Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l16c16s16r-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_6f6m7l16c16s16r__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 > 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 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
92 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
93
94
95 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
96 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
97 i0 += 16;
98
99 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
100 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
101
102 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
103 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
104
105 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
106 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
107 i1 += 16;
108
109 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
110 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
111
112 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
113 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
114
115 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
116 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
117 i2 += 16;
118
119 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
120 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
121
122 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
123 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
124
125 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
126 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
127 i3 += 16;
128
129 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
130 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
131
132 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
133 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
134
135 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
136 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
137 i4 += 16;
138
139 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
140 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
141
142 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
143 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
144
145 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
146 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
147 i5 += 16;
148
149 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
150 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
151
152 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
153 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
154
155 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t));
156
157 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
158 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
159
160 _mm256_storeu_si256((__m256i*) b, vacc01234567);
161 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
162 b += 16;
163 }
164
165 assert(c == 0);
166 }
167
168 // Middle pass to process 6 inputs in each iteration.
169 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
170 int32_t* b = buffer;
171
172 const int8_t* i0 = input[0];
173 assert(i0 != NULL);
174 if XNN_UNPREDICTABLE(i0 != zero) {
175 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
176 }
177 const int8_t* i1 = input[1];
178 assert(i1 != NULL);
179 if XNN_UNPREDICTABLE(i1 != zero) {
180 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
181 }
182 const int8_t* i2 = input[2];
183 assert(i2 != NULL);
184 if XNN_UNPREDICTABLE(i2 != zero) {
185 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
186 }
187 const int8_t* i3 = input[3];
188 assert(i3 != NULL);
189 if XNN_UNPREDICTABLE(i3 != zero) {
190 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
191 }
192 const int8_t* i4 = input[4];
193 assert(i4 != NULL);
194 if XNN_UNPREDICTABLE(i4 != zero) {
195 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
196 }
197 const int8_t* i5 = input[5];
198 assert(i5 != NULL);
199 if XNN_UNPREDICTABLE(i5 != zero) {
200 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
201 }
202 input += 6;
203
204 size_t c = round_up_po2(channels, 16);
205
206 for (; c >= 16; c -= 16) {
207 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
208 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
209
210 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
211 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
212
213
214 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
215 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
216 i0 += 16;
217
218 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
219 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
220
221 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
222 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
223
224 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
225 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
226 i1 += 16;
227
228 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
229 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
230
231 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
232 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
233
234 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
235 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
236 i2 += 16;
237
238 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
239 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
240
241 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
242 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
243
244 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
245 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
246 i3 += 16;
247
248 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
249 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
250
251 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
252 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
253
254 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
255 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
256 i4 += 16;
257
258 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
259 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
260
261 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
262 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
263
264 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
265 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
266 i5 += 16;
267
268 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
269 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
270
271 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
272 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
273
274 w = (const void*) ((uintptr_t) w + 96 * sizeof(int8_t));
275
276 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
277 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
278
279 _mm256_storeu_si256((__m256i*) b, vacc01234567);
280 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
281 b += 16;
282 }
283
284 assert(c == 0);
285 }
286
287
288 // Last pass to process up to 7 inputs.
289 {
290 const int32_t* b = buffer;
291
292 const int8_t* i0 = input[0];
293 assert(i0 != NULL);
294 if XNN_UNPREDICTABLE(i0 != zero) {
295 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
296 }
297 const int8_t* i1 = input[1];
298 assert(i1 != NULL);
299 if XNN_UNPREDICTABLE(i1 != zero) {
300 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
301 }
302 const int8_t* i2 = input[2];
303 assert(i2 != NULL);
304 if XNN_UNPREDICTABLE(i2 != zero) {
305 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
306 }
307 const int8_t* i3 = input[3];
308 assert(i3 != NULL);
309 if XNN_UNPREDICTABLE(i3 != zero) {
310 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
311 }
312 const int8_t* i4 = input[4];
313 assert(i4 != NULL);
314 if XNN_UNPREDICTABLE(i4 != zero) {
315 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
316 }
317 const int8_t* i5 = input[5];
318 assert(i5 != NULL);
319 if XNN_UNPREDICTABLE(i5 != zero) {
320 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
321 }
322 const int8_t* i6 = input[6];
323 assert(i6 != NULL);
324 if XNN_UNPREDICTABLE(i6 != zero) {
325 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
326 }
327
328 size_t c = channels;
329
330 for (; c >= 16; c -= 16) {
331 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
332 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
333 b += 16;
334
335 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
336 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
337
338
339 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
340 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
341 i0 += 16;
342
343 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
344 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
345
346 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
347 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
348
349 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
350 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
351 i1 += 16;
352
353 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
354 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
355
356 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
357 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
358
359 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
360 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
361 i2 += 16;
362
363 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
364 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
365
366 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
367 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
368
369 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
370 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
371 i3 += 16;
372
373 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
374 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
375
376 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
377 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
378
379 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
380 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
381 i4 += 16;
382
383 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
384 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
385
386 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
387 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
388
389 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
390 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
391 i5 += 16;
392
393 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
394 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
395
396 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
397 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
398
399 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
400 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
401 i6 += 16;
402
403 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
404 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
405
406 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
407 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
408
409 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
410 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
411
412 w = (const void*) ((uintptr_t) w + 112 * sizeof(int8_t));
413
414 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
415 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
416
417 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
418 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
419
420 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
421 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
422
423 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
424 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
425
426 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
427
428 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
429
430 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
431
432 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
433 output += 16;
434 }
435
436 if XNN_UNLIKELY(c != 0) {
437 {
438 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
439 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
440 b += 16;
441
442 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
443 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
444
445
446 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
447 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w)));
448
449 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
450 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
451
452 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
453 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
454
455 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
456 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
457
458 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
459 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
460
461 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
462 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
463
464 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
465 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
466
467 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
468 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
469
470 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
471 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
472
473 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
474 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
475
476 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
477 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
478
479 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
480 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
481
482 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
483 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
484
485 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
486 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
487
488 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
489 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
490
491 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
492 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
493
494 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
495 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
496
497 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
498 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
499
500 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
501 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
502
503 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
504 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
505
506 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
507 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
508
509 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
510 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
511
512 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
513 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
514
515 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
516 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
517
518 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
519 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
520
521 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
522 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
523
524 w = (void*) ((uintptr_t) w + 112 * sizeof(int8_t));
525
526 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
527 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point));
528
529 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
530 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
531
532 if (c & 8) {
533 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
534 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
535 output += 8;
536 }
537 if (c & 4) {
538 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
539 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
540 output += 4;
541 }
542 if (c & 2) {
543 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
544 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
545 output += 2;
546 }
547 if (c & 1) {
548 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
549 output += 1;
550 }
551 }
552 }
553 }
554
555
556 input = (const int8_t**) ((uintptr_t) input + input_stride);
557 output = (int8_t*) ((uintptr_t) output + output_increment);
558 } while (--output_width != 0);
559 }