Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l16c16s16r-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_5f5m5l16c16s16r__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 >= 16; c -= 16) {
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
86
87 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
88 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
89 i0 += 16;
90
91 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
92 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
93 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
94 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
95
96 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
97 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
98 i1 += 16;
99
100 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
101 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
102 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
103 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
104
105 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
106 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
107 i2 += 16;
108
109 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
110 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
111 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
112 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
113
114 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
115 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
116 i3 += 16;
117
118 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
119 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
120 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
121 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
122
123 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
124 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
125 i4 += 16;
126
127 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
128 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
129 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
130 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
131
132 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t));
133
134 _mm256_storeu_si256((__m256i*) b, vacc01234567);
135 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
136 b += 16;
137 }
138
139 assert(c == 0);
140 }
141
142 // Middle pass to process 5 inputs in each iteration.
143 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
144 int32_t* b = buffer;
145
146 const int8_t* i0 = input[0];
147 assert(i0 != NULL);
148 if XNN_UNPREDICTABLE(i0 != zero) {
149 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
150 }
151 const int8_t* i1 = input[1];
152 assert(i1 != NULL);
153 if XNN_UNPREDICTABLE(i1 != zero) {
154 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
155 }
156 const int8_t* i2 = input[2];
157 assert(i2 != NULL);
158 if XNN_UNPREDICTABLE(i2 != zero) {
159 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
160 }
161 const int8_t* i3 = input[3];
162 assert(i3 != NULL);
163 if XNN_UNPREDICTABLE(i3 != zero) {
164 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
165 }
166 const int8_t* i4 = input[4];
167 assert(i4 != NULL);
168 if XNN_UNPREDICTABLE(i4 != zero) {
169 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
170 }
171 input += 5;
172
173 size_t c = round_up_po2(channels, 16);
174
175 for (; c >= 16; c -= 16) {
176 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
177 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
178
179
180 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
181 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
182 i0 += 16;
183
184 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
185 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
186 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
187 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
188
189 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
190 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
191 i1 += 16;
192
193 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
194 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
195 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
196 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
197
198 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
199 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
200 i2 += 16;
201
202 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
203 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
204 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
205 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
206
207 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
208 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
209 i3 += 16;
210
211 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
212 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
213 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
214 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
215
216 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
217 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
218 i4 += 16;
219
220 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
221 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
222 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
223 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
224
225 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
226
227 _mm256_storeu_si256((__m256i*) b, vacc01234567);
228 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
229 b += 16;
230 }
231
232 assert(c == 0);
233 }
234
235
236 // Last pass to process up to 5 inputs.
237 {
238 const int32_t* b = buffer;
239
240 const int8_t* i0 = input[0];
241 assert(i0 != NULL);
242 if XNN_UNPREDICTABLE(i0 != zero) {
243 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
244 }
245 const int8_t* i1 = input[1];
246 assert(i1 != NULL);
247 if XNN_UNPREDICTABLE(i1 != zero) {
248 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
249 }
250 const int8_t* i2 = input[2];
251 assert(i2 != NULL);
252 if XNN_UNPREDICTABLE(i2 != zero) {
253 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
254 }
255 const int8_t* i3 = input[3];
256 assert(i3 != NULL);
257 if XNN_UNPREDICTABLE(i3 != zero) {
258 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
259 }
260 const int8_t* i4 = input[4];
261 assert(i4 != NULL);
262 if XNN_UNPREDICTABLE(i4 != zero) {
263 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
264 }
265
266 size_t c = channels;
267
268 for (; c >= 16; c -= 16) {
269 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
270 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
271 b += 16;
272
273
274 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
275 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
276 i0 += 16;
277
278 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
279 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
280 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
281 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
282
283 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
284 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
285 i1 += 16;
286
287 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
288 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
289 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
290 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
291
292 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
293 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
294 i2 += 16;
295
296 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
297 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
298 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
299 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
300
301 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
302 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
303 i3 += 16;
304
305 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
306 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
307 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
308 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
309
310 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
311 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
312 i4 += 16;
313
314 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
315 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
316 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
317 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
318
319 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
320
321 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
322 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
323
324 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
325 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
326
327 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
328 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
329
330 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
331 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
332
333 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
334
335 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
336
337 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
338
339 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
340 output += 16;
341 }
342
343 if XNN_UNLIKELY(c != 0) {
344 {
345 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
346 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
347 b += 16;
348
349
350 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
351 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w)));
352
353 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
354 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
355 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
356 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
357
358 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
359 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
360
361 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
362 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
363 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
364 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
365
366 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
367 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
368
369 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
370 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
371 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
372 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
373
374 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
375 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
376
377 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
378 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
379 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
380 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
381
382 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
383 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
384
385 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
386 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
387 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
388 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
389
390 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
391 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
392
393 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
394 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
395
396 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
397 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
398
399 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
400 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
401
402 w = (void*) ((uintptr_t) w + 80 * sizeof(int8_t));
403
404 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
405 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point));
406
407 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
408 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
409
410 if (c & 8) {
411 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
412 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
413 output += 8;
414 }
415 if (c & 4) {
416 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
417 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
418 output += 4;
419 }
420 if (c & 2) {
421 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
422 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
423 output += 2;
424 }
425 if (c & 1) {
426 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
427 output += 1;
428 }
429 }
430 }
431 }
432
433
434 input = (const int8_t**) ((uintptr_t) input + input_stride);
435 output = (int8_t*) ((uintptr_t) output + output_increment);
436 } while (--output_width != 0);
437 }