Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l16c16s16r-minmax-fp32-avx2-mul16-add16-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_8f8m9l16c16s16r__avx2_mul16_add16_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 > 8);
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 8 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 const int8_t* i6 = input[6];
84 assert(i6 != NULL);
85 if XNN_UNPREDICTABLE(i6 != zero) {
86 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
87 }
88 const int8_t* i7 = input[7];
89 assert(i7 != NULL);
90 if XNN_UNPREDICTABLE(i7 != zero) {
91 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
92 }
93 input += 8;
94
95 size_t c = round_up_po2(channels, 16);
96
97 for (; c >= 16; c -= 16) {
98 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
99 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
100
101 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
102 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
103
104
105 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
106 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
107 i0 += 16;
108
109 __m256i vacc0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
110
111
112 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
113 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
114 i1 += 16;
115
116 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
117
118 __m256i vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
119 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
120 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
121
122 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
123 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
124 i2 += 16;
125
126 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
127
128
129 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
130 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
131 i3 += 16;
132
133 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
134
135 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
136 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
137 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
138
139 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
140 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
141 i4 += 16;
142
143 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
144
145
146 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
147 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
148 i5 += 16;
149
150 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
151
152 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
153 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
154 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
155
156 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
157 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
158 i6 += 16;
159
160 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
161
162
163 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
164 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
165 i7 += 16;
166
167 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
168
169 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
170 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
171 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
172
173 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t));
174
175 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
176 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
177
178 _mm256_storeu_si256((__m256i*) b, vacc01234567);
179 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
180 b += 16;
181 }
182
183 assert(c == 0);
184 }
185
186 // Middle pass to process 8 inputs in each iteration.
187 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
188 int32_t* b = buffer;
189
190 const int8_t* i0 = input[0];
191 assert(i0 != NULL);
192 if XNN_UNPREDICTABLE(i0 != zero) {
193 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
194 }
195 const int8_t* i1 = input[1];
196 assert(i1 != NULL);
197 if XNN_UNPREDICTABLE(i1 != zero) {
198 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
199 }
200 const int8_t* i2 = input[2];
201 assert(i2 != NULL);
202 if XNN_UNPREDICTABLE(i2 != zero) {
203 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
204 }
205 const int8_t* i3 = input[3];
206 assert(i3 != NULL);
207 if XNN_UNPREDICTABLE(i3 != zero) {
208 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
209 }
210 const int8_t* i4 = input[4];
211 assert(i4 != NULL);
212 if XNN_UNPREDICTABLE(i4 != zero) {
213 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
214 }
215 const int8_t* i5 = input[5];
216 assert(i5 != NULL);
217 if XNN_UNPREDICTABLE(i5 != zero) {
218 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
219 }
220 const int8_t* i6 = input[6];
221 assert(i6 != NULL);
222 if XNN_UNPREDICTABLE(i6 != zero) {
223 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
224 }
225 const int8_t* i7 = input[7];
226 assert(i7 != NULL);
227 if XNN_UNPREDICTABLE(i7 != zero) {
228 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
229 }
230 input += 8;
231
232 size_t c = round_up_po2(channels, 16);
233
234 for (; c >= 16; c -= 16) {
235 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
236 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
237
238 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
239 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
240
241
242 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
243 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
244 i0 += 16;
245
246 __m256i vacc0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
247
248
249 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
250 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
251 i1 += 16;
252
253 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
254
255 __m256i vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
256 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
257 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
258
259 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
260 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
261 i2 += 16;
262
263 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
264
265
266 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
267 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
268 i3 += 16;
269
270 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
271
272 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
273 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
274 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
275
276 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
277 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
278 i4 += 16;
279
280 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
281
282
283 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
284 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
285 i5 += 16;
286
287 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
288
289 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
290 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
291 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
292
293 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
294 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
295 i6 += 16;
296
297 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
298
299
300 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
301 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
302 i7 += 16;
303
304 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
305
306 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
307 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
308 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
309
310 w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
311
312 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
313 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
314
315 _mm256_storeu_si256((__m256i*) b, vacc01234567);
316 _mm256_storeu_si256((__m256i*) (b + 8), vacc89ABCDEF);
317 b += 16;
318 }
319
320 assert(c == 0);
321 }
322
323
324 // Last pass to process up to 9 inputs.
325 {
326 const int32_t* b = buffer;
327
328 const int8_t* i0 = input[0];
329 assert(i0 != NULL);
330 if XNN_UNPREDICTABLE(i0 != zero) {
331 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
332 }
333 const int8_t* i1 = input[1];
334 assert(i1 != NULL);
335 if XNN_UNPREDICTABLE(i1 != zero) {
336 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
337 }
338 const int8_t* i2 = input[2];
339 assert(i2 != NULL);
340 if XNN_UNPREDICTABLE(i2 != zero) {
341 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
342 }
343 const int8_t* i3 = input[3];
344 assert(i3 != NULL);
345 if XNN_UNPREDICTABLE(i3 != zero) {
346 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
347 }
348 const int8_t* i4 = input[4];
349 assert(i4 != NULL);
350 if XNN_UNPREDICTABLE(i4 != zero) {
351 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
352 }
353 const int8_t* i5 = input[5];
354 assert(i5 != NULL);
355 if XNN_UNPREDICTABLE(i5 != zero) {
356 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
357 }
358 const int8_t* i6 = input[6];
359 assert(i6 != NULL);
360 if XNN_UNPREDICTABLE(i6 != zero) {
361 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
362 }
363 const int8_t* i7 = input[7];
364 assert(i7 != NULL);
365 if XNN_UNPREDICTABLE(i7 != zero) {
366 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
367 }
368 const int8_t* i8 = input[8];
369 assert(i8 != NULL);
370 if XNN_UNPREDICTABLE(i8 != zero) {
371 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
372 }
373
374 size_t c = channels;
375
376 for (; c >= 16; c -= 16) {
377 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
378 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
379 b += 16;
380
381 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
382 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
383
384
385 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
386 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
387 i0 += 16;
388
389 __m256i vacc0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
390
391
392 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
393 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
394 i1 += 16;
395
396 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
397
398 __m256i vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
399 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
400 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
401
402 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
403 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
404 i2 += 16;
405
406 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
407
408
409 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
410 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
411 i3 += 16;
412
413 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
414
415 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
416 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
417 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
418
419 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
420 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
421 i4 += 16;
422
423 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
424
425
426 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
427 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
428 i5 += 16;
429
430 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
431
432 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
433 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
434 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
435
436 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
437 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
438 i6 += 16;
439
440 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
441
442
443 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
444 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
445 i7 += 16;
446
447 vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
448
449 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
450 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
451 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
452
453 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
454 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
455 i8 += 16;
456
457 vacc0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
458
459 vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
460 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
461 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
462
463 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
464 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
465
466 w = (const void*) ((uintptr_t) w + 144 * sizeof(int8_t));
467
468 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
469 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
470
471 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
472 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
473
474 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
475 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
476
477 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
478 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
479
480 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
481
482 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
483
484 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
485
486 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
487 output += 16;
488 }
489
490 if XNN_UNLIKELY(c != 0) {
491 {
492 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
493 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) (b + 8));
494 b += 16;
495
496 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
497 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
498
499
500 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
501 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w)));
502
503 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
504 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
505
506 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
507 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
508
509 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
510 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
511
512 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
513 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
514
515 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
516 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
517
518 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
519 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
520
521 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
522 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
523
524 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
525 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
526
527 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
528 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
529
530 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
531 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
532
533 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
534 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
535
536 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
537 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
538
539 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
540 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
541
542 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
543 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
544
545 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
546 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
547
548 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
549 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
550
551 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
552 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
553
554 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
555 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
556
557 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
558 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
559
560 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
561 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
562
563 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
564 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
565
566 const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
567 const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
568
569 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
570 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
571
572 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
573 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
574
575 const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
576 const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
577
578 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
579 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
580
581 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
582 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
583
584 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
585 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
586
587 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
588 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
589
590 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
591 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
592
593 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
594 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
595
596 w = (void*) ((uintptr_t) w + 144 * sizeof(int8_t));
597
598 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
599 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), _mm256_castsi256_si128(voutput_zero_point));
600
601 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
602 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
603
604 if (c & 8) {
605 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
606 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
607 output += 8;
608 }
609 if (c & 4) {
610 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
611 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
612 output += 4;
613 }
614 if (c & 2) {
615 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
616 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
617 output += 2;
618 }
619 if (c & 1) {
620 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
621 output += 1;
622 }
623 }
624 }
625 }
626
627
628 input = (const int8_t**) ((uintptr_t) input + input_stride);
629 output = (int8_t*) ((uintptr_t) output + output_increment);
630 } while (--output_width != 0);
631 }