Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l32c16s16r-minmax-fp32-avx2-mul16-vpmovsx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-avx2-mul16-vpmovsx.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/dwconv.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l32c16s16r__avx2_mul16_vpmovsx(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 5);
36
37 const __m256 vscale = _mm256_set1_ps(params->fp32_scalar.scale);
38 XNN_FORCE_REALIZATION(vscale);
39 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);
40 const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point);
41 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
42 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
43 XNN_FORCE_REALIZATION(voutput_zero_point);
44 XNN_FORCE_REALIZATION(voutput_min);
45
46 do {
47 const void* w = weights;
48
49 // First pass to process 5 inputs.
50 {
51 int32_t* b = buffer;
52
53 const int8_t* i0 = input[0];
54 assert(i0 != NULL);
55 if XNN_UNPREDICTABLE(i0 != zero) {
56 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
57 }
58 const int8_t* i1 = input[1];
59 assert(i1 != NULL);
60 if XNN_UNPREDICTABLE(i1 != zero) {
61 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
62 }
63 const int8_t* i2 = input[2];
64 assert(i2 != NULL);
65 if XNN_UNPREDICTABLE(i2 != zero) {
66 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
67 }
68 const int8_t* i3 = input[3];
69 assert(i3 != NULL);
70 if XNN_UNPREDICTABLE(i3 != zero) {
71 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
72 }
73 const int8_t* i4 = input[4];
74 assert(i4 != NULL);
75 if XNN_UNPREDICTABLE(i4 != zero) {
76 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
77 }
78 input += 5;
79
80 size_t c = round_up_po2(channels, 16);
81
82 for (; c >= 32; c -= 32) {
83 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
84 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
85 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t)));
86 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 24 * sizeof(int32_t)));
87
88
89 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
90 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
91 const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16)));
92 const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
93 i0 += 32;
94
95 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
96 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
97 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
98 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
99 const __m256i vprod0xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
100 const __m128i vprod0xOPQRSTUV = _mm256_extracti128_si256(vprod0xGHIJKLMNOPQRSTUV, 1);
101 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0xGHIJKLMNOPQRSTUV)));
102 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV));
103
104 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
105 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
106 const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
107 const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
108 i1 += 32;
109
110 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
111 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
112 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
113 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
114 const __m256i vprod1xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV);
115 const __m128i vprod1xOPQRSTUV = _mm256_extracti128_si256(vprod1xGHIJKLMNOPQRSTUV, 1);
116 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1xGHIJKLMNOPQRSTUV)));
117 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV));
118
119 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
120 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
121 const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
122 const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
123 i2 += 32;
124
125 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
126 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
127 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
128 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
129 const __m256i vprod2xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
130 const __m128i vprod2xOPQRSTUV = _mm256_extracti128_si256(vprod2xGHIJKLMNOPQRSTUV, 1);
131 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2xGHIJKLMNOPQRSTUV)));
132 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod2xOPQRSTUV));
133
134 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
135 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
136 const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
137 const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
138 i3 += 32;
139
140 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
141 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
142 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
143 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
144 const __m256i vprod3xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV);
145 const __m128i vprod3xOPQRSTUV = _mm256_extracti128_si256(vprod3xGHIJKLMNOPQRSTUV, 1);
146 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3xGHIJKLMNOPQRSTUV)));
147 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod3xOPQRSTUV));
148
149 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
150 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
151 const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
152 const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
153 i4 += 32;
154
155 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
156 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
157 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
158 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
159 const __m256i vprod4xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
160 const __m128i vprod4xOPQRSTUV = _mm256_extracti128_si256(vprod4xGHIJKLMNOPQRSTUV, 1);
161 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4xGHIJKLMNOPQRSTUV)));
162 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod4xOPQRSTUV));
163
164 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t));
165
166 _mm256_storeu_si256((__m256i*) b, vacc01234567);
167 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
168 _mm256_storeu_si256((__m256i*) (b + 16), vaccGHIJKLMN);
169 _mm256_storeu_si256((__m256i*) (b + 24), vaccOPQRSTUV);
170 b += 32;
171 }
172
173 if XNN_UNLIKELY(c != 0) {
174 do {
175 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
176 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
177
178
179 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
180 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t))));
181 i0 += 16;
182
183 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
184 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
185 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
186 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
187
188 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
189 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
190 i1 += 16;
191
192 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
193 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
194 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
195 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
196
197 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
198 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
199 i2 += 16;
200
201 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
202 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
203 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
204 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
205
206 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
207 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
208 i3 += 16;
209
210 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
211 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
212 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
213 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
214
215 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
216 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
217 i4 += 16;
218
219 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
220 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
221 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
222 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
223
224 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t));
225
226 _mm256_storeu_si256((__m256i*) b, vacc01234567);
227 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
228 b += 16;
229 c -= 16;
230 } while (c != 0);
231
232 }
233 }
234
235 // Middle pass to process 5 inputs in each iteration.
236 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
237 int32_t* b = buffer;
238
239 const int8_t* i0 = input[0];
240 assert(i0 != NULL);
241 if XNN_UNPREDICTABLE(i0 != zero) {
242 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
243 }
244 const int8_t* i1 = input[1];
245 assert(i1 != NULL);
246 if XNN_UNPREDICTABLE(i1 != zero) {
247 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
248 }
249 const int8_t* i2 = input[2];
250 assert(i2 != NULL);
251 if XNN_UNPREDICTABLE(i2 != zero) {
252 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
253 }
254 const int8_t* i3 = input[3];
255 assert(i3 != NULL);
256 if XNN_UNPREDICTABLE(i3 != zero) {
257 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
258 }
259 const int8_t* i4 = input[4];
260 assert(i4 != NULL);
261 if XNN_UNPREDICTABLE(i4 != zero) {
262 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
263 }
264 input += 5;
265
266 size_t c = round_up_po2(channels, 16);
267
268 for (; c >= 32; c -= 32) {
269 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
270 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
271 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) (b + 16));
272 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) (b + 24));
273
274
275 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
276 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
277 const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16)));
278 const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
279 i0 += 32;
280
281 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
282 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
283 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
284 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
285 const __m256i vprod0xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
286 const __m128i vprod0xOPQRSTUV = _mm256_extracti128_si256(vprod0xGHIJKLMNOPQRSTUV, 1);
287 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0xGHIJKLMNOPQRSTUV)));
288 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV));
289
290 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
291 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
292 const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
293 const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
294 i1 += 32;
295
296 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
297 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
298 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
299 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
300 const __m256i vprod1xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV);
301 const __m128i vprod1xOPQRSTUV = _mm256_extracti128_si256(vprod1xGHIJKLMNOPQRSTUV, 1);
302 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1xGHIJKLMNOPQRSTUV)));
303 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV));
304
305 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
306 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
307 const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
308 const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
309 i2 += 32;
310
311 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
312 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
313 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
314 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
315 const __m256i vprod2xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
316 const __m128i vprod2xOPQRSTUV = _mm256_extracti128_si256(vprod2xGHIJKLMNOPQRSTUV, 1);
317 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2xGHIJKLMNOPQRSTUV)));
318 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod2xOPQRSTUV));
319
320 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
321 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
322 const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
323 const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
324 i3 += 32;
325
326 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
327 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
328 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
329 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
330 const __m256i vprod3xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV);
331 const __m128i vprod3xOPQRSTUV = _mm256_extracti128_si256(vprod3xGHIJKLMNOPQRSTUV, 1);
332 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3xGHIJKLMNOPQRSTUV)));
333 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod3xOPQRSTUV));
334
335 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
336 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
337 const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
338 const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 144 * sizeof(int8_t))));
339 i4 += 32;
340
341 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
342 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
343 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
344 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
345 const __m256i vprod4xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
346 const __m128i vprod4xOPQRSTUV = _mm256_extracti128_si256(vprod4xGHIJKLMNOPQRSTUV, 1);
347 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4xGHIJKLMNOPQRSTUV)));
348 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod4xOPQRSTUV));
349
350 w = (const void*) ((uintptr_t) w + 160 * sizeof(int8_t));
351
352 _mm256_storeu_si256((__m256i*) b, vacc01234567);
353 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
354 _mm256_storeu_si256((__m256i*) (b + 16), vaccGHIJKLMN);
355 _mm256_storeu_si256((__m256i*) (b + 24), vaccOPQRSTUV);
356 b += 32;
357 }
358
359 if XNN_UNLIKELY(c != 0) {
360 do {
361 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
362 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
363
364
365 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
366 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w)));
367 i0 += 16;
368
369 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
370 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
371 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
372 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
373
374 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
375 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
376 i1 += 16;
377
378 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
379 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
380 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
381 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
382
383 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
384 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
385 i2 += 16;
386
387 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
388 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
389 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
390 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
391
392 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
393 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
394 i3 += 16;
395
396 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
397 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
398 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
399 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
400
401 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
402 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
403 i4 += 16;
404
405 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
406 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
407 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
408 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
409
410 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
411
412 _mm256_storeu_si256((__m256i*) b, vacc01234567);
413 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
414 b += 16;
415 c -= 16;
416 } while (c != 0);
417 }
418 }
419
420
421 // Last pass to process up to 5 inputs.
422 {
423 const int32_t* b = buffer;
424
425 const int8_t* i0 = input[0];
426 assert(i0 != NULL);
427 if XNN_UNPREDICTABLE(i0 != zero) {
428 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
429 }
430 const int8_t* i1 = input[1];
431 assert(i1 != NULL);
432 if XNN_UNPREDICTABLE(i1 != zero) {
433 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
434 }
435 const int8_t* i2 = input[2];
436 assert(i2 != NULL);
437 if XNN_UNPREDICTABLE(i2 != zero) {
438 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
439 }
440 const int8_t* i3 = input[3];
441 assert(i3 != NULL);
442 if XNN_UNPREDICTABLE(i3 != zero) {
443 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
444 }
445 const int8_t* i4 = input[4];
446 assert(i4 != NULL);
447 if XNN_UNPREDICTABLE(i4 != zero) {
448 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
449 }
450
451 size_t c = channels;
452
453 for (; c >= 32; c -= 32) {
454 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
455 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
456 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) (b + 16));
457 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) (b + 24));
458 b += 32;
459
460
461 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
462 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
463 const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16)));
464 const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
465 i0 += 32;
466
467 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
468 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
469 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
470 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
471 const __m256i vprod0xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
472 const __m128i vprod0xOPQRSTUV = _mm256_extracti128_si256(vprod0xGHIJKLMNOPQRSTUV, 1);
473 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0xGHIJKLMNOPQRSTUV)));
474 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV));
475
476 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
477 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
478 const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
479 const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
480 i1 += 32;
481
482 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
483 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
484 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
485 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
486 const __m256i vprod1xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV);
487 const __m128i vprod1xOPQRSTUV = _mm256_extracti128_si256(vprod1xGHIJKLMNOPQRSTUV, 1);
488 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1xGHIJKLMNOPQRSTUV)));
489 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV));
490
491 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
492 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
493 const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
494 const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
495 i2 += 32;
496
497 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
498 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
499 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
500 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
501 const __m256i vprod2xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
502 const __m128i vprod2xOPQRSTUV = _mm256_extracti128_si256(vprod2xGHIJKLMNOPQRSTUV, 1);
503 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2xGHIJKLMNOPQRSTUV)));
504 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod2xOPQRSTUV));
505
506 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
507 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
508 const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
509 const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
510 i3 += 32;
511
512 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
513 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
514 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
515 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
516 const __m256i vprod3xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV);
517 const __m128i vprod3xOPQRSTUV = _mm256_extracti128_si256(vprod3xGHIJKLMNOPQRSTUV, 1);
518 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3xGHIJKLMNOPQRSTUV)));
519 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod3xOPQRSTUV));
520
521 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
522 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
523 const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
524 const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 144 * sizeof(int8_t))));
525 i4 += 32;
526
527 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
528 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
529 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
530 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
531 const __m256i vprod4xGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
532 const __m128i vprod4xOPQRSTUV = _mm256_extracti128_si256(vprod4xGHIJKLMNOPQRSTUV, 1);
533 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4xGHIJKLMNOPQRSTUV)));
534 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod4xOPQRSTUV));
535
536 w = (const void*) ((uintptr_t) w + 160 * sizeof(int8_t));
537
538 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
539 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
540 __m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
541 __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
542
543 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
544 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
545 vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscale);
546 vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscale);
547
548 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
549 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
550 vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point);
551 vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point);
552
553 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
554 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
555 vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN);
556 vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV);
557
558 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
559 const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
560
561 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
562 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
563
564 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
565 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
566
567 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
568 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
569 output += 32;
570 }
571
572 if XNN_UNLIKELY(c != 0) {
573 do {
574 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
575 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
576 b += 16;
577
578
579 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
580 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w)));
581 i0 += 16;
582
583 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
584 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
585 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
586 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
587
588 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
589 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
590 i1 += 16;
591
592 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
593 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
594 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
595 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
596
597 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
598 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
599 i2 += 16;
600
601 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
602 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
603 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
604 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
605
606 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
607 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
608 i3 += 16;
609
610 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
611 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
612 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
613 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
614
615 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
616 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
617 i4 += 16;
618
619 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
620 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
621 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
622 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
623
624 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
625 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
626
627 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
628 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
629
630 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
631 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
632
633 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
634 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
635
636 w = (void*) ((uintptr_t) w + 80 * sizeof(int8_t));
637
638 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
639 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point));
640
641 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
642 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
643
644 if XNN_LIKELY(c >= 16) {
645 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
646 output += 16;
647 c -= 16;
648 } else {
649 if (c & 8) {
650 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
651 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
652 output += 8;
653 }
654 if (c & 4) {
655 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
656 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
657 output += 4;
658 }
659 if (c & 2) {
660 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
661 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
662 output += 2;
663 }
664 if (c & 1) {
665 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
666 output += 1;
667 }
668 c = 0;
669 }
670 } while (c != 0);
671 }
672 }
673
674
675 input = (const int8_t**) ((uintptr_t) input + input_stride);
676 output = (int8_t*) ((uintptr_t) output + output_increment);
677 } while (--output_width != 0);
678 }