Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c8s8r-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_6f6m7l8c8s8r__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 > 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
101 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
102 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
103 i1 += 8;
104
105 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
106
107 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
108 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
109
110 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
111 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
112 i2 += 8;
113
114 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
115
116
117 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
118 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
119 i3 += 8;
120
121 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
122
123 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
124 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
125
126 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
127 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
128 i4 += 8;
129
130 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
131
132
133 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
134 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
135 i5 += 8;
136
137 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
138
139 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
140 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
141
142 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t));
143
144 wasm_v128_store(b, vacc0123);
145 wasm_v128_store(b + 4, vacc4567);
146 b += 8;
147 }
148
149 assert(c == 0);
150 }
151
152
153 // Middle pass to process 6 inputs in each iteration.
154 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
155 int32_t* b = buffer;
156 const int8_t* i0 = input[0];
157 assert(i0 != NULL);
158 if XNN_UNPREDICTABLE(i0 != zero) {
159 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
160 }
161 const int8_t* i1 = input[1];
162 assert(i1 != NULL);
163 if XNN_UNPREDICTABLE(i1 != zero) {
164 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
165 }
166 const int8_t* i2 = input[2];
167 assert(i2 != NULL);
168 if XNN_UNPREDICTABLE(i2 != zero) {
169 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
170 }
171 const int8_t* i3 = input[3];
172 assert(i3 != NULL);
173 if XNN_UNPREDICTABLE(i3 != zero) {
174 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
175 }
176 const int8_t* i4 = input[4];
177 assert(i4 != NULL);
178 if XNN_UNPREDICTABLE(i4 != zero) {
179 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
180 }
181 const int8_t* i5 = input[5];
182 assert(i5 != NULL);
183 if XNN_UNPREDICTABLE(i5 != zero) {
184 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
185 }
186 input += 6;
187
188 size_t c = round_up_po2(channels, 8);
189
190 for (; c >= 8; c -= 8) {
191 v128_t vacc0123 = wasm_v128_load(b);
192 v128_t vacc4567 = wasm_v128_load(b + 4);
193
194 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
195 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
196 i0 += 8;
197
198 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
199
200
201 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
202 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
203 i1 += 8;
204
205 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
206
207 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
208 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
209
210 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
211 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
212 i2 += 8;
213
214 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
215
216
217 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
218 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
219 i3 += 8;
220
221 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
222
223 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
224 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
225
226 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
227 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
228 i4 += 8;
229
230 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
231
232
233 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
234 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
235 i5 += 8;
236
237 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
238
239 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
240 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
241
242 w = (const void*) ((uintptr_t) w + 48 * sizeof(int8_t));
243
244 wasm_v128_store(b, vacc0123);
245 wasm_v128_store(b + 4, vacc4567);
246 b += 8;
247 }
248
249 assert(c == 0);
250 }
251
252 // Last pass to process up to 7 inputs.
253 {
254 const int32_t* b = buffer;
255 const int8_t* i0 = input[0];
256 assert(i0 != NULL);
257 if XNN_UNPREDICTABLE(i0 != zero) {
258 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
259 }
260 const int8_t* i1 = input[1];
261 assert(i1 != NULL);
262 if XNN_UNPREDICTABLE(i1 != zero) {
263 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
264 }
265 const int8_t* i2 = input[2];
266 assert(i2 != NULL);
267 if XNN_UNPREDICTABLE(i2 != zero) {
268 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
269 }
270 const int8_t* i3 = input[3];
271 assert(i3 != NULL);
272 if XNN_UNPREDICTABLE(i3 != zero) {
273 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
274 }
275 const int8_t* i4 = input[4];
276 assert(i4 != NULL);
277 if XNN_UNPREDICTABLE(i4 != zero) {
278 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
279 }
280 const int8_t* i5 = input[5];
281 assert(i5 != NULL);
282 if XNN_UNPREDICTABLE(i5 != zero) {
283 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
284 }
285 const int8_t* i6 = input[6];
286 assert(i6 != NULL);
287 if XNN_UNPREDICTABLE(i6 != zero) {
288 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
289 }
290
291 size_t c = channels;
292
293 for (; c >= 8; c -= 8) {
294 v128_t vacc0123 = wasm_v128_load(b);
295 v128_t vacc4567 = wasm_v128_load(b + 4);
296 b += 8;
297
298 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
299 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
300 i0 += 8;
301
302 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
303
304
305 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
306 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
307 i1 += 8;
308
309 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
310
311 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
312 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
313
314 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
315 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
316 i2 += 8;
317
318 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
319
320
321 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
322 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
323 i3 += 8;
324
325 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
326
327 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
328 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
329
330 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
331 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
332 i4 += 8;
333
334 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
335
336
337 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
338 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
339 i5 += 8;
340
341 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
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 vi6x01234567 = wasm_i16x8_load8x8(i6);
347 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
348 i6 += 8;
349
350 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
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
356 w = (const void*) ((uintptr_t) w + 56 * sizeof(int8_t));
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
375 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
376
377 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
378
379 wasm_v128_store64_lane(output, vout0123456701234567, 0);
380 output += 8;
381 }
382
383 if XNN_UNLIKELY(c != 0) {
384 {
385 v128_t vacc0123 = wasm_v128_load(b);
386 v128_t vacc4567 = wasm_v128_load(b + 4);
387
388 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
389 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w));
390
391 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
392
393
394 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
395 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
396
397 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
398
399 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
400 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
401
402 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
403 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
404
405 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
406
407
408 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
409 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
410
411 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
412
413 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
414 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
415
416 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
417 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
418
419 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
420
421
422 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
423 v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
424
425 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
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 vi6x01234567 = wasm_i16x8_load8x8(i6);
431 v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
432
433 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
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
439 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
440 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
441
442 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
443 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
444
445
446 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
447 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
448
449 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
450 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
451
452 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
453 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
454
455 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
456 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
457
458 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
459
460 if (c & 4) {
461 wasm_v128_store32_lane(output, vout0123456701234567, 0);
462 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
463 output += 4;
464 }
465 if (c & 2) {
466 wasm_v128_store16_lane(output, vout0123456701234567, 0);
467 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
468 output += 2;
469 }
470 if (c & 1) {
471 wasm_v128_store8_lane(output, vout0123456701234567, 0);
472 output += 1;
473 }
474 }
475 }
476 }
477
478 input = (const int8_t**) ((uintptr_t) input + input_stride);
479 output = (int8_t*) ((uintptr_t) output + output_increment);
480 } while (--output_width != 0);
481 }