Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c8s8r-minmax-fp32-wasmsimd-mul16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-wasmsimd-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_6f6m7l8c8s8r__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 size_t kernel_size,
29 int32_t* buffer,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(channels != 0);
33 assert(output_width != 0);
34 assert(kernel_size > 6);
35
36
37 const v128_t vscale = wasm_v128_load32_splat(&params->fp32_scalar.scale);
38 XNN_FORCE_REALIZATION(vscale);
39
40 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
41 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
42 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
43 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
44 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
45 XNN_FORCE_REALIZATION(vmagic_bias);
46 XNN_FORCE_REALIZATION(vmagic_min);
47 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
48 XNN_FORCE_REALIZATION(voutput_max);
49
50 do {
51 const void* w = weights;
52
53 // First pass to process 6 inputs.
54 {
55 int32_t* b = buffer;
56 const int8_t* i0 = input[0];
57 assert(i0 != NULL);
58 if XNN_UNPREDICTABLE(i0 != zero) {
59 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
60 }
61 const int8_t* i1 = input[1];
62 assert(i1 != NULL);
63 if XNN_UNPREDICTABLE(i1 != zero) {
64 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
65 }
66 const int8_t* i2 = input[2];
67 assert(i2 != NULL);
68 if XNN_UNPREDICTABLE(i2 != zero) {
69 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
70 }
71 const int8_t* i3 = input[3];
72 assert(i3 != NULL);
73 if XNN_UNPREDICTABLE(i3 != zero) {
74 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
75 }
76 const int8_t* i4 = input[4];
77 assert(i4 != NULL);
78 if XNN_UNPREDICTABLE(i4 != zero) {
79 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
80 }
81 const int8_t* i5 = input[5];
82 assert(i5 != NULL);
83 if XNN_UNPREDICTABLE(i5 != zero) {
84 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
85 }
86 input += 6;
87
88 size_t c = round_up_po2(channels, 8);
89
90 for (; c >= 8; c -= 8) {
91 v128_t vacc0123 = wasm_v128_load(w);
92 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
93
94 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
95 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
96 i0 += 8;
97
98 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
99
100 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
101 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
102
103 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
104 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
105 i1 += 8;
106
107 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
108
109 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
110 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
111
112 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
113 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
114 i2 += 8;
115
116 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
117
118 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
119 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
120
121 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
122 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
123 i3 += 8;
124
125 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
126
127 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
128 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
129
130 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
131 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
132 i4 += 8;
133
134 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
135
136 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
137 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
138
139 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
140 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
141 i5 += 8;
142
143 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
144
145 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
146 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
147
148 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t));
149
150 wasm_v128_store(b, vacc0123);
151 wasm_v128_store(b + 4, vacc4567);
152 b += 8;
153 }
154
155 assert(c == 0);
156 }
157
158
159 // Middle pass to process 6 inputs in each iteration.
160 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
161 int32_t* b = buffer;
162 const int8_t* i0 = input[0];
163 assert(i0 != NULL);
164 if XNN_UNPREDICTABLE(i0 != zero) {
165 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
166 }
167 const int8_t* i1 = input[1];
168 assert(i1 != NULL);
169 if XNN_UNPREDICTABLE(i1 != zero) {
170 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
171 }
172 const int8_t* i2 = input[2];
173 assert(i2 != NULL);
174 if XNN_UNPREDICTABLE(i2 != zero) {
175 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
176 }
177 const int8_t* i3 = input[3];
178 assert(i3 != NULL);
179 if XNN_UNPREDICTABLE(i3 != zero) {
180 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
181 }
182 const int8_t* i4 = input[4];
183 assert(i4 != NULL);
184 if XNN_UNPREDICTABLE(i4 != zero) {
185 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
186 }
187 const int8_t* i5 = input[5];
188 assert(i5 != NULL);
189 if XNN_UNPREDICTABLE(i5 != zero) {
190 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
191 }
192 input += 6;
193
194 size_t c = round_up_po2(channels, 8);
195
196 for (; c >= 8; c -= 8) {
197 v128_t vacc0123 = wasm_v128_load(b);
198 v128_t vacc4567 = wasm_v128_load(b + 4);
199
200 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
201 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
202 i0 += 8;
203
204 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
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
209 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
210 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
211 i1 += 8;
212
213 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
214
215 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
216 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
217
218 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
219 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
220 i2 += 8;
221
222 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
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
227 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
228 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
229 i3 += 8;
230
231 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
232
233 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
234 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
235
236 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
237 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
238 i4 += 8;
239
240 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
241
242 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
243 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
244
245 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
246 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
247 i5 += 8;
248
249 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
250
251 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
252 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
253
254 w = (const void*) ((uintptr_t) w + 48 * sizeof(int8_t));
255
256 wasm_v128_store(b, vacc0123);
257 wasm_v128_store(b + 4, vacc4567);
258 b += 8;
259 }
260
261 assert(c == 0);
262 }
263
264 // Last pass to process up to 7 inputs.
265 {
266 const int32_t* b = buffer;
267 const int8_t* i0 = input[0];
268 assert(i0 != NULL);
269 if XNN_UNPREDICTABLE(i0 != zero) {
270 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
271 }
272 const int8_t* i1 = input[1];
273 assert(i1 != NULL);
274 if XNN_UNPREDICTABLE(i1 != zero) {
275 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
276 }
277 const int8_t* i2 = input[2];
278 assert(i2 != NULL);
279 if XNN_UNPREDICTABLE(i2 != zero) {
280 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
281 }
282 const int8_t* i3 = input[3];
283 assert(i3 != NULL);
284 if XNN_UNPREDICTABLE(i3 != zero) {
285 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
286 }
287 const int8_t* i4 = input[4];
288 assert(i4 != NULL);
289 if XNN_UNPREDICTABLE(i4 != zero) {
290 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
291 }
292 const int8_t* i5 = input[5];
293 assert(i5 != NULL);
294 if XNN_UNPREDICTABLE(i5 != zero) {
295 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
296 }
297 const int8_t* i6 = input[6];
298 assert(i6 != NULL);
299 if XNN_UNPREDICTABLE(i6 != zero) {
300 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
301 }
302
303 size_t c = channels;
304
305 for (; c >= 8; c -= 8) {
306 v128_t vacc0123 = wasm_v128_load(b);
307 v128_t vacc4567 = wasm_v128_load(b + 4);
308 b += 8;
309
310 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
311 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
312 i0 += 8;
313
314 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
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 vi1x01234567 = wasm_i16x8_load8x8(i1);
320 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
321 i1 += 8;
322
323 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
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 vi2x01234567 = wasm_i16x8_load8x8(i2);
329 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
330 i2 += 8;
331
332 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
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 vi3x01234567 = wasm_i16x8_load8x8(i3);
338 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
339 i3 += 8;
340
341 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
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 vi4x01234567 = wasm_i16x8_load8x8(i4);
347 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
348 i4 += 8;
349
350 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
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 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
356 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
357 i5 += 8;
358
359 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
360
361 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
362 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
363
364 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
365 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
366 i6 += 8;
367
368 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
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
373
374 w = (const void*) ((uintptr_t) w + 56 * sizeof(int8_t));
375
376 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
377 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
378
379 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
380 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
381
382 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
383 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
384
385 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
386 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
387
388 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
389 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
390
391 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
392
393 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
394
395 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
396
397 wasm_v128_store64_lane(output, vout0123456701234567, 0);
398 output += 8;
399 }
400
401 if XNN_UNLIKELY(c != 0) {
402 {
403 v128_t vacc0123 = wasm_v128_load(b);
404 v128_t vacc4567 = wasm_v128_load(b + 4);
405
406 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
407 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w));
408
409 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
410
411 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
412 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
413
414 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
415 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
416
417 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
418
419 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
420 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
421
422 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
423 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
424
425 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
426
427 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
428 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
429
430 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
431 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
432
433 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
434
435 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
436 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
437
438 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
439 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
440
441 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
442
443 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
444 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
445
446 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
447 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
448
449 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
450
451 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
452 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
453
454 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
455 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
456
457 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
458
459 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
460 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
461
462
463 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
464 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
465
466 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
467 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
468
469
470 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
471 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
472
473 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
474 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
475
476 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
477 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
478
479 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
480 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
481
482 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
483
484 if (c & 4) {
485 wasm_v128_store32_lane(output, vout0123456701234567, 0);
486 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
487 output += 4;
488 }
489 if (c & 2) {
490 wasm_v128_store16_lane(output, vout0123456701234567, 0);
491 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
492 output += 2;
493 }
494 if (c & 1) {
495 wasm_v128_store8_lane(output, vout0123456701234567, 0);
496 output += 1;
497 }
498 }
499 }
500 }
501
502 input = (const int8_t**) ((uintptr_t) input + input_stride);
503 output = (int8_t*) ((uintptr_t) output + output_increment);
504 } while (--output_width != 0);
505 }