Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l4c1s1r-minmax-fp32-wasm-fmagic.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-scalar.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 #include <stddef.h>
12 #include <stdint.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/microparams.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l4c1s1r__wasm_fmagic(
22 size_t channels,
23 size_t output_width,
24 const int8_t** input,
25 const void* weights,
26 int8_t* output,
27 intptr_t input_stride,
28 size_t output_increment,
29 size_t input_offset,
30 const int8_t* zero,
31 size_t kernel_size,
32 int32_t* buffer,
33 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37 assert(kernel_size > 5);
38
39 const float vscale = params->fp32_scalar.scale;
40 const float voutput_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
41 const float voutput_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point;
42 const float vmagic_bias = 12582912.0f;
43 const int32_t vmagic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point;
44 do {
45 const void* w = weights;
46
47 // First pass to process 5 inputs.
48 {
49 int32_t* b = buffer;
50 const int8_t* i0 = input[0];
51 assert(i0 != NULL);
52 if XNN_UNPREDICTABLE(i0 != zero) {
53 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
54 }
55 const int8_t* i1 = input[1];
56 assert(i1 != NULL);
57 if XNN_UNPREDICTABLE(i1 != zero) {
58 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
59 }
60 const int8_t* i2 = input[2];
61 assert(i2 != NULL);
62 if XNN_UNPREDICTABLE(i2 != zero) {
63 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
64 }
65 const int8_t* i3 = input[3];
66 assert(i3 != NULL);
67 if XNN_UNPREDICTABLE(i3 != zero) {
68 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
69 }
70 const int8_t* i4 = input[4];
71 assert(i4 != NULL);
72 if XNN_UNPREDICTABLE(i4 != zero) {
73 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
74 }
75 input += 5;
76
77 size_t c = channels;
78 for (; c >= 4; c -= 4) {
79 int32_t vacc0 = ((const int32_t*) w)[0];
80 int32_t vacc1 = ((const int32_t*) w)[1];
81 int32_t vacc2 = ((const int32_t*) w)[2];
82 int32_t vacc3 = ((const int32_t*) w)[3];
83
84 const int32_t vi0x0 = (int32_t) i0[0];
85 const int32_t vi0x1 = (int32_t) i0[1];
86 const int32_t vi0x2 = (int32_t) i0[2];
87 const int32_t vi0x3 = (int32_t) i0[3];
88 i0 += 4;
89
90 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0];
91 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1];
92 const int32_t vk0x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2];
93 const int32_t vk0x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3];
94
95 vacc0 += vi0x0 * vk0x0;
96 vacc1 += vi0x1 * vk0x1;
97 vacc2 += vi0x2 * vk0x2;
98 vacc3 += vi0x3 * vk0x3;
99
100 const int32_t vi1x0 = (int32_t) i1[0];
101 const int32_t vi1x1 = (int32_t) i1[1];
102 const int32_t vi1x2 = (int32_t) i1[2];
103 const int32_t vi1x3 = (int32_t) i1[3];
104 i1 += 4;
105
106 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4];
107 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5];
108 const int32_t vk1x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6];
109 const int32_t vk1x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7];
110
111 vacc0 += vi1x0 * vk1x0;
112 vacc1 += vi1x1 * vk1x1;
113 vacc2 += vi1x2 * vk1x2;
114 vacc3 += vi1x3 * vk1x3;
115
116 const int32_t vi2x0 = (int32_t) i2[0];
117 const int32_t vi2x1 = (int32_t) i2[1];
118 const int32_t vi2x2 = (int32_t) i2[2];
119 const int32_t vi2x3 = (int32_t) i2[3];
120 i2 += 4;
121
122 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8];
123 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[9];
124 const int32_t vk2x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[10];
125 const int32_t vk2x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[11];
126
127 vacc0 += vi2x0 * vk2x0;
128 vacc1 += vi2x1 * vk2x1;
129 vacc2 += vi2x2 * vk2x2;
130 vacc3 += vi2x3 * vk2x3;
131
132 const int32_t vi3x0 = (int32_t) i3[0];
133 const int32_t vi3x1 = (int32_t) i3[1];
134 const int32_t vi3x2 = (int32_t) i3[2];
135 const int32_t vi3x3 = (int32_t) i3[3];
136 i3 += 4;
137
138 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[12];
139 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[13];
140 const int32_t vk3x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[14];
141 const int32_t vk3x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[15];
142
143 vacc0 += vi3x0 * vk3x0;
144 vacc1 += vi3x1 * vk3x1;
145 vacc2 += vi3x2 * vk3x2;
146 vacc3 += vi3x3 * vk3x3;
147
148 const int32_t vi4x0 = (int32_t) i4[0];
149 const int32_t vi4x1 = (int32_t) i4[1];
150 const int32_t vi4x2 = (int32_t) i4[2];
151 const int32_t vi4x3 = (int32_t) i4[3];
152 i4 += 4;
153
154 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[16];
155 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[17];
156 const int32_t vk4x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[18];
157 const int32_t vk4x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[19];
158
159 vacc0 += vi4x0 * vk4x0;
160 vacc1 += vi4x1 * vk4x1;
161 vacc2 += vi4x2 * vk4x2;
162 vacc3 += vi4x3 * vk4x3;
163
164 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 20 * sizeof(int8_t));
165 b[0] = vacc0;
166 b[1] = vacc1;
167 b[2] = vacc2;
168 b[3] = vacc3;
169 b += 4;
170 }
171 if XNN_UNLIKELY(c != 0) {
172 do {
173 int32_t vacc = unaligned_load_s32(w);
174 const int32_t vi0 = (int32_t) *i0++;
175 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[0];
176 vacc += vi0 * vk0;
177 const int32_t vi1 = (int32_t) *i1++;
178 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[1];
179 vacc += vi1 * vk1;
180 const int32_t vi2 = (int32_t) *i2++;
181 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[2];
182 vacc += vi2 * vk2;
183 const int32_t vi3 = (int32_t) *i3++;
184 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[3];
185 vacc += vi3 * vk3;
186 const int32_t vi4 = (int32_t) *i4++;
187 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[4];
188 vacc += vi4 * vk4;
189
190 w = (const void*) ((uintptr_t) w + 1 * sizeof(int32_t) + 5 * sizeof(int8_t));
191 *b++ = vacc;
192 } while (--c != 0);
193 }
194 }
195
196 // Middle pass to process 5 inputs in each iteration.
197 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
198 int32_t* b = buffer;
199 const int8_t* i0 = input[0];
200 assert(i0 != NULL);
201 if XNN_UNPREDICTABLE(i0 != zero) {
202 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
203 }
204 const int8_t* i1 = input[1];
205 assert(i1 != NULL);
206 if XNN_UNPREDICTABLE(i1 != zero) {
207 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
208 }
209 const int8_t* i2 = input[2];
210 assert(i2 != NULL);
211 if XNN_UNPREDICTABLE(i2 != zero) {
212 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
213 }
214 const int8_t* i3 = input[3];
215 assert(i3 != NULL);
216 if XNN_UNPREDICTABLE(i3 != zero) {
217 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
218 }
219 const int8_t* i4 = input[4];
220 assert(i4 != NULL);
221 if XNN_UNPREDICTABLE(i4 != zero) {
222 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
223 }
224 input += 5;
225
226 size_t c = channels;
227 for (; c >= 4; c -= 4) {
228 int32_t vacc0 = b[0];
229 int32_t vacc1 = b[1];
230 int32_t vacc2 = b[2];
231 int32_t vacc3 = b[3];
232
233 const int32_t vi0x0 = (int32_t) i0[0];
234 const int32_t vi0x1 = (int32_t) i0[1];
235 const int32_t vi0x2 = (int32_t) i0[2];
236 const int32_t vi0x3 = (int32_t) i0[3];
237 i0 += 4;
238
239 const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0];
240 const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1];
241 const int32_t vk0x2 = (int32_t) ((const int8_t*) w)[2];
242 const int32_t vk0x3 = (int32_t) ((const int8_t*) w)[3];
243
244 vacc0 += vi0x0 * vk0x0;
245 vacc1 += vi0x1 * vk0x1;
246 vacc2 += vi0x2 * vk0x2;
247 vacc3 += vi0x3 * vk0x3;
248
249 const int32_t vi1x0 = (int32_t) i1[0];
250 const int32_t vi1x1 = (int32_t) i1[1];
251 const int32_t vi1x2 = (int32_t) i1[2];
252 const int32_t vi1x3 = (int32_t) i1[3];
253 i1 += 4;
254
255 const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[4];
256 const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[5];
257 const int32_t vk1x2 = (int32_t) ((const int8_t*) w)[6];
258 const int32_t vk1x3 = (int32_t) ((const int8_t*) w)[7];
259
260 vacc0 += vi1x0 * vk1x0;
261 vacc1 += vi1x1 * vk1x1;
262 vacc2 += vi1x2 * vk1x2;
263 vacc3 += vi1x3 * vk1x3;
264
265 const int32_t vi2x0 = (int32_t) i2[0];
266 const int32_t vi2x1 = (int32_t) i2[1];
267 const int32_t vi2x2 = (int32_t) i2[2];
268 const int32_t vi2x3 = (int32_t) i2[3];
269 i2 += 4;
270
271 const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[8];
272 const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[9];
273 const int32_t vk2x2 = (int32_t) ((const int8_t*) w)[10];
274 const int32_t vk2x3 = (int32_t) ((const int8_t*) w)[11];
275
276 vacc0 += vi2x0 * vk2x0;
277 vacc1 += vi2x1 * vk2x1;
278 vacc2 += vi2x2 * vk2x2;
279 vacc3 += vi2x3 * vk2x3;
280
281 const int32_t vi3x0 = (int32_t) i3[0];
282 const int32_t vi3x1 = (int32_t) i3[1];
283 const int32_t vi3x2 = (int32_t) i3[2];
284 const int32_t vi3x3 = (int32_t) i3[3];
285 i3 += 4;
286
287 const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[12];
288 const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[13];
289 const int32_t vk3x2 = (int32_t) ((const int8_t*) w)[14];
290 const int32_t vk3x3 = (int32_t) ((const int8_t*) w)[15];
291
292 vacc0 += vi3x0 * vk3x0;
293 vacc1 += vi3x1 * vk3x1;
294 vacc2 += vi3x2 * vk3x2;
295 vacc3 += vi3x3 * vk3x3;
296
297 const int32_t vi4x0 = (int32_t) i4[0];
298 const int32_t vi4x1 = (int32_t) i4[1];
299 const int32_t vi4x2 = (int32_t) i4[2];
300 const int32_t vi4x3 = (int32_t) i4[3];
301 i4 += 4;
302
303 const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[16];
304 const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[17];
305 const int32_t vk4x2 = (int32_t) ((const int8_t*) w)[18];
306 const int32_t vk4x3 = (int32_t) ((const int8_t*) w)[19];
307
308 vacc0 += vi4x0 * vk4x0;
309 vacc1 += vi4x1 * vk4x1;
310 vacc2 += vi4x2 * vk4x2;
311 vacc3 += vi4x3 * vk4x3;
312
313 w = (const void*) ((uintptr_t) w + 20 * sizeof(int8_t));
314 b[0] = vacc0;
315 b[1] = vacc1;
316 b[2] = vacc2;
317 b[3] = vacc3;
318 b += 4;
319 }
320 if XNN_UNLIKELY(c != 0) {
321 do {
322 int32_t vacc = *b;
323 const int32_t vi0 = (int32_t) *i0++;
324 const int32_t vk0 = (int32_t) ((const int8_t*) w)[0];
325 vacc += vi0 * vk0;
326 const int32_t vi1 = (int32_t) *i1++;
327 const int32_t vk1 = (int32_t) ((const int8_t*) w)[1];
328 vacc += vi1 * vk1;
329 const int32_t vi2 = (int32_t) *i2++;
330 const int32_t vk2 = (int32_t) ((const int8_t*) w)[2];
331 vacc += vi2 * vk2;
332 const int32_t vi3 = (int32_t) *i3++;
333 const int32_t vk3 = (int32_t) ((const int8_t*) w)[3];
334 vacc += vi3 * vk3;
335 const int32_t vi4 = (int32_t) *i4++;
336 const int32_t vk4 = (int32_t) ((const int8_t*) w)[4];
337 vacc += vi4 * vk4;
338
339 w = (const void*) ((uintptr_t) w + 5 * sizeof(int8_t));
340 *b++ = vacc;
341 } while (--c != 0);
342 }
343 }
344
345 // Last pass to process up to 5 inputs.
346 {
347 const int32_t* b = buffer;
348 const int8_t* i0 = input[0];
349 assert(i0 != NULL);
350 if XNN_UNPREDICTABLE(i0 != zero) {
351 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
352 }
353 const int8_t* i1 = input[1];
354 assert(i1 != NULL);
355 if XNN_UNPREDICTABLE(i1 != zero) {
356 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
357 }
358 const int8_t* i2 = input[2];
359 assert(i2 != NULL);
360 if XNN_UNPREDICTABLE(i2 != zero) {
361 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
362 }
363 const int8_t* i3 = input[3];
364 assert(i3 != NULL);
365 if XNN_UNPREDICTABLE(i3 != zero) {
366 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
367 }
368 const int8_t* i4 = input[4];
369 assert(i4 != NULL);
370 if XNN_UNPREDICTABLE(i4 != zero) {
371 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
372 }
373
374 size_t c = channels;
375 for (; c >= 4; c -= 4) {
376 int32_t vacc0 = b[0];
377 int32_t vacc1 = b[1];
378 int32_t vacc2 = b[2];
379 int32_t vacc3 = b[3];
380 b += 4;
381
382 const int32_t vi0x0 = (int32_t) i0[0];
383 const int32_t vi0x1 = (int32_t) i0[1];
384 const int32_t vi0x2 = (int32_t) i0[2];
385 const int32_t vi0x3 = (int32_t) i0[3];
386 i0 += 4;
387
388 const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0];
389 const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1];
390 const int32_t vk0x2 = (int32_t) ((const int8_t*) w)[2];
391 const int32_t vk0x3 = (int32_t) ((const int8_t*) w)[3];
392
393 vacc0 += vi0x0 * vk0x0;
394 vacc1 += vi0x1 * vk0x1;
395 vacc2 += vi0x2 * vk0x2;
396 vacc3 += vi0x3 * vk0x3;
397
398 const int32_t vi1x0 = (int32_t) i1[0];
399 const int32_t vi1x1 = (int32_t) i1[1];
400 const int32_t vi1x2 = (int32_t) i1[2];
401 const int32_t vi1x3 = (int32_t) i1[3];
402 i1 += 4;
403
404 const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[4];
405 const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[5];
406 const int32_t vk1x2 = (int32_t) ((const int8_t*) w)[6];
407 const int32_t vk1x3 = (int32_t) ((const int8_t*) w)[7];
408
409 vacc0 += vi1x0 * vk1x0;
410 vacc1 += vi1x1 * vk1x1;
411 vacc2 += vi1x2 * vk1x2;
412 vacc3 += vi1x3 * vk1x3;
413
414 const int32_t vi2x0 = (int32_t) i2[0];
415 const int32_t vi2x1 = (int32_t) i2[1];
416 const int32_t vi2x2 = (int32_t) i2[2];
417 const int32_t vi2x3 = (int32_t) i2[3];
418 i2 += 4;
419
420 const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[8];
421 const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[9];
422 const int32_t vk2x2 = (int32_t) ((const int8_t*) w)[10];
423 const int32_t vk2x3 = (int32_t) ((const int8_t*) w)[11];
424
425 vacc0 += vi2x0 * vk2x0;
426 vacc1 += vi2x1 * vk2x1;
427 vacc2 += vi2x2 * vk2x2;
428 vacc3 += vi2x3 * vk2x3;
429
430 const int32_t vi3x0 = (int32_t) i3[0];
431 const int32_t vi3x1 = (int32_t) i3[1];
432 const int32_t vi3x2 = (int32_t) i3[2];
433 const int32_t vi3x3 = (int32_t) i3[3];
434 i3 += 4;
435
436 const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[12];
437 const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[13];
438 const int32_t vk3x2 = (int32_t) ((const int8_t*) w)[14];
439 const int32_t vk3x3 = (int32_t) ((const int8_t*) w)[15];
440
441 vacc0 += vi3x0 * vk3x0;
442 vacc1 += vi3x1 * vk3x1;
443 vacc2 += vi3x2 * vk3x2;
444 vacc3 += vi3x3 * vk3x3;
445
446 const int32_t vi4x0 = (int32_t) i4[0];
447 const int32_t vi4x1 = (int32_t) i4[1];
448 const int32_t vi4x2 = (int32_t) i4[2];
449 const int32_t vi4x3 = (int32_t) i4[3];
450 i4 += 4;
451
452 const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[16];
453 const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[17];
454 const int32_t vk4x2 = (int32_t) ((const int8_t*) w)[18];
455 const int32_t vk4x3 = (int32_t) ((const int8_t*) w)[19];
456
457 vacc0 += vi4x0 * vk4x0;
458 vacc1 += vi4x1 * vk4x1;
459 vacc2 += vi4x2 * vk4x2;
460 vacc3 += vi4x3 * vk4x3;
461
462 w = (const void*) ((uintptr_t) w + 20 * sizeof(int8_t));
463
464 float vfpacc0 = (float) vacc0;
465 float vfpacc1 = (float) vacc1;
466 float vfpacc2 = (float) vacc2;
467 float vfpacc3 = (float) vacc3;
468
469 vfpacc0 *= vscale;
470 vfpacc1 *= vscale;
471 vfpacc2 *= vscale;
472 vfpacc3 *= vscale;
473
474 vfpacc0 = __builtin_wasm_max_f32(vfpacc0, voutput_min_less_zero_point);
475 vfpacc1 = __builtin_wasm_max_f32(vfpacc1, voutput_min_less_zero_point);
476 vfpacc2 = __builtin_wasm_max_f32(vfpacc2, voutput_min_less_zero_point);
477 vfpacc3 = __builtin_wasm_max_f32(vfpacc3, voutput_min_less_zero_point);
478
479 vfpacc0 = __builtin_wasm_min_f32(vfpacc0, voutput_max_less_zero_point);
480 vfpacc1 = __builtin_wasm_min_f32(vfpacc1, voutput_max_less_zero_point);
481 vfpacc2 = __builtin_wasm_min_f32(vfpacc2, voutput_max_less_zero_point);
482 vfpacc3 = __builtin_wasm_min_f32(vfpacc3, voutput_max_less_zero_point);
483
484 vfpacc0 += vmagic_bias;
485 vfpacc1 += vmagic_bias;
486 vfpacc2 += vmagic_bias;
487 vfpacc3 += vmagic_bias;
488
489 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point;
490 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point;
491 int32_t vout2 = (int32_t) float_as_uint32(vfpacc2) - vmagic_bias_less_output_zero_point;
492 int32_t vout3 = (int32_t) float_as_uint32(vfpacc3) - vmagic_bias_less_output_zero_point;
493
494 output[0] = (int8_t) vout0;
495 output[1] = (int8_t) vout1;
496 output[2] = (int8_t) vout2;
497 output[3] = (int8_t) vout3;
498 output += 4;
499 }
500 if XNN_UNLIKELY(c != 0) {
501 do {
502 int32_t vacc = *b++;
503 const int32_t vi0 = (int32_t) *i0++;
504 const int32_t vk0 = (int32_t) ((const int8_t*) w)[0];
505 vacc += vi0 * vk0;
506 const int32_t vi1 = (int32_t) *i1++;
507 const int32_t vk1 = (int32_t) ((const int8_t*) w)[1];
508 vacc += vi1 * vk1;
509 const int32_t vi2 = (int32_t) *i2++;
510 const int32_t vk2 = (int32_t) ((const int8_t*) w)[2];
511 vacc += vi2 * vk2;
512 const int32_t vi3 = (int32_t) *i3++;
513 const int32_t vk3 = (int32_t) ((const int8_t*) w)[3];
514 vacc += vi3 * vk3;
515 const int32_t vi4 = (int32_t) *i4++;
516 const int32_t vk4 = (int32_t) ((const int8_t*) w)[4];
517 vacc += vi4 * vk4;
518
519 w = (const void*) ((uintptr_t) w + 5 * sizeof(int8_t));
520
521 float vfpacc = (float) vacc * vscale;
522
523 vfpacc = __builtin_wasm_max_f32(vfpacc, voutput_min_less_zero_point);
524 vfpacc = __builtin_wasm_min_f32(vfpacc, voutput_max_less_zero_point);
525 vfpacc += vmagic_bias;
526 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
527
528 *output++ = (int8_t) vout;
529 } while (--c != 0);
530 }
531 }
532
533 input = (const int8_t**) ((uintptr_t) input + input_stride);
534 output = (int8_t*) ((uintptr_t) output + output_increment);
535 } while (--output_width != 0);
536 }