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
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32(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_up8x25__avx2_mul32(
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)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 const int8_t* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
82 }
83 const int8_t* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
87 }
88 const int8_t* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
92 }
93 const int8_t* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
97 }
98 const int8_t* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
102 }
103 const int8_t* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
107 }
108 const int8_t* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
112 }
113 const int8_t* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
117 }
118 const int8_t* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
122 }
123 const int8_t* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
127 }
128 const int8_t* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
132 }
133 const int8_t* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
137 }
138 const int8_t* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
142 }
143 const int8_t* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
147 }
148 const int8_t* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
152 }
153 const int8_t* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const int8_t**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const void* w = weights;
162 for (; c >= 8; c -= 8) {
163 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
164
165
166 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
167 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
168 i0 += 8;
169
170 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
171
172 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
173 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
174 i1 += 8;
175
176 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
177
178 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
179 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
180 i2 += 8;
181
182 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
183
184 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
185 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
186 i3 += 8;
187
188 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
189
190 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
191 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
192 i4 += 8;
193
194 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
195
196 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
197 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
198 i5 += 8;
199
200 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
201
202 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
203 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))));
204 i6 += 8;
205
206 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
207
208 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
209 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))));
210 i7 += 8;
211
212 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
213
214 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
215 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))));
216 i8 += 8;
217
218 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
219
220 const __m256i vi9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i9));
221 const __m256i vk9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))));
222 i9 += 8;
223
224 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
225
226 const __m256i vi10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i10));
227 const __m256i vk10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))));
228 i10 += 8;
229
230 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
231
232 const __m256i vi11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i11));
233 const __m256i vk11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))));
234 i11 += 8;
235
236 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
237
238 const __m256i vi12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i12));
239 const __m256i vk12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))));
240 i12 += 8;
241
242 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
243
244 const __m256i vi13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i13));
245 const __m256i vk13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))));
246 i13 += 8;
247
248 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
249
250 const __m256i vi14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i14));
251 const __m256i vk14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))));
252 i14 += 8;
253
254 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
255
256 const __m256i vi15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i15));
257 const __m256i vk15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))));
258 i15 += 8;
259
260 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
261
262 const __m256i vi16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i16));
263 const __m256i vk16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))));
264 i16 += 8;
265
266 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
267
268 const __m256i vi17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i17));
269 const __m256i vk17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))));
270 i17 += 8;
271
272 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
273
274 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18));
275 const __m256i vk18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))));
276 i18 += 8;
277
278 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
279
280 const __m256i vi19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i19));
281 const __m256i vk19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))));
282 i19 += 8;
283
284 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
285
286 const __m256i vi20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i20));
287 const __m256i vk20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))));
288 i20 += 8;
289
290 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
291
292 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21));
293 const __m256i vk21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))));
294 i21 += 8;
295
296 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
297
298 const __m256i vi22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i22));
299 const __m256i vk22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))));
300 i22 += 8;
301
302 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
303
304 const __m256i vi23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i23));
305 const __m256i vk23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))));
306 i23 += 8;
307
308 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
309
310 const __m256i vi24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i24));
311 const __m256i vk24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))));
312 i24 += 8;
313
314 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
315
316 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
317
318 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
319
320 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
321 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
322
323 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
324 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
325
326 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
327
328 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
329 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
330
331 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
332
333 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
334 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
335
336 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
337 output += 8;
338 }
339 if XNN_UNLIKELY(c != 0) {
340 {
341 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
342
343
344 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
345 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
346
347 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
348
349 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
350 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
351
352 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
353
354 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
355 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
356
357 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
358
359 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
360 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
361
362 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
363
364 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
365 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
366
367 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
368
369 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
370 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
371
372 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
373
374 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
375 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))));
376
377 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
378
379 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
380 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))));
381
382 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
383
384 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
385 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))));
386
387 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
388
389 const __m256i vi9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i9));
390 const __m256i vk9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))));
391
392 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
393
394 const __m256i vi10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i10));
395 const __m256i vk10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))));
396
397 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
398
399 const __m256i vi11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i11));
400 const __m256i vk11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))));
401
402 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
403
404 const __m256i vi12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i12));
405 const __m256i vk12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))));
406
407 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
408
409 const __m256i vi13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i13));
410 const __m256i vk13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))));
411
412 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
413
414 const __m256i vi14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i14));
415 const __m256i vk14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))));
416
417 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
418
419 const __m256i vi15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i15));
420 const __m256i vk15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))));
421
422 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
423
424 const __m256i vi16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i16));
425 const __m256i vk16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))));
426
427 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
428
429 const __m256i vi17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i17));
430 const __m256i vk17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))));
431
432 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
433
434 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18));
435 const __m256i vk18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))));
436
437 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
438
439 const __m256i vi19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i19));
440 const __m256i vk19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))));
441
442 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
443
444 const __m256i vi20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i20));
445 const __m256i vk20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))));
446
447 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
448
449 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21));
450 const __m256i vk21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))));
451
452 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
453
454 const __m256i vi22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i22));
455 const __m256i vk22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))));
456
457 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
458
459 const __m256i vi23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i23));
460 const __m256i vk23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))));
461
462 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
463
464 const __m256i vi24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i24));
465 const __m256i vk24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))));
466
467 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
468
469
470 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
471 vscaled01234567 = _mm256_mul_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.scale));
472 vscaled01234567 = _mm256_min_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point));
473 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
474
475
476 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
477 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
478
479 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
480
481 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
482 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
483
484 if (c & 4) {
485 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
486 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
487 output += 4;
488 }
489 if (c & 2) {
490 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
491 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
492 output += 2;
493 }
494 if (c & 1) {
495 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
496 output += 1;
497 }
498 }
499 }
500
501 output = (int8_t*) ((uintptr_t) output + output_increment);
502 } while (--output_width != 0);
503 }
504