Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l16c16s16r-minmax-fp32-avx2-mul16-vpunpck.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-avx2-mul16-vpunpck.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_vpunpck(
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 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
87 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
88
89
90 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
91 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
92 i0 += 16;
93
94 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
95 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
96
97 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
98 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
99
100 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
101 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
102 i1 += 16;
103
104 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
105 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
106
107 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
108 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
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 vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
115 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
116
117 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
118 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
119
120 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
121 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
122 i3 += 16;
123
124 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
125 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
126
127 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
128 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
129
130 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
131 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
132 i4 += 16;
133
134 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
135 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
136
137 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
138 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
139
140 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t));
141
142 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
143 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
144
145 _mm256_storeu_si256((__m256i*) b, vacc01234567);
146 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
147 b += 16;
148 }
149
150 assert(c == 0);
151 }
152
153 // Middle pass to process 5 inputs in each iteration.
154 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
155 int32_t* b = buffer;
156
157 const int8_t* i0 = input[0];
158 assert(i0 != NULL);
159 if XNN_UNPREDICTABLE(i0 != zero) {
160 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
161 }
162 const int8_t* i1 = input[1];
163 assert(i1 != NULL);
164 if XNN_UNPREDICTABLE(i1 != zero) {
165 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
166 }
167 const int8_t* i2 = input[2];
168 assert(i2 != NULL);
169 if XNN_UNPREDICTABLE(i2 != zero) {
170 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
171 }
172 const int8_t* i3 = input[3];
173 assert(i3 != NULL);
174 if XNN_UNPREDICTABLE(i3 != zero) {
175 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
176 }
177 const int8_t* i4 = input[4];
178 assert(i4 != NULL);
179 if XNN_UNPREDICTABLE(i4 != zero) {
180 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
181 }
182 input += 5;
183
184 size_t c = round_up_po2(channels, 16);
185
186 for (; c >= 16; c -= 16) {
187 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
188 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
189
190 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
191 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
192
193
194 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
195 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
196 i0 += 16;
197
198 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
199 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
200
201 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
202 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
203
204 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
205 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
206 i1 += 16;
207
208 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
209 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
210
211 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
212 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
213
214 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
215 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
216 i2 += 16;
217
218 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
219 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
220
221 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
222 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
223
224 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
225 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
226 i3 += 16;
227
228 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
229 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
230
231 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
232 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
233
234 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
235 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
236 i4 += 16;
237
238 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
239 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
240
241 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
242 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
243
244 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
245
246 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
247 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
248
249 _mm256_storeu_si256((__m256i*) b, vacc01234567);
250 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
251 b += 16;
252 }
253
254 assert(c == 0);
255 }
256
257
258 // Last pass to process up to 5 inputs.
259 {
260 const int32_t* b = buffer;
261
262 const int8_t* i0 = input[0];
263 assert(i0 != NULL);
264 if XNN_UNPREDICTABLE(i0 != zero) {
265 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
266 }
267 const int8_t* i1 = input[1];
268 assert(i1 != NULL);
269 if XNN_UNPREDICTABLE(i1 != zero) {
270 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
271 }
272 const int8_t* i2 = input[2];
273 assert(i2 != NULL);
274 if XNN_UNPREDICTABLE(i2 != zero) {
275 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
276 }
277 const int8_t* i3 = input[3];
278 assert(i3 != NULL);
279 if XNN_UNPREDICTABLE(i3 != zero) {
280 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
281 }
282 const int8_t* i4 = input[4];
283 assert(i4 != NULL);
284 if XNN_UNPREDICTABLE(i4 != zero) {
285 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
286 }
287
288 size_t c = channels;
289
290 for (; c >= 16; c -= 16) {
291 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
292 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
293 b += 16;
294
295 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
296 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
297
298
299 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
300 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
301 i0 += 16;
302
303 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
304 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
305
306 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
307 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
308
309 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
310 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
311 i1 += 16;
312
313 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
314 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
315
316 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
317 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
318
319 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
320 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
321 i2 += 16;
322
323 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
324 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
325
326 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
327 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
328
329 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
330 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
331 i3 += 16;
332
333 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
334 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
335
336 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
337 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
338
339 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
340 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
341 i4 += 16;
342
343 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
344 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
345
346 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
347 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
348
349 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
350 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
351
352 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
353
354 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
355 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
356
357 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
358 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
359
360 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
361 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
362
363 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
364 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
365
366 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
367
368 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
369
370 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
371
372 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
373 output += 16;
374 }
375
376 if XNN_UNLIKELY(c != 0) {
377 {
378 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
379 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
380 b += 16;
381
382 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
383 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
384
385
386 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
387 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w)));
388
389 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
390 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
391
392 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
393 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
394
395 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
396 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
397
398 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
399 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
400
401 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
402 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
403
404 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
405 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
406
407 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
408 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
409
410 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
411 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
412
413 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
414 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
415
416 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
417 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
418
419 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
420 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
421
422 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
423 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
424
425 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
426 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
427
428 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
429 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
430
431 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
432 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
433
434 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
435 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
436
437 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
438 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
439
440 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
441 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
442
443 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
444 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
445
446 w = (void*) ((uintptr_t) w + 80 * sizeof(int8_t));
447
448 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
449 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point));
450
451 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
452 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
453
454 if (c & 8) {
455 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
456 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
457 output += 8;
458 }
459 if (c & 4) {
460 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
461 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
462 output += 4;
463 }
464 if (c & 2) {
465 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
466 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
467 output += 2;
468 }
469 if (c & 1) {
470 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
471 output += 1;
472 }
473 }
474 }
475 }
476
477
478 input = (const int8_t**) ((uintptr_t) input + input_stride);
479 output = (int8_t*) ((uintptr_t) output + output_increment);
480 } while (--output_width != 0);
481 }