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_qs8_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float vscale = params->fp32_scalar_imagic.scale;
33 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
34 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
35 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
36 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 const int8_t* i9 = input[9];
84 assert(i9 != NULL);
85 if XNN_UNPREDICTABLE(i9 != zero) {
86 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
87 }
88 const int8_t* i10 = input[10];
89 assert(i10 != NULL);
90 if XNN_UNPREDICTABLE(i10 != zero) {
91 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
92 }
93 const int8_t* i11 = input[11];
94 assert(i11 != NULL);
95 if XNN_UNPREDICTABLE(i11 != zero) {
96 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
97 }
98 const int8_t* i12 = input[12];
99 assert(i12 != NULL);
100 if XNN_UNPREDICTABLE(i12 != zero) {
101 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
102 }
103 const int8_t* i13 = input[13];
104 assert(i13 != NULL);
105 if XNN_UNPREDICTABLE(i13 != zero) {
106 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
107 }
108 const int8_t* i14 = input[14];
109 assert(i14 != NULL);
110 if XNN_UNPREDICTABLE(i14 != zero) {
111 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
112 }
113 const int8_t* i15 = input[15];
114 assert(i15 != NULL);
115 if XNN_UNPREDICTABLE(i15 != zero) {
116 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
117 }
118 const int8_t* i16 = input[16];
119 assert(i16 != NULL);
120 if XNN_UNPREDICTABLE(i16 != zero) {
121 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
122 }
123 const int8_t* i17 = input[17];
124 assert(i17 != NULL);
125 if XNN_UNPREDICTABLE(i17 != zero) {
126 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
127 }
128 const int8_t* i18 = input[18];
129 assert(i18 != NULL);
130 if XNN_UNPREDICTABLE(i18 != zero) {
131 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
132 }
133 const int8_t* i19 = input[19];
134 assert(i19 != NULL);
135 if XNN_UNPREDICTABLE(i19 != zero) {
136 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
137 }
138 const int8_t* i20 = input[20];
139 assert(i20 != NULL);
140 if XNN_UNPREDICTABLE(i20 != zero) {
141 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
142 }
143 const int8_t* i21 = input[21];
144 assert(i21 != NULL);
145 if XNN_UNPREDICTABLE(i21 != zero) {
146 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
147 }
148 const int8_t* i22 = input[22];
149 assert(i22 != NULL);
150 if XNN_UNPREDICTABLE(i22 != zero) {
151 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
152 }
153 const int8_t* i23 = input[23];
154 assert(i23 != NULL);
155 if XNN_UNPREDICTABLE(i23 != zero) {
156 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
157 }
158 const int8_t* i24 = input[24];
159 assert(i24 != NULL);
160 if XNN_UNPREDICTABLE(i24 != zero) {
161 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
162 }
163 input = (const int8_t**) ((uintptr_t) input + input_stride);
164
165 size_t c = channels;
166 const void* w = weights;
167 for (; c >= 2; c -= 2) {
168 int32_t vacc0 = unaligned_indexed_load_s32(w, 0);
169 int32_t vacc1 = unaligned_indexed_load_s32(w, 1);
170
171
172 const int32_t vi0x0 = (int32_t) i0[0];
173 const int32_t vi0x1 = (int32_t) i0[1];
174 i0 += 2;
175
176 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
177 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1];
178
179 vacc0 += vi0x0 * vk0x0;
180 vacc1 += vi0x1 * vk0x1;
181
182 const int32_t vi1x0 = (int32_t) i1[0];
183 const int32_t vi1x1 = (int32_t) i1[1];
184 i1 += 2;
185
186 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
187 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3];
188
189 vacc0 += vi1x0 * vk1x0;
190 vacc1 += vi1x1 * vk1x1;
191
192 const int32_t vi2x0 = (int32_t) i2[0];
193 const int32_t vi2x1 = (int32_t) i2[1];
194 i2 += 2;
195
196 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
197 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5];
198
199 vacc0 += vi2x0 * vk2x0;
200 vacc1 += vi2x1 * vk2x1;
201
202 const int32_t vi3x0 = (int32_t) i3[0];
203 const int32_t vi3x1 = (int32_t) i3[1];
204 i3 += 2;
205
206 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
207 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7];
208
209 vacc0 += vi3x0 * vk3x0;
210 vacc1 += vi3x1 * vk3x1;
211
212 const int32_t vi4x0 = (int32_t) i4[0];
213 const int32_t vi4x1 = (int32_t) i4[1];
214 i4 += 2;
215
216 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
217 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9];
218
219 vacc0 += vi4x0 * vk4x0;
220 vacc1 += vi4x1 * vk4x1;
221
222 const int32_t vi5x0 = (int32_t) i5[0];
223 const int32_t vi5x1 = (int32_t) i5[1];
224 i5 += 2;
225
226 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
227 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11];
228
229 vacc0 += vi5x0 * vk5x0;
230 vacc1 += vi5x1 * vk5x1;
231
232 const int32_t vi6x0 = (int32_t) i6[0];
233 const int32_t vi6x1 = (int32_t) i6[1];
234 i6 += 2;
235
236 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
237 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13];
238
239 vacc0 += vi6x0 * vk6x0;
240 vacc1 += vi6x1 * vk6x1;
241
242 const int32_t vi7x0 = (int32_t) i7[0];
243 const int32_t vi7x1 = (int32_t) i7[1];
244 i7 += 2;
245
246 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
247 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15];
248
249 vacc0 += vi7x0 * vk7x0;
250 vacc1 += vi7x1 * vk7x1;
251
252 const int32_t vi8x0 = (int32_t) i8[0];
253 const int32_t vi8x1 = (int32_t) i8[1];
254 i8 += 2;
255
256 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
257 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17];
258
259 vacc0 += vi8x0 * vk8x0;
260 vacc1 += vi8x1 * vk8x1;
261
262 const int32_t vi9x0 = (int32_t) i9[0];
263 const int32_t vi9x1 = (int32_t) i9[1];
264 i9 += 2;
265
266 const int32_t vk9x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[18];
267 const int32_t vk9x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[19];
268
269 vacc0 += vi9x0 * vk9x0;
270 vacc1 += vi9x1 * vk9x1;
271
272 const int32_t vi10x0 = (int32_t) i10[0];
273 const int32_t vi10x1 = (int32_t) i10[1];
274 i10 += 2;
275
276 const int32_t vk10x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[20];
277 const int32_t vk10x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[21];
278
279 vacc0 += vi10x0 * vk10x0;
280 vacc1 += vi10x1 * vk10x1;
281
282 const int32_t vi11x0 = (int32_t) i11[0];
283 const int32_t vi11x1 = (int32_t) i11[1];
284 i11 += 2;
285
286 const int32_t vk11x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[22];
287 const int32_t vk11x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[23];
288
289 vacc0 += vi11x0 * vk11x0;
290 vacc1 += vi11x1 * vk11x1;
291
292 const int32_t vi12x0 = (int32_t) i12[0];
293 const int32_t vi12x1 = (int32_t) i12[1];
294 i12 += 2;
295
296 const int32_t vk12x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[24];
297 const int32_t vk12x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[25];
298
299 vacc0 += vi12x0 * vk12x0;
300 vacc1 += vi12x1 * vk12x1;
301
302 const int32_t vi13x0 = (int32_t) i13[0];
303 const int32_t vi13x1 = (int32_t) i13[1];
304 i13 += 2;
305
306 const int32_t vk13x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[26];
307 const int32_t vk13x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[27];
308
309 vacc0 += vi13x0 * vk13x0;
310 vacc1 += vi13x1 * vk13x1;
311
312 const int32_t vi14x0 = (int32_t) i14[0];
313 const int32_t vi14x1 = (int32_t) i14[1];
314 i14 += 2;
315
316 const int32_t vk14x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[28];
317 const int32_t vk14x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[29];
318
319 vacc0 += vi14x0 * vk14x0;
320 vacc1 += vi14x1 * vk14x1;
321
322 const int32_t vi15x0 = (int32_t) i15[0];
323 const int32_t vi15x1 = (int32_t) i15[1];
324 i15 += 2;
325
326 const int32_t vk15x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[30];
327 const int32_t vk15x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[31];
328
329 vacc0 += vi15x0 * vk15x0;
330 vacc1 += vi15x1 * vk15x1;
331
332 const int32_t vi16x0 = (int32_t) i16[0];
333 const int32_t vi16x1 = (int32_t) i16[1];
334 i16 += 2;
335
336 const int32_t vk16x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[32];
337 const int32_t vk16x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[33];
338
339 vacc0 += vi16x0 * vk16x0;
340 vacc1 += vi16x1 * vk16x1;
341
342 const int32_t vi17x0 = (int32_t) i17[0];
343 const int32_t vi17x1 = (int32_t) i17[1];
344 i17 += 2;
345
346 const int32_t vk17x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[34];
347 const int32_t vk17x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[35];
348
349 vacc0 += vi17x0 * vk17x0;
350 vacc1 += vi17x1 * vk17x1;
351
352 const int32_t vi18x0 = (int32_t) i18[0];
353 const int32_t vi18x1 = (int32_t) i18[1];
354 i18 += 2;
355
356 const int32_t vk18x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[36];
357 const int32_t vk18x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[37];
358
359 vacc0 += vi18x0 * vk18x0;
360 vacc1 += vi18x1 * vk18x1;
361
362 const int32_t vi19x0 = (int32_t) i19[0];
363 const int32_t vi19x1 = (int32_t) i19[1];
364 i19 += 2;
365
366 const int32_t vk19x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[38];
367 const int32_t vk19x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[39];
368
369 vacc0 += vi19x0 * vk19x0;
370 vacc1 += vi19x1 * vk19x1;
371
372 const int32_t vi20x0 = (int32_t) i20[0];
373 const int32_t vi20x1 = (int32_t) i20[1];
374 i20 += 2;
375
376 const int32_t vk20x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[40];
377 const int32_t vk20x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[41];
378
379 vacc0 += vi20x0 * vk20x0;
380 vacc1 += vi20x1 * vk20x1;
381
382 const int32_t vi21x0 = (int32_t) i21[0];
383 const int32_t vi21x1 = (int32_t) i21[1];
384 i21 += 2;
385
386 const int32_t vk21x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[42];
387 const int32_t vk21x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[43];
388
389 vacc0 += vi21x0 * vk21x0;
390 vacc1 += vi21x1 * vk21x1;
391
392 const int32_t vi22x0 = (int32_t) i22[0];
393 const int32_t vi22x1 = (int32_t) i22[1];
394 i22 += 2;
395
396 const int32_t vk22x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[44];
397 const int32_t vk22x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[45];
398
399 vacc0 += vi22x0 * vk22x0;
400 vacc1 += vi22x1 * vk22x1;
401
402 const int32_t vi23x0 = (int32_t) i23[0];
403 const int32_t vi23x1 = (int32_t) i23[1];
404 i23 += 2;
405
406 const int32_t vk23x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[46];
407 const int32_t vk23x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[47];
408
409 vacc0 += vi23x0 * vk23x0;
410 vacc1 += vi23x1 * vk23x1;
411
412 const int32_t vi24x0 = (int32_t) i24[0];
413 const int32_t vi24x1 = (int32_t) i24[1];
414 i24 += 2;
415
416 const int32_t vk24x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[48];
417 const int32_t vk24x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[49];
418
419 vacc0 += vi24x0 * vk24x0;
420 vacc1 += vi24x1 * vk24x1;
421
422 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 50 * sizeof(int8_t));
423
424 float vfpacc0 = (float) vacc0;
425 float vfpacc1 = (float) vacc1;
426
427 vfpacc0 *= vscale;
428 vfpacc1 *= vscale;
429
430 vfpacc0 += vmagic_bias;
431 vfpacc1 += vmagic_bias;
432
433 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0);
434 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1);
435
436 vout0 = math_max_s32(vout0, vmagic_min);
437 vout1 = math_max_s32(vout1, vmagic_min);
438
439 vout0 = math_min_s32(vout0, vmagic_max);
440 vout1 = math_min_s32(vout1, vmagic_max);
441
442 vout0 -= vmagic_bias_less_zero_point;
443 vout1 -= vmagic_bias_less_zero_point;
444
445 output[0] = (int8_t) vout0;
446 output[1] = (int8_t) vout1;
447 output += 2;
448 }
449 if XNN_UNLIKELY(c != 0) {
450 int32_t vacc = unaligned_load_s32(w);
451
452 const int32_t vi0 = (int32_t) *i0;
453 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
454 vacc += vi0 * vk0;
455 const int32_t vi1 = (int32_t) *i1;
456 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
457 vacc += vi1 * vk1;
458 const int32_t vi2 = (int32_t) *i2;
459 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
460 vacc += vi2 * vk2;
461 const int32_t vi3 = (int32_t) *i3;
462 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
463 vacc += vi3 * vk3;
464 const int32_t vi4 = (int32_t) *i4;
465 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
466 vacc += vi4 * vk4;
467 const int32_t vi5 = (int32_t) *i5;
468 const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
469 vacc += vi5 * vk5;
470 const int32_t vi6 = (int32_t) *i6;
471 const int32_t vk6 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
472 vacc += vi6 * vk6;
473 const int32_t vi7 = (int32_t) *i7;
474 const int32_t vk7 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
475 vacc += vi7 * vk7;
476 const int32_t vi8 = (int32_t) *i8;
477 const int32_t vk8 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
478 vacc += vi8 * vk8;
479 const int32_t vi9 = (int32_t) *i9;
480 const int32_t vk9 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[18];
481 vacc += vi9 * vk9;
482 const int32_t vi10 = (int32_t) *i10;
483 const int32_t vk10 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[20];
484 vacc += vi10 * vk10;
485 const int32_t vi11 = (int32_t) *i11;
486 const int32_t vk11 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[22];
487 vacc += vi11 * vk11;
488 const int32_t vi12 = (int32_t) *i12;
489 const int32_t vk12 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[24];
490 vacc += vi12 * vk12;
491 const int32_t vi13 = (int32_t) *i13;
492 const int32_t vk13 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[26];
493 vacc += vi13 * vk13;
494 const int32_t vi14 = (int32_t) *i14;
495 const int32_t vk14 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[28];
496 vacc += vi14 * vk14;
497 const int32_t vi15 = (int32_t) *i15;
498 const int32_t vk15 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[30];
499 vacc += vi15 * vk15;
500 const int32_t vi16 = (int32_t) *i16;
501 const int32_t vk16 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[32];
502 vacc += vi16 * vk16;
503 const int32_t vi17 = (int32_t) *i17;
504 const int32_t vk17 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[34];
505 vacc += vi17 * vk17;
506 const int32_t vi18 = (int32_t) *i18;
507 const int32_t vk18 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[36];
508 vacc += vi18 * vk18;
509 const int32_t vi19 = (int32_t) *i19;
510 const int32_t vk19 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[38];
511 vacc += vi19 * vk19;
512 const int32_t vi20 = (int32_t) *i20;
513 const int32_t vk20 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[40];
514 vacc += vi20 * vk20;
515 const int32_t vi21 = (int32_t) *i21;
516 const int32_t vk21 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[42];
517 vacc += vi21 * vk21;
518 const int32_t vi22 = (int32_t) *i22;
519 const int32_t vk22 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[44];
520 vacc += vi22 * vk22;
521 const int32_t vi23 = (int32_t) *i23;
522 const int32_t vk23 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[46];
523 vacc += vi23 * vk23;
524 const int32_t vi24 = (int32_t) *i24;
525 const int32_t vk24 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[48];
526 vacc += vi24 * vk24;
527
528 float vfpacc = (float) vacc * vscale;
529
530 vfpacc += vmagic_bias;
531 int32_t vout = (int32_t) float_as_uint32(vfpacc);
532 vout = math_max_s32(vout, vmagic_min);
533 vout = math_min_s32(vout, vmagic_max);
534 vout -= vmagic_bias_less_zero_point;
535
536 *output++ = (int8_t) vout;
537 }
538
539 output = (int8_t*) ((uintptr_t) output + output_increment);
540 } while (--output_width != 0);
541 }
542