Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p16c-minmax-fp32-wasmsimd-mul16.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_9p16c__wasmsimd_mul16(
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 input = (const int8_t**) ((uintptr_t) input + input_stride);
93
94 size_t c = channels;
95 const void* w = weights;
96 for (; c >= 16; c -= 16) {
97 v128_t vacc0123 = wasm_v128_load(w);
98 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
99 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
100 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
101
102
103 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
104 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
105 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
106 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
107 i0 += 16;
108
109 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
110 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
111
112 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
113 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
114 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
115 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
116
117 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
118 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
119 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
120 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
121 i1 += 16;
122
123 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
124 vprod89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF);
125
126 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
127 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
128 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
129 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
130
131 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
132 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
133 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
134 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
135 i2 += 16;
136
137 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
138 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
139
140 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
141 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
142 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
143 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
144
145 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
146 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
147 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
148 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
149 i3 += 16;
150
151 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
152 vprod89ABCDEF = wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF);
153
154 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
155 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
156 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
157 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
158
159 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
160 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
161 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
162 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
163 i4 += 16;
164
165 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
166 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
167
168 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
169 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
170 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
171 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
172
173 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
174 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
175 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
176 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
177 i5 += 16;
178
179 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
180 vprod89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF);
181
182 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
183 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
184 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
185 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
186
187 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
188 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
189 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
190 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
191 i6 += 16;
192
193 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
194 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
195
196 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
197 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
198 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
199 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
200
201 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
202 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
203 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
204 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
205 i7 += 16;
206
207 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
208 vprod89ABCDEF = wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF);
209
210 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
211 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
212 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
213 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
214
215 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
216 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
217 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
218 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
219 i8 += 16;
220
221 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
222 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
223
224 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
225 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
226 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
227 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
228
229
230 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
231
232 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
233 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
234 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
235 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
236
237 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
238 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
239 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
240 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
241
242 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
243 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
244 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
245 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
246
247 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
248 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
249 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
250 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
251
252 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
253 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
254 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
255 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
256
257 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
258 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
259
260 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
261
262 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
263
264 wasm_v128_store(output, vout0123456789ABCDEF);
265 output += 16;
266 }
267 if XNN_UNLIKELY(c != 0) {
268 const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
269 do {
270 v128_t vacc0123 = wasm_v128_load(w);
271 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
272
273
274 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
275 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
276 i0 += 8;
277
278 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
279
280 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
281 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
282
283 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
284 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
285 i1 += 8;
286
287 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
288
289 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
290 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
291
292 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
293 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
294 i2 += 8;
295
296 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
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
301 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
302 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
303 i3 += 8;
304
305 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
306
307 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
308 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
309
310 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
311 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 64));
312 i4 += 8;
313
314 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
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
319 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
320 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 80));
321 i5 += 8;
322
323 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
324
325 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
326 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
327
328 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
329 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
330 i6 += 8;
331
332 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
333
334 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
335 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
336
337 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
338 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 112));
339 i7 += 8;
340
341 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
342
343 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
344 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
345
346 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
347 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 128));
348 i8 += 8;
349
350 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
351
352 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
353 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
354
355 k += 8;
356
357
358 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
359 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
360
361 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
362 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
363
364 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
365 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
366
367 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
368 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
369
370 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
371 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
372
373 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
374 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
375
376 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
377
378 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
379
380 if XNN_LIKELY(c >= 8) {
381 wasm_v128_store64_lane(output, vout0123456701234567, 0);
382 output += 8;
383 c -= 8;
384 } else {
385 if (c & 4) {
386 wasm_v128_store32_lane(output, vout0123456701234567, 0);
387 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
388 output += 4;
389 }
390 if (c & 2) {
391 wasm_v128_store16_lane(output, vout0123456701234567, 0);
392 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
393 output += 2;
394 }
395 if (c & 1) {
396 wasm_v128_store8_lane(output, vout0123456701234567, 0);
397 output += 1;
398 }
399 c = 0;
400 }
401 } while (c != 0);
402 }
403
404 output = (int8_t*) ((uintptr_t) output + output_increment);
405 } while (--output_width != 0);
406 }