Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l16c16s1r-minmax-fp32-avx512skx-mul32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-avx512skx-mul32.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/intrinsics-polyfill.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_8f8m9l16c16s1r__avx512skx_mul32(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 size_t kernel_size,
29 int32_t* buffer,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(channels != 0);
33 assert(output_width != 0);
34 assert(kernel_size > 8);
35
36
37 const __m512 vscale = _mm512_set1_ps(params->fp32_scalar.scale);
38 // XNN_FORCE_REALIZATION(vscale);
39 const __m512 voutput_max_less_zero_point = _mm512_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((int16_t) 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 const int8_t* i0 = input[0];
53 assert(i0 != NULL);
54 if XNN_UNPREDICTABLE(i0 != zero) {
55 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
56 }
57 const int8_t* i1 = input[1];
58 assert(i1 != NULL);
59 if XNN_UNPREDICTABLE(i1 != zero) {
60 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
61 }
62 const int8_t* i2 = input[2];
63 assert(i2 != NULL);
64 if XNN_UNPREDICTABLE(i2 != zero) {
65 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
66 }
67 const int8_t* i3 = input[3];
68 assert(i3 != NULL);
69 if XNN_UNPREDICTABLE(i3 != zero) {
70 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
71 }
72 const int8_t* i4 = input[4];
73 assert(i4 != NULL);
74 if XNN_UNPREDICTABLE(i4 != zero) {
75 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
76 }
77 const int8_t* i5 = input[5];
78 assert(i5 != NULL);
79 if XNN_UNPREDICTABLE(i5 != zero) {
80 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
81 }
82 const int8_t* i6 = input[6];
83 assert(i6 != NULL);
84 if XNN_UNPREDICTABLE(i6 != zero) {
85 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
86 }
87 const int8_t* i7 = input[7];
88 assert(i7 != NULL);
89 if XNN_UNPREDICTABLE(i7 != zero) {
90 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
91 }
92 input += 8;
93
94 size_t c = channels;
95
96 for (; c >= 16; c -= 16) {
97 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
98
99 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
100 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
101 i0 += 16;
102
103 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
104
105 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
106 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
107 i1 += 16;
108
109 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
110
111 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
112 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
113 i2 += 16;
114
115 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
116
117 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
118 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
119 i3 += 16;
120
121 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
122
123 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
124 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
125 i4 += 16;
126
127 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
128
129 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
130 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
131 i5 += 16;
132
133 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
134
135 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
136 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
137 i6 += 16;
138
139 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
140
141 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
142 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
143 i7 += 16;
144
145 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
146
147 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t));
148
149 _mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF);
150 b += 16;
151 }
152
153
154 if (c != 0) {
155 assert(c >= 1);
156 assert(c <= 15);
157 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
158
159 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
160 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
161 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
162
163 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
164 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
165 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
166
167 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
168 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
169 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
170
171 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
172 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
173 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
174
175 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
176 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
177 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
178
179 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
180 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
181 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
182
183 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
184 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
185 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
186
187 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
188 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
189 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
190
191 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t));
192
193 _mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF);
194 }
195 }
196
197 // Middle pass to process 8 inputs in each iteration.
198 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
199 int32_t* b = buffer;
200 const int8_t* i0 = input[0];
201 assert(i0 != NULL);
202 if XNN_UNPREDICTABLE(i0 != zero) {
203 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
204 }
205 const int8_t* i1 = input[1];
206 assert(i1 != NULL);
207 if XNN_UNPREDICTABLE(i1 != zero) {
208 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
209 }
210 const int8_t* i2 = input[2];
211 assert(i2 != NULL);
212 if XNN_UNPREDICTABLE(i2 != zero) {
213 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
214 }
215 const int8_t* i3 = input[3];
216 assert(i3 != NULL);
217 if XNN_UNPREDICTABLE(i3 != zero) {
218 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
219 }
220 const int8_t* i4 = input[4];
221 assert(i4 != NULL);
222 if XNN_UNPREDICTABLE(i4 != zero) {
223 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
224 }
225 const int8_t* i5 = input[5];
226 assert(i5 != NULL);
227 if XNN_UNPREDICTABLE(i5 != zero) {
228 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
229 }
230 const int8_t* i6 = input[6];
231 assert(i6 != NULL);
232 if XNN_UNPREDICTABLE(i6 != zero) {
233 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
234 }
235 const int8_t* i7 = input[7];
236 assert(i7 != NULL);
237 if XNN_UNPREDICTABLE(i7 != zero) {
238 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
239 }
240 input += 8;
241
242 size_t c = channels;
243
244 for (; c >= 16; c -= 16) {
245 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b);
246
247 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
248 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
249 i0 += 16;
250
251 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
252
253 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
254 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
255 i1 += 16;
256
257 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
258
259 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
260 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
261 i2 += 16;
262
263 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
264
265 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
266 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
267 i3 += 16;
268
269 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
270
271 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
272 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
273 i4 += 16;
274
275 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
276
277 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
278 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
279 i5 += 16;
280
281 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
282
283 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
284 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
285 i6 += 16;
286
287 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
288
289 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
290 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
291 i7 += 16;
292
293 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
294
295 w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
296
297 _mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF);
298 b += 16;
299 }
300
301
302 if (c != 0) {
303 assert(c >= 1);
304 assert(c <= 15);
305 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b);
306
307 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
308 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
309 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
310
311 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
312 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
313 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
314
315 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
316 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
317 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
318
319 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
320 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
321 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
322
323 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
324 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
325 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
326
327 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
328 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
329 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
330
331 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
332 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
333 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
334
335 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
336 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
337 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
338
339 w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
340
341 _mm512_storeu_si512((__m512i*) b, vacc0123456789ABCDEF);
342 }
343 }
344
345 // Last pass to process up to 9 inputs.
346 {
347 const int32_t* b = buffer;
348 const int8_t* i0 = input[0];
349 assert(i0 != NULL);
350 if XNN_UNPREDICTABLE(i0 != zero) {
351 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
352 }
353 const int8_t* i1 = input[1];
354 assert(i1 != NULL);
355 if XNN_UNPREDICTABLE(i1 != zero) {
356 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
357 }
358 const int8_t* i2 = input[2];
359 assert(i2 != NULL);
360 if XNN_UNPREDICTABLE(i2 != zero) {
361 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
362 }
363 const int8_t* i3 = input[3];
364 assert(i3 != NULL);
365 if XNN_UNPREDICTABLE(i3 != zero) {
366 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
367 }
368 const int8_t* i4 = input[4];
369 assert(i4 != NULL);
370 if XNN_UNPREDICTABLE(i4 != zero) {
371 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
372 }
373 const int8_t* i5 = input[5];
374 assert(i5 != NULL);
375 if XNN_UNPREDICTABLE(i5 != zero) {
376 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
377 }
378 const int8_t* i6 = input[6];
379 assert(i6 != NULL);
380 if XNN_UNPREDICTABLE(i6 != zero) {
381 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
382 }
383 const int8_t* i7 = input[7];
384 assert(i7 != NULL);
385 if XNN_UNPREDICTABLE(i7 != zero) {
386 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
387 }
388 const int8_t* i8 = input[8];
389 assert(i8 != NULL);
390 if XNN_UNPREDICTABLE(i8 != zero) {
391 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
392 }
393
394 size_t c = channels;
395
396 for (; c >= 16; c -= 16) {
397 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b);
398 b += 16;
399
400 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
401 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
402 i0 += 16;
403
404 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
405
406 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
407 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
408 i1 += 16;
409
410 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
411
412 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
413 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
414 i2 += 16;
415
416 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
417
418 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
419 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
420 i3 += 16;
421
422 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
423
424 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
425 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
426 i4 += 16;
427
428 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
429
430 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
431 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
432 i5 += 16;
433
434 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
435
436 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
437 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
438 i6 += 16;
439
440 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
441
442 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
443 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
444 i7 += 16;
445
446 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
447
448 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
449 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
450 i8 += 16;
451
452 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
453
454 w = (const void*) ((uintptr_t) w + 144 * sizeof(int8_t));
455
456 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
457
458 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
459
460 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
461
462 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
463
464 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
465
466 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
467 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
468 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
469
470 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
471
472 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
473 output += 16;
474 }
475
476 if XNN_UNLIKELY(c != 0) {
477 // Prepare mask for valid 8-bit elements (depends on nc).
478 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
479 {
480 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(b);
481 b += 16;
482
483 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
484 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
485
486 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
487
488 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
489 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
490
491 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
492
493 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
494 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
495
496 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
497
498 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
499 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
500
501 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
502
503 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
504 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
505
506 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
507
508 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
509 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t))));
510
511 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
512
513 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
514 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
515
516 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
517
518 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
519 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t))));
520
521 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
522
523 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
524 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t))));
525
526 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
527
528 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
529 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
530 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
531 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
532
533
534 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
535
536 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
537 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
538 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
539 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
540
541
542 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
543 output = (int8_t*) ((uintptr_t) output + c);
544 }
545 }
546 }
547
548 input = (const int8_t**) ((uintptr_t) input + input_stride);
549 output = (int8_t*) ((uintptr_t) output + output_increment);
550 } while (--output_width != 0);
551 }