Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l8c8s8r-minmax-fp32-wasmsimd-mul16-add16.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_8f8m9l8c8s8r__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 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 > 8);
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 8 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 const int8_t* i6 = input[6];
87 assert(i6 != NULL);
88 if XNN_UNPREDICTABLE(i6 != zero) {
89 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
90 }
91 const int8_t* i7 = input[7];
92 assert(i7 != NULL);
93 if XNN_UNPREDICTABLE(i7 != zero) {
94 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
95 }
96 input += 8;
97
98 size_t c = round_up_po2(channels, 8);
99
100 for (; c >= 8; c -= 8) {
101 v128_t vacc0123 = wasm_v128_load(w);
102 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
103
104 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
105 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
106 i0 += 8;
107
108 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
109
110
111 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
112 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
113 i1 += 8;
114
115 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
116
117 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
118 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
119
120 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
121 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
122 i2 += 8;
123
124 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
125
126
127 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
128 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
129 i3 += 8;
130
131 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
132
133 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
134 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
135
136 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
137 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
138 i4 += 8;
139
140 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
141
142
143 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
144 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
145 i5 += 8;
146
147 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
148
149 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
150 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
151
152 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
153 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
154 i6 += 8;
155
156 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
157
158
159 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
160 v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
161 i7 += 8;
162
163 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
164
165 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
166 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
167
168 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t));
169
170 wasm_v128_store(b, vacc0123);
171 wasm_v128_store(b + 4, vacc4567);
172 b += 8;
173 }
174
175 assert(c == 0);
176 }
177
178
179 // Middle pass to process 8 inputs in each iteration.
180 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
181 int32_t* b = buffer;
182 const int8_t* i0 = input[0];
183 assert(i0 != NULL);
184 if XNN_UNPREDICTABLE(i0 != zero) {
185 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
186 }
187 const int8_t* i1 = input[1];
188 assert(i1 != NULL);
189 if XNN_UNPREDICTABLE(i1 != zero) {
190 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
191 }
192 const int8_t* i2 = input[2];
193 assert(i2 != NULL);
194 if XNN_UNPREDICTABLE(i2 != zero) {
195 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
196 }
197 const int8_t* i3 = input[3];
198 assert(i3 != NULL);
199 if XNN_UNPREDICTABLE(i3 != zero) {
200 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
201 }
202 const int8_t* i4 = input[4];
203 assert(i4 != NULL);
204 if XNN_UNPREDICTABLE(i4 != zero) {
205 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
206 }
207 const int8_t* i5 = input[5];
208 assert(i5 != NULL);
209 if XNN_UNPREDICTABLE(i5 != zero) {
210 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
211 }
212 const int8_t* i6 = input[6];
213 assert(i6 != NULL);
214 if XNN_UNPREDICTABLE(i6 != zero) {
215 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
216 }
217 const int8_t* i7 = input[7];
218 assert(i7 != NULL);
219 if XNN_UNPREDICTABLE(i7 != zero) {
220 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
221 }
222 input += 8;
223
224 size_t c = round_up_po2(channels, 8);
225
226 for (; c >= 8; c -= 8) {
227 v128_t vacc0123 = wasm_v128_load(b);
228 v128_t vacc4567 = wasm_v128_load(b + 4);
229
230 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
231 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
232 i0 += 8;
233
234 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
235
236
237 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
238 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
239 i1 += 8;
240
241 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
242
243 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
244 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
245
246 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
247 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
248 i2 += 8;
249
250 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
251
252
253 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
254 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
255 i3 += 8;
256
257 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
258
259 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
260 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
261
262 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
263 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
264 i4 += 8;
265
266 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
267
268
269 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
270 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
271 i5 += 8;
272
273 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
274
275 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
276 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
277
278 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
279 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
280 i6 += 8;
281
282 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
283
284
285 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
286 v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 56 * sizeof(int8_t)));
287 i7 += 8;
288
289 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
290
291 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
292 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
293
294 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
295
296 wasm_v128_store(b, vacc0123);
297 wasm_v128_store(b + 4, vacc4567);
298 b += 8;
299 }
300
301 assert(c == 0);
302 }
303
304 // Last pass to process up to 9 inputs.
305 {
306 const int32_t* b = buffer;
307 const int8_t* i0 = input[0];
308 assert(i0 != NULL);
309 if XNN_UNPREDICTABLE(i0 != zero) {
310 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
311 }
312 const int8_t* i1 = input[1];
313 assert(i1 != NULL);
314 if XNN_UNPREDICTABLE(i1 != zero) {
315 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
316 }
317 const int8_t* i2 = input[2];
318 assert(i2 != NULL);
319 if XNN_UNPREDICTABLE(i2 != zero) {
320 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
321 }
322 const int8_t* i3 = input[3];
323 assert(i3 != NULL);
324 if XNN_UNPREDICTABLE(i3 != zero) {
325 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
326 }
327 const int8_t* i4 = input[4];
328 assert(i4 != NULL);
329 if XNN_UNPREDICTABLE(i4 != zero) {
330 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
331 }
332 const int8_t* i5 = input[5];
333 assert(i5 != NULL);
334 if XNN_UNPREDICTABLE(i5 != zero) {
335 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
336 }
337 const int8_t* i6 = input[6];
338 assert(i6 != NULL);
339 if XNN_UNPREDICTABLE(i6 != zero) {
340 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
341 }
342 const int8_t* i7 = input[7];
343 assert(i7 != NULL);
344 if XNN_UNPREDICTABLE(i7 != zero) {
345 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
346 }
347 const int8_t* i8 = input[8];
348 assert(i8 != NULL);
349 if XNN_UNPREDICTABLE(i8 != zero) {
350 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
351 }
352
353 size_t c = channels;
354
355 for (; c >= 8; c -= 8) {
356 v128_t vacc0123 = wasm_v128_load(b);
357 v128_t vacc4567 = wasm_v128_load(b + 4);
358 b += 8;
359
360 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
361 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
362 i0 += 8;
363
364 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
365
366
367 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
368 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
369 i1 += 8;
370
371 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
372
373 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
374 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
375
376 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
377 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
378 i2 += 8;
379
380 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
381
382
383 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
384 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
385 i3 += 8;
386
387 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
388
389 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
390 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
391
392 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
393 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
394 i4 += 8;
395
396 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
397
398
399 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
400 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
401 i5 += 8;
402
403 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
404
405 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
406 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
407
408 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
409 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
410 i6 += 8;
411
412 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
413
414
415 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
416 v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 56 * sizeof(int8_t)));
417 i7 += 8;
418
419 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
420
421 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
422 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
423
424 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
425 v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 64 * sizeof(int8_t)));
426 i8 += 8;
427
428 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
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
433
434 w = (const void*) ((uintptr_t) w + 72 * sizeof(int8_t));
435
436 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
437 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
438
439 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
440 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
441
442 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
443 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
444
445 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
446 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
447
448 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
449 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
450
451 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
452
453 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
454
455 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
456
457 wasm_v128_store64_lane(output, vout0123456701234567, 0);
458 output += 8;
459 }
460
461 if XNN_UNLIKELY(c != 0) {
462 {
463 v128_t vacc0123 = wasm_v128_load(b);
464 v128_t vacc4567 = wasm_v128_load(b + 4);
465
466 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
467 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w));
468
469 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
470
471
472 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
473 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
474
475 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
476
477 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
478 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
479
480 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
481 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
482
483 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
484
485
486 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
487 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
488
489 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
490
491 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
492 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
493
494 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
495 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
496
497 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
498
499
500 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
501 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
502
503 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
504
505 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
506 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
507
508 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
509 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
510
511 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
512
513
514 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
515 v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 56 * sizeof(int8_t)));
516
517 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
518
519 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
520 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
521
522 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
523 v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 64 * sizeof(int8_t)));
524
525 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
526
527 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
528 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
529
530
531 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
532 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
533
534 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
535 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
536
537
538 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
539 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
540
541 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
542 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
543
544 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
545 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
546
547 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
548 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
549
550 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
551
552 if (c & 4) {
553 wasm_v128_store32_lane(output, vout0123456701234567, 0);
554 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
555 output += 4;
556 }
557 if (c & 2) {
558 wasm_v128_store16_lane(output, vout0123456701234567, 0);
559 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
560 output += 2;
561 }
562 if (c & 1) {
563 wasm_v128_store8_lane(output, vout0123456701234567, 0);
564 output += 1;
565 }
566 }
567 }
568 }
569
570 input = (const int8_t**) ((uintptr_t) input + input_stride);
571 output = (int8_t*) ((uintptr_t) output + output_increment);
572 } while (--output_width != 0);
573 }