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
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__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 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_MSAN
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->fp32_avx512.output_max_less_zero_point);
34 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx512.output_zero_point);
35 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx512.output_min);
36
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 const int8_t* i9 = input[9];
84 assert(i9 != NULL);
85 if XNN_UNPREDICTABLE(i9 != zero) {
86 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
87 }
88 const int8_t* i10 = input[10];
89 assert(i10 != NULL);
90 if XNN_UNPREDICTABLE(i10 != zero) {
91 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
92 }
93 const int8_t* i11 = input[11];
94 assert(i11 != NULL);
95 if XNN_UNPREDICTABLE(i11 != zero) {
96 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
97 }
98 const int8_t* i12 = input[12];
99 assert(i12 != NULL);
100 if XNN_UNPREDICTABLE(i12 != zero) {
101 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
102 }
103 const int8_t* i13 = input[13];
104 assert(i13 != NULL);
105 if XNN_UNPREDICTABLE(i13 != zero) {
106 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
107 }
108 const int8_t* i14 = input[14];
109 assert(i14 != NULL);
110 if XNN_UNPREDICTABLE(i14 != zero) {
111 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
112 }
113 const int8_t* i15 = input[15];
114 assert(i15 != NULL);
115 if XNN_UNPREDICTABLE(i15 != zero) {
116 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
117 }
118 const int8_t* i16 = input[16];
119 assert(i16 != NULL);
120 if XNN_UNPREDICTABLE(i16 != zero) {
121 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
122 }
123 const int8_t* i17 = input[17];
124 assert(i17 != NULL);
125 if XNN_UNPREDICTABLE(i17 != zero) {
126 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
127 }
128 const int8_t* i18 = input[18];
129 assert(i18 != NULL);
130 if XNN_UNPREDICTABLE(i18 != zero) {
131 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
132 }
133 const int8_t* i19 = input[19];
134 assert(i19 != NULL);
135 if XNN_UNPREDICTABLE(i19 != zero) {
136 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
137 }
138 const int8_t* i20 = input[20];
139 assert(i20 != NULL);
140 if XNN_UNPREDICTABLE(i20 != zero) {
141 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
142 }
143 const int8_t* i21 = input[21];
144 assert(i21 != NULL);
145 if XNN_UNPREDICTABLE(i21 != zero) {
146 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
147 }
148 const int8_t* i22 = input[22];
149 assert(i22 != NULL);
150 if XNN_UNPREDICTABLE(i22 != zero) {
151 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
152 }
153 const int8_t* i23 = input[23];
154 assert(i23 != NULL);
155 if XNN_UNPREDICTABLE(i23 != zero) {
156 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
157 }
158 const int8_t* i24 = input[24];
159 assert(i24 != NULL);
160 if XNN_UNPREDICTABLE(i24 != zero) {
161 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
162 }
163 input = (const int8_t**) ((uintptr_t) input + input_stride);
164
165 size_t c = channels;
166 const void* w = weights;
167 for (; c >= 16; c -= 16) {
168 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
169
170
171 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
172 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
173 i0 += 16;
174
175 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
176
177 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
178 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
179 i1 += 16;
180
181 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
182
183 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
184 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
185 i2 += 16;
186
187 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
188
189 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
190 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
191 i3 += 16;
192
193 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
194
195 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
196 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
197 i4 += 16;
198
199 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
200
201 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
202 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
203 i5 += 16;
204
205 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
206
207 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
208 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
209 i6 += 16;
210
211 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
212
213 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
214 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
215 i7 += 16;
216
217 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
218
219 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
220 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
221 i8 += 16;
222
223 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
224
225 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9));
226 const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t))));
227 i9 += 16;
228
229 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
230
231 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10));
232 const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t))));
233 i10 += 16;
234
235 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
236
237 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11));
238 const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t))));
239 i11 += 16;
240
241 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
242
243 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12));
244 const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t))));
245 i12 += 16;
246
247 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
248
249 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13));
250 const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t))));
251 i13 += 16;
252
253 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
254
255 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14));
256 const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t))));
257 i14 += 16;
258
259 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
260
261 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15));
262 const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t))));
263 i15 += 16;
264
265 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
266
267 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16));
268 const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t))));
269 i16 += 16;
270
271 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
272
273 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17));
274 const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t))));
275 i17 += 16;
276
277 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
278
279 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18));
280 const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t))));
281 i18 += 16;
282
283 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
284
285 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19));
286 const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t))));
287 i19 += 16;
288
289 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
290
291 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20));
292 const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t))));
293 i20 += 16;
294
295 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
296
297 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21));
298 const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t))));
299 i21 += 16;
300
301 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
302
303 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22));
304 const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t))));
305 i22 += 16;
306
307 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
308
309 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23));
310 const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t))));
311 i23 += 16;
312
313 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
314
315 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24));
316 const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t))));
317 i24 += 16;
318
319 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
320
321 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t));
322
323 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
324
325 const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps(w);
326 w = (const void*) ((uintptr_t) w + 16 * sizeof(float));
327 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale0123456789ABCDEF);
328
329 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
330
331 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
332
333 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
334
335 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
336 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
337 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
338
339 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
340
341 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
342 output += 16;
343 }
344 if XNN_UNLIKELY(c != 0) {
345 // Prepare mask for valid 8-bit elements (depends on nc).
346 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
347 {
348 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
349
350
351 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
352 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
353
354 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
355
356 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
357 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
358
359 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
360
361 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
362 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
363
364 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
365
366 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
367 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
368
369 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
370
371 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
372 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
373
374 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
375
376 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
377 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
378
379 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
380
381 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
382 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
383
384 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
385
386 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
387 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
388
389 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
390
391 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
392 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
393
394 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
395
396 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9));
397 const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t))));
398
399 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
400
401 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10));
402 const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(int8_t))));
403
404 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
405
406 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11));
407 const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(int8_t))));
408
409 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
410
411 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12));
412 const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(int8_t))));
413
414 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
415
416 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13));
417 const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(int8_t))));
418
419 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
420
421 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14));
422 const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(int8_t))));
423
424 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
425
426 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15));
427 const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(int8_t))));
428
429 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
430
431 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16));
432 const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(int8_t))));
433
434 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
435
436 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17));
437 const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(int8_t))));
438
439 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
440
441 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18));
442 const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(int8_t))));
443
444 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
445
446 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19));
447 const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(int8_t))));
448
449 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
450
451 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20));
452 const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(int8_t))));
453
454 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
455
456 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21));
457 const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(int8_t))));
458
459 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
460
461 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22));
462 const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(int8_t))));
463
464 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
465
466 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23));
467 const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(int8_t))));
468
469 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
470
471 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24));
472 const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(int8_t))));
473
474 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
475
476
477 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
478 const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(int8_t)));
479 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale0123456789ABCDEF);
480 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
481 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
482
483
484 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
485
486 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
487 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
488 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
489 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
490
491 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
492 output = (int8_t*) ((uintptr_t) output + c);
493 }
494 }
495
496 output = (int8_t*) ((uintptr_t) output + output_increment);
497 } while (--output_width != 0);
498 }
499