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_qs8_dwconv_minmax_fp32_ukernel_up32x25__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__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_qs8_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 vscale = _mm512_load_ps(params->fp32_avx512.scale);
34 const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->fp32_avx512.output_max_less_zero_point);
35 const __m512i voutput_zero_point = _mm512_load_si512(params->fp32_avx512.output_zero_point);
36 const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->fp32_avx512.output_min);
37 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 5, 1, 6, 2, 4, 0);
38
39 do {
40 const int8_t* i0 = input[0];
41 assert(i0 != NULL);
42 if XNN_UNPREDICTABLE(i0 != zero) {
43 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44 }
45 const int8_t* i1 = input[1];
46 assert(i1 != NULL);
47 if XNN_UNPREDICTABLE(i1 != zero) {
48 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
49 }
50 const int8_t* i2 = input[2];
51 assert(i2 != NULL);
52 if XNN_UNPREDICTABLE(i2 != zero) {
53 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
54 }
55 const int8_t* i3 = input[3];
56 assert(i3 != NULL);
57 if XNN_UNPREDICTABLE(i3 != zero) {
58 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
59 }
60 const int8_t* i4 = input[4];
61 assert(i4 != NULL);
62 if XNN_UNPREDICTABLE(i4 != zero) {
63 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
64 }
65 const int8_t* i5 = input[5];
66 assert(i5 != NULL);
67 if XNN_UNPREDICTABLE(i5 != zero) {
68 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
69 }
70 const int8_t* i6 = input[6];
71 assert(i6 != NULL);
72 if XNN_UNPREDICTABLE(i6 != zero) {
73 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
74 }
75 const int8_t* i7 = input[7];
76 assert(i7 != NULL);
77 if XNN_UNPREDICTABLE(i7 != zero) {
78 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
79 }
80 const int8_t* i8 = input[8];
81 assert(i8 != NULL);
82 if XNN_UNPREDICTABLE(i8 != zero) {
83 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
84 }
85 const int8_t* i9 = input[9];
86 assert(i9 != NULL);
87 if XNN_UNPREDICTABLE(i9 != zero) {
88 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
89 }
90 const int8_t* i10 = input[10];
91 assert(i10 != NULL);
92 if XNN_UNPREDICTABLE(i10 != zero) {
93 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
94 }
95 const int8_t* i11 = input[11];
96 assert(i11 != NULL);
97 if XNN_UNPREDICTABLE(i11 != zero) {
98 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
99 }
100 const int8_t* i12 = input[12];
101 assert(i12 != NULL);
102 if XNN_UNPREDICTABLE(i12 != zero) {
103 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
104 }
105 const int8_t* i13 = input[13];
106 assert(i13 != NULL);
107 if XNN_UNPREDICTABLE(i13 != zero) {
108 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
109 }
110 const int8_t* i14 = input[14];
111 assert(i14 != NULL);
112 if XNN_UNPREDICTABLE(i14 != zero) {
113 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
114 }
115 const int8_t* i15 = input[15];
116 assert(i15 != NULL);
117 if XNN_UNPREDICTABLE(i15 != zero) {
118 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
119 }
120 const int8_t* i16 = input[16];
121 assert(i16 != NULL);
122 if XNN_UNPREDICTABLE(i16 != zero) {
123 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
124 }
125 const int8_t* i17 = input[17];
126 assert(i17 != NULL);
127 if XNN_UNPREDICTABLE(i17 != zero) {
128 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
129 }
130 const int8_t* i18 = input[18];
131 assert(i18 != NULL);
132 if XNN_UNPREDICTABLE(i18 != zero) {
133 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
134 }
135 const int8_t* i19 = input[19];
136 assert(i19 != NULL);
137 if XNN_UNPREDICTABLE(i19 != zero) {
138 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
139 }
140 const int8_t* i20 = input[20];
141 assert(i20 != NULL);
142 if XNN_UNPREDICTABLE(i20 != zero) {
143 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
144 }
145 const int8_t* i21 = input[21];
146 assert(i21 != NULL);
147 if XNN_UNPREDICTABLE(i21 != zero) {
148 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
149 }
150 const int8_t* i22 = input[22];
151 assert(i22 != NULL);
152 if XNN_UNPREDICTABLE(i22 != zero) {
153 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
154 }
155 const int8_t* i23 = input[23];
156 assert(i23 != NULL);
157 if XNN_UNPREDICTABLE(i23 != zero) {
158 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
159 }
160 const int8_t* i24 = input[24];
161 assert(i24 != NULL);
162 if XNN_UNPREDICTABLE(i24 != zero) {
163 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
164 }
165 input = (const int8_t**) ((uintptr_t) input + input_stride);
166
167 size_t c = channels;
168 const void* w = weights;
169 for (; c >= 32; c -= 32) {
170 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
171 __m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((uintptr_t) w + 16 * sizeof(int32_t)));
172
173
174 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
175 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
176 const __m512i vi0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i0 + 16)));
177 const __m512i vk0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
178 i0 += 32;
179
180 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
181 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV));
182
183 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
184 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
185 const __m512i vi1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i1 + 16)));
186 const __m512i vk1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
187 i1 += 32;
188
189 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
190 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV));
191
192 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
193 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
194 const __m512i vi2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i2 + 16)));
195 const __m512i vk2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
196 i2 += 32;
197
198 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
199 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV));
200
201 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
202 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
203 const __m512i vi3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i3 + 16)));
204 const __m512i vk3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
205 i3 += 32;
206
207 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
208 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV));
209
210 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
211 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
212 const __m512i vi4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i4 + 16)));
213 const __m512i vk4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
214 i4 += 32;
215
216 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
217 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV));
218
219 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
220 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
221 const __m512i vi5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i5 + 16)));
222 const __m512i vk5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
223 i5 += 32;
224
225 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
226 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV));
227
228 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
229 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
230 const __m512i vi6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i6 + 16)));
231 const __m512i vk6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
232 i6 += 32;
233
234 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
235 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV));
236
237 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
238 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
239 const __m512i vi7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i7 + 16)));
240 const __m512i vk7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
241 i7 += 32;
242
243 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
244 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV));
245
246 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
247 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
248 const __m512i vi8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i8 + 16)));
249 const __m512i vk8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
250 i8 += 32;
251
252 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
253 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV));
254
255 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9));
256 const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t))));
257 const __m512i vi9xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i9 + 16)));
258 const __m512i vk9xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 304 * sizeof(int8_t))));
259 i9 += 32;
260
261 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
262 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV));
263
264 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10));
265 const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 320 * sizeof(int8_t))));
266 const __m512i vi10xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i10 + 16)));
267 const __m512i vk10xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 336 * sizeof(int8_t))));
268 i10 += 32;
269
270 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
271 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV));
272
273 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11));
274 const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 352 * sizeof(int8_t))));
275 const __m512i vi11xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i11 + 16)));
276 const __m512i vk11xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 368 * sizeof(int8_t))));
277 i11 += 32;
278
279 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
280 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV));
281
282 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12));
283 const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 384 * sizeof(int8_t))));
284 const __m512i vi12xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i12 + 16)));
285 const __m512i vk12xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 400 * sizeof(int8_t))));
286 i12 += 32;
287
288 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
289 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV));
290
291 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13));
292 const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 416 * sizeof(int8_t))));
293 const __m512i vi13xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i13 + 16)));
294 const __m512i vk13xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 432 * sizeof(int8_t))));
295 i13 += 32;
296
297 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
298 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV));
299
300 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14));
301 const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 448 * sizeof(int8_t))));
302 const __m512i vi14xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i14 + 16)));
303 const __m512i vk14xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 464 * sizeof(int8_t))));
304 i14 += 32;
305
306 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
307 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV));
308
309 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15));
310 const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 480 * sizeof(int8_t))));
311 const __m512i vi15xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i15 + 16)));
312 const __m512i vk15xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 496 * sizeof(int8_t))));
313 i15 += 32;
314
315 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
316 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV));
317
318 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16));
319 const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 512 * sizeof(int8_t))));
320 const __m512i vi16xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i16 + 16)));
321 const __m512i vk16xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 528 * sizeof(int8_t))));
322 i16 += 32;
323
324 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
325 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV));
326
327 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17));
328 const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 544 * sizeof(int8_t))));
329 const __m512i vi17xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i17 + 16)));
330 const __m512i vk17xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 560 * sizeof(int8_t))));
331 i17 += 32;
332
333 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
334 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV));
335
336 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18));
337 const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 576 * sizeof(int8_t))));
338 const __m512i vi18xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i18 + 16)));
339 const __m512i vk18xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 592 * sizeof(int8_t))));
340 i18 += 32;
341
342 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
343 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV));
344
345 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19));
346 const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 608 * sizeof(int8_t))));
347 const __m512i vi19xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i19 + 16)));
348 const __m512i vk19xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 624 * sizeof(int8_t))));
349 i19 += 32;
350
351 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
352 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV));
353
354 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20));
355 const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 640 * sizeof(int8_t))));
356 const __m512i vi20xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i20 + 16)));
357 const __m512i vk20xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 656 * sizeof(int8_t))));
358 i20 += 32;
359
360 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
361 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV));
362
363 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21));
364 const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 672 * sizeof(int8_t))));
365 const __m512i vi21xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i21 + 16)));
366 const __m512i vk21xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 688 * sizeof(int8_t))));
367 i21 += 32;
368
369 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
370 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV));
371
372 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22));
373 const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 704 * sizeof(int8_t))));
374 const __m512i vi22xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i22 + 16)));
375 const __m512i vk22xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 720 * sizeof(int8_t))));
376 i22 += 32;
377
378 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
379 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV));
380
381 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23));
382 const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 736 * sizeof(int8_t))));
383 const __m512i vi23xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i23 + 16)));
384 const __m512i vk23xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 752 * sizeof(int8_t))));
385 i23 += 32;
386
387 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
388 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV));
389
390 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24));
391 const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 768 * sizeof(int8_t))));
392 const __m512i vi24xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i24 + 16)));
393 const __m512i vk24xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 784 * sizeof(int8_t))));
394 i24 += 32;
395
396 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
397 vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV));
398
399 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t));
400
401 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
402 __m512 vscaledGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vaccGHIJKLMNOPQRSTUV);
403
404 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
405 vscaledGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaledGHIJKLMNOPQRSTUV, vscale);
406
407 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
408 vscaledGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaledGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
409
410 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
411 vaccGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaledGHIJKLMNOPQRSTUV);
412
413 __m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point);
414 __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vaccGHIJKLMNOPQRSTUV), _mm512_extracti32x8_epi32(vaccGHIJKLMNOPQRSTUV, 1)), _mm512_castsi512_si256(voutput_zero_point));
415
416 const __m256i vout0123GHIJ4567KLMN = _mm512_castsi512_si256(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV);
417 const __m256i vout89ABOPQRCDEFSTUV = _mm512_extracti32x8_epi32(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV, 1);
418 const __m256i vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV = _mm256_packs_epi16(vout0123GHIJ4567KLMN, vout89ABOPQRCDEFSTUV);
419 __m256i vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_permutevar8x32_epi32(vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV, vpermute_mask);
420 const __m128i voutGHIJOPQR = _mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV);
421 const __m128i voutKLMNSTUV = _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1);
422 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(voutGHIJOPQR, voutKLMNSTUV), _MM_SHUFFLE(3, 1, 2, 0));
423
424 vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_max_epi8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_min);
425 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, _mm256_castsi256_si128(voutput_min));
426
427 _mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV);
428 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
429 output += 32;
430 }
431 if XNN_UNLIKELY(c != 0) {
432 // Prepare mask for valid 8-bit elements (depends on nc).
433 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
434 const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
435 do {
436 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
437
438
439 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
440 const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) k));
441 i0 += 16;
442
443 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
444
445 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
446 const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 32)));
447 i1 += 16;
448
449 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
450
451 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
452 const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 64)));
453 i2 += 16;
454
455 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
456
457 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
458 const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 96)));
459 i3 += 16;
460
461 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
462
463 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
464 const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 128)));
465 i4 += 16;
466
467 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
468
469 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
470 const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 160)));
471 i5 += 16;
472
473 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
474
475 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
476 const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 192)));
477 i6 += 16;
478
479 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
480
481 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
482 const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 224)));
483 i7 += 16;
484
485 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
486
487 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
488 const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 256)));
489 i8 += 16;
490
491 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
492
493 const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9));
494 const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 288)));
495 i9 += 16;
496
497 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
498
499 const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10));
500 const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 320)));
501 i10 += 16;
502
503 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF));
504
505 const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11));
506 const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 352)));
507 i11 += 16;
508
509 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
510
511 const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12));
512 const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 384)));
513 i12 += 16;
514
515 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF));
516
517 const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13));
518 const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 416)));
519 i13 += 16;
520
521 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
522
523 const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14));
524 const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 448)));
525 i14 += 16;
526
527 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF));
528
529 const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15));
530 const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 480)));
531 i15 += 16;
532
533 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
534
535 const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16));
536 const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 512)));
537 i16 += 16;
538
539 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF));
540
541 const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17));
542 const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 544)));
543 i17 += 16;
544
545 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
546
547 const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18));
548 const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 576)));
549 i18 += 16;
550
551 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF));
552
553 const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19));
554 const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 608)));
555 i19 += 16;
556
557 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
558
559 const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20));
560 const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 640)));
561 i20 += 16;
562
563 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF));
564
565 const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21));
566 const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 672)));
567 i21 += 16;
568
569 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
570
571 const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22));
572 const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 704)));
573 i22 += 16;
574
575 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF));
576
577 const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23));
578 const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 736)));
579 i23 += 16;
580
581 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
582
583 const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24));
584 const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 768)));
585 i24 += 16;
586
587 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF));
588
589 k += 16;
590
591 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
592 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
593 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
594 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
595
596 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
597
598 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point));
599
600 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
601 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
602 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
603 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_min));
604
605 if XNN_LIKELY(c >= 16) {
606 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
607 output += 16;
608 c -= 16;
609 } else {
610 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
611 output = (int8_t*) ((uintptr_t) output + c);
612 c = 0;
613 }
614 } while (c != 0);
615 }
616
617 output = (int8_t*) ((uintptr_t) output + output_increment);
618 } while (--output_width != 0);
619 }
620