1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <xnnpack/dwconv.h>
13 #include <xnnpack/math.h>
14 #include <xnnpack/unaligned.h>
15
16
xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
33 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
34 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
35 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
36 do {
37 const int8_t* i0 = input[0];
38 assert(i0 != NULL);
39 if XNN_UNPREDICTABLE(i0 != zero) {
40 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41 }
42 const int8_t* i1 = input[1];
43 assert(i1 != NULL);
44 if XNN_UNPREDICTABLE(i1 != zero) {
45 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46 }
47 const int8_t* i2 = input[2];
48 assert(i2 != NULL);
49 if XNN_UNPREDICTABLE(i2 != zero) {
50 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51 }
52 const int8_t* i3 = input[3];
53 assert(i3 != NULL);
54 if XNN_UNPREDICTABLE(i3 != zero) {
55 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56 }
57 const int8_t* i4 = input[4];
58 assert(i4 != NULL);
59 if XNN_UNPREDICTABLE(i4 != zero) {
60 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61 }
62 const int8_t* i5 = input[5];
63 assert(i5 != NULL);
64 if XNN_UNPREDICTABLE(i5 != zero) {
65 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66 }
67 const int8_t* i6 = input[6];
68 assert(i6 != NULL);
69 if XNN_UNPREDICTABLE(i6 != zero) {
70 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71 }
72 const int8_t* i7 = input[7];
73 assert(i7 != NULL);
74 if XNN_UNPREDICTABLE(i7 != zero) {
75 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76 }
77 const int8_t* i8 = input[8];
78 assert(i8 != NULL);
79 if XNN_UNPREDICTABLE(i8 != zero) {
80 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81 }
82 const int8_t* i9 = input[9];
83 assert(i9 != NULL);
84 if XNN_UNPREDICTABLE(i9 != zero) {
85 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
86 }
87 const int8_t* i10 = input[10];
88 assert(i10 != NULL);
89 if XNN_UNPREDICTABLE(i10 != zero) {
90 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
91 }
92 const int8_t* i11 = input[11];
93 assert(i11 != NULL);
94 if XNN_UNPREDICTABLE(i11 != zero) {
95 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
96 }
97 const int8_t* i12 = input[12];
98 assert(i12 != NULL);
99 if XNN_UNPREDICTABLE(i12 != zero) {
100 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
101 }
102 const int8_t* i13 = input[13];
103 assert(i13 != NULL);
104 if XNN_UNPREDICTABLE(i13 != zero) {
105 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
106 }
107 const int8_t* i14 = input[14];
108 assert(i14 != NULL);
109 if XNN_UNPREDICTABLE(i14 != zero) {
110 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
111 }
112 const int8_t* i15 = input[15];
113 assert(i15 != NULL);
114 if XNN_UNPREDICTABLE(i15 != zero) {
115 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
116 }
117 const int8_t* i16 = input[16];
118 assert(i16 != NULL);
119 if XNN_UNPREDICTABLE(i16 != zero) {
120 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
121 }
122 const int8_t* i17 = input[17];
123 assert(i17 != NULL);
124 if XNN_UNPREDICTABLE(i17 != zero) {
125 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
126 }
127 const int8_t* i18 = input[18];
128 assert(i18 != NULL);
129 if XNN_UNPREDICTABLE(i18 != zero) {
130 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
131 }
132 const int8_t* i19 = input[19];
133 assert(i19 != NULL);
134 if XNN_UNPREDICTABLE(i19 != zero) {
135 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
136 }
137 const int8_t* i20 = input[20];
138 assert(i20 != NULL);
139 if XNN_UNPREDICTABLE(i20 != zero) {
140 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
141 }
142 const int8_t* i21 = input[21];
143 assert(i21 != NULL);
144 if XNN_UNPREDICTABLE(i21 != zero) {
145 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
146 }
147 const int8_t* i22 = input[22];
148 assert(i22 != NULL);
149 if XNN_UNPREDICTABLE(i22 != zero) {
150 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
151 }
152 const int8_t* i23 = input[23];
153 assert(i23 != NULL);
154 if XNN_UNPREDICTABLE(i23 != zero) {
155 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
156 }
157 const int8_t* i24 = input[24];
158 assert(i24 != NULL);
159 if XNN_UNPREDICTABLE(i24 != zero) {
160 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
161 }
162 input = (const int8_t**) ((uintptr_t) input + input_stride);
163
164 size_t c = channels;
165 const void* w = weights;
166 for (; c >= 2; c -= 2) {
167 int32_t vacc0 = unaligned_indexed_load_s32(w, 0);
168 int32_t vacc1 = unaligned_indexed_load_s32(w, 1);
169
170
171 const int32_t vi0x0 = (int32_t) i0[0];
172 const int32_t vi0x1 = (int32_t) i0[1];
173 i0 += 2;
174
175 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
176 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1];
177
178 vacc0 += vi0x0 * vk0x0;
179 vacc1 += vi0x1 * vk0x1;
180
181 const int32_t vi1x0 = (int32_t) i1[0];
182 const int32_t vi1x1 = (int32_t) i1[1];
183 i1 += 2;
184
185 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
186 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3];
187
188 vacc0 += vi1x0 * vk1x0;
189 vacc1 += vi1x1 * vk1x1;
190
191 const int32_t vi2x0 = (int32_t) i2[0];
192 const int32_t vi2x1 = (int32_t) i2[1];
193 i2 += 2;
194
195 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
196 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5];
197
198 vacc0 += vi2x0 * vk2x0;
199 vacc1 += vi2x1 * vk2x1;
200
201 const int32_t vi3x0 = (int32_t) i3[0];
202 const int32_t vi3x1 = (int32_t) i3[1];
203 i3 += 2;
204
205 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
206 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7];
207
208 vacc0 += vi3x0 * vk3x0;
209 vacc1 += vi3x1 * vk3x1;
210
211 const int32_t vi4x0 = (int32_t) i4[0];
212 const int32_t vi4x1 = (int32_t) i4[1];
213 i4 += 2;
214
215 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
216 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9];
217
218 vacc0 += vi4x0 * vk4x0;
219 vacc1 += vi4x1 * vk4x1;
220
221 const int32_t vi5x0 = (int32_t) i5[0];
222 const int32_t vi5x1 = (int32_t) i5[1];
223 i5 += 2;
224
225 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
226 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11];
227
228 vacc0 += vi5x0 * vk5x0;
229 vacc1 += vi5x1 * vk5x1;
230
231 const int32_t vi6x0 = (int32_t) i6[0];
232 const int32_t vi6x1 = (int32_t) i6[1];
233 i6 += 2;
234
235 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
236 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13];
237
238 vacc0 += vi6x0 * vk6x0;
239 vacc1 += vi6x1 * vk6x1;
240
241 const int32_t vi7x0 = (int32_t) i7[0];
242 const int32_t vi7x1 = (int32_t) i7[1];
243 i7 += 2;
244
245 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
246 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15];
247
248 vacc0 += vi7x0 * vk7x0;
249 vacc1 += vi7x1 * vk7x1;
250
251 const int32_t vi8x0 = (int32_t) i8[0];
252 const int32_t vi8x1 = (int32_t) i8[1];
253 i8 += 2;
254
255 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
256 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17];
257
258 vacc0 += vi8x0 * vk8x0;
259 vacc1 += vi8x1 * vk8x1;
260
261 const int32_t vi9x0 = (int32_t) i9[0];
262 const int32_t vi9x1 = (int32_t) i9[1];
263 i9 += 2;
264
265 const int32_t vk9x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[18];
266 const int32_t vk9x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[19];
267
268 vacc0 += vi9x0 * vk9x0;
269 vacc1 += vi9x1 * vk9x1;
270
271 const int32_t vi10x0 = (int32_t) i10[0];
272 const int32_t vi10x1 = (int32_t) i10[1];
273 i10 += 2;
274
275 const int32_t vk10x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[20];
276 const int32_t vk10x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[21];
277
278 vacc0 += vi10x0 * vk10x0;
279 vacc1 += vi10x1 * vk10x1;
280
281 const int32_t vi11x0 = (int32_t) i11[0];
282 const int32_t vi11x1 = (int32_t) i11[1];
283 i11 += 2;
284
285 const int32_t vk11x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[22];
286 const int32_t vk11x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[23];
287
288 vacc0 += vi11x0 * vk11x0;
289 vacc1 += vi11x1 * vk11x1;
290
291 const int32_t vi12x0 = (int32_t) i12[0];
292 const int32_t vi12x1 = (int32_t) i12[1];
293 i12 += 2;
294
295 const int32_t vk12x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[24];
296 const int32_t vk12x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[25];
297
298 vacc0 += vi12x0 * vk12x0;
299 vacc1 += vi12x1 * vk12x1;
300
301 const int32_t vi13x0 = (int32_t) i13[0];
302 const int32_t vi13x1 = (int32_t) i13[1];
303 i13 += 2;
304
305 const int32_t vk13x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[26];
306 const int32_t vk13x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[27];
307
308 vacc0 += vi13x0 * vk13x0;
309 vacc1 += vi13x1 * vk13x1;
310
311 const int32_t vi14x0 = (int32_t) i14[0];
312 const int32_t vi14x1 = (int32_t) i14[1];
313 i14 += 2;
314
315 const int32_t vk14x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[28];
316 const int32_t vk14x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[29];
317
318 vacc0 += vi14x0 * vk14x0;
319 vacc1 += vi14x1 * vk14x1;
320
321 const int32_t vi15x0 = (int32_t) i15[0];
322 const int32_t vi15x1 = (int32_t) i15[1];
323 i15 += 2;
324
325 const int32_t vk15x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[30];
326 const int32_t vk15x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[31];
327
328 vacc0 += vi15x0 * vk15x0;
329 vacc1 += vi15x1 * vk15x1;
330
331 const int32_t vi16x0 = (int32_t) i16[0];
332 const int32_t vi16x1 = (int32_t) i16[1];
333 i16 += 2;
334
335 const int32_t vk16x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[32];
336 const int32_t vk16x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[33];
337
338 vacc0 += vi16x0 * vk16x0;
339 vacc1 += vi16x1 * vk16x1;
340
341 const int32_t vi17x0 = (int32_t) i17[0];
342 const int32_t vi17x1 = (int32_t) i17[1];
343 i17 += 2;
344
345 const int32_t vk17x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[34];
346 const int32_t vk17x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[35];
347
348 vacc0 += vi17x0 * vk17x0;
349 vacc1 += vi17x1 * vk17x1;
350
351 const int32_t vi18x0 = (int32_t) i18[0];
352 const int32_t vi18x1 = (int32_t) i18[1];
353 i18 += 2;
354
355 const int32_t vk18x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[36];
356 const int32_t vk18x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[37];
357
358 vacc0 += vi18x0 * vk18x0;
359 vacc1 += vi18x1 * vk18x1;
360
361 const int32_t vi19x0 = (int32_t) i19[0];
362 const int32_t vi19x1 = (int32_t) i19[1];
363 i19 += 2;
364
365 const int32_t vk19x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[38];
366 const int32_t vk19x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[39];
367
368 vacc0 += vi19x0 * vk19x0;
369 vacc1 += vi19x1 * vk19x1;
370
371 const int32_t vi20x0 = (int32_t) i20[0];
372 const int32_t vi20x1 = (int32_t) i20[1];
373 i20 += 2;
374
375 const int32_t vk20x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[40];
376 const int32_t vk20x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[41];
377
378 vacc0 += vi20x0 * vk20x0;
379 vacc1 += vi20x1 * vk20x1;
380
381 const int32_t vi21x0 = (int32_t) i21[0];
382 const int32_t vi21x1 = (int32_t) i21[1];
383 i21 += 2;
384
385 const int32_t vk21x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[42];
386 const int32_t vk21x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[43];
387
388 vacc0 += vi21x0 * vk21x0;
389 vacc1 += vi21x1 * vk21x1;
390
391 const int32_t vi22x0 = (int32_t) i22[0];
392 const int32_t vi22x1 = (int32_t) i22[1];
393 i22 += 2;
394
395 const int32_t vk22x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[44];
396 const int32_t vk22x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[45];
397
398 vacc0 += vi22x0 * vk22x0;
399 vacc1 += vi22x1 * vk22x1;
400
401 const int32_t vi23x0 = (int32_t) i23[0];
402 const int32_t vi23x1 = (int32_t) i23[1];
403 i23 += 2;
404
405 const int32_t vk23x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[46];
406 const int32_t vk23x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[47];
407
408 vacc0 += vi23x0 * vk23x0;
409 vacc1 += vi23x1 * vk23x1;
410
411 const int32_t vi24x0 = (int32_t) i24[0];
412 const int32_t vi24x1 = (int32_t) i24[1];
413 i24 += 2;
414
415 const int32_t vk24x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[48];
416 const int32_t vk24x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[49];
417
418 vacc0 += vi24x0 * vk24x0;
419 vacc1 += vi24x1 * vk24x1;
420
421 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 50 * sizeof(int8_t));
422
423 float vfpacc0 = (float) vacc0;
424 float vfpacc1 = (float) vacc1;
425
426 const float vscale0 = unaligned_indexed_load_f32(w, 0);
427 const float vscale1 = unaligned_indexed_load_f32(w, 1);
428 w = (const void*) ((const float*) w + 2);
429
430 vfpacc0 *= vscale0;
431 vfpacc1 *= vscale1;
432
433 vfpacc0 += vmagic_bias;
434 vfpacc1 += vmagic_bias;
435
436 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0);
437 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1);
438
439 vout0 = math_max_s32(vout0, vmagic_min);
440 vout1 = math_max_s32(vout1, vmagic_min);
441
442 vout0 = math_min_s32(vout0, vmagic_max);
443 vout1 = math_min_s32(vout1, vmagic_max);
444
445 vout0 -= vmagic_bias_less_zero_point;
446 vout1 -= vmagic_bias_less_zero_point;
447
448 output[0] = (int8_t) vout0;
449 output[1] = (int8_t) vout1;
450 output += 2;
451 }
452 if XNN_UNLIKELY(c != 0) {
453 int32_t vacc = unaligned_load_s32(w);
454
455 const int32_t vi0 = (int32_t) *i0;
456 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
457 vacc += vi0 * vk0;
458 const int32_t vi1 = (int32_t) *i1;
459 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
460 vacc += vi1 * vk1;
461 const int32_t vi2 = (int32_t) *i2;
462 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
463 vacc += vi2 * vk2;
464 const int32_t vi3 = (int32_t) *i3;
465 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
466 vacc += vi3 * vk3;
467 const int32_t vi4 = (int32_t) *i4;
468 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
469 vacc += vi4 * vk4;
470 const int32_t vi5 = (int32_t) *i5;
471 const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
472 vacc += vi5 * vk5;
473 const int32_t vi6 = (int32_t) *i6;
474 const int32_t vk6 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
475 vacc += vi6 * vk6;
476 const int32_t vi7 = (int32_t) *i7;
477 const int32_t vk7 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
478 vacc += vi7 * vk7;
479 const int32_t vi8 = (int32_t) *i8;
480 const int32_t vk8 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
481 vacc += vi8 * vk8;
482 const int32_t vi9 = (int32_t) *i9;
483 const int32_t vk9 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[18];
484 vacc += vi9 * vk9;
485 const int32_t vi10 = (int32_t) *i10;
486 const int32_t vk10 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[20];
487 vacc += vi10 * vk10;
488 const int32_t vi11 = (int32_t) *i11;
489 const int32_t vk11 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[22];
490 vacc += vi11 * vk11;
491 const int32_t vi12 = (int32_t) *i12;
492 const int32_t vk12 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[24];
493 vacc += vi12 * vk12;
494 const int32_t vi13 = (int32_t) *i13;
495 const int32_t vk13 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[26];
496 vacc += vi13 * vk13;
497 const int32_t vi14 = (int32_t) *i14;
498 const int32_t vk14 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[28];
499 vacc += vi14 * vk14;
500 const int32_t vi15 = (int32_t) *i15;
501 const int32_t vk15 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[30];
502 vacc += vi15 * vk15;
503 const int32_t vi16 = (int32_t) *i16;
504 const int32_t vk16 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[32];
505 vacc += vi16 * vk16;
506 const int32_t vi17 = (int32_t) *i17;
507 const int32_t vk17 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[34];
508 vacc += vi17 * vk17;
509 const int32_t vi18 = (int32_t) *i18;
510 const int32_t vk18 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[36];
511 vacc += vi18 * vk18;
512 const int32_t vi19 = (int32_t) *i19;
513 const int32_t vk19 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[38];
514 vacc += vi19 * vk19;
515 const int32_t vi20 = (int32_t) *i20;
516 const int32_t vk20 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[40];
517 vacc += vi20 * vk20;
518 const int32_t vi21 = (int32_t) *i21;
519 const int32_t vk21 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[42];
520 vacc += vi21 * vk21;
521 const int32_t vi22 = (int32_t) *i22;
522 const int32_t vk22 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[44];
523 vacc += vi22 * vk22;
524 const int32_t vi23 = (int32_t) *i23;
525 const int32_t vk23 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[46];
526 vacc += vi23 * vk23;
527 const int32_t vi24 = (int32_t) *i24;
528 const int32_t vk24 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[48];
529 vacc += vi24 * vk24;
530
531 typedef XNN_UNALIGNED float unaligned_float;
532 const float vscale = *((const unaligned_float*) ((uintptr_t) w + 2 * sizeof(int32_t) + 50 * sizeof(int8_t)));
533 float vfpacc = (float) vacc * vscale;
534
535 vfpacc += vmagic_bias;
536 int32_t vout = (int32_t) float_as_uint32(vfpacc);
537 vout = math_max_s32(vout, vmagic_min);
538 vout = math_min_s32(vout, vmagic_max);
539 vout -= vmagic_bias_less_zero_point;
540
541 *output++ = (int8_t) vout;
542 }
543
544 output = (int8_t*) ((uintptr_t) output + output_increment);
545 } while (--output_width != 0);
546 }
547