Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p16c-minmax-fp32-avx512skx-mul32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx512skx-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_25p16c__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 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m512 vscale = _mm512_set1_ps(params->fp32_scalar.scale);
34 XNN_FORCE_REALIZATION(vscale);
35 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);
36 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
37 const __m256i voutput_zero_point = _mm256_set1_epi16((int16_t) params->fp32_scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
39 // XNN_FORCE_REALIZATION(voutput_zero_point);
40 // XNN_FORCE_REALIZATION(voutput_min);
41
42 do {
43 const int8_t* i0 = input[0];
44 assert(i0 != NULL);
45 if XNN_UNPREDICTABLE(i0 != zero) {
46 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
47 }
48 const int8_t* i1 = input[1];
49 assert(i1 != NULL);
50 if XNN_UNPREDICTABLE(i1 != zero) {
51 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
52 }
53 const int8_t* i2 = input[2];
54 assert(i2 != NULL);
55 if XNN_UNPREDICTABLE(i2 != zero) {
56 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
57 }
58 const int8_t* i3 = input[3];
59 assert(i3 != NULL);
60 if XNN_UNPREDICTABLE(i3 != zero) {
61 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
62 }
63 const int8_t* i4 = input[4];
64 assert(i4 != NULL);
65 if XNN_UNPREDICTABLE(i4 != zero) {
66 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
67 }
68 const int8_t* i5 = input[5];
69 assert(i5 != NULL);
70 if XNN_UNPREDICTABLE(i5 != zero) {
71 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
72 }
73 const int8_t* i6 = input[6];
74 assert(i6 != NULL);
75 if XNN_UNPREDICTABLE(i6 != zero) {
76 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
77 }
78 const int8_t* i7 = input[7];
79 assert(i7 != NULL);
80 if XNN_UNPREDICTABLE(i7 != zero) {
81 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
82 }
83 const int8_t* i8 = input[8];
84 assert(i8 != NULL);
85 if XNN_UNPREDICTABLE(i8 != zero) {
86 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
87 }
88 const int8_t* i9 = input[9];
89 assert(i9 != NULL);
90 if XNN_UNPREDICTABLE(i9 != zero) {
91 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
92 }
93 const int8_t* i10 = input[10];
94 assert(i10 != NULL);
95 if XNN_UNPREDICTABLE(i10 != zero) {
96 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
97 }
98 const int8_t* i11 = input[11];
99 assert(i11 != NULL);
100 if XNN_UNPREDICTABLE(i11 != zero) {
101 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
102 }
103 const int8_t* i12 = input[12];
104 assert(i12 != NULL);
105 if XNN_UNPREDICTABLE(i12 != zero) {
106 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
107 }
108 const int8_t* i13 = input[13];
109 assert(i13 != NULL);
110 if XNN_UNPREDICTABLE(i13 != zero) {
111 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
112 }
113 const int8_t* i14 = input[14];
114 assert(i14 != NULL);
115 if XNN_UNPREDICTABLE(i14 != zero) {
116 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
117 }
118 const int8_t* i15 = input[15];
119 assert(i15 != NULL);
120 if XNN_UNPREDICTABLE(i15 != zero) {
121 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
122 }
123 const int8_t* i16 = input[16];
124 assert(i16 != NULL);
125 if XNN_UNPREDICTABLE(i16 != zero) {
126 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
127 }
128 const int8_t* i17 = input[17];
129 assert(i17 != NULL);
130 if XNN_UNPREDICTABLE(i17 != zero) {
131 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
132 }
133 const int8_t* i18 = input[18];
134 assert(i18 != NULL);
135 if XNN_UNPREDICTABLE(i18 != zero) {
136 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
137 }
138 const int8_t* i19 = input[19];
139 assert(i19 != NULL);
140 if XNN_UNPREDICTABLE(i19 != zero) {
141 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
142 }
143 const int8_t* i20 = input[20];
144 assert(i20 != NULL);
145 if XNN_UNPREDICTABLE(i20 != zero) {
146 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
147 }
148 const int8_t* i21 = input[21];
149 assert(i21 != NULL);
150 if XNN_UNPREDICTABLE(i21 != zero) {
151 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
152 }
153 const int8_t* i22 = input[22];
154 assert(i22 != NULL);
155 if XNN_UNPREDICTABLE(i22 != zero) {
156 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
157 }
158 const int8_t* i23 = input[23];
159 assert(i23 != NULL);
160 if XNN_UNPREDICTABLE(i23 != zero) {
161 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
162 }
163 const int8_t* i24 = input[24];
164 assert(i24 != NULL);
165 if XNN_UNPREDICTABLE(i24 != zero) {
166 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
167 }
168 input = (const int8_t**) ((uintptr_t) input + input_stride);
169
170 size_t c = channels;
171 const void* w = weights;
172 for (; c >= 16; c -= 16) {
173 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
174
175
176 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
177 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
178 i0 += 16;
179
180 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
181
182 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
183 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
184 i1 += 16;
185
186 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
187
188 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
189 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
190 i2 += 16;
191
192 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
193
194 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
195 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
196 i3 += 16;
197
198 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
199
200 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
201 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
202 i4 += 16;
203
204 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
205
206 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
207 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
208 i5 += 16;
209
210 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
211
212 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
213 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
214 i6 += 16;
215
216 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
217
218 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
219 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
220 i7 += 16;
221
222 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
223
224 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
225 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
226 i8 += 16;
227
228 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
229
230 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9));
231 const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t))));
232 i9 += 16;
233
234 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
235
236 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10));
237 const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t))));
238 i10 += 16;
239
240 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
241
242 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11));
243 const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t))));
244 i11 += 16;
245
246 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
247
248 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12));
249 const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t))));
250 i12 += 16;
251
252 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
253
254 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13));
255 const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t))));
256 i13 += 16;
257
258 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
259
260 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14));
261 const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t))));
262 i14 += 16;
263
264 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
265
266 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15));
267 const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t))));
268 i15 += 16;
269
270 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
271
272 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16));
273 const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t))));
274 i16 += 16;
275
276 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
277
278 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17));
279 const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t))));
280 i17 += 16;
281
282 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
283
284 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18));
285 const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t))));
286 i18 += 16;
287
288 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
289
290 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19));
291 const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t))));
292 i19 += 16;
293
294 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
295
296 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20));
297 const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t))));
298 i20 += 16;
299
300 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
301
302 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21));
303 const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t))));
304 i21 += 16;
305
306 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
307
308 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22));
309 const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t))));
310 i22 += 16;
311
312 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
313
314 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23));
315 const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t))));
316 i23 += 16;
317
318 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
319
320 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24));
321 const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t))));
322 i24 += 16;
323
324 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
325
326 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
327
328 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
329
330 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
331
332 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
333
334 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
335
336 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
337
338 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
339 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
340 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
341
342 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
343
344 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
345 output += 16;
346 }
347 if XNN_UNLIKELY(c != 0) {
348 // Prepare mask for valid 8-bit elements (depends on nc).
349 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
350 {
351 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
352
353
354 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
355 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
356
357 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
358
359 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
360 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
361
362 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
363
364 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
365 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
366
367 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
368
369 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
370 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
371
372 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
373
374 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
375 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
376
377 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
378
379 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
380 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
381
382 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
383
384 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
385 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
386
387 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
388
389 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
390 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
391
392 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
393
394 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
395 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
396
397 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
398
399 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9));
400 const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t))));
401
402 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
403
404 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10));
405 const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t))));
406
407 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
408
409 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11));
410 const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t))));
411
412 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
413
414 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12));
415 const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t))));
416
417 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
418
419 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13));
420 const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t))));
421
422 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
423
424 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14));
425 const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t))));
426
427 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
428
429 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15));
430 const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t))));
431
432 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
433
434 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16));
435 const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t))));
436
437 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
438
439 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17));
440 const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t))));
441
442 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
443
444 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18));
445 const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t))));
446
447 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
448
449 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19));
450 const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t))));
451
452 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
453
454 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20));
455 const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t))));
456
457 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
458
459 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21));
460 const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t))));
461
462 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
463
464 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22));
465 const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t))));
466
467 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
468
469 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23));
470 const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t))));
471
472 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
473
474 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24));
475 const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t))));
476
477 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
478
479
480 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
481 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
482 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
483 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
484
485
486 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
487
488 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
489 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
490 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
491 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
492
493 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
494 output = (int8_t*) ((uintptr_t) output + c);
495 }
496 }
497
498 output = (int8_t*) ((uintptr_t) output + output_increment);
499 } while (--output_width != 0);
500 }