Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p16c-minmax-fp32-wasmsimd-mul16-add16.c

raw
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 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_25p16c__wasmsimd_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33
34 const v128_t vscale = wasm_v128_load32_splat(&params->fp32_scalar.scale);
35 XNN_FORCE_REALIZATION(vscale);
36 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
37 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
38 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
39 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
40 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
41 XNN_FORCE_REALIZATION(vmagic_bias);
42 XNN_FORCE_REALIZATION(vmagic_min);
43 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
44 XNN_FORCE_REALIZATION(voutput_max);
45
46 do {
47 const int8_t* i0 = input[0];
48 assert(i0 != NULL);
49 if XNN_UNPREDICTABLE(i0 != zero) {
50 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
51 }
52 const int8_t* i1 = input[1];
53 assert(i1 != NULL);
54 if XNN_UNPREDICTABLE(i1 != zero) {
55 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
56 }
57 const int8_t* i2 = input[2];
58 assert(i2 != NULL);
59 if XNN_UNPREDICTABLE(i2 != zero) {
60 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
61 }
62 const int8_t* i3 = input[3];
63 assert(i3 != NULL);
64 if XNN_UNPREDICTABLE(i3 != zero) {
65 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
66 }
67 const int8_t* i4 = input[4];
68 assert(i4 != NULL);
69 if XNN_UNPREDICTABLE(i4 != zero) {
70 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
71 }
72 const int8_t* i5 = input[5];
73 assert(i5 != NULL);
74 if XNN_UNPREDICTABLE(i5 != zero) {
75 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
76 }
77 const int8_t* i6 = input[6];
78 assert(i6 != NULL);
79 if XNN_UNPREDICTABLE(i6 != zero) {
80 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
81 }
82 const int8_t* i7 = input[7];
83 assert(i7 != NULL);
84 if XNN_UNPREDICTABLE(i7 != zero) {
85 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
86 }
87 const int8_t* i8 = input[8];
88 assert(i8 != NULL);
89 if XNN_UNPREDICTABLE(i8 != zero) {
90 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
91 }
92 const int8_t* i9 = input[9];
93 assert(i9 != NULL);
94 if XNN_UNPREDICTABLE(i9 != zero) {
95 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
96 }
97 const int8_t* i10 = input[10];
98 assert(i10 != NULL);
99 if XNN_UNPREDICTABLE(i10 != zero) {
100 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
101 }
102 const int8_t* i11 = input[11];
103 assert(i11 != NULL);
104 if XNN_UNPREDICTABLE(i11 != zero) {
105 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
106 }
107 const int8_t* i12 = input[12];
108 assert(i12 != NULL);
109 if XNN_UNPREDICTABLE(i12 != zero) {
110 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
111 }
112 const int8_t* i13 = input[13];
113 assert(i13 != NULL);
114 if XNN_UNPREDICTABLE(i13 != zero) {
115 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
116 }
117 const int8_t* i14 = input[14];
118 assert(i14 != NULL);
119 if XNN_UNPREDICTABLE(i14 != zero) {
120 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
121 }
122 const int8_t* i15 = input[15];
123 assert(i15 != NULL);
124 if XNN_UNPREDICTABLE(i15 != zero) {
125 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
126 }
127 const int8_t* i16 = input[16];
128 assert(i16 != NULL);
129 if XNN_UNPREDICTABLE(i16 != zero) {
130 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
131 }
132 const int8_t* i17 = input[17];
133 assert(i17 != NULL);
134 if XNN_UNPREDICTABLE(i17 != zero) {
135 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
136 }
137 const int8_t* i18 = input[18];
138 assert(i18 != NULL);
139 if XNN_UNPREDICTABLE(i18 != zero) {
140 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
141 }
142 const int8_t* i19 = input[19];
143 assert(i19 != NULL);
144 if XNN_UNPREDICTABLE(i19 != zero) {
145 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
146 }
147 const int8_t* i20 = input[20];
148 assert(i20 != NULL);
149 if XNN_UNPREDICTABLE(i20 != zero) {
150 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
151 }
152 const int8_t* i21 = input[21];
153 assert(i21 != NULL);
154 if XNN_UNPREDICTABLE(i21 != zero) {
155 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
156 }
157 const int8_t* i22 = input[22];
158 assert(i22 != NULL);
159 if XNN_UNPREDICTABLE(i22 != zero) {
160 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
161 }
162 const int8_t* i23 = input[23];
163 assert(i23 != NULL);
164 if XNN_UNPREDICTABLE(i23 != zero) {
165 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
166 }
167 const int8_t* i24 = input[24];
168 assert(i24 != NULL);
169 if XNN_UNPREDICTABLE(i24 != zero) {
170 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
171 }
172 input = (const int8_t**) ((uintptr_t) input + input_stride);
173
174 size_t c = channels;
175 const void* w = weights;
176 for (; c >= 16; c -= 16) {
177 v128_t vacc0123 = wasm_v128_load(w);
178 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
179 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
180 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
181
182
183 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
184 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
185 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
186 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
187 i0 += 16;
188
189 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
190 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
191
192
193 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
194 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
195 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
196 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
197 i1 += 16;
198
199 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
200 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
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
207 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
208 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
209 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
210 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
211 i2 += 16;
212
213 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
214 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
215
216
217 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
218 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
219 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
220 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
221 i3 += 16;
222
223 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
224 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
225
226 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
227 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
228 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
229 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
230
231 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
232 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
233 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
234 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
235 i4 += 16;
236
237 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
238 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
239
240
241 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
242 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
243 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
244 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
245 i5 += 16;
246
247 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
248 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF));
249
250 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
251 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
252 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
253 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
254
255 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
256 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
257 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
258 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
259 i6 += 16;
260
261 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
262 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
263
264
265 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
266 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
267 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
268 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
269 i7 += 16;
270
271 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
272 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF));
273
274 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
275 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
276 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
277 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
278
279 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
280 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
281 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
282 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
283 i8 += 16;
284
285 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
286 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
287
288
289 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
290 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
291 const v128_t vi9x89ABCDEF = wasm_i16x8_load8x8(i9 + 8);
292 const v128_t vk9x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(int8_t)));
293 i9 += 16;
294
295 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi9x01234567, vk9x01234567));
296 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi9x89ABCDEF, vk9x89ABCDEF));
297
298 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
299 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
300 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
301 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
302
303 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
304 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t)));
305 const v128_t vi10x89ABCDEF = wasm_i16x8_load8x8(i10 + 8);
306 const v128_t vk10x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(int8_t)));
307 i10 += 16;
308
309 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
310 vprod89ABCDEF = wasm_i16x8_mul(vi10x89ABCDEF, vk10x89ABCDEF);
311
312
313 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
314 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t)));
315 const v128_t vi11x89ABCDEF = wasm_i16x8_load8x8(i11 + 8);
316 const v128_t vk11x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(int8_t)));
317 i11 += 16;
318
319 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi11x01234567, vk11x01234567));
320 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi11x89ABCDEF, vk11x89ABCDEF));
321
322 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
323 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
324 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
325 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
326
327 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
328 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t)));
329 const v128_t vi12x89ABCDEF = wasm_i16x8_load8x8(i12 + 8);
330 const v128_t vk12x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(int8_t)));
331 i12 += 16;
332
333 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
334 vprod89ABCDEF = wasm_i16x8_mul(vi12x89ABCDEF, vk12x89ABCDEF);
335
336
337 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
338 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t)));
339 const v128_t vi13x89ABCDEF = wasm_i16x8_load8x8(i13 + 8);
340 const v128_t vk13x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(int8_t)));
341 i13 += 16;
342
343 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi13x01234567, vk13x01234567));
344 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi13x89ABCDEF, vk13x89ABCDEF));
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 vi14x01234567 = wasm_i16x8_load8x8(i14);
352 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t)));
353 const v128_t vi14x89ABCDEF = wasm_i16x8_load8x8(i14 + 8);
354 const v128_t vk14x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(int8_t)));
355 i14 += 16;
356
357 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
358 vprod89ABCDEF = wasm_i16x8_mul(vi14x89ABCDEF, vk14x89ABCDEF);
359
360
361 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
362 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t)));
363 const v128_t vi15x89ABCDEF = wasm_i16x8_load8x8(i15 + 8);
364 const v128_t vk15x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(int8_t)));
365 i15 += 16;
366
367 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi15x01234567, vk15x01234567));
368 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi15x89ABCDEF, vk15x89ABCDEF));
369
370 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
371 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
372 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
373 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
374
375 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
376 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t)));
377 const v128_t vi16x89ABCDEF = wasm_i16x8_load8x8(i16 + 8);
378 const v128_t vk16x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(int8_t)));
379 i16 += 16;
380
381 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
382 vprod89ABCDEF = wasm_i16x8_mul(vi16x89ABCDEF, vk16x89ABCDEF);
383
384
385 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
386 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t)));
387 const v128_t vi17x89ABCDEF = wasm_i16x8_load8x8(i17 + 8);
388 const v128_t vk17x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(int8_t)));
389 i17 += 16;
390
391 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi17x01234567, vk17x01234567));
392 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi17x89ABCDEF, vk17x89ABCDEF));
393
394 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
395 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
396 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
397 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
398
399 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
400 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t)));
401 const v128_t vi18x89ABCDEF = wasm_i16x8_load8x8(i18 + 8);
402 const v128_t vk18x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(int8_t)));
403 i18 += 16;
404
405 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
406 vprod89ABCDEF = wasm_i16x8_mul(vi18x89ABCDEF, vk18x89ABCDEF);
407
408
409 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
410 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t)));
411 const v128_t vi19x89ABCDEF = wasm_i16x8_load8x8(i19 + 8);
412 const v128_t vk19x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(int8_t)));
413 i19 += 16;
414
415 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi19x01234567, vk19x01234567));
416 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi19x89ABCDEF, vk19x89ABCDEF));
417
418 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
419 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
420 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
421 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
422
423 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
424 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t)));
425 const v128_t vi20x89ABCDEF = wasm_i16x8_load8x8(i20 + 8);
426 const v128_t vk20x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(int8_t)));
427 i20 += 16;
428
429 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
430 vprod89ABCDEF = wasm_i16x8_mul(vi20x89ABCDEF, vk20x89ABCDEF);
431
432
433 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
434 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t)));
435 const v128_t vi21x89ABCDEF = wasm_i16x8_load8x8(i21 + 8);
436 const v128_t vk21x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(int8_t)));
437 i21 += 16;
438
439 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi21x01234567, vk21x01234567));
440 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi21x89ABCDEF, vk21x89ABCDEF));
441
442 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
443 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
444 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
445 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
446
447 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
448 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t)));
449 const v128_t vi22x89ABCDEF = wasm_i16x8_load8x8(i22 + 8);
450 const v128_t vk22x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(int8_t)));
451 i22 += 16;
452
453 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
454 vprod89ABCDEF = wasm_i16x8_mul(vi22x89ABCDEF, vk22x89ABCDEF);
455
456
457 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
458 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t)));
459 const v128_t vi23x89ABCDEF = wasm_i16x8_load8x8(i23 + 8);
460 const v128_t vk23x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(int8_t)));
461 i23 += 16;
462
463 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi23x01234567, vk23x01234567));
464 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF));
465
466 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
467 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
468 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
469 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
470
471 const v128_t vi24x01234567 = wasm_i16x8_load8x8(i24);
472 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t)));
473 const v128_t vi24x89ABCDEF = wasm_i16x8_load8x8(i24 + 8);
474 const v128_t vk24x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(int8_t)));
475 i24 += 16;
476
477 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
478 vprod89ABCDEF = wasm_i16x8_mul(vi24x89ABCDEF, vk24x89ABCDEF);
479
480 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
481 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
482 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
483 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
484
485
486 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
487
488 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
489 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
490 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
491 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
492
493 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
494 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
495 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
496 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
497
498 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
499 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
500 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
501 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
502
503 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
504 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
505 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
506 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
507
508 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
509 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
510 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
511 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
512
513 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
514 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
515
516 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
517
518 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
519
520 wasm_v128_store(output, vout0123456789ABCDEF);
521 output += 16;
522 }
523 if XNN_UNLIKELY(c != 0) {
524 const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
525 do {
526 v128_t vacc0123 = wasm_v128_load(w);
527 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
528
529
530 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
531 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
532 i0 += 8;
533
534 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
535
536
537 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
538 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
539 i1 += 8;
540
541 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
542
543 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
544 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
545
546 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
547 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
548 i2 += 8;
549
550 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
551
552
553 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
554 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
555 i3 += 8;
556
557 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
558
559 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
560 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
561
562 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
563 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 64));
564 i4 += 8;
565
566 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
567
568
569 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
570 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 80));
571 i5 += 8;
572
573 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
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 vi6x01234567 = wasm_i16x8_load8x8(i6);
579 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
580 i6 += 8;
581
582 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
583
584
585 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
586 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 112));
587 i7 += 8;
588
589 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
590
591 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
592 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
593
594 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
595 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 128));
596 i8 += 8;
597
598 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
599
600
601 const v128_t vi9x01234567 = wasm_i16x8_load8x8(i9);
602 const v128_t vk9x01234567 = wasm_i16x8_load8x8((const void*) (k + 144));
603 i9 += 8;
604
605 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi9x01234567, vk9x01234567));
606
607 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
608 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
609
610 const v128_t vi10x01234567 = wasm_i16x8_load8x8(i10);
611 const v128_t vk10x01234567 = wasm_i16x8_load8x8((const void*) (k + 160));
612 i10 += 8;
613
614 vprod01234567 = wasm_i16x8_mul(vi10x01234567, vk10x01234567);
615
616
617 const v128_t vi11x01234567 = wasm_i16x8_load8x8(i11);
618 const v128_t vk11x01234567 = wasm_i16x8_load8x8((const void*) (k + 176));
619 i11 += 8;
620
621 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi11x01234567, vk11x01234567));
622
623 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
624 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
625
626 const v128_t vi12x01234567 = wasm_i16x8_load8x8(i12);
627 const v128_t vk12x01234567 = wasm_i16x8_load8x8((const void*) (k + 192));
628 i12 += 8;
629
630 vprod01234567 = wasm_i16x8_mul(vi12x01234567, vk12x01234567);
631
632
633 const v128_t vi13x01234567 = wasm_i16x8_load8x8(i13);
634 const v128_t vk13x01234567 = wasm_i16x8_load8x8((const void*) (k + 208));
635 i13 += 8;
636
637 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi13x01234567, vk13x01234567));
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
642 const v128_t vi14x01234567 = wasm_i16x8_load8x8(i14);
643 const v128_t vk14x01234567 = wasm_i16x8_load8x8((const void*) (k + 224));
644 i14 += 8;
645
646 vprod01234567 = wasm_i16x8_mul(vi14x01234567, vk14x01234567);
647
648
649 const v128_t vi15x01234567 = wasm_i16x8_load8x8(i15);
650 const v128_t vk15x01234567 = wasm_i16x8_load8x8((const void*) (k + 240));
651 i15 += 8;
652
653 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi15x01234567, vk15x01234567));
654
655 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
656 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
657
658 const v128_t vi16x01234567 = wasm_i16x8_load8x8(i16);
659 const v128_t vk16x01234567 = wasm_i16x8_load8x8((const void*) (k + 256));
660 i16 += 8;
661
662 vprod01234567 = wasm_i16x8_mul(vi16x01234567, vk16x01234567);
663
664
665 const v128_t vi17x01234567 = wasm_i16x8_load8x8(i17);
666 const v128_t vk17x01234567 = wasm_i16x8_load8x8((const void*) (k + 272));
667 i17 += 8;
668
669 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi17x01234567, vk17x01234567));
670
671 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
672 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
673
674 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18);
675 const v128_t vk18x01234567 = wasm_i16x8_load8x8((const void*) (k + 288));
676 i18 += 8;
677
678 vprod01234567 = wasm_i16x8_mul(vi18x01234567, vk18x01234567);
679
680
681 const v128_t vi19x01234567 = wasm_i16x8_load8x8(i19);
682 const v128_t vk19x01234567 = wasm_i16x8_load8x8((const void*) (k + 304));
683 i19 += 8;
684
685 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi19x01234567, vk19x01234567));
686
687 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
688 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
689
690 const v128_t vi20x01234567 = wasm_i16x8_load8x8(i20);
691 const v128_t vk20x01234567 = wasm_i16x8_load8x8((const void*) (k + 320));
692 i20 += 8;
693
694 vprod01234567 = wasm_i16x8_mul(vi20x01234567, vk20x01234567);
695
696
697 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21);
698 const v128_t vk21x01234567 = wasm_i16x8_load8x8((const void*) (k + 336));
699 i21 += 8;
700
701 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi21x01234567, vk21x01234567));
702
703 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
704 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
705
706 const v128_t vi22x01234567 = wasm_i16x8_load8x8(i22);
707 const v128_t vk22x01234567 = wasm_i16x8_load8x8((const void*) (k + 352));
708 i22 += 8;
709
710 vprod01234567 = wasm_i16x8_mul(vi22x01234567, vk22x01234567);
711
712
713 const v128_t vi23x01234567 = wasm_i16x8_load8x8(i23);
714 const v128_t vk23x01234567 = wasm_i16x8_load8x8((const void*) (k + 368));
715 i23 += 8;
716
717 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi23x01234567, vk23x01234567));
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 vi24x01234567 = wasm_i16x8_load8x8(i24);
723 const v128_t vk24x01234567 = wasm_i16x8_load8x8((const void*) (k + 384));
724 i24 += 8;
725
726 vprod01234567 = wasm_i16x8_mul(vi24x01234567, vk24x01234567);
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 k += 8;
732
733
734 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
735 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
736
737 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
738 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
739
740 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
741 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
742
743 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
744 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
745
746 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
747 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
748
749 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
750 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
751
752 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
753
754 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
755
756 if XNN_LIKELY(c >= 8) {
757 wasm_v128_store64_lane(output, vout0123456701234567, 0);
758 output += 8;
759 c -= 8;
760 } else {
761 if (c & 4) {
762 wasm_v128_store32_lane(output, vout0123456701234567, 0);
763 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
764 output += 4;
765 }
766 if (c & 2) {
767 wasm_v128_store16_lane(output, vout0123456701234567, 0);
768 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
769 output += 2;
770 }
771 if (c & 1) {
772 wasm_v128_store8_lane(output, vout0123456701234567, 0);
773 output += 1;
774 }
775 c = 0;
776 }
777 } while (c != 0);
778 }
779
780 output = (int8_t*) ((uintptr_t) output + output_increment);
781 } while (--output_width != 0);
782 }