Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l16c16s16r-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_6f6m7l16c16s16r__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 > 6);
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 6 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 const int8_t* i5 = input[5];
79 assert(i5 != NULL);
80 if XNN_UNPREDICTABLE(i5 != zero) {
81 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
82 }
83 input += 6;
84
85 size_t c = round_up_po2(channels, 16);
86
87 for (; c >= 16; c -= 16) {
88 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
89 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
90
91
92 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
93 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
94 i0 += 16;
95
96 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
97 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
98 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
99 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
100
101 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
102 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
103 i1 += 16;
104
105 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
106 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
107 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
108 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
109
110 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
111 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
112 i2 += 16;
113
114 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
115 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
116 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
117 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
118
119 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
120 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
121 i3 += 16;
122
123 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
124 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
125 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
126 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
127
128 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
129 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
130 i4 += 16;
131
132 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
133 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
134 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
135 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
136
137 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
138 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
139 i5 += 16;
140
141 const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
142 const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1);
143 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF)));
144 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF));
145
146 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t));
147
148 _mm256_storeu_si256((__m256i*) b, vacc01234567);
149 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
150 b += 16;
151 }
152
153 assert(c == 0);
154 }
155
156 // Middle pass to process 6 inputs in each iteration.
157 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
158 int32_t* b = buffer;
159
160 const int8_t* i0 = input[0];
161 assert(i0 != NULL);
162 if XNN_UNPREDICTABLE(i0 != zero) {
163 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
164 }
165 const int8_t* i1 = input[1];
166 assert(i1 != NULL);
167 if XNN_UNPREDICTABLE(i1 != zero) {
168 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
169 }
170 const int8_t* i2 = input[2];
171 assert(i2 != NULL);
172 if XNN_UNPREDICTABLE(i2 != zero) {
173 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
174 }
175 const int8_t* i3 = input[3];
176 assert(i3 != NULL);
177 if XNN_UNPREDICTABLE(i3 != zero) {
178 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
179 }
180 const int8_t* i4 = input[4];
181 assert(i4 != NULL);
182 if XNN_UNPREDICTABLE(i4 != zero) {
183 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
184 }
185 const int8_t* i5 = input[5];
186 assert(i5 != NULL);
187 if XNN_UNPREDICTABLE(i5 != zero) {
188 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
189 }
190 input += 6;
191
192 size_t c = round_up_po2(channels, 16);
193
194 for (; c >= 16; c -= 16) {
195 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
196 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
197
198
199 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
200 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
201 i0 += 16;
202
203 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
204 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
205 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
206 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
207
208 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
209 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
210 i1 += 16;
211
212 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
213 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
214 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
215 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
216
217 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
218 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
219 i2 += 16;
220
221 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
222 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
223 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
224 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
225
226 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
227 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
228 i3 += 16;
229
230 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
231 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
232 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
233 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
234
235 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
236 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
237 i4 += 16;
238
239 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
240 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
241 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
242 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
243
244 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
245 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
246 i5 += 16;
247
248 const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
249 const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1);
250 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF)));
251 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF));
252
253 w = (const void*) ((uintptr_t) w + 96 * sizeof(int8_t));
254
255 _mm256_storeu_si256((__m256i*) b, vacc01234567);
256 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
257 b += 16;
258 }
259
260 assert(c == 0);
261 }
262
263
264 // Last pass to process up to 7 inputs.
265 {
266 const int32_t* b = buffer;
267
268 const int8_t* i0 = input[0];
269 assert(i0 != NULL);
270 if XNN_UNPREDICTABLE(i0 != zero) {
271 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
272 }
273 const int8_t* i1 = input[1];
274 assert(i1 != NULL);
275 if XNN_UNPREDICTABLE(i1 != zero) {
276 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
277 }
278 const int8_t* i2 = input[2];
279 assert(i2 != NULL);
280 if XNN_UNPREDICTABLE(i2 != zero) {
281 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
282 }
283 const int8_t* i3 = input[3];
284 assert(i3 != NULL);
285 if XNN_UNPREDICTABLE(i3 != zero) {
286 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
287 }
288 const int8_t* i4 = input[4];
289 assert(i4 != NULL);
290 if XNN_UNPREDICTABLE(i4 != zero) {
291 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
292 }
293 const int8_t* i5 = input[5];
294 assert(i5 != NULL);
295 if XNN_UNPREDICTABLE(i5 != zero) {
296 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
297 }
298 const int8_t* i6 = input[6];
299 assert(i6 != NULL);
300 if XNN_UNPREDICTABLE(i6 != zero) {
301 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
302 }
303
304 size_t c = channels;
305
306 for (; c >= 16; c -= 16) {
307 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
308 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
309 b += 16;
310
311
312 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
313 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
314 i0 += 16;
315
316 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
317 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
318 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
319 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
320
321 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
322 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
323 i1 += 16;
324
325 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
326 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
327 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
328 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
329
330 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
331 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
332 i2 += 16;
333
334 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
335 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
336 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
337 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
338
339 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
340 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
341 i3 += 16;
342
343 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
344 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
345 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
346 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
347
348 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
349 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
350 i4 += 16;
351
352 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
353 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
354 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
355 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
356
357 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
358 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
359 i5 += 16;
360
361 const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
362 const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1);
363 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF)));
364 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF));
365
366 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
367 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
368 i6 += 16;
369
370 const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
371 const __m128i vprod6x89ABCDEF = _mm256_extracti128_si256(vprod6x0123456789ABCDEF, 1);
372 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod6x0123456789ABCDEF)));
373 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod6x89ABCDEF));
374
375 w = (const void*) ((uintptr_t) w + 112 * sizeof(int8_t));
376
377 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
378 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
379
380 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
381 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
382
383 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
384 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
385
386 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
387 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
388
389 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
390
391 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
392
393 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
394
395 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
396 output += 16;
397 }
398
399 if XNN_UNLIKELY(c != 0) {
400 {
401 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
402 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
403 b += 16;
404
405
406 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
407 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w)));
408
409 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
410 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
411 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
412 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
413
414 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
415 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
416
417 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
418 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
419 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
420 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
421
422 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
423 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
424
425 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
426 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
427 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
428 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
429
430 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
431 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
432
433 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
434 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
435 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
436 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
437
438 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
439 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
440
441 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
442 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
443 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
444 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
445
446 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
447 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
448
449 const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
450 const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1);
451 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF)));
452 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF));
453
454 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
455 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
456
457 const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
458 const __m128i vprod6x89ABCDEF = _mm256_extracti128_si256(vprod6x0123456789ABCDEF, 1);
459 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod6x0123456789ABCDEF)));
460 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod6x89ABCDEF));
461
462 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
463 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
464
465 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
466 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
467
468 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
469 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
470
471 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
472 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
473
474 w = (void*) ((uintptr_t) w + 112 * sizeof(int8_t));
475
476 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
477 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point));
478
479 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
480 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
481
482 if (c & 8) {
483 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
484 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
485 output += 8;
486 }
487 if (c & 4) {
488 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
489 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
490 output += 4;
491 }
492 if (c & 2) {
493 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
494 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
495 output += 2;
496 }
497 if (c & 1) {
498 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
499 output += 1;
500 }
501 }
502 }
503 }
504
505
506 input = (const int8_t**) ((uintptr_t) input + input_stride);
507 output = (int8_t*) ((uintptr_t) output + output_increment);
508 } while (--output_width != 0);
509 }