Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l32c8s8r-minmax-fp32-avx2-mul32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-avx2-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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/common.h"
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qs8_dwconv_minmax_fp32_ukernel_8f8m9l32c8s8r__avx2_mul32(
21 size_t channels,
22 size_t output_width,
23 const int8_t** input,
24 const void* weights,
25 int8_t* output,
26 intptr_t input_stride,
27 size_t output_increment,
28 size_t input_offset,
29 const int8_t* zero,
30 size_t kernel_size,
31 int32_t* buffer,
32 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(channels != 0);
35 assert(output_width != 0);
36 assert(kernel_size > 8);
37
38
39 const __m256 vscale = _mm256_set1_ps(params->fp32_scalar.scale);
40 XNN_FORCE_REALIZATION(vscale);
41 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);
42 const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point);
43 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
44 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47
48 do {
49 const void* w = weights;
50
51 // First pass to process 8 inputs.
52 {
53 int32_t* b = buffer;
54 const int8_t* i0 = input[0];
55 assert(i0 != NULL);
56 if XNN_UNPREDICTABLE(i0 != zero) {
57 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
58 }
59 const int8_t* i1 = input[1];
60 assert(i1 != NULL);
61 if XNN_UNPREDICTABLE(i1 != zero) {
62 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
63 }
64 const int8_t* i2 = input[2];
65 assert(i2 != NULL);
66 if XNN_UNPREDICTABLE(i2 != zero) {
67 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
68 }
69 const int8_t* i3 = input[3];
70 assert(i3 != NULL);
71 if XNN_UNPREDICTABLE(i3 != zero) {
72 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
73 }
74 const int8_t* i4 = input[4];
75 assert(i4 != NULL);
76 if XNN_UNPREDICTABLE(i4 != zero) {
77 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
78 }
79 const int8_t* i5 = input[5];
80 assert(i5 != NULL);
81 if XNN_UNPREDICTABLE(i5 != zero) {
82 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
83 }
84 const int8_t* i6 = input[6];
85 assert(i6 != NULL);
86 if XNN_UNPREDICTABLE(i6 != zero) {
87 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
88 }
89 const int8_t* i7 = input[7];
90 assert(i7 != NULL);
91 if XNN_UNPREDICTABLE(i7 != zero) {
92 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
93 }
94 input += 8;
95
96 size_t c = round_up_po2(channels, 8);
97
98 for (; c >= 32; c -= 32) {
99 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
100 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 8));
101 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 16));
102 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 24));
103
104
105 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
106 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
107 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
108 const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 8 * sizeof(int8_t))));
109 const __m256i vi0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 16)));
110 const __m256i vk0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
111 const __m256i vi0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 24)));
112 const __m256i vk0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 24 * sizeof(int8_t))));
113 i0 += 32;
114
115 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
116 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
117 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi0xGHIJKLMN, vk0xGHIJKLMN));
118 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi0xOPQRSTUV, vk0xOPQRSTUV));
119
120 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
121 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
122 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
123 const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 40 * sizeof(int8_t))));
124 const __m256i vi1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 16)));
125 const __m256i vk1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
126 const __m256i vi1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 24)));
127 const __m256i vk1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 56 * sizeof(int8_t))));
128 i1 += 32;
129
130 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
131 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
132 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi1xGHIJKLMN, vk1xGHIJKLMN));
133 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi1xOPQRSTUV, vk1xOPQRSTUV));
134
135 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
136 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
137 const __m256i vi2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
138 const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 72 * sizeof(int8_t))));
139 const __m256i vi2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 16)));
140 const __m256i vk2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
141 const __m256i vi2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 24)));
142 const __m256i vk2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 88 * sizeof(int8_t))));
143 i2 += 32;
144
145 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
146 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
147 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi2xGHIJKLMN, vk2xGHIJKLMN));
148 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi2xOPQRSTUV, vk2xOPQRSTUV));
149
150 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
151 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
152 const __m256i vi3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
153 const __m256i vk3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 104 * sizeof(int8_t))));
154 const __m256i vi3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 16)));
155 const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
156 const __m256i vi3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 24)));
157 const __m256i vk3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 120 * sizeof(int8_t))));
158 i3 += 32;
159
160 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
161 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi3x89ABCDEF, vk3x89ABCDEF));
162 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN));
163 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi3xOPQRSTUV, vk3xOPQRSTUV));
164
165 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
166 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
167 const __m256i vi4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
168 const __m256i vk4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 136 * sizeof(int8_t))));
169 const __m256i vi4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 16)));
170 const __m256i vk4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
171 const __m256i vi4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 24)));
172 const __m256i vk4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 152 * sizeof(int8_t))));
173 i4 += 32;
174
175 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
176 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF));
177 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN));
178 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV));
179
180 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
181 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
182 const __m256i vi5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
183 const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 168 * sizeof(int8_t))));
184 const __m256i vi5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 16)));
185 const __m256i vk5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
186 const __m256i vi5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 24)));
187 const __m256i vk5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 184 * sizeof(int8_t))));
188 i5 += 32;
189
190 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
191 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF));
192 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi5xGHIJKLMN, vk5xGHIJKLMN));
193 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi5xOPQRSTUV, vk5xOPQRSTUV));
194
195 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
196 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
197 const __m256i vi6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
198 const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 200 * sizeof(int8_t))));
199 const __m256i vi6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 16)));
200 const __m256i vk6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
201 const __m256i vi6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 24)));
202 const __m256i vk6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 216 * sizeof(int8_t))));
203 i6 += 32;
204
205 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
206 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF));
207 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi6xGHIJKLMN, vk6xGHIJKLMN));
208 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi6xOPQRSTUV, vk6xOPQRSTUV));
209
210 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
211 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
212 const __m256i vi7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 8)));
213 const __m256i vk7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 232 * sizeof(int8_t))));
214 const __m256i vi7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 16)));
215 const __m256i vk7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
216 const __m256i vi7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 24)));
217 const __m256i vk7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 248 * sizeof(int8_t))));
218 i7 += 32;
219
220 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
221 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi7x89ABCDEF, vk7x89ABCDEF));
222 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN));
223 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi7xOPQRSTUV, vk7xOPQRSTUV));
224
225 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t));
226
227 _mm256_storeu_si256((__m256i*) b, vacc01234567);
228 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
229 _mm256_storeu_si256((__m256i*) (b + 16), vaccGHIJKLMN);
230 _mm256_storeu_si256((__m256i*) (b + 24), vaccOPQRSTUV);
231 b += 32;
232 }
233
234 if XNN_UNLIKELY(c != 0) {
235 do {
236 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
237
238 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
239 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t))));
240 i0 += 8;
241
242 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
243
244 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
245 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
246 i1 += 8;
247
248 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
249
250 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
251 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
252 i2 += 8;
253
254 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
255
256 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
257 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
258 i3 += 8;
259
260 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
261
262 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
263 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
264 i4 += 8;
265
266 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
267
268 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
269 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
270 i5 += 8;
271
272 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
273
274 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
275 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))));
276 i6 += 8;
277
278 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
279
280 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
281 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))));
282 i7 += 8;
283
284 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
285
286 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t));
287
288 _mm256_storeu_si256((__m256i*) b, vacc01234567);
289 b += 8;
290 c -= 8;
291 } while (c != 0);
292 }
293 }
294
295 // Middle pass to process 8 inputs in each iteration.
296 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
297 int32_t* b = buffer;
298 const int8_t* i0 = input[0];
299 assert(i0 != NULL);
300 if XNN_UNPREDICTABLE(i0 != zero) {
301 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
302 }
303 const int8_t* i1 = input[1];
304 assert(i1 != NULL);
305 if XNN_UNPREDICTABLE(i1 != zero) {
306 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
307 }
308 const int8_t* i2 = input[2];
309 assert(i2 != NULL);
310 if XNN_UNPREDICTABLE(i2 != zero) {
311 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
312 }
313 const int8_t* i3 = input[3];
314 assert(i3 != NULL);
315 if XNN_UNPREDICTABLE(i3 != zero) {
316 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
317 }
318 const int8_t* i4 = input[4];
319 assert(i4 != NULL);
320 if XNN_UNPREDICTABLE(i4 != zero) {
321 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
322 }
323 const int8_t* i5 = input[5];
324 assert(i5 != NULL);
325 if XNN_UNPREDICTABLE(i5 != zero) {
326 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
327 }
328 const int8_t* i6 = input[6];
329 assert(i6 != NULL);
330 if XNN_UNPREDICTABLE(i6 != zero) {
331 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
332 }
333 const int8_t* i7 = input[7];
334 assert(i7 != NULL);
335 if XNN_UNPREDICTABLE(i7 != zero) {
336 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
337 }
338 input += 8;
339
340 size_t c = round_up_po2(channels, 8);
341
342 for (; c >= 32; c -= 32) {
343 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
344 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
345 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) (b + 16));
346 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) (b + 24));
347
348
349 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
350 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
351 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
352 const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
353 const __m256i vi0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 16)));
354 const __m256i vk0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
355 const __m256i vi0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 24)));
356 const __m256i vk0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
357 i0 += 32;
358
359 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
360 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
361 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi0xGHIJKLMN, vk0xGHIJKLMN));
362 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi0xOPQRSTUV, vk0xOPQRSTUV));
363
364 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
365 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
366 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
367 const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t))));
368 const __m256i vi1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 16)));
369 const __m256i vk1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
370 const __m256i vi1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 24)));
371 const __m256i vk1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t))));
372 i1 += 32;
373
374 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
375 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
376 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi1xGHIJKLMN, vk1xGHIJKLMN));
377 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi1xOPQRSTUV, vk1xOPQRSTUV));
378
379 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
380 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
381 const __m256i vi2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
382 const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t))));
383 const __m256i vi2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 16)));
384 const __m256i vk2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
385 const __m256i vi2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 24)));
386 const __m256i vk2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88 * sizeof(int8_t))));
387 i2 += 32;
388
389 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
390 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
391 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi2xGHIJKLMN, vk2xGHIJKLMN));
392 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi2xOPQRSTUV, vk2xOPQRSTUV));
393
394 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
395 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
396 const __m256i vi3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
397 const __m256i vk3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 104 * sizeof(int8_t))));
398 const __m256i vi3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 16)));
399 const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
400 const __m256i vi3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 24)));
401 const __m256i vk3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 120 * sizeof(int8_t))));
402 i3 += 32;
403
404 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
405 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi3x89ABCDEF, vk3x89ABCDEF));
406 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN));
407 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi3xOPQRSTUV, vk3xOPQRSTUV));
408
409 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
410 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
411 const __m256i vi4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
412 const __m256i vk4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 136 * sizeof(int8_t))));
413 const __m256i vi4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 16)));
414 const __m256i vk4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 144 * sizeof(int8_t))));
415 const __m256i vi4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 24)));
416 const __m256i vk4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 152 * sizeof(int8_t))));
417 i4 += 32;
418
419 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
420 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF));
421 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN));
422 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV));
423
424 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
425 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 160 * sizeof(int8_t))));
426 const __m256i vi5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
427 const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 168 * sizeof(int8_t))));
428 const __m256i vi5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 16)));
429 const __m256i vk5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 176 * sizeof(int8_t))));
430 const __m256i vi5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 24)));
431 const __m256i vk5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 184 * sizeof(int8_t))));
432 i5 += 32;
433
434 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
435 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF));
436 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi5xGHIJKLMN, vk5xGHIJKLMN));
437 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi5xOPQRSTUV, vk5xOPQRSTUV));
438
439 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
440 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 192 * sizeof(int8_t))));
441 const __m256i vi6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
442 const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 200 * sizeof(int8_t))));
443 const __m256i vi6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 16)));
444 const __m256i vk6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 208 * sizeof(int8_t))));
445 const __m256i vi6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 24)));
446 const __m256i vk6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 216 * sizeof(int8_t))));
447 i6 += 32;
448
449 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
450 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF));
451 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi6xGHIJKLMN, vk6xGHIJKLMN));
452 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi6xOPQRSTUV, vk6xOPQRSTUV));
453
454 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
455 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 224 * sizeof(int8_t))));
456 const __m256i vi7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 8)));
457 const __m256i vk7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 232 * sizeof(int8_t))));
458 const __m256i vi7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 16)));
459 const __m256i vk7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 240 * sizeof(int8_t))));
460 const __m256i vi7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 24)));
461 const __m256i vk7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 248 * sizeof(int8_t))));
462 i7 += 32;
463
464 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
465 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi7x89ABCDEF, vk7x89ABCDEF));
466 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN));
467 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi7xOPQRSTUV, vk7xOPQRSTUV));
468
469 w = (const void*) ((uintptr_t) w + 256 * sizeof(int8_t));
470
471 _mm256_storeu_si256((__m256i*) b, vacc01234567);
472 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
473 _mm256_storeu_si256((__m256i*) (b + 16), vaccGHIJKLMN);
474 _mm256_storeu_si256((__m256i*) (b + 24), vaccOPQRSTUV);
475 b += 32;
476 }
477
478 if XNN_UNLIKELY(c != 0) {
479 do {
480 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
481
482 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
483 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w)));
484 i0 += 8;
485
486 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
487
488 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
489 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
490 i1 += 8;
491
492 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
493
494 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
495 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
496 i2 += 8;
497
498 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
499
500 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
501 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
502 i3 += 8;
503
504 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
505
506 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
507 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
508 i4 += 8;
509
510 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
511
512 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
513 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t))));
514 i5 += 8;
515
516 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
517
518 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
519 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
520 i6 += 8;
521
522 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
523
524 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
525 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t))));
526 i7 += 8;
527
528 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
529
530 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
531
532 _mm256_storeu_si256((__m256i*) b, vacc01234567);
533 b += 8;
534 c -= 8;
535 } while (c != 0);
536 }
537 }
538
539 // Last pass to process up to 9 inputs.
540 {
541 const int32_t* b = buffer;
542 const int8_t* i0 = input[0];
543 assert(i0 != NULL);
544 if XNN_UNPREDICTABLE(i0 != zero) {
545 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
546 }
547 const int8_t* i1 = input[1];
548 assert(i1 != NULL);
549 if XNN_UNPREDICTABLE(i1 != zero) {
550 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
551 }
552 const int8_t* i2 = input[2];
553 assert(i2 != NULL);
554 if XNN_UNPREDICTABLE(i2 != zero) {
555 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
556 }
557 const int8_t* i3 = input[3];
558 assert(i3 != NULL);
559 if XNN_UNPREDICTABLE(i3 != zero) {
560 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
561 }
562 const int8_t* i4 = input[4];
563 assert(i4 != NULL);
564 if XNN_UNPREDICTABLE(i4 != zero) {
565 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
566 }
567 const int8_t* i5 = input[5];
568 assert(i5 != NULL);
569 if XNN_UNPREDICTABLE(i5 != zero) {
570 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
571 }
572 const int8_t* i6 = input[6];
573 assert(i6 != NULL);
574 if XNN_UNPREDICTABLE(i6 != zero) {
575 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
576 }
577 const int8_t* i7 = input[7];
578 assert(i7 != NULL);
579 if XNN_UNPREDICTABLE(i7 != zero) {
580 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
581 }
582 const int8_t* i8 = input[8];
583 assert(i8 != NULL);
584 if XNN_UNPREDICTABLE(i8 != zero) {
585 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
586 }
587
588 size_t c = channels;
589
590 for (; c >= 32; c -= 32) {
591 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
592 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
593 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) (b + 16));
594 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) (b + 24));
595 b += 32;
596
597
598 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
599 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
600 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
601 const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
602 const __m256i vi0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 16)));
603 const __m256i vk0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
604 const __m256i vi0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 24)));
605 const __m256i vk0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
606 i0 += 32;
607
608 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
609 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
610 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi0xGHIJKLMN, vk0xGHIJKLMN));
611 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi0xOPQRSTUV, vk0xOPQRSTUV));
612
613 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
614 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
615 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
616 const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t))));
617 const __m256i vi1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 16)));
618 const __m256i vk1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
619 const __m256i vi1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 24)));
620 const __m256i vk1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t))));
621 i1 += 32;
622
623 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
624 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
625 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi1xGHIJKLMN, vk1xGHIJKLMN));
626 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi1xOPQRSTUV, vk1xOPQRSTUV));
627
628 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
629 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
630 const __m256i vi2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
631 const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t))));
632 const __m256i vi2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 16)));
633 const __m256i vk2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
634 const __m256i vi2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 24)));
635 const __m256i vk2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88 * sizeof(int8_t))));
636 i2 += 32;
637
638 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
639 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
640 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi2xGHIJKLMN, vk2xGHIJKLMN));
641 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi2xOPQRSTUV, vk2xOPQRSTUV));
642
643 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
644 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
645 const __m256i vi3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
646 const __m256i vk3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 104 * sizeof(int8_t))));
647 const __m256i vi3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 16)));
648 const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
649 const __m256i vi3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 24)));
650 const __m256i vk3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 120 * sizeof(int8_t))));
651 i3 += 32;
652
653 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
654 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi3x89ABCDEF, vk3x89ABCDEF));
655 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN));
656 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi3xOPQRSTUV, vk3xOPQRSTUV));
657
658 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
659 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
660 const __m256i vi4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
661 const __m256i vk4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 136 * sizeof(int8_t))));
662 const __m256i vi4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 16)));
663 const __m256i vk4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 144 * sizeof(int8_t))));
664 const __m256i vi4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 24)));
665 const __m256i vk4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 152 * sizeof(int8_t))));
666 i4 += 32;
667
668 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
669 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF));
670 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN));
671 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV));
672
673 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
674 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 160 * sizeof(int8_t))));
675 const __m256i vi5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
676 const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 168 * sizeof(int8_t))));
677 const __m256i vi5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 16)));
678 const __m256i vk5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 176 * sizeof(int8_t))));
679 const __m256i vi5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 24)));
680 const __m256i vk5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 184 * sizeof(int8_t))));
681 i5 += 32;
682
683 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
684 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF));
685 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi5xGHIJKLMN, vk5xGHIJKLMN));
686 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi5xOPQRSTUV, vk5xOPQRSTUV));
687
688 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
689 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 192 * sizeof(int8_t))));
690 const __m256i vi6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
691 const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 200 * sizeof(int8_t))));
692 const __m256i vi6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 16)));
693 const __m256i vk6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 208 * sizeof(int8_t))));
694 const __m256i vi6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 24)));
695 const __m256i vk6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 216 * sizeof(int8_t))));
696 i6 += 32;
697
698 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
699 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF));
700 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi6xGHIJKLMN, vk6xGHIJKLMN));
701 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi6xOPQRSTUV, vk6xOPQRSTUV));
702
703 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
704 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 224 * sizeof(int8_t))));
705 const __m256i vi7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 8)));
706 const __m256i vk7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 232 * sizeof(int8_t))));
707 const __m256i vi7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 16)));
708 const __m256i vk7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 240 * sizeof(int8_t))));
709 const __m256i vi7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 24)));
710 const __m256i vk7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 248 * sizeof(int8_t))));
711 i7 += 32;
712
713 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
714 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi7x89ABCDEF, vk7x89ABCDEF));
715 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN));
716 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi7xOPQRSTUV, vk7xOPQRSTUV));
717
718 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
719 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 256 * sizeof(int8_t))));
720 const __m256i vi8x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 8)));
721 const __m256i vk8x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 264 * sizeof(int8_t))));
722 const __m256i vi8xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 16)));
723 const __m256i vk8xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 272 * sizeof(int8_t))));
724 const __m256i vi8xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 24)));
725 const __m256i vk8xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 280 * sizeof(int8_t))));
726 i8 += 32;
727
728 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
729 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi8x89ABCDEF, vk8x89ABCDEF));
730 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi8xGHIJKLMN, vk8xGHIJKLMN));
731 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi8xOPQRSTUV, vk8xOPQRSTUV));
732
733 w = (const void*) ((uintptr_t) w + 288 * sizeof(int8_t));
734
735 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
736 __m256 vscaled89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
737 __m256 vscaledGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
738 __m256 vscaledOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
739
740 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
741 vscaled89ABCDEF = _mm256_mul_ps(vscaled89ABCDEF, vscale);
742 vscaledGHIJKLMN = _mm256_mul_ps(vscaledGHIJKLMN, vscale);
743 vscaledOPQRSTUV = _mm256_mul_ps(vscaledOPQRSTUV, vscale);
744
745 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
746 vscaled89ABCDEF = _mm256_min_ps(vscaled89ABCDEF, voutput_max_less_zero_point);
747 vscaledGHIJKLMN = _mm256_min_ps(vscaledGHIJKLMN, voutput_max_less_zero_point);
748 vscaledOPQRSTUV = _mm256_min_ps(vscaledOPQRSTUV, voutput_max_less_zero_point);
749
750 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
751 vacc89ABCDEF = _mm256_cvtps_epi32(vscaled89ABCDEF);
752 vaccGHIJKLMN = _mm256_cvtps_epi32(vscaledGHIJKLMN);
753 vaccOPQRSTUV = _mm256_cvtps_epi32(vscaledOPQRSTUV);
754
755 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
756 __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
757
758 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
759 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
760
761 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
762 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
763
764 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
765 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
766 output += 32;
767 }
768
769 if XNN_UNLIKELY(c != 0) {
770 do {
771 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
772 b += 8;
773
774 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
775 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
776 i0 += 8;
777
778 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
779
780 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
781 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
782 i1 += 8;
783
784 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
785
786 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
787 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
788 i2 += 8;
789
790 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
791
792 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
793 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
794 i3 += 8;
795
796 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
797
798 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
799 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
800 i4 += 8;
801
802 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
803
804 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
805 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t))));
806 i5 += 8;
807
808 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
809
810 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
811 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
812 i6 += 8;
813
814 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
815
816 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
817 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t))));
818 i7 += 8;
819
820 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
821
822 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
823 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
824 i8 += 8;
825
826 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
827
828 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
829 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
830 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
831 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
832
833 w = (void*) ((uintptr_t) w + 72 * sizeof(int8_t));
834
835 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
836
837 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
838
839 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
840
841 if XNN_LIKELY(c >= 8) {
842 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
843 output += 8;
844 c -= 8;
845 } else {
846 if (c & 4) {
847 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
848 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
849 output += 4;
850 }
851 if (c & 2) {
852 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
853 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
854 output += 2;
855 }
856 if (c & 1) {
857 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
858 output += 1;
859 }
860 c = 0;
861 }
862 } while (c != 0);
863 }
864 }
865
866 input = (const int8_t**) ((uintptr_t) input + input_stride);
867 output = (int8_t*) ((uintptr_t) output + output_increment);
868 } while (--output_width != 0);
869 }