Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p8c-minmax-fp32-avx2-mul32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx2-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <immintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_25p8c__avx2_mul32(
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 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33
34 const __m256 vscale = _mm256_set1_ps(params->fp32_scalar.scale);
35 XNN_FORCE_REALIZATION(vscale);
36
37 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
38 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
39 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
40 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
41 XNN_FORCE_REALIZATION(voutput_zero_point);
42 XNN_FORCE_REALIZATION(voutput_min);
43
44 do {
45 const int8_t* i0 = input[0];
46 assert(i0 != NULL);
47 if XNN_UNPREDICTABLE(i0 != zero) {
48 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
49 }
50 const int8_t* i1 = input[1];
51 assert(i1 != NULL);
52 if XNN_UNPREDICTABLE(i1 != zero) {
53 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
54 }
55 const int8_t* i2 = input[2];
56 assert(i2 != NULL);
57 if XNN_UNPREDICTABLE(i2 != zero) {
58 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
59 }
60 const int8_t* i3 = input[3];
61 assert(i3 != NULL);
62 if XNN_UNPREDICTABLE(i3 != zero) {
63 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
64 }
65 const int8_t* i4 = input[4];
66 assert(i4 != NULL);
67 if XNN_UNPREDICTABLE(i4 != zero) {
68 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
69 }
70 const int8_t* i5 = input[5];
71 assert(i5 != NULL);
72 if XNN_UNPREDICTABLE(i5 != zero) {
73 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
74 }
75 const int8_t* i6 = input[6];
76 assert(i6 != NULL);
77 if XNN_UNPREDICTABLE(i6 != zero) {
78 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
79 }
80 const int8_t* i7 = input[7];
81 assert(i7 != NULL);
82 if XNN_UNPREDICTABLE(i7 != zero) {
83 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
84 }
85 const int8_t* i8 = input[8];
86 assert(i8 != NULL);
87 if XNN_UNPREDICTABLE(i8 != zero) {
88 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
89 }
90 const int8_t* i9 = input[9];
91 assert(i9 != NULL);
92 if XNN_UNPREDICTABLE(i9 != zero) {
93 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
94 }
95 const int8_t* i10 = input[10];
96 assert(i10 != NULL);
97 if XNN_UNPREDICTABLE(i10 != zero) {
98 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
99 }
100 const int8_t* i11 = input[11];
101 assert(i11 != NULL);
102 if XNN_UNPREDICTABLE(i11 != zero) {
103 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
104 }
105 const int8_t* i12 = input[12];
106 assert(i12 != NULL);
107 if XNN_UNPREDICTABLE(i12 != zero) {
108 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
109 }
110 const int8_t* i13 = input[13];
111 assert(i13 != NULL);
112 if XNN_UNPREDICTABLE(i13 != zero) {
113 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
114 }
115 const int8_t* i14 = input[14];
116 assert(i14 != NULL);
117 if XNN_UNPREDICTABLE(i14 != zero) {
118 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
119 }
120 const int8_t* i15 = input[15];
121 assert(i15 != NULL);
122 if XNN_UNPREDICTABLE(i15 != zero) {
123 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
124 }
125 const int8_t* i16 = input[16];
126 assert(i16 != NULL);
127 if XNN_UNPREDICTABLE(i16 != zero) {
128 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
129 }
130 const int8_t* i17 = input[17];
131 assert(i17 != NULL);
132 if XNN_UNPREDICTABLE(i17 != zero) {
133 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
134 }
135 const int8_t* i18 = input[18];
136 assert(i18 != NULL);
137 if XNN_UNPREDICTABLE(i18 != zero) {
138 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
139 }
140 const int8_t* i19 = input[19];
141 assert(i19 != NULL);
142 if XNN_UNPREDICTABLE(i19 != zero) {
143 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
144 }
145 const int8_t* i20 = input[20];
146 assert(i20 != NULL);
147 if XNN_UNPREDICTABLE(i20 != zero) {
148 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
149 }
150 const int8_t* i21 = input[21];
151 assert(i21 != NULL);
152 if XNN_UNPREDICTABLE(i21 != zero) {
153 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
154 }
155 const int8_t* i22 = input[22];
156 assert(i22 != NULL);
157 if XNN_UNPREDICTABLE(i22 != zero) {
158 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
159 }
160 const int8_t* i23 = input[23];
161 assert(i23 != NULL);
162 if XNN_UNPREDICTABLE(i23 != zero) {
163 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
164 }
165 const int8_t* i24 = input[24];
166 assert(i24 != NULL);
167 if XNN_UNPREDICTABLE(i24 != zero) {
168 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
169 }
170 input = (const int8_t**) ((uintptr_t) input + input_stride);
171
172 size_t c = channels;
173 const void* w = weights;
174 for (; c >= 8; c -= 8) {
175 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
176
177
178 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
179 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
180 i0 += 8;
181
182 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
183
184 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
185 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
186 i1 += 8;
187
188 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
189
190 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
191 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
192 i2 += 8;
193
194 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
195
196 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
197 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
198 i3 += 8;
199
200 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
201
202 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
203 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
204 i4 += 8;
205
206 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
207
208 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
209 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
210 i5 += 8;
211
212 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
213
214 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
215 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))));
216 i6 += 8;
217
218 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
219
220 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
221 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))));
222 i7 += 8;
223
224 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
225
226 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
227 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))));
228 i8 += 8;
229
230 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
231
232 const __m256i vi9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i9));
233 const __m256i vk9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))));
234 i9 += 8;
235
236 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
237
238 const __m256i vi10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i10));
239 const __m256i vk10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))));
240 i10 += 8;
241
242 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
243
244 const __m256i vi11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i11));
245 const __m256i vk11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))));
246 i11 += 8;
247
248 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
249
250 const __m256i vi12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i12));
251 const __m256i vk12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))));
252 i12 += 8;
253
254 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
255
256 const __m256i vi13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i13));
257 const __m256i vk13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))));
258 i13 += 8;
259
260 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
261
262 const __m256i vi14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i14));
263 const __m256i vk14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))));
264 i14 += 8;
265
266 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
267
268 const __m256i vi15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i15));
269 const __m256i vk15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))));
270 i15 += 8;
271
272 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
273
274 const __m256i vi16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i16));
275 const __m256i vk16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))));
276 i16 += 8;
277
278 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
279
280 const __m256i vi17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i17));
281 const __m256i vk17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))));
282 i17 += 8;
283
284 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
285
286 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18));
287 const __m256i vk18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))));
288 i18 += 8;
289
290 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
291
292 const __m256i vi19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i19));
293 const __m256i vk19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))));
294 i19 += 8;
295
296 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
297
298 const __m256i vi20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i20));
299 const __m256i vk20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))));
300 i20 += 8;
301
302 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
303
304 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21));
305 const __m256i vk21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))));
306 i21 += 8;
307
308 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
309
310 const __m256i vi22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i22));
311 const __m256i vk22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))));
312 i22 += 8;
313
314 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
315
316 const __m256i vi23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i23));
317 const __m256i vk23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))));
318 i23 += 8;
319
320 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
321
322 const __m256i vi24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i24));
323 const __m256i vk24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))));
324 i24 += 8;
325
326 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
327
328 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
329
330 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
331
332 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
333
334 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
335
336 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
337
338 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
339
340 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
341
342 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
343
344 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
345 output += 8;
346 }
347 if XNN_UNLIKELY(c != 0) {
348 {
349 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
350
351
352 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
353 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
354
355 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
356
357 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
358 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
359
360 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
361
362 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
363 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
364
365 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
366
367 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
368 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
369
370 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
371
372 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
373 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
374
375 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
376
377 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
378 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
379
380 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
381
382 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
383 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))));
384
385 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
386
387 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
388 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))));
389
390 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
391
392 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
393 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))));
394
395 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
396
397 const __m256i vi9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i9));
398 const __m256i vk9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))));
399
400 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
401
402 const __m256i vi10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i10));
403 const __m256i vk10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))));
404
405 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
406
407 const __m256i vi11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i11));
408 const __m256i vk11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))));
409
410 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
411
412 const __m256i vi12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i12));
413 const __m256i vk12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))));
414
415 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
416
417 const __m256i vi13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i13));
418 const __m256i vk13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))));
419
420 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
421
422 const __m256i vi14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i14));
423 const __m256i vk14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))));
424
425 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
426
427 const __m256i vi15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i15));
428 const __m256i vk15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))));
429
430 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
431
432 const __m256i vi16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i16));
433 const __m256i vk16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))));
434
435 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
436
437 const __m256i vi17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i17));
438 const __m256i vk17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))));
439
440 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
441
442 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18));
443 const __m256i vk18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))));
444
445 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
446
447 const __m256i vi19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i19));
448 const __m256i vk19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))));
449
450 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
451
452 const __m256i vi20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i20));
453 const __m256i vk20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))));
454
455 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
456
457 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21));
458 const __m256i vk21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))));
459
460 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
461
462 const __m256i vi22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i22));
463 const __m256i vk22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))));
464
465 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
466
467 const __m256i vi23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i23));
468 const __m256i vk23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))));
469
470 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
471
472 const __m256i vi24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i24));
473 const __m256i vk24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))));
474
475 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
476
477
478 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
479 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
480 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
481 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
482
483 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
484
485 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
486
487 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
488
489 if (c & 4) {
490 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
491 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
492 output += 4;
493 }
494 if (c & 2) {
495 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
496 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
497 output += 2;
498 }
499 if (c & 1) {
500 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
501 output += 1;
502 }
503 }
504 }
505
506 output = (int8_t*) ((uintptr_t) output + output_increment);
507 } while (--output_width != 0);
508 }