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
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m256i vk_zero_point = _mm256_cvtepu16_epi32(_mm_load_si128((const __m128i*) params->fp32_avx2.kernel_zero_point));
34 do {
35 const uint8_t* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
39 }
40 const uint8_t* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
44 }
45 const uint8_t* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
49 }
50 const uint8_t* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
54 }
55 const uint8_t* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
59 }
60 const uint8_t* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
64 }
65 const uint8_t* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
69 }
70 const uint8_t* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
74 }
75 const uint8_t* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
79 }
80 const uint8_t* i9 = input[9];
81 assert(i9 != NULL);
82 if XNN_UNPREDICTABLE(i9 != zero) {
83 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
84 }
85 const uint8_t* i10 = input[10];
86 assert(i10 != NULL);
87 if XNN_UNPREDICTABLE(i10 != zero) {
88 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
89 }
90 const uint8_t* i11 = input[11];
91 assert(i11 != NULL);
92 if XNN_UNPREDICTABLE(i11 != zero) {
93 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
94 }
95 const uint8_t* i12 = input[12];
96 assert(i12 != NULL);
97 if XNN_UNPREDICTABLE(i12 != zero) {
98 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
99 }
100 const uint8_t* i13 = input[13];
101 assert(i13 != NULL);
102 if XNN_UNPREDICTABLE(i13 != zero) {
103 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
104 }
105 const uint8_t* i14 = input[14];
106 assert(i14 != NULL);
107 if XNN_UNPREDICTABLE(i14 != zero) {
108 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
109 }
110 const uint8_t* i15 = input[15];
111 assert(i15 != NULL);
112 if XNN_UNPREDICTABLE(i15 != zero) {
113 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
114 }
115 const uint8_t* i16 = input[16];
116 assert(i16 != NULL);
117 if XNN_UNPREDICTABLE(i16 != zero) {
118 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
119 }
120 const uint8_t* i17 = input[17];
121 assert(i17 != NULL);
122 if XNN_UNPREDICTABLE(i17 != zero) {
123 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
124 }
125 const uint8_t* i18 = input[18];
126 assert(i18 != NULL);
127 if XNN_UNPREDICTABLE(i18 != zero) {
128 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
129 }
130 const uint8_t* i19 = input[19];
131 assert(i19 != NULL);
132 if XNN_UNPREDICTABLE(i19 != zero) {
133 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
134 }
135 const uint8_t* i20 = input[20];
136 assert(i20 != NULL);
137 if XNN_UNPREDICTABLE(i20 != zero) {
138 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
139 }
140 const uint8_t* i21 = input[21];
141 assert(i21 != NULL);
142 if XNN_UNPREDICTABLE(i21 != zero) {
143 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
144 }
145 const uint8_t* i22 = input[22];
146 assert(i22 != NULL);
147 if XNN_UNPREDICTABLE(i22 != zero) {
148 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
149 }
150 const uint8_t* i23 = input[23];
151 assert(i23 != NULL);
152 if XNN_UNPREDICTABLE(i23 != zero) {
153 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
154 }
155 const uint8_t* i24 = input[24];
156 assert(i24 != NULL);
157 if XNN_UNPREDICTABLE(i24 != zero) {
158 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
159 }
160 input = (const uint8_t**) ((uintptr_t) input + input_stride);
161
162 size_t c = channels;
163 const void* w = weights;
164 for (; c >= 8; c -= 8) {
165 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
166
167
168 const __m256i vi0x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i0));
169 const __m256i vk0x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
170 i0 += 8;
171
172 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
173
174 const __m256i vi1x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i1));
175 const __m256i vk1x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)))), vk_zero_point);
176 i1 += 8;
177
178 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
179
180 const __m256i vi2x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i2));
181 const __m256i vk2x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
182 i2 += 8;
183
184 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
185
186 const __m256i vi3x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i3));
187 const __m256i vk3x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)))), vk_zero_point);
188 i3 += 8;
189
190 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
191
192 const __m256i vi4x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i4));
193 const __m256i vk4x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
194 i4 += 8;
195
196 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
197
198 const __m256i vi5x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i5));
199 const __m256i vk5x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)))), vk_zero_point);
200 i5 += 8;
201
202 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
203
204 const __m256i vi6x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i6));
205 const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
206 i6 += 8;
207
208 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
209
210 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7));
211 const __m256i vk7x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)))), vk_zero_point);
212 i7 += 8;
213
214 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
215
216 const __m256i vi8x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i8));
217 const __m256i vk8x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
218 i8 += 8;
219
220 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
221
222 const __m256i vi9x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i9));
223 const __m256i vk9x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)))), vk_zero_point);
224 i9 += 8;
225
226 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
227
228 const __m256i vi10x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i10));
229 const __m256i vk10x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
230 i10 += 8;
231
232 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
233
234 const __m256i vi11x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i11));
235 const __m256i vk11x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)))), vk_zero_point);
236 i11 += 8;
237
238 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
239
240 const __m256i vi12x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i12));
241 const __m256i vk12x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
242 i12 += 8;
243
244 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
245
246 const __m256i vi13x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i13));
247 const __m256i vk13x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)))), vk_zero_point);
248 i13 += 8;
249
250 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
251
252 const __m256i vi14x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i14));
253 const __m256i vk14x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
254 i14 += 8;
255
256 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
257
258 const __m256i vi15x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i15));
259 const __m256i vk15x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)))), vk_zero_point);
260 i15 += 8;
261
262 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
263
264 const __m256i vi16x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i16));
265 const __m256i vk16x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
266 i16 += 8;
267
268 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
269
270 const __m256i vi17x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i17));
271 const __m256i vk17x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)))), vk_zero_point);
272 i17 += 8;
273
274 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
275
276 const __m256i vi18x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i18));
277 const __m256i vk18x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)))), vk_zero_point);
278 i18 += 8;
279
280 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
281
282 const __m256i vi19x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i19));
283 const __m256i vk19x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)))), vk_zero_point);
284 i19 += 8;
285
286 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
287
288 const __m256i vi20x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i20));
289 const __m256i vk20x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)))), vk_zero_point);
290 i20 += 8;
291
292 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
293
294 const __m256i vi21x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i21));
295 const __m256i vk21x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)))), vk_zero_point);
296 i21 += 8;
297
298 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
299
300 const __m256i vi22x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i22));
301 const __m256i vk22x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)))), vk_zero_point);
302 i22 += 8;
303
304 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
305
306 const __m256i vi23x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i23));
307 const __m256i vk23x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)))), vk_zero_point);
308 i23 += 8;
309
310 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
311
312 const __m256i vi24x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i24));
313 const __m256i vk24x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)))), vk_zero_point);
314 i24 += 8;
315
316 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
317
318 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
319
320 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
321
322 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
323 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
324
325 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
326 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
327
328 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
329
330 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
331 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
332
333 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
334
335 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
336 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
337
338 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
339 output += 8;
340 }
341 if XNN_UNLIKELY(c != 0) {
342 {
343 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
344
345
346 const __m256i vi0x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i0));
347 const __m256i vk0x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
348
349 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
350
351 const __m256i vi1x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i1));
352 const __m256i vk1x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)))), vk_zero_point);
353
354 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
355
356 const __m256i vi2x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i2));
357 const __m256i vk2x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
358
359 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
360
361 const __m256i vi3x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i3));
362 const __m256i vk3x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)))), vk_zero_point);
363
364 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
365
366 const __m256i vi4x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i4));
367 const __m256i vk4x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
368
369 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
370
371 const __m256i vi5x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i5));
372 const __m256i vk5x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)))), vk_zero_point);
373
374 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
375
376 const __m256i vi6x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i6));
377 const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
378
379 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
380
381 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7));
382 const __m256i vk7x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)))), vk_zero_point);
383
384 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
385
386 const __m256i vi8x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i8));
387 const __m256i vk8x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
388
389 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
390
391 const __m256i vi9x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i9));
392 const __m256i vk9x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)))), vk_zero_point);
393
394 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
395
396 const __m256i vi10x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i10));
397 const __m256i vk10x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
398
399 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
400
401 const __m256i vi11x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i11));
402 const __m256i vk11x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)))), vk_zero_point);
403
404 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
405
406 const __m256i vi12x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i12));
407 const __m256i vk12x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
408
409 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
410
411 const __m256i vi13x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i13));
412 const __m256i vk13x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)))), vk_zero_point);
413
414 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
415
416 const __m256i vi14x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i14));
417 const __m256i vk14x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
418
419 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
420
421 const __m256i vi15x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i15));
422 const __m256i vk15x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)))), vk_zero_point);
423
424 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
425
426 const __m256i vi16x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i16));
427 const __m256i vk16x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
428
429 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
430
431 const __m256i vi17x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i17));
432 const __m256i vk17x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)))), vk_zero_point);
433
434 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
435
436 const __m256i vi18x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i18));
437 const __m256i vk18x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)))), vk_zero_point);
438
439 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
440
441 const __m256i vi19x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i19));
442 const __m256i vk19x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)))), vk_zero_point);
443
444 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
445
446 const __m256i vi20x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i20));
447 const __m256i vk20x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)))), vk_zero_point);
448
449 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
450
451 const __m256i vi21x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i21));
452 const __m256i vk21x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)))), vk_zero_point);
453
454 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
455
456 const __m256i vi22x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i22));
457 const __m256i vk22x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)))), vk_zero_point);
458
459 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
460
461 const __m256i vi23x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i23));
462 const __m256i vk23x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)))), vk_zero_point);
463
464 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
465
466 const __m256i vi24x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i24));
467 const __m256i vk24x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)))), vk_zero_point);
468
469 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
470
471
472 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
473 vscaled01234567 = _mm256_mul_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.scale));
474 vscaled01234567 = _mm256_min_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point));
475 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
476
477
478 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
479 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
480
481 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
482
483 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
484 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
485
486 if (c & 4) {
487 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
488 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
489 output += 4;
490 }
491 if (c & 2) {
492 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
493 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
494 output += 2;
495 }
496 if (c & 1) {
497 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
498 output += 1;
499 }
500 }
501 }
502
503 output = (uint8_t*) ((uintptr_t) output + output_increment);
504 } while (--output_width != 0);
505 }
506