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_qs8_dwconv_minmax_fp32_ukernel_up16x25__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__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_qs8_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 >= 16; c -= 16) {
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
168
169 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
170 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
171 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
172 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
173 i0 += 16;
174
175 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
176 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
177
178 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
179 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
180 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
181 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
182
183 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
184 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
185 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
186 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
187 i1 += 16;
188
189 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
190 vprod89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF);
191
192 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
193 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
194 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
195 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
196
197 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
198 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
199 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
200 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
201 i2 += 16;
202
203 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
204 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
205
206 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
207 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
208 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
209 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
210
211 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
212 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
213 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
214 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
215 i3 += 16;
216
217 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
218 vprod89ABCDEF = wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF);
219
220 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
221 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
222 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
223 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
224
225 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
226 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
227 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
228 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
229 i4 += 16;
230
231 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
232 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
233
234 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
235 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
236 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
237 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
238
239 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
240 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
241 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
242 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
243 i5 += 16;
244
245 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
246 vprod89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF);
247
248 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
249 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
250 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
251 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
252
253 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
254 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
255 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
256 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
257 i6 += 16;
258
259 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
260 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
261
262 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
263 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
264 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
265 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
266
267 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
268 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
269 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
270 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
271 i7 += 16;
272
273 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
274 vprod89ABCDEF = wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF);
275
276 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
277 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
278 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
279 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
280
281 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
282 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
283 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
284 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
285 i8 += 16;
286
287 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
288 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
289
290 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
291 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
292 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
293 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
294
295 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
296 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
297 const v128_t vi9x89ABCDEF = wasm_i16x8_load8x8(i9 + 8);
298 const v128_t vk9x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
299 i9 += 16;
300
301 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
302 vprod89ABCDEF = wasm_i16x8_mul(vi9x89ABCDEF, vk9x89ABCDEF);
303
304 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
305 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
306 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
307 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
308
309 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
310 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
311 const v128_t vi10x89ABCDEF = wasm_i16x8_load8x8(i10 + 8);
312 const v128_t vk10x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
313 i10 += 16;
314
315 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
316 vprod89ABCDEF = wasm_i16x8_mul(vi10x89ABCDEF, vk10x89ABCDEF);
317
318 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
319 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
320 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
321 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
322
323 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
324 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
325 const v128_t vi11x89ABCDEF = wasm_i16x8_load8x8(i11 + 8);
326 const v128_t vk11x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
327 i11 += 16;
328
329 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
330 vprod89ABCDEF = wasm_i16x8_mul(vi11x89ABCDEF, vk11x89ABCDEF);
331
332 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
333 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
334 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
335 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
336
337 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
338 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
339 const v128_t vi12x89ABCDEF = wasm_i16x8_load8x8(i12 + 8);
340 const v128_t vk12x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
341 i12 += 16;
342
343 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
344 vprod89ABCDEF = wasm_i16x8_mul(vi12x89ABCDEF, vk12x89ABCDEF);
345
346 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
347 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
348 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
349 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
350
351 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
352 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
353 const v128_t vi13x89ABCDEF = wasm_i16x8_load8x8(i13 + 8);
354 const v128_t vk13x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
355 i13 += 16;
356
357 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
358 vprod89ABCDEF = wasm_i16x8_mul(vi13x89ABCDEF, vk13x89ABCDEF);
359
360 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
361 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
362 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
363 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
364
365 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
366 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
367 const v128_t vi14x89ABCDEF = wasm_i16x8_load8x8(i14 + 8);
368 const v128_t vk14x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
369 i14 += 16;
370
371 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
372 vprod89ABCDEF = wasm_i16x8_mul(vi14x89ABCDEF, vk14x89ABCDEF);
373
374 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
375 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
376 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
377 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
378
379 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
380 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
381 const v128_t vi15x89ABCDEF = wasm_i16x8_load8x8(i15 + 8);
382 const v128_t vk15x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
383 i15 += 16;
384
385 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
386 vprod89ABCDEF = wasm_i16x8_mul(vi15x89ABCDEF, vk15x89ABCDEF);
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
393 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
394 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
395 const v128_t vi16x89ABCDEF = wasm_i16x8_load8x8(i16 + 8);
396 const v128_t vk16x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
397 i16 += 16;
398
399 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
400 vprod89ABCDEF = wasm_i16x8_mul(vi16x89ABCDEF, vk16x89ABCDEF);
401
402 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
403 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
404 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
405 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
406
407 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
408 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
409 const v128_t vi17x89ABCDEF = wasm_i16x8_load8x8(i17 + 8);
410 const v128_t vk17x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
411 i17 += 16;
412
413 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
414 vprod89ABCDEF = wasm_i16x8_mul(vi17x89ABCDEF, vk17x89ABCDEF);
415
416 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
417 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
418 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
419 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
420
421 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
422 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
423 const v128_t vi18x89ABCDEF = wasm_i16x8_load8x8(i18 + 8);
424 const v128_t vk18x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
425 i18 += 16;
426
427 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
428 vprod89ABCDEF = wasm_i16x8_mul(vi18x89ABCDEF, vk18x89ABCDEF);
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
435 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
436 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
437 const v128_t vi19x89ABCDEF = wasm_i16x8_load8x8(i19 + 8);
438 const v128_t vk19x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
439 i19 += 16;
440
441 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
442 vprod89ABCDEF = wasm_i16x8_mul(vi19x89ABCDEF, vk19x89ABCDEF);
443
444 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
445 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
446 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
447 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
448
449 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
450 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
451 const v128_t vi20x89ABCDEF = wasm_i16x8_load8x8(i20 + 8);
452 const v128_t vk20x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
453 i20 += 16;
454
455 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
456 vprod89ABCDEF = wasm_i16x8_mul(vi20x89ABCDEF, vk20x89ABCDEF);
457
458 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
459 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
460 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
461 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
462
463 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
464 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
465 const v128_t vi21x89ABCDEF = wasm_i16x8_load8x8(i21 + 8);
466 const v128_t vk21x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
467 i21 += 16;
468
469 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
470 vprod89ABCDEF = wasm_i16x8_mul(vi21x89ABCDEF, vk21x89ABCDEF);
471
472 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
473 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
474 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
475 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
476
477 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
478 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
479 const v128_t vi22x89ABCDEF = wasm_i16x8_load8x8(i22 + 8);
480 const v128_t vk22x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
481 i22 += 16;
482
483 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
484 vprod89ABCDEF = wasm_i16x8_mul(vi22x89ABCDEF, vk22x89ABCDEF);
485
486 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
487 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
488 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
489 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
490
491 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
492 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
493 const v128_t vi23x89ABCDEF = wasm_i16x8_load8x8(i23 + 8);
494 const v128_t vk23x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
495 i23 += 16;
496
497 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
498 vprod89ABCDEF = wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF);
499
500 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
501 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
502 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
503 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
504
505 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
506 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
507 const v128_t vi24x89ABCDEF = wasm_i16x8_load8x8(i24 + 8);
508 const v128_t vk24x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
509 i24 += 16;
510
511 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
512 vprod89ABCDEF = wasm_i16x8_mul(vi24x89ABCDEF, vk24x89ABCDEF);
513
514 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
515 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
516 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
517 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
518
519
520 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
521
522 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
523 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
524 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
525 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
526
527 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
528 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
529 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
530 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
531 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
532
533 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
534 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
535 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
536 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
537 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
538
539 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
540 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
541 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
542 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
543 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
544
545 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
546 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
547 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
548 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
549 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
550
551 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
552 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
553
554 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
555
556 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
557 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
558
559 wasm_v128_store(output, vout0123456789ABCDEF);
560 output += 16;
561 }
562 if XNN_UNLIKELY(c != 0) {
563 const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
564 do {
565 v128_t vacc0123 = wasm_v128_load(w);
566 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
567
568
569 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
570 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
571 i0 += 8;
572
573 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
574
575 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
576 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
577
578 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
579 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
580 i1 += 8;
581
582 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
583
584 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
585 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
586
587 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
588 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
589 i2 += 8;
590
591 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
592
593 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
594 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
595
596 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
597 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
598 i3 += 8;
599
600 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
601
602 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
603 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
604
605 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
606 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 64));
607 i4 += 8;
608
609 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
610
611 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
612 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
613
614 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
615 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 80));
616 i5 += 8;
617
618 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
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
623 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
624 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
625 i6 += 8;
626
627 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
628
629 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
630 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
631
632 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
633 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 112));
634 i7 += 8;
635
636 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
637
638 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
639 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
640
641 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
642 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 128));
643 i8 += 8;
644
645 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
646
647 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
648 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
649
650 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
651 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) (k + 144));
652 i9 += 8;
653
654 vprod01234567 = wasm_i16x8_mul(vi9x01234567, vk9x01234567);
655
656 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
657 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
658
659 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
660 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) (k + 160));
661 i10 += 8;
662
663 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
664
665 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
666 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
667
668 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
669 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) (k + 176));
670 i11 += 8;
671
672 vprod01234567 = wasm_i16x8_mul(vi11x01234567, vk11x01234567);
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 vi12x01234567 = wasm_i16x8_load8x8(i12);
678 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) (k + 192));
679 i12 += 8;
680
681 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
682
683 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
684 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
685
686 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
687 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) (k + 208));
688 i13 += 8;
689
690 vprod01234567 = wasm_i16x8_mul(vi13x01234567, vk13x01234567);
691
692 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
693 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
694
695 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
696 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) (k + 224));
697 i14 += 8;
698
699 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
700
701 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
702 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
703
704 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
705 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) (k + 240));
706 i15 += 8;
707
708 vprod01234567 = wasm_i16x8_mul(vi15x01234567, vk15x01234567);
709
710 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
711 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
712
713 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
714 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) (k + 256));
715 i16 += 8;
716
717 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
718
719 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
720 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
721
722 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
723 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) (k + 272));
724 i17 += 8;
725
726 vprod01234567 = wasm_i16x8_mul(vi17x01234567, vk17x01234567);
727
728 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
729 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
730
731 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
732 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) (k + 288));
733 i18 += 8;
734
735 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
736
737 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
738 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
739
740 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
741 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) (k + 304));
742 i19 += 8;
743
744 vprod01234567 = wasm_i16x8_mul(vi19x01234567, vk19x01234567);
745
746 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
747 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
748
749 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
750 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) (k + 320));
751 i20 += 8;
752
753 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
754
755 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
756 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
757
758 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
759 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) (k + 336));
760 i21 += 8;
761
762 vprod01234567 = wasm_i16x8_mul(vi21x01234567, vk21x01234567);
763
764 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
765 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
766
767 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
768 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) (k + 352));
769 i22 += 8;
770
771 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
772
773 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
774 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
775
776 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
777 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) (k + 368));
778 i23 += 8;
779
780 vprod01234567 = wasm_i16x8_mul(vi23x01234567, vk23x01234567);
781
782 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
783 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
784
785 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
786 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) (k + 384));
787 i24 += 8;
788
789 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
790
791 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
792 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
793
794 k += 8;
795
796
797 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
798 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
799
800 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
801 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
802 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
803
804 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
805 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
806 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
807
808 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
809 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
810 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
811
812 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
813 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
814 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
815
816 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
817 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
818
819 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
820 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
821
822 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
823
824 if XNN_LIKELY(c >= 8) {
825 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
826 output += 8;
827 c -= 8;
828 } else {
829 if (c & 4) {
830 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
831 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
832 output += 4;
833 }
834 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
835 if (c & 2) {
836 *((uint16_t*) output) = (uint16_t) vout0123;
837 vout0123 >>= 16;
838 output += 2;
839 }
840 if (c & 1) {
841 *output = (int8_t) vout0123;
842 output += 1;
843 }
844 c = 0;
845 }
846 } while (c != 0);
847 }
848
849 output = (int8_t*) ((uintptr_t) output + output_increment);
850 } while (--output_width != 0);
851 }
852