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_up16x25__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_up16x25__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 >= 16; c -= 16) {
165 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
166 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 8));
167
168
169 const __m256i vi0x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i0));
170 const __m256i vk0x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
171 const __m256i vi0x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
172 const __m256i vk0x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)))), vk_zero_point);
173 i0 += 16;
174
175 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
176 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
177
178 const __m256i vi1x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i1));
179 const __m256i vk1x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
180 const __m256i vi1x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
181 const __m256i vk1x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)))), vk_zero_point);
182 i1 += 16;
183
184 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
185 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
186
187 const __m256i vi2x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i2));
188 const __m256i vk2x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
189 const __m256i vi2x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
190 const __m256i vk2x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)))), vk_zero_point);
191 i2 += 16;
192
193 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
194 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
195
196 const __m256i vi3x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i3));
197 const __m256i vk3x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
198 const __m256i vi3x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
199 const __m256i vk3x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)))), vk_zero_point);
200 i3 += 16;
201
202 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
203 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi3x89ABCDEF, vk3x89ABCDEF));
204
205 const __m256i vi4x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i4));
206 const __m256i vk4x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
207 const __m256i vi4x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
208 const __m256i vk4x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)))), vk_zero_point);
209 i4 += 16;
210
211 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
212 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF));
213
214 const __m256i vi5x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i5));
215 const __m256i vk5x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
216 const __m256i vi5x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
217 const __m256i vk5x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)))), vk_zero_point);
218 i5 += 16;
219
220 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
221 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF));
222
223 const __m256i vi6x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i6));
224 const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
225 const __m256i vi6x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
226 const __m256i vk6x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)))), vk_zero_point);
227 i6 += 16;
228
229 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
230 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF));
231
232 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7));
233 const __m256i vk7x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
234 const __m256i vi7x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 8)));
235 const __m256i vk7x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)))), vk_zero_point);
236 i7 += 16;
237
238 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
239 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi7x89ABCDEF, vk7x89ABCDEF));
240
241 const __m256i vi8x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i8));
242 const __m256i vk8x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
243 const __m256i vi8x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 8)));
244 const __m256i vk8x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)))), vk_zero_point);
245 i8 += 16;
246
247 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
248 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi8x89ABCDEF, vk8x89ABCDEF));
249
250 const __m256i vi9x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i9));
251 const __m256i vk9x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t)))), vk_zero_point);
252 const __m256i vi9x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i9 + 8)));
253 const __m256i vk9x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(uint8_t)))), vk_zero_point);
254 i9 += 16;
255
256 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
257 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi9x89ABCDEF, vk9x89ABCDEF));
258
259 const __m256i vi10x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i10));
260 const __m256i vk10x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(uint8_t)))), vk_zero_point);
261 const __m256i vi10x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i10 + 8)));
262 const __m256i vk10x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(uint8_t)))), vk_zero_point);
263 i10 += 16;
264
265 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
266 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi10x89ABCDEF, vk10x89ABCDEF));
267
268 const __m256i vi11x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i11));
269 const __m256i vk11x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(uint8_t)))), vk_zero_point);
270 const __m256i vi11x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i11 + 8)));
271 const __m256i vk11x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(uint8_t)))), vk_zero_point);
272 i11 += 16;
273
274 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
275 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi11x89ABCDEF, vk11x89ABCDEF));
276
277 const __m256i vi12x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i12));
278 const __m256i vk12x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(uint8_t)))), vk_zero_point);
279 const __m256i vi12x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i12 + 8)));
280 const __m256i vk12x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(uint8_t)))), vk_zero_point);
281 i12 += 16;
282
283 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
284 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi12x89ABCDEF, vk12x89ABCDEF));
285
286 const __m256i vi13x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i13));
287 const __m256i vk13x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(uint8_t)))), vk_zero_point);
288 const __m256i vi13x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i13 + 8)));
289 const __m256i vk13x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(uint8_t)))), vk_zero_point);
290 i13 += 16;
291
292 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
293 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi13x89ABCDEF, vk13x89ABCDEF));
294
295 const __m256i vi14x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i14));
296 const __m256i vk14x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(uint8_t)))), vk_zero_point);
297 const __m256i vi14x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i14 + 8)));
298 const __m256i vk14x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(uint8_t)))), vk_zero_point);
299 i14 += 16;
300
301 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
302 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi14x89ABCDEF, vk14x89ABCDEF));
303
304 const __m256i vi15x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i15));
305 const __m256i vk15x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(uint8_t)))), vk_zero_point);
306 const __m256i vi15x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i15 + 8)));
307 const __m256i vk15x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(uint8_t)))), vk_zero_point);
308 i15 += 16;
309
310 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
311 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi15x89ABCDEF, vk15x89ABCDEF));
312
313 const __m256i vi16x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i16));
314 const __m256i vk16x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(uint8_t)))), vk_zero_point);
315 const __m256i vi16x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i16 + 8)));
316 const __m256i vk16x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(uint8_t)))), vk_zero_point);
317 i16 += 16;
318
319 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
320 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi16x89ABCDEF, vk16x89ABCDEF));
321
322 const __m256i vi17x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i17));
323 const __m256i vk17x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(uint8_t)))), vk_zero_point);
324 const __m256i vi17x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i17 + 8)));
325 const __m256i vk17x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(uint8_t)))), vk_zero_point);
326 i17 += 16;
327
328 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
329 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi17x89ABCDEF, vk17x89ABCDEF));
330
331 const __m256i vi18x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i18));
332 const __m256i vk18x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(uint8_t)))), vk_zero_point);
333 const __m256i vi18x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i18 + 8)));
334 const __m256i vk18x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(uint8_t)))), vk_zero_point);
335 i18 += 16;
336
337 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
338 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi18x89ABCDEF, vk18x89ABCDEF));
339
340 const __m256i vi19x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i19));
341 const __m256i vk19x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(uint8_t)))), vk_zero_point);
342 const __m256i vi19x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i19 + 8)));
343 const __m256i vk19x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(uint8_t)))), vk_zero_point);
344 i19 += 16;
345
346 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
347 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi19x89ABCDEF, vk19x89ABCDEF));
348
349 const __m256i vi20x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i20));
350 const __m256i vk20x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(uint8_t)))), vk_zero_point);
351 const __m256i vi20x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i20 + 8)));
352 const __m256i vk20x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(uint8_t)))), vk_zero_point);
353 i20 += 16;
354
355 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
356 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi20x89ABCDEF, vk20x89ABCDEF));
357
358 const __m256i vi21x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i21));
359 const __m256i vk21x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(uint8_t)))), vk_zero_point);
360 const __m256i vi21x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i21 + 8)));
361 const __m256i vk21x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(uint8_t)))), vk_zero_point);
362 i21 += 16;
363
364 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
365 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi21x89ABCDEF, vk21x89ABCDEF));
366
367 const __m256i vi22x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i22));
368 const __m256i vk22x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(uint8_t)))), vk_zero_point);
369 const __m256i vi22x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i22 + 8)));
370 const __m256i vk22x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(uint8_t)))), vk_zero_point);
371 i22 += 16;
372
373 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
374 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi22x89ABCDEF, vk22x89ABCDEF));
375
376 const __m256i vi23x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i23));
377 const __m256i vk23x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(uint8_t)))), vk_zero_point);
378 const __m256i vi23x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i23 + 8)));
379 const __m256i vk23x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(uint8_t)))), vk_zero_point);
380 i23 += 16;
381
382 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
383 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi23x89ABCDEF, vk23x89ABCDEF));
384
385 const __m256i vi24x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i24));
386 const __m256i vk24x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(uint8_t)))), vk_zero_point);
387 const __m256i vi24x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i24 + 8)));
388 const __m256i vk24x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(uint8_t)))), vk_zero_point);
389 i24 += 16;
390
391 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
392 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi24x89ABCDEF, vk24x89ABCDEF));
393
394 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(uint8_t));
395
396 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
397 __m256 vscaled89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
398
399 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
400 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
401 vscaled89ABCDEF = _mm256_mul_ps(vscaled89ABCDEF, vscale);
402
403 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
404 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
405 vscaled89ABCDEF = _mm256_min_ps(vscaled89ABCDEF, voutput_max_less_zero_point);
406
407 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
408 vacc89ABCDEF = _mm256_cvtps_epi32(vscaled89ABCDEF);
409
410 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
411 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
412
413 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
414
415 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
416 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
417
418 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
419 output += 16;
420 }
421 if XNN_UNLIKELY(c != 0) {
422 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
423 do {
424 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
425
426
427 const __m256i vi0x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i0));
428 const __m256i vk0x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) k)), vk_zero_point);
429 i0 += 8;
430
431 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
432
433 const __m256i vi1x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i1));
434 const __m256i vk1x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 16))), vk_zero_point);
435 i1 += 8;
436
437 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
438
439 const __m256i vi2x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i2));
440 const __m256i vk2x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 32))), vk_zero_point);
441 i2 += 8;
442
443 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
444
445 const __m256i vi3x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i3));
446 const __m256i vk3x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 48))), vk_zero_point);
447 i3 += 8;
448
449 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
450
451 const __m256i vi4x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i4));
452 const __m256i vk4x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 64))), vk_zero_point);
453 i4 += 8;
454
455 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
456
457 const __m256i vi5x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i5));
458 const __m256i vk5x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 80))), vk_zero_point);
459 i5 += 8;
460
461 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
462
463 const __m256i vi6x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i6));
464 const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 96))), vk_zero_point);
465 i6 += 8;
466
467 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
468
469 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7));
470 const __m256i vk7x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 112))), vk_zero_point);
471 i7 += 8;
472
473 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
474
475 const __m256i vi8x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i8));
476 const __m256i vk8x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 128))), vk_zero_point);
477 i8 += 8;
478
479 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
480
481 const __m256i vi9x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i9));
482 const __m256i vk9x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 144))), vk_zero_point);
483 i9 += 8;
484
485 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
486
487 const __m256i vi10x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i10));
488 const __m256i vk10x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 160))), vk_zero_point);
489 i10 += 8;
490
491 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
492
493 const __m256i vi11x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i11));
494 const __m256i vk11x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 176))), vk_zero_point);
495 i11 += 8;
496
497 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
498
499 const __m256i vi12x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i12));
500 const __m256i vk12x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 192))), vk_zero_point);
501 i12 += 8;
502
503 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
504
505 const __m256i vi13x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i13));
506 const __m256i vk13x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 208))), vk_zero_point);
507 i13 += 8;
508
509 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
510
511 const __m256i vi14x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i14));
512 const __m256i vk14x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 224))), vk_zero_point);
513 i14 += 8;
514
515 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
516
517 const __m256i vi15x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i15));
518 const __m256i vk15x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 240))), vk_zero_point);
519 i15 += 8;
520
521 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
522
523 const __m256i vi16x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i16));
524 const __m256i vk16x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 256))), vk_zero_point);
525 i16 += 8;
526
527 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
528
529 const __m256i vi17x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i17));
530 const __m256i vk17x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 272))), vk_zero_point);
531 i17 += 8;
532
533 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
534
535 const __m256i vi18x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i18));
536 const __m256i vk18x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 288))), vk_zero_point);
537 i18 += 8;
538
539 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
540
541 const __m256i vi19x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i19));
542 const __m256i vk19x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 304))), vk_zero_point);
543 i19 += 8;
544
545 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
546
547 const __m256i vi20x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i20));
548 const __m256i vk20x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 320))), vk_zero_point);
549 i20 += 8;
550
551 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
552
553 const __m256i vi21x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i21));
554 const __m256i vk21x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 336))), vk_zero_point);
555 i21 += 8;
556
557 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
558
559 const __m256i vi22x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i22));
560 const __m256i vk22x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 352))), vk_zero_point);
561 i22 += 8;
562
563 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
564
565 const __m256i vi23x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i23));
566 const __m256i vk23x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 368))), vk_zero_point);
567 i23 += 8;
568
569 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
570
571 const __m256i vi24x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i24));
572 const __m256i vk24x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 384))), vk_zero_point);
573 i24 += 8;
574
575 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
576
577 k += 8;
578
579 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
580 vscaled01234567 = _mm256_mul_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.scale));
581 vscaled01234567 = _mm256_min_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point));
582 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
583
584 w = (const void*) ((const int32_t*) w + 8);
585
586 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
587 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
588
589 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
590
591 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
592 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
593
594 if XNN_LIKELY(c >= 8) {
595 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
596 output += 8;
597 c -= 8;
598 } else {
599 if (c & 4) {
600 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
601 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
602 output += 4;
603 }
604 if (c & 2) {
605 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
606 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
607 output += 2;
608 }
609 if (c & 1) {
610 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
611 output += 1;
612 }
613 c = 0;
614 }
615 } while (c != 0);
616 }
617
618 output = (uint8_t*) ((uintptr_t) output + output_increment);
619 } while (--output_width != 0);
620 }
621