
671 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
672 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
673 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
674 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
675 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
676 vaccGHIJ = wasm_f32x4_add(vaccGHIJ, vmagic_bias);
677 vaccKLMN = wasm_f32x4_add(vaccKLMN, vmagic_bias);
678
679 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
680 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
681 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
682 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
683 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
684 vaccGHIJ = wasm_i32x4_max(vaccGHIJ, vmagic_min);
685 vaccKLMN = wasm_i32x4_max(vaccKLMN, vmagic_min);
686
687 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
688 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
689 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
690 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
691 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
692 vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, vmagic_bias_less_output_zero_point);
693 vaccKLMN = wasm_i32x4_sub(vaccKLMN, vmagic_bias_less_output_zero_point);
694
695 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
696 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
697 v128_t voutGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
698
699 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
700 v128_t voutGHIJKLMNGHIJKLMN = wasm_i8x16_narrow_i16x8(voutGHIJKLMN, voutGHIJKLMN);
701
702 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
703 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
704 voutGHIJKLMNGHIJKLMN = wasm_i8x16_min(voutGHIJKLMNGHIJKLMN, voutput_max);
705
706 wasm_v128_store(output, vout0123456789ABCDEF);
707 *((double*) (output + 16)) = wasm_f64x2_extract_lane(voutGHIJKLMNGHIJKLMN, 0);
708 output += 24;
709 }
710 if XNN_UNLIKELY(c != 0) {
711 const int8_t* k = (const int8_t*) ((uintptr_t) w + 24 * sizeof(int32_t));
712 do {
713 v128_t vacc0123 = wasm_v128_load(w);
714 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
715
716
717 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
718 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
719 i0 += 8;
720
721 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
722
723 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
724 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
725
726 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
727 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 24));
728 i1 += 8;
729
730 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
731
732 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
733 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
734
735 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
736 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
737 i2 += 8;
738
739 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
740
741 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
742 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
743
744 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
745 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 72));
746 i3 += 8;
747
748 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
749
750 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
751 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
752
753 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
754 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
755 i4 += 8;
756
757 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
758
759 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
760 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
761
762 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
763 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 120));
764 i5 += 8;
765
766 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
767
768 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
769 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
770
771 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
772 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 144));
773 i6 += 8;
774
775 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
776
777 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
778 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
779
780 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
781 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 168));
782 i7 += 8;
783
784 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
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 vi8x01234567 = wasm_i16x8_load8x8(i8);
790 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 192));
791 i8 += 8;
792
793 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
794
795 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
796 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
797
798 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
799 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) (k + 216));
800 i9 += 8;
801
802 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
803
804 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
805 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
806
807 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
808 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) (k + 240));
809 i10 += 8;
810
811 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
812
813 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
814 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
815
816 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
817 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) (k + 264));
818 i11 += 8;
819
820 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
821
822 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
823 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
824
825 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
826 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) (k + 288));
827 i12 += 8;
828
829 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
830
831 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
832 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
833
834 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
835 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) (k + 312));
836 i13 += 8;
837
838 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
839
840 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
841 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
842
843 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
844 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) (k + 336));
845 i14 += 8;
846
847 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
848
849 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
850 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
851
852 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
853 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) (k + 360));
854 i15 += 8;
855
856 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
857
858 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
859 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
860
861 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
862 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) (k + 384));
863 i16 += 8;
864
865 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
866
867 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
868 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
869
870 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
871 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) (k + 408));
872 i17 += 8;
873
874 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
875
876 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
877 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
878
879 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
880 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) (k + 432));
881 i18 += 8;
882
883 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
884
885 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
886 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
887
888 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
889 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) (k + 456));
890 i19 += 8;
891
892 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
893
894 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
895 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
896
897 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
898 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) (k + 480));
899 i20 += 8;
900
901 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
902
903 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
904 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
905
906 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
907 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) (k + 504));
908 i21 += 8;
909
910 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
911
912 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
913 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
914
915 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
916 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) (k + 528));
917 i22 += 8;
918
919 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
920
921 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
922 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
923
924 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
925 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) (k + 552));
926 i23 += 8;
927
928 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
929
930 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
931 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
932
933 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
934 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) (k + 576));
935 i24 += 8;
936
937 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
938
939 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
940 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
941
942 k += 8;
943
944
945 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
946 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
947
948 const v128_t vscale0123 = wasm_v128_load((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t)));
949 const v128_t vscale4567 = wasm_v128_load((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 600 * sizeof(int8_t) + 4 * sizeof(float)));
950
951 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
952 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
953
954 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
955 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
956 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
957
958 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
959 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
960 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
961
962 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
963 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
964 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
965
966 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
967 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
968
969 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
970 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
971
972 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
973
974 if XNN_LIKELY(c >= 8) {
975 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
976 output += 8;
977 c -= 8;
978 } else {
979 if (c & 4) {
980 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
981 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
982 output += 4;
983 }
984 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
985 if (c & 2) {
986 *((uint16_t*) output) = (uint16_t) vout0123;
987 vout0123 >>= 16;
988 output += 2;
989 }
990 if (c & 1) {
991 *output = (int8_t) vout0123;
992 output += 1;
993 }
994 c = 0;
995 }
996 } while (c != 0);
997 }
998
999 output = (int8_t*) ((uintptr_t) output + output_increment);
1000 } while (--output_width != 0);
1001 }
1002